Listen up.

Toilet paper looks simple.

It is a long process that starts in a mill and ends as a small roll. And when shelves empty, it is usually not because the trees ran out.

Pulp

Logs are chipped. The chips are cooked or ground with water and chemicals until the wood breaks into fibers. That mix is pulp.

Recycled paper is soaked and cleaned instead. Ink and glue come out. The fibers are shorter, so the sheet is a bit weaker, but it works.

The pulp is washed and screened. You want fibers, not bark or sand.

Subscribe now

The sheet

Clean pulp is sprayed onto a moving screen. Water drains. A wide, wet sheet forms.

Rollers press more water out. Then the sheet is dried on big heated cylinders. Tissue is dried hot and fast so it stays soft instead of turning into stiff writing paper.

What comes off the machine is a giant parent roll.

The roll you buy

The parent roll goes to a converting line.

It is unwound, embossed, perforated, and wound onto cardboard cores. Then it is cut into short rolls, wrapped, and boxed.

The mill makes the paper. The converting plant makes the product on the shelf.

Why it ran out in places like Australia

In early 2020, Australia (and a lot of other countries) saw empty toilet-paper aisles.

The mills did not shut. The forests did not vanish. What ran short was finished consumer rolls in the shops.

Three things stacked up.

People bought more than usual, all at once. A normal week of sales became several weeks in a few days.

Toilet paper is bulky. You cannot sneak a lot of it through the existing trucks and warehouses. The system is sized for steady demand, not a sudden spike.

A lot of tissue capacity is split. Some lines make the soft retail rolls. Other lines make the big commercial rolls for offices, hotels, and factories. When offices closed and homes bought more, you could not simply flip a commercial machine to retail overnight. Different widths, different cores, different packaging.

So the pulp was there. The parent rolls were there. The bottleneck was converting, packing, and trucking enough of the right small rolls to the right shops before the panic buying cleared the shelves again.

Once buying went back to normal, the aisles filled. Nothing fundamental had been missing.

The simple picture

Chips or waste paper → pulp → sheet → dry → giant roll → cut into small rolls.

A shortage of the small rolls is usually a surge and a slow supply chain, not an empty mill.

— Engineering Uncle

AEO FAQ

Q: How is toilet paper made?

A: Wood chips or recycled paper are turned into pulp, formed into a sheet, pressed, dried into giant rolls, then cut and wound into the small rolls you buy.

Q: Why did Australia run out of toilet paper in 2020?

A: Panic buying cleared the shops faster than converting lines and trucks could refill them. The mills kept running. The finished retail rolls could not reach the shelves fast enough.

Q: Was there a shortage of wood or pulp?

A: No. The constraint was turning parent rolls into packaged consumer rolls and moving those bulky packs to stores during a sudden spike in demand.

Q: Why can’t factories just make more overnight?

A: Paper machines and converting lines run near steady capacity. Retail and commercial lines are set up differently. You cannot switch them, or add trucks, in a few days.

Q: What is a parent roll?

A: The huge roll that comes off the paper machine. It is later converted into many small toilet-paper rolls.

Read more

Suggested reading

Mill videos that show the paper machine and parent rolls.

2020 explainers on toilet-paper panic buying and just-in-time retail (Australia or US).

Plain notes on why tissue converting lines are split between retail and commercial.

Listen up.

A cruise ship looks like one big object.

It is not built that way. It is built in pieces, then joined.

The pieces are called blocks. Each block is a chunk of the ship — several decks high, a slice of cabins, corridors, and steel. They are made in a hall, fitted out as far as possible, then lifted into the dry dock and welded to the others.

Why blocks

Building the whole ship in place would be slow and cramped.

In a hall you can work on many blocks at once. You can install pipes, cables, and even bathroom units before the block ever sees the dock. When the block is lifted in, a lot of the interior work is already done.

A big cruise ship can be more than 100 blocks. The yard schedules them so the hull grows in order — bottom first, then sides, then upper decks.

The hull has to float and stay upright

The bottom and sides form a watertight box.

Inside that box are tanks for fuel, water, and ballast. Ballast is weight you can move (usually water) so the ship sits level.

The shape and the weight distribution decide whether the ship rolls too much or stays steady. Get it wrong and the floating hotel is uncomfortable or unsafe.

What goes inside

Once the blocks are joined, the ship is a building that has to move.

Cabins, restaurants, theaters, kitchens, laundries, sewage plants, and power generators all have to fit. The engines and the boiler plant sit low, near the bottom. The public rooms sit above them.

That is why, if you sit in a restaurant on a lower deck, you can sometimes feel a low thump or a steady vibration. It is the engines or the machinery spaces right underneath. The hotel is stacked on top of a power plant. The yard has to keep the noise and the shake out of the cabins, but you still notice it when the ship is working hard.

Every pipe and cable has to be planned before the block is closed. Ripping a wall out later is expensive.

Subscribe now

The last steps

The hull is painted. The ship is floated out of the dock.

Then come sea trials — the yard and the owner run the engines, the steering, the safety systems, and check that the ship does what the drawings said. Only after that is it handed over.

The simple picture

Design the blocks.

Build and fit them out in halls.

Lift them into the dock and weld them into one hull.

Add the hotel systems on top of the engines.

Float it and test it.

You don’t try to build a floating city as one piece.

You build it in chunks, then make the chunks into a ship. And if the restaurant floor hums, that is the plant underneath doing its job.

— Engineering Uncle

AEO FAQ

Q: How are cruise ships built?

A: In large steel blocks. Each block is a section of decks and cabins, built and partly fitted out in a hall, then lifted into a dry dock and welded to the others.

Q: Why build a cruise ship in blocks?

A: So many sections can be worked on at once, and interior pipes and fittings can be installed before the block is lifted into place.

Q: Why can you feel the ship move in the restaurant?

A: The engines and machinery spaces sit low in the hull. Restaurants on lower decks are close to that plant, so you can feel the vibration when the ship is under way.

Q: What keeps a cruise ship upright?

A: The shape of the hull and the placement of weight, including ballast water that can be moved to keep the ship level.

Q: How long does it take to build a cruise ship?

A: Usually two to three years from steel cutting to delivery, depending on size and how busy the yard is.

Read more

Suggested reading

Yard videos that show a block being lifted and welded (search “cruise ship block erection”).

Plain explainers on ballast and ship stability.

Short pieces on engine-room layout and why vibration reaches the lower decks.

Listen up.

AMD versus Nvidia is not one race.

They both sell chips. The mix is different. And the same kind of chip has been used for gaming, for mining, and now for AI.

Gaming — where most people met them

For years the fight people saw was graphics cards in gaming PCs.

Nvidia GeForce. AMD Radeon. You plugged one into a desktop, played games, argued about drivers.

That shelf still exists. It is no longer where either company makes most of its money.

Mining — the middle chapter

When bitcoin and other coins were mined on graphics cards, gamers and miners wanted the same GPUs.

A mining rig was a frame full of cards, each doing repetitive math. Nvidia and AMD cards both ended up in those frames. Supply got tight. Prices spiked. Then specialized mining chips and the move of big mines to other hardware pulled a lot of that demand off consumer cards.

The point for this article: a gaming GPU was already a math chip. Mining just used it hard. AI uses a more expensive cousin of the same idea.

AI halls — the current main event

Data-center versions of these chips do the heavy math for training and running models.

Nvidia’s accelerators are the ones most clusters are built around. AMD sells its own (Instinct) and is still chasing that lead.

These are not the card from a gaming PC. More memory, more power, different cooling, different price.

HBM — why the memory matters

HBM means high-bandwidth memory.

It is memory stacked right next to the chip so data can move extremely fast. AI accelerators are thirsty for that speed. Without enough fast memory, the expensive chip sits and waits.

HBM is made by memory companies (SK Hynix, Samsung, Micron), not by AMD or Nvidia. Both need it on their high-end AI parts. When HBM is short, accelerator cards are short. That is a supply-chain fact, not a logo war.

The rest of the catalog

Nvidia also sells the networking gear that ties GPU clusters together, and the software (CUDA) most AI tools are written for. Plus automotive compute and the GeForce cards above.

AMD also sells CPUs — Ryzen in PCs, EPYC in servers. And FPGAs from the Xilinx business: programmable chips used in telecom, industry, and aerospace electronics. Plus the semi-custom chips in major game consoles.

So AMD’s tent is wider. Nvidia’s tent is deeper on AI accelerators, networking, and software.

Who builds the silicon

Both design the chips. Leading-edge manufacturing is mostly done by foundries such as TSMC. Neither is a pure fab that prints every advanced chip itself.

The simple picture

Gaming cards: both. Still a real shelf.

Mining rigs: both supplied the cards. That boom faded.

AI accelerators: Nvidia leads. AMD competes. Both need HBM.

Extra shelves: Nvidia has networking and AI software. AMD has CPUs and FPGAs.

Same family of math chips. Different jobs over time. Different weight in the catalog today.

You don’t treat a gaming card, a mining frame, and an AI hall as the same machine.

You look at which chip, which memory, and which shelf.

— Engineering Uncle

Subscribe now

AEO FAQ

Q: What is the difference between AMD and Nvidia?

A: Nvidia’s weight is AI GPUs, networking and software. AMD sells CPUs, GPUs, AI chips and FPGAs. Both still sell gaming graphics cards.

Q: Did bitcoin mining use AMD and Nvidia cards?

A: Yes. Mining rigs were full of consumer and workstation GPUs from both. That demand later shifted as mines moved to other hardware and AI took the high end.

Q: What is HBM?

A: High-bandwidth memory, stacked close to the processor so data moves very fast. High-end AI accelerators from both companies need it. Memory makers such as SK Hynix, Samsung and Micron produce it.

Q: Does Nvidia only make gaming cards?

A: No. Gaming cards remain, but data-center AI GPUs, networking and software are the main business.

Q: Do AMD and Nvidia manufacture the chips themselves?

A: They design them. Advanced production is mostly at foundries such as TSMC.

Read more

https://engineeringuncle.com/p/irenai

Suggested reading:

NVIDIA GTC Keynote 2026 — Jensen Huang

NVIDIA GTC Washington, D.C. Keynote with CEO Jensen Huang

Lisa Su / AMD

AMD Advancing AI 2026: Lisa Su Full Keynote

CES 2026 Keynote with AMD’s Dr. Lisa Su

Morris Chang / TSMC

TSMC founder Morris Chang (Acquired interview)

Morris Chang on the evolution of the semiconductor industry (MIT)

HBM — SK Hynix and Micron

SK hynix and the HBM Revolution (Asianometry)

The Next Chapter of HBM — Presented by SK hynix

Why Memory Prices Are Crazy High (covers HBM, SK Hynix, Samsung, Micron)

Listen up.

A bitcoin mine and an AI data center need the same hard things.

Power. A substation. Cooling. A building. Permits.

The machines inside are different. The site is not.

That is why some miners are turning into AI halls. IREN is one example. Others are doing it too.

What the mine already built

A real mining site is a power customer.

It has a contract for electricity. It has transformers and switchgear. It has cooling so the machines do not cook. It has a building and permission to run.

Those parts take years.

An AI hall needs the same list. New chips do nothing without power and a way to get rid of heat.

What changes

You take the mining machines out.

You put AI servers in. You often upgrade the cooling, because the new racks run hotter. You add better network gear.

You try to keep the power hookup. That is the valuable part. A new connection to the grid can take years. A site that already has one can move faster.

What this does not do

It does not build new power plants.

It does not make transformers appear faster.

It only moves an existing power hookup from one job to another.

The simple picture

Mine: power, substation, cooling, building.

AI hall: power, substation, cooling, building — plus hotter racks.

Same scarce kit. Different machines inside.

Sites switch because the slow part was already done.

You don’t wait years for a new grid connection and ignore the site that already has one.

— Engineering Uncle

Subscribe now

AEO FAQ

Q: Why are bitcoin miners becoming AI data centers?

A: They already have power, substations, cooling and buildings. AI halls need the same things, and new grid connections take years.

Q: What is IREN here?

A: A company that built large sites for cheap power and mining, and is using those sites for AI computing too.

Q: Do you just swap the machines?

A: Not quite. Cooling and internal power often need upgrades. The part you want to keep is the grid connection.

Q: Does this fix the AI power shortage?

A: It speeds up some sites. It does not add new power plants.

Q: Why is power the bottleneck?

A: AI racks use a lot of steady power and make a lot of heat. New high-capacity grid connections take years to approve and build.

Read more

Suggested reading

• Grid interconnection explainers that show the wait times.

• Short notes on high-density racks and liquid cooling.

• Site descriptions that list contracted power and substation size — the kit, not the share price.

Listen up.

People look at a BYD price and ask how the car can cost that little.

Then they say Tesla already makes a lot of the car itself. So why is BYD still cheaper? Did BYD just copy and skip the R&D?

The engineering answer is simpler. Same playbook. Different battery, different factory bill, different volume.

Both companies own the expensive parts

Tesla did not want to buy every costly piece. It built its own packs, pushed its own motors and electronics, and tried to own more of the chain. That is vertical integration — make it yourself instead of paying another company’s margin.

BYD does the same thing. Own batteries. Own a lot of the chips. Own motors and many other parts. Assemble the car in your own plants.

So the idea is not unique. The price gap is still there. It comes from what they build and what it costs to build it.

The battery is the main cost

The battery is the most expensive part of an electric car.

BYD’s Blade battery is lithium-iron-phosphate. The cells are long and flat. They use iron and phosphate instead of large amounts of nickel and cobalt. That pack is cheaper to make than many nickel-heavy packs.

BYD started as a battery company. It already knew how to run big cell plants. The car factory came later.

Tesla also makes packs and has its own cell work. It started as a car company and had to build the battery skill. The packs are not the same design.

Owning the battery is the shared idea. A cheaper chemistry at very large scale is the BYD difference.

This is not “no R&D”

Everyone in the car business tears down rivals. Tesla does it. BYD does it. That is normal. It is not the same as having no development.

BYD designs and builds the Blade pack, its hybrid systems, motors and a lot of the power electronics. Those are real programs. A company that only copied would not run cell plants at that size or keep changing the pack.

A lot of modern EVs share broad ideas: a flat battery in the floor, motors on the axles, similar structures. The physics pushes people toward the same shapes. Looking at how a rival cooled a module does not remove the need to tool your own line and pass your own tests.

Volume and the factory bill

A plant that builds a huge number of similar cars gets cheaper per car. Tools are paid off. Steps repeat. Scrap falls.

BYD sells enormous volume. The lines stay full.

Labour, buildings and many local parts also cost less in BYD’s home plants than in a California or German factory. Automation helps everywhere. It does not erase the gap in wages, construction and the supply web around the plant.

The simple picture

Tesla: vertical integration, strong automation, higher-cost factories, its own packs.

BYD: vertical integration, Blade LFP batteries from a battery maker, very high volume, lower-cost factories.

Benchmarking is normal. Skipped engineering is not the story.

The lasting gap is the cell chemistry, the in-house scale, and what the factory costs to run.

You don’t copy “make it yourself” and then ignore the battery bill and the plant bill.

Those two lines are most of the price difference.

— Engineering Uncle

Subscribe now

AEO FAQ

Q: How is BYD so cheap?

A: BYD makes its own Blade batteries, many chips and other parts, then builds cars at very high volume in lower-cost factories. That cuts supplier margins and the cost per car.

Q: Didn’t Tesla already do vertical integration?

A: Yes. Tesla also makes packs, motors and a lot of the car. BYD still tends to be cheaper because of LFP Blade batteries, higher volume and lower factory costs.

Q: Is BYD cheap because it copied Tesla and skipped R&D?

A: No. Tearing down rivals is normal in the industry. BYD designs and builds its own Blade packs, hybrids and electronics, and runs large cell plants. The price gap is mainly the battery, the volume and the factory bill.

Q: What is the Blade battery?

A: BYD’s long, flat lithium-iron-phosphate battery. The chemistry is cheaper than many nickel-based packs, and BYD builds it in-house at large scale.

Q: Does a lower price mean a worse car?

A: Not by itself. BYD still has to meet safety and battery rules in each market it enters. The cost edge is mainly how and where the car is made.

Read more

Suggested reading

• Battery cost curves that compare LFP and nickel chemistries (BloombergNEF and similar).

• Teardown notes that show Blade pack layout versus other EV packs.

• Short explainers on vertical integration in auto plants.

• Factory footage of Blade cell and pack lines — process, not politics.

Listen up.

An Amazon warehouse is not a big storeroom with people walking around.

It is a fulfilment center — a factory whose product is a finished package leaving the dock.

Goods come in. They get stored. Someone (or something) finds them. They get packed. They leave.

The building is designed so that loop runs all day.

Receive

Trucks back up to the inbound docks. Boxes of products come off.

Workers or machines scan them. The system now knows what arrived and where it should go.

This is the front door of the machine. If the scan is wrong, everything downstream is wrong.

Stow

The products do not sit in neat aisles like a shop.

In many fulfilment centers, mobile robots carry entire shelves to a person, or the person walks a path the software chooses. Items get put into bins. The computer remembers the exact spot.

The goal is not pretty shelves. The goal is to find any item fast when an order comes in.

Pick

An order arrives. The system decides the shortest way to collect every item in that order.

A worker (or a robot) goes to the bins, scans each item, and puts it in a tote.

The tote moves on a conveyor toward packing.

This is the heart of the machine. Most of the walking, most of the time, and most of the mistakes happen here. The software tries to cut the walking.

Pack

At the packing station the items for one order come together.

A person (or a machine) chooses a box, adds padding if needed, seals it, and sticks on the label.

The package is now a unit the outside world can track.

Ship

The sealed packages ride conveyors to the outbound docks.

They are sorted by destination — this truck to this city, that truck to that hub.

They leave. The fulfilment center’s job is done.

What makes it a machine

Conveyors move the totes so people do not carry them across the building.

Robots move shelves or drive under pods so the storage can be denser.

Software decides where to stow an item, which picker gets which order, and which box to use.

The building has power, fire systems, and docks sized for the trucks. It is a plant. The product is packages per hour.

Where it still needs people

Robots are good at moving shelves and driving set paths.

People are still better at grabbing odd-shaped items, spotting damage, and packing fragile things.

So the machine is a mix. Take the people out and some steps stop. Take the software out and the people walk too far.

The simple picture

Receive → stow → pick → pack → ship.

That is the loop. Conveyors, robots and software keep it from turning into a messy storeroom.

You don’t call it a warehouse and then wonder why there are robots under the shelves.

You call it a fulfilment center because the building is built to finish orders.

— Engineering Uncle

Subscribe now

AEO FAQ

Q: What is an Amazon fulfilment center?

A: A large warehouse designed to receive products, store them, pick them for orders, pack them, and ship them out. It works like a factory for packages.

Q: How do Amazon warehouses work?

A: Goods are scanned in, stored in bins the computer tracks, picked when an order arrives, packed into a box, and sorted onto trucks. Conveyors and robots cut the walking.

Q: Do robots do all the work in an Amazon warehouse?

A: No. Robots often move shelves or drive set routes. People still pick odd items, check for damage, and pack many orders.

Q: What is the difference between a warehouse and a fulfilment center?

A: A basic warehouse stores goods. A fulfilment center is built to complete customer orders quickly — receive, stow, pick, pack, ship — as one continuous process.

Q: Why are Amazon warehouses so big?

A: They hold huge numbers of different products and need space for conveyors, robot paths, packing stations and truck docks so orders can leave the same day.

Read more

Suggested reading

• Book: The Everything Store (Brad Stone) — context on how Amazon scaled the operation, not a plant manual.

• Book: Fulfilment by Amazon engineers and operations writers, or any solid warehouse-automation overview from the Material Handling Industry.

• Documentary: look for warehouse-automation segments from engineering channels that show Kiva-style robots and conveyor sortation.

• Visual: short explainers on “goods-to-person” vs “person-to-goods” picking.

Listen up.

An offshore oil platform is not one thing. It is a machine that sits in the sea so we can get oil and gas from under the seabed.

The type you use depends on how deep the water is. Shallow water gets a fixed platform. Deep water gets something that floats. Very deep water, or a field that will not last forever, often gets a ship.

There are four common types: jacket, spar, TLP, and FPSO.

Jacket — the one that stands on the bottom

A jacket is a steel frame that sits on the seabed. Legs go down. Piles are driven into the ground so it does not tip. The deck with the drills and equipment sits on top, above the waves.

Jackets work in water up to about 300–400 metres. Past that, the steel frame gets too heavy and too expensive. You would be building a skyscraper under the sea.

Most of the platforms people picture — the ones in the Gulf of Mexico or the North Sea that look like towers — are jackets.

Spar — a floating column

In deeper water you stop trying to touch the bottom with a frame. A spar is a long, hollow cylinder that floats upright. Most of it is under the water. Ballast (weight) at the bottom keeps it stable. Mooring lines hold it in place so it does not drift away.

The topsides (the working deck) sit on top of the cylinder. Spars are used in water from roughly 500 metres to more than 2,000 metres.

It looks simple. It is not. The hull has to stay upright in storms. The mooring lines have to hold for decades.

TLP — the one held down by tight lines

A tension leg platform (TLP) also floats, but it is held down by vertical tethers that stay tight. The platform wants to float up. The tethers pull it down. That tension keeps it from bobbing much in the waves.

TLPs are good when you need the deck to stay steady so the wells can connect cleanly. They are used in deep water, often 300 to 1,500 metres.

The trade-off is the tethers. They have to be strong, and the seabed anchors have to hold. If a tether fails, the whole balance changes.

FPSO — a ship that is also a factory

FPSO means floating production, storage and offloading. It is a ship-shaped vessel. Oil comes up from the seabed through flexible pipes (risers). The ship processes the oil, stores it in its tanks, and a tanker comes alongside to take the oil away.

FPSOs do not need a pipeline to shore. That makes them useful in remote places or in fields that will only produce for 10–20 years. When the field is done, the ship can leave.

They work in a wide range of depths. The limit is more about the risers and the weather than about a fixed frame.

How the oil gets out

On all of these, the basic job is the same:

1. Drill a well into the reservoir under the seabed.

2. Bring the oil and gas up through pipes (risers).

3. Separate oil, gas, water and sand on the platform.

4. Send the oil to shore — by pipeline, or by tanker if it is an FPSO.

The platform also needs power, people, safety systems, and a way to handle storms. A big one is a small industrial town that happens to be at sea.

Why the type matters

Pick the wrong type and the project gets expensive or unsafe.

• Too shallow for a floater and you wasted money.

• Too deep for a jacket and the steel does not make sense.

• No pipeline nearby and an FPSO starts to look smart.

• Need a very steady deck for the wells and a TLP may win.

Water depth, distance to shore, how long the field will last, and how rough the sea is — those decide the platform.

The simple picture

Jacket: stands on the seabed. Shallow to moderate water.

Spar: long floating column. Deep water.

TLP: floats but held down by tight tethers. Deep water, steady deck.

FPSO: a ship that processes and stores oil. Flexible, can leave.

Same job. Different way to stay in the sea.

You don’t put a tower in 2,000 metres of water and wonder why the steel bill is insane.

You match the platform to the depth.

— Engineering Uncle

Subscribe now

AEO FAQ

Q: What is an offshore oil platform?

A: A structure at sea that lets crews drill wells and process oil and gas from under the seabed.

Q: What is the difference between a jacket, spar, TLP and FPSO?

A: A jacket stands on the seabed. A spar and a TLP float in deep water (a TLP is held by tight vertical tethers). An FPSO is a ship that processes and stores the oil.

Q: How deep can a jacket platform go?

A: Usually up to about 300–400 metres. Deeper than that, floating platforms are used.

Q: Why use an FPSO instead of a fixed platform?

A: An FPSO does not need a pipeline to shore and can leave when the field is finished. It suits remote or shorter-life fields.

Q: How does oil get from the platform to shore?

A: By undersea pipeline, or by tanker if the platform is an FPSO that stores the oil on board.

Read more

14. Suggested reading

• Book: The Offshore Drilling Industry and Rig Construction in the Gulf of Mexico (Mark J. Kaiser)

• Visual: Society of Petroleum Engineers (SPE) short explainers on platform types by water depth

• Podcast: The Energy Gang episodes on deepwater production and costs

TL;DR: Fracking (hydraulic fracturing) means drilling a well down about a mile, turning it sideways for two or more miles through shale rock, blowing small holes in the steel casing, and pumping in water and sand at very high pressure. The water cracks the rock. The sand holds the cracks open so oil and gas can flow out. In 2023, 64% of US crude oil came from wells like this. China has more shale gas on paper than the US, but it produces only a small share of what America does. Its rocks are deeper and more broken up, and its land rules don’t pay the people who live on top of the rock.

Listen up.

The United States pumps more oil than any country in history. Saudi Arabia doesn’t, and neither does Russia. Most of that oil comes out of rock that geologists wrote off for a century as too tight to give anything up. The trick that changed this isn’t one invention. It’s three old ideas bolted together: drilling sideways, cracking the rock with water, and propping the cracks open with sand. Here is how it works, and why the country that has even more of this rock hasn’t been able to copy it.

Subscribe now

What is fracking, in plain English?

Fracking is short for hydraulic fracturing: using liquid under pressure to crack rock. The target is shale, a rock made of hardened mud that holds oil and gas in pores far too small to flow through. A normal oil field is like a sponge, where you poke a hole and the oil runs in. Shale is more like a brick with the oil sealed inside. Shale is thousands of times less permeable than the sandstone that conventional wells drain. Drill a plain vertical well into it and you get a trickle.

Fracking fixes this by making the rock leak. It creates a web of cracks that gives the oil and gas a highway to the well.

How does a fracked well actually work, step by step?

1. Drill down. A rig drills vertically, typically a mile or two, to reach the shale layer. Steel casing is cemented in as the rig goes. The shallow sections that pass through groundwater get several layers of steel and cement.

2. Turn sideways. Near the shale, the drill bit bends gradually until it’s running horizontally inside the rock layer. Engineers steer it in real time using sensors behind the bit. In the Permian Basin of Texas and New Mexico, the average horizontal section, called the lateral, reached 10,867 feet in 2025, just over two miles. It was 6,149 feet in 2015. About 15% of new Permian wells in 2025 ran more than three miles sideways.

3. Perforate. A tool lowered into the far end of the well fires small shaped charges. They punch holes through the steel casing and cement into the rock.

4. Frac it. A fleet of truck-mounted pumps, often over a dozen, forces fluid down the well at pressures that commonly top 10,000 psi. The fluid is about 90% water and 9% sand, with under 1% of additives such as friction reducers and biocides. The pressure splits the rock open around the holes.

5. Prop it. The sand, called proppant, gets carried into the fresh cracks. When the pumps stop and the pressure drops, the cracks try to close, but the sand grains hold them open like pebbles jammed in a door.

6. Plug and repeat. A plug seals off that section, and the crew moves back toward the surface and does it again. A long lateral is fracked in dozens of these stages, one after another.

7. Drill out and flow. The plugs are drilled out, some of the water flows back up, and oil and gas follow it.

One well can use anywhere from 1.5 million to 16 million gallons of water, according to the US Geological Survey. The low end fills about two Olympic swimming pools, and the high end fills more than twenty.

Why do fracked wells decline so fast?

A fracked well drains only the rock near its cracks. Output starts high and then falls steeply, and a shale well commonly loses more than half its production in the first year. So shale is a treadmill: to hold production flat, companies have to keep drilling new wells. That is why US output reacts to oil prices within months, while a giant conventional field takes a decade to build.

The treadmill has gotten faster. The Permian completes roughly 6,000 wells a year, about the same as in 2022. Yet its output grew from 2.9 million barrels of oil equivalent per day in 2015 to 11.2 million in 2025, because each well is longer and better placed.

Subscribe now

Who invented fracking?

Nobody invented it in one go. Engineers were cracking rock with fluid in the 1940s, and horizontal drilling existed for decades. The breakthrough was stubbornness. George P. Mitchell’s company, Mitchell Energy, drilled its first Barnett Shale well near Fort Worth, Texas, in 1981. It spent almost two decades tweaking the process on rock the rest of the industry laughed at. In 1997–98 its engineers switched from thick gel fluids to cheap “slickwater”, which cut treatment costs to about a fifth. Pair that with horizontal wells in the 2000s and the shale boom was on. Mitchell Energy was sold to Devon Energy in 2002, and the playbook spread to the Marcellus, the Bakken, the Eagle Ford and the Permian.

Does fracking cause earthquakes or poison water?

This part is uncomfortable, so here are the plain facts.

• Earthquakes: Mostly not from fracking itself. The USGS says most induced earthquakes come from injecting wastewater deep underground for disposal, not from the frac. In Oklahoma, the US state with the most induced quakes, about 2% are linked directly to fracking. The largest known fracking-induced quake was magnitude 4.0, in Texas in 2018. Wastewater disposal, though, is part of the same industry, and it has caused real damage.

• Water: The frac itself happens a mile or more below drinking-water aquifers. The documented problems usually come from failed casing or cement, surface spills, and handling the salty, chemical-laden water that flows back up. That is a construction and operations problem, not a physics one, which means it’s fixable but only if it’s enforced.

• Methane: Leaks from wells and pipelines erode the climate benefit of gas over coal. This is where regulation matters most.

Why did fracking succeed in the US but not in China?

This is the interesting part. On paper, China should be winning. The US EIA’s resource estimates gave China about 1,115 trillion cubic feet of technically recoverable shale gas, nearly double the US figure. Beijing wanted the energy security badly. Yet China produced only about 0.9 trillion cubic feet of shale gas in 2024, against total gas use of about 15 trillion. On the oil side, China’s shale oil passed 8.5 million tonnes in 2025, roughly 170,000 barrels a day. US tight oil ran at about 8.3 million barrels a day in 2023, around fifty times more.

Four reasons explain the gap.

1. The rock is harder. China’s best gas shales, in the Sichuan Basin, are thinner, older, deeper and more faulted than the big US plays. Chinese operators are drilling to 3,500–4,500 metres, roughly twice the depth of a typical Marcellus well. Deeper means hotter, higher pressure and more expensive wells. The shale sits under mountains and farmland instead of west Texas scrub.

2. The landowner gets nothing. This is the big one. In the United States, a private landowner can own the minerals under their land. An oil company has to lease them, paying a signing bonus and a royalty, often 12.5% to 25% of what the well produces. A Texas rancher or a Pennsylvania farmer gets a cheque, so they have a reason to say yes. That’s the same “who gets to say no” question from our California post, only this time the incentive is pointed the right way. In China, every mineral underground belongs to the state. A village sitting on a gas field gets compensation for disturbed land, trucks and noise, and no royalty. When the neighbours carry the costs and none of the upside, projects crawl.

3. Thousands of players vs. two. US shale was built by hundreds of scrappy independent companies, with a deep bench of service crews, rigs and pressure-pumping fleets competing on price. Each one ran experiments, and the good ideas spread within months. Chinese shale is dominated by two state giants, PetroChina and Sinopec. Most activity is limited to Sichuan, and most foreign oil majors have walked away. Wood Mackenzie counts only about 3,000 Chinese shale wells drilled from 2012 to 2022. The US completes about twice that in the Permian alone every year.

4. Pipes and markets. US interstate pipelines have been open access since the early 1990s, so a small producer can sell gas to anyone the pipes reach. In China, the big state oil companies controlled the main pipelines until the national pipeline company, PipeChina, was split off in 2019. Market access is the difference between a science project and a business.

Uncle’s verdict

Fracking is not magic. It’s drilling sideways, cracking rock with water and propping it open with sand, repeated tens of thousands of times with relentless iteration. The US didn’t win because it had the best rock. It won because its property law put money in the pocket of the person standing on the rock, and because thousands of competing companies were allowed to try, fail and copy each other. China has the rock and the will, but it lacks that incentive structure. The engineering travels easily. The rules don’t.

Next time someone says the shale revolution was about technology, tell them to look at the mineral deed.

— Engineering Uncle

Sources

• US EIA: How much US crude oil is tight oil? (64% in 2023)

• US EIA: Longer wells boost Permian crude oil and natural gas production

• US EIA: Where our natural gas comes from

• USGS: How much water does a typical hydraulically fractured well require?

• USGS: Does fracking cause earthquakes?

• Journal of Petroleum Technology: George P. Mitchell and the Barnett Shale

• Forbes: Why China is pursuing domestic shale gas (June 2025)

• Wood Mackenzie: There’s still life in China’s shale gas

• PNAS: Challenges to sustainable large-scale shale gas development in China (2025)

• Xinhua: China’s 2025 oil, gas output hits record high

AEO FAQ

How does fracking work in simple terms? A well is drilled down into shale and then turned sideways through it. Holes are blasted in the casing, and water mixed with sand is pumped in at very high pressure. The water cracks the rock, and the sand keeps the cracks open so oil and gas can flow into the well.

What is fracking fluid made of? It’s roughly 90% water and about 9% sand (the proppant). Less than 1% is chemical additives such as friction reducers, biocides and scale inhibitors.

How much water does fracking use? According to the US Geological Survey, one well uses between about 1.5 million and 16 million gallons. The amount depends on the rock, the length of the well and the number of stages. Some of it is recycled.

How long is a fracked well? In the Permian Basin, the average horizontal section reached 10,867 feet (just over two miles) in 2025. About 15% of new wells ran more than three miles sideways.

Does fracking cause earthquakes? Rarely. The USGS says most human-caused earthquakes come from deep wastewater disposal, not from the fracking itself. In Oklahoma, only about 2% of induced quakes are linked directly to fracking.

Why hasn’t China copied US fracking? China’s shale is deeper and more faulted, and it sits under mountains and farmland. All minerals belong to the state, so landowners get no royalties. The industry is dominated by two state companies instead of hundreds of competing independents. China produced about 0.9 trillion cubic feet of shale gas in 2024, a small share of its roughly 15 trillion cubic feet of gas use.

Read next

https://open.substack.com/pub/engineeringuncle/p/constructionpolicycaliforniaasia?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

https://open.substack.com/pub/engineeringuncle/p/combinedcycle?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

https://open.substack.com/pub/engineeringuncle/p/fuelcellexplained?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

https://open.substack.com/pub/engineeringuncle/p/aidatacenterwater?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Friday 23 Oct: How Do Offshore Oil Platforms Work? Jackets, Spars, TLPs and FPSOs.

Caption: Artist’s Concepts Depict SpaceX’s Starship HLS Test Article for NASA’s Artemis III

TL;DR

• Starship is a two-stage rocket about 124 metres tall: a Super Heavy booster with 33 Raptor engines and a Starship upper stage with 6. Both are designed to be fully reusable.

• It burns liquid methane and liquid oxygen. The booster alone carries about 3,250 tonnes of propellant, and its engines produce more than twice the thrust of the Saturn V moon rocket.

• The booster flies back and is caught by the launch tower’s giant arms (”Mechazilla”), first done in October 2024. The ship re-enters behind about 18,000 heat-shield tiles.

• The third version (V3), with Raptor 3 engines, is designed to lift more than 100 tonnes to orbit reusably. Its July 2026 test flight reached space and deployed 20 Starlink satellites, but the booster had engine problems coming down.

• For NASA’s moon landing, a lander Starship must be refuelled in orbit by about ten tanker flights. The first propellant transfer demonstration is planned for late 2026 at the earliest, and SpaceX is more than a year behind its original schedule.

Listen up.

Short answer first: Starship works like any rocket, burning fuel to throw itself upward, but every piece is designed to come back and fly again. A giant booster lifts it off the pad, separates and flies back to be caught by the tower. The upper stage, the ship itself, reaches orbit, does its job, then falls back through the atmosphere belly-first and lands. Refuel, stack, fly again. That’s the whole idea.

The reason nobody did this before isn’t that nobody thought of it. It’s that every part of it is brutally hard.

What are the two parts of Starship?

Starship is two rockets stacked on top of each other, about 124 metres tall in its current (V3) form, taller than the Statue of Liberty with its pedestal.

1. Super Heavy, the booster. The bottom stage. It has 33 Raptor engines and carries about 3,250 tonnes of propellant. At full power, its engines produce around 89 meganewtons of thrust, more than twice the Saturn V that took Apollo to the Moon. It does the heavy lifting for the first couple of minutes, then separates.

2. Starship, the ship. The top stage, which people also call Starship. It has 6 Raptor engines: three tuned for the atmosphere and three with large nozzles for the vacuum of space. It carries about 1,500 tonnes of propellant, plus the payload: satellites, cargo, or eventually people.

Why is Starship made of stainless steel?

Most rockets are made of aluminium or carbon fibre, which are light. Starship is stainless steel, which is heavier. It sounds backwards, but it makes sense:

• Steel gets stronger when it’s very cold, and the tanks hold propellant at around -180°C.

• Steel handles heat much better than aluminium or carbon composites, which matters on re-entry.

• Steel is cheap and easy to weld. SpaceX builds Starships in tents and hangars at Starbase, Texas, and can fix or change a design quickly.

When you plan to build a lot of rockets and crash some of them while learning, cheap and weldable beats light and precious.

Subscribe now

What are Raptor engines, and why methane?

Raptor is SpaceX’s engine. It burns liquid methane with liquid oxygen.

Methane was chosen for three reasons:

• It burns cleaner than kerosene, leaving little soot in the engine, which helps reuse.

• It’s cheap and easy to store at rocket temperatures.

• It can be made on Mars from carbon dioxide and water (the Sabatier reaction), which matters if you ever want to fly home.

Raptor uses a very efficient but complicated design called full-flow staged combustion, running at extremely high pressure. The latest version, Raptor 3, removes many external pipes and parts to make the engine simpler, lighter and more reliable, and it’s what powers Starship V3.

How does the launch tower catch the booster?

This is the part that makes people’s jaws drop.

After separating, Super Heavy flips around, relights some of its engines to slow down and fly back to the launch site, then hovers beside the tower. Two huge arms on the tower, nicknamed “Mechazilla” or the chopsticks, close around it and catch it by small pins near the top.

Why not land on legs like Falcon 9? Legs are heavy, and every kilogram of legs is a kilogram less payload. Catching the booster on the tower also puts it right back where it needs to be for the next launch.

The first catch happened on 13 October 2024. It has been repeated since. For the first V3 flights in 2026, SpaceX chose to splash the booster down in the Gulf of Mexico instead, to be safe with the new design.

How does Starship survive re-entry?

Coming back from orbit, the ship hits the atmosphere at about 27,000 km/h. The air in front of it turns into glowing plasma.

Starship falls belly-first, like a skydiver, using its broad side as a brake. The belly is covered in about 18,000 black hexagonal heat-shield tiles, each able to handle around 1,400°C. Four flaps, two near the nose and two near the tail, steer it through the fall.

Just before landing, it does the famous “belly flop” flip: relighting its engines, swinging vertical and slowing to land. Making the heat shield reusable without heavy repairs between flights is still one of the biggest unsolved problems.

Why does Starship need to refuel in orbit?

A Starship reaching orbit has used almost all its propellant. To go further, to the Moon or Mars, it needs to fill up again in space.

The plan: launch tanker Starships that carry propellant instead of cargo, and transfer it into a waiting ship in orbit, like aerial refuelling for planes, but with super-cold liquids that slowly boil away.

For NASA’s moon landing, SpaceX’s lunar lander version, Starship HLS, needs about ten tanker launches to fill up. The first ship-to-ship propellant transfer demonstration is planned for late 2026 at the earliest.

Subscribe now

Where is Starship in 2026?

Starship had flown 13 test flights by the end of July 2026: eight successes and five failures, including three failures in a row in 2025.

The new, bigger V3 first flew in 2026. Its July 2026 flight was mostly a success: all 33 booster engines lit, the ship reached space with all six engines running, deployed 20 Starlink satellites, restarted an engine in space, re-entered and splashed down intact near Australia. The weak spot was the booster: only 10 of 13 engines relit for its landing burn, and it hit the Gulf of Mexico harder than planned.

Next up: the first truly orbital flight, more booster catches with V3, and propellant transfer. NASA’s crewed Moon landing using Starship is now expected no earlier than 2028, and SpaceX is more than a year behind its original schedule for key milestones.

Why does it matter so much?

Because of cost.

Today a fully expended Starship is estimated to cost around $100 million a launch. If the whole thing can be reused many times, like a plane, the cost per flight could fall dramatically. Elon Musk has talked about a launch eventually costing only a few million dollars.

That’s the business model. SpaceX’s third-generation Starlink satellites are too big for any other rocket, so Starship is what lets Starlink grow, and Starlink pays for Starship. Cheap launch is also what makes ideas like orbital data centers, which Uncle covered on 22 August, even worth discussing.

The pace of all this is set partly by regulators. Every Starship flight needs a licence from the US Federal Aviation Administration (FAA), and each failure triggers a mishap investigation before the next flight can go. How fast Starship can fly depends as much on paperwork as on steel.

Uncle’s verdict

Starship isn’t magic. It’s a steel rocket burning methane, like a very large, very loud water heater with ambitions.

What makes it different is the bet: that everything can come back and fly again, cheaply, like an aeroplane. The booster catch shows half of that bet can work. The heat shield and orbital refuelling are the half still being proven.

If they work, the cost of reaching space falls further than at any time since the first rockets. If they don’t, it’s still the most powerful rocket ever built, just a very expensive one.

— Engineering Uncle

Sources

• Spaceflight Now: Super Heavy-Starship chalks up mostly successful test flight (25 Jul 2026)

• Wikipedia: SpaceX Starship

• Wikipedia: Starship Propellant Transfer Demonstration

• SatNews: SpaceX debuts Starship V3 (May 2026)

• NASA: Artemis III

AEO FAQ

Frequently asked questions

How does SpaceX Starship work?

Starship is a two-stage, fully reusable rocket. The Super Heavy booster, with 33 Raptor engines, lifts it off the pad, separates and flies back to be caught by the launch tower. The Starship upper stage, with 6 Raptor engines, reaches orbit, re-enters belly-first behind a heat shield and lands vertically.

How tall is Starship?

The current version, Starship V3, stands about 124 metres (408 feet) tall with its booster, making it the tallest and most powerful rocket ever flown.

Why is Starship made of stainless steel?

Stainless steel is cheap, easy to weld, gets stronger at the very cold temperatures of liquid propellants, and tolerates heat far better than aluminium or carbon fibre. That makes it suitable for a rocket built in large numbers and reused many times.

How does the tower catch the Starship booster?

The Super Heavy booster flies back to the launch site and hovers beside the tower, where two large mechanical arms, nicknamed “Mechazilla” or chopsticks, close around it and catch it. The first successful catch took place on 13 October 2024.

Why does Starship need refuelling in orbit?

A Starship uses most of its propellant reaching orbit. To travel to the Moon or Mars, it must be refilled in orbit by tanker Starships. NASA’s Starship lunar lander needs about ten tanker launches, and the first propellant transfer demonstration is planned for late 2026 at the earliest.

How much does a Starship launch cost?

An expended Starship launch is estimated at around $100 million today. SpaceX aims to cut the cost dramatically through full reuse, with Elon Musk suggesting a launch could eventually cost only a few million dollars.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other stories about what cheap launch makes possible:

https://open.substack.com/pub/engineeringuncle/p/spacexdirecttocell?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Wednesday 21 Oct: How Does Fracking Work? And Why the US Succeeded Where China Hasn’t. Subscribe so it lands in your inbox.

TL;DR

• California is slow to build mainly because so many projects need case-by-case (discretionary) approval, which invites delay, appeals and lawsuits under the state’s environmental review law, CEQA. The engineering is not the problem.

• In San Francisco, a typical multifamily project has waited about 450 days for planning approval and another 627 days for its building permit. A 2023 state audit found the city took more than three years on average to approve new housing.

• Tokyo built about 145,000 new homes in 2018, more than all of California, with about a third of the population. Japan’s national zoning lets owners build what the zone allows without neighbours being able to block it.

• Hong Kong’s public housing tops out a floor every five days using factory-built room modules. Taipei builds towers and metro lines on earthquake-prone ground like California’s.

• California is changing: faster “by-right” approvals under SB 35 cut approval times by roughly 60% in Los Angeles and San Francisco, and June 2025 reforms (AB 130 and SB 131) exempted many infill housing projects from CEQA.

Listen up.

Short answer first: California can’t build quickly because almost every project has to ask permission, one at a time, and almost anyone can say no. In Tokyo, if the zoning allows your building, you build it. In Hong Kong, the government owns the land and decides in advance what goes where. In California, projects go through case-by-case review, public hearings and, often, lawsuits. The concrete and steel are the same. The system around them is not.

Uncle writes this from Taipei, which sits on active faults just like Los Angeles. So let’s get the most common excuse out of the way first.

Subscribe now

Is it earthquakes?

No. California’s earthquake codes are strict, but so are Japan’s and Taiwan’s, and both keep building.

Japan has some of the toughest seismic rules in the world, and Tokyo still builds more housing every year than the whole state of California. Taiwan tightened its building codes after the 1999 Chi-Chi earthquake, which killed more than 2,400 people, and Taipei has kept adding high-rises and metro lines since.

Earthquakes make buildings more expensive: deeper foundations, more steel, dampers and base isolation (Uncle covered sliding “shoes” on 22 September). They don’t explain waiting years for a permit.

Why is it so hard to build housing in California?

Four things, and all of them happen before construction.

1. Discretionary approval. In much of California, a project that meets the rules can still be sent through case-by-case review, where planning commissions and councils can add conditions, demand changes or say no. Every step is a chance for delay.

2. CEQA. The California Environmental Quality Act requires environmental review for many projects. It was meant to protect nature, but anyone can sue under it, including neighbours who simply don’t want new housing. Even the threat of a lawsuit adds months and cost.

3. Local control. Hundreds of cities set their own zoning, fees and processes. Much of the land in California’s big cities is zoned only for single-family homes.

4. Slow permits after approval. Even a fully approved project then waits for building permits. In San Francisco, a typical multifamily project has taken about 450 days for planning approval and another 627 days for its building permit, according to data compiled by the San Francisco Chronicle. A 2023 state audit found the city took more than three years on average to approve new housing.

None of those four is an engineering problem. They are process problems.

Subscribe now

How does Tokyo build so much?

Tokyo is the favourite example for housing economists, and for good reason. In 2018, the Tokyo metropolitan area (the prefecture) built about 145,000 new homes, more than all of California, with roughly a third of the population. Rents and prices have stayed remarkably stable for decades.

The reason is how Japan zones land:

• Zoning is set nationally, with 12 zone types for the whole country. Cities choose which zone applies where, but they can’t invent their own rules.

• Building is “as-of-right”. If your plan fits the zone, you get the permit. As one Tokyo planner put it, neighbours basically have no right to stop development.

• Buildings aren’t treated as permanent. Houses are routinely rebuilt every few decades, so the city keeps renewing and densifying.

No neighbourhood veto, no lawsuit-by-default, no case-by-case negotiation.

How does Hong Kong build so fast?

Hong Kong is different again. The government owns almost all the land and sells long leases with the rules largely set before the sale. It also builds a huge share of housing itself: roughly half of residents live in public or subsidised housing.

On site, it is a factory:

• Repetition. Estates are towers of 40 near-identical floors. The crew gets faster with every floor.

• Modular Integrated Construction (MiC). Whole rooms, already fitted out, are built in factories and lifted into place. According to a May 2025 Legislative Council paper, Hong Kong’s public housing reached one floor every five days this way, and is aiming for four.

• Rail plus Property. The rail operator, MTR, builds homes and malls above new stations, and the property profits help pay for the railway. The builder makes more money by opening the line early.

What about Taipei?

Taipei is the closest match to California: a wealthy, democratic place on active faults, with strong seismic codes. It builds dense mid- and high-rise housing and keeps expanding its metro.

It isn’t perfect. Taipei has its own housing affordability problem, with prices high relative to local incomes. But the city’s default is dense, mixed-use building near transit, not single-family zoning with case-by-case review. New towers and metro works are a constant part of the Taipei skyline.

Is building fast always good?

No, and Hong Kong shows why.

In November 2025, a fire at Wang Fuk Court (Wang Fuk Yuen) in Tai Po killed 168 people, Hong Kong’s deadliest fire in decades. It wasn’t a new building. The estate’s eight towers were wrapped in scaffolding and green protective netting for a HK$330 million exterior renovation. Tests later found that some of that netting failed fire-safety standards, reportedly with compliant netting hung only where inspectors took samples, and flammable foam boards had been fixed over the windows. The fire raced up the netting and jumped from tower to tower.

The lesson isn’t that checks slow things down. It’s that the checks have to be real. California doesn’t need fewer inspections. It needs faster paperwork.

Is California getting faster?

Yes, slowly, and the results are measurable.

• SB 35 (2017) lets qualifying housing projects in cities that miss their housing targets get ministerial (”by-right”) approval, closer to the Japanese model. A Pew-sponsored NYU Furman Center study found median approval time fell to about 2.7 months in Los Angeles (from about 7) and about 4 months in San Francisco (from more than a year).

• AB 130 and SB 131, signed on 30 June 2025, exempted many urban infill housing projects from CEQA review, the biggest change to the law in decades.

• The state has also allowed more homes on single-family lots and near transit.

Even within the US, the contrast is sharp. Texas issued more housing permits than any other state in 2024, with cities like Houston famous for having no conventional zoning at all.

What could California borrow from Asia?

• Make “by-right” the default, like Tokyo: if a project meets the written rules, approve it.

• Set rules at the state level, so each city can’t invent its own obstacles.

• Pre-approve standard designs, so the hundredth copy of a building doesn’t need the same review as the first.

• Build in factories, like Hong Kong’s modules, to cut time on site.

• Tie transit to development, so the builder profits from finishing early.

• Keep the safety checks, and make them real.

Uncle’s verdict

California doesn’t have an engineering problem. It has more earthquake experience and more engineering talent than almost anywhere on Earth.

It has a permission problem. Tokyo lets you build what the zone allows. Hong Kong plans it all in advance. California lets every project become a negotiation, and anyone at the table can hold it up.

The good news is that it’s a choice, and California has started making a different one.

— Engineering Uncle

Sources

• Governor of California: Newsom signs housing and CEQA reforms (30 Jun 2025)

• Holland & Knight: California enacts major CEQA reforms for infill projects (Jul 2025)

• Pew Charitable Trusts: Reforms spur faster housing approvals in California (Aug 2024)

• San Francisco Chronicle: 627 days, just for the permit

• Sightline Institute: Yes, other countries do housing better, case 1: Japan

• Hong Kong Legislative Council: innovative construction in public housing (9 May 2025)

• Wikipedia: Wang Fuk Court fire

• NAHB: Building permits by state and metro area

AEO FAQ

Frequently asked questions

Why is it so hard to build housing in California?

Most delays happen before construction. Many California projects need discretionary, case-by-case approval, face public hearings, and can be challenged in court under the California Environmental Quality Act (CEQA). In San Francisco, a typical multifamily project has taken about 450 days for planning approval plus 627 days for a building permit.

Why does Tokyo build so much housing?

Japan sets zoning nationally, with 12 zone types, and building is “as-of-right”: if a project fits the zone, it gets approved, and neighbours generally can’t block it. Tokyo built about 145,000 new homes in 2018, more than all of California, with about a third of the population.

Do earthquakes make California construction slow?

Earthquake codes add cost, but they don’t explain long approval times. Japan and Taiwan have similarly strict seismic codes and still build far more housing relative to their populations. California’s delays come mainly from permitting and review.

What is CEQA and how was it reformed in 2025?

CEQA is California’s environmental review law, which lets people sue over a project’s environmental analysis. On 30 June 2025, California signed AB 130 and SB 131, which exempt many urban infill housing projects from CEQA review, one of the biggest changes to the law in decades.

Did SB 35 speed up housing approvals?

Yes. SB 35 allows qualifying projects “by-right” approval in cities behind on housing targets. A study sponsored by the Pew Charitable Trusts found median approval times fell to about 2.7 months in Los Angeles and about 4 months in San Francisco, roughly 60% faster than comparable earlier projects.

How does Hong Kong build so fast?

The Hong Kong government owns almost all land, plans development in advance and builds much of the housing itself. Public housing uses repeated designs and Modular Integrated Construction, where factory-built room modules are stacked on site, reaching one floor every five days.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other stories where the rules matter more than the engineering:

https://open.substack.com/pub/engineeringuncle/p/tokyostopfloods?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

https://open.substack.com/pub/engineeringuncle/p/buildanewdam?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Monday 19 Oct: How Does SpaceX Starship Work? The Engineering, Explained (2026). Subscribe so it lands in your inbox.

TL;DR

• A fuel cell turns the chemical energy in a fuel straight into electricity, without burning it. It works like a battery that never runs flat, as long as fuel and air keep flowing in.

• Hydrogen fuel cells (PEM type) run at about 80°C, need platinum and produce only water. They power cars like the Toyota Mirai, buses and forklifts.

• Solid oxide fuel cells (SOFC), like Bloom Energy’s, run at several hundred to about 1,000°C on a ceramic electrolyte. They can use natural gas directly and need no platinum.

• A Bloom Energy Server 6.5 makes 325 kW in a unit about 9 m long, at 53–65% electrical efficiency, with almost no NOx and no water use, but still emits about 308–378 kg of CO₂ per MWh on natural gas.

• AI data centers want them because the grid can take years to connect. Brookfield expanded its financing partnership with Bloom from $5 billion to $25 billion in June 2026, and AEP has secured up to 1 GW of Bloom fuel cells.

Listen up.

Short answer first: a fuel cell makes electricity by letting a fuel and oxygen react on opposite sides of a thin barrier, instead of burning them together. The barrier lets only charged particles (ions) through, so the electrons are forced to go the long way, through a wire. That flow of electrons is your electricity. As long as fuel and air keep coming, it keeps going.

Think of it as a battery that you refuel instead of recharge.

Subscribe now

How does a fuel cell work?

Every fuel cell has the same three layers:

1. The anode, where the fuel goes in.

2. The electrolyte, a thin layer in the middle that lets certain ions through but blocks electrons.

3. The cathode, where air (oxygen) goes in.

Here is the trick. The fuel wants to react with oxygen, just as it would in a flame. But the electrolyte sits in the way. Only ions can cross it. The electrons that the reaction releases have to travel around the outside, through a wire and whatever you plug in, before they can finish the job on the other side.

No flame. No turbine. No moving parts. Just chemistry and a wire.

One cell makes less than a volt, so real fuel cells stack hundreds of thin cells together, like slices in a loaf, to reach useful power.

What is the difference between a hydrogen fuel cell and a solid oxide fuel cell?

There are several types, but two matter most today.

Hydrogen fuel cells (PEM, proton exchange membrane)

• The electrolyte is a thin plastic membrane that lets hydrogen ions (protons) through.

• They run warm, about 80°C, and start in seconds.

• They need platinum as a catalyst, which is expensive.

• They need fairly pure hydrogen as fuel, and their only exhaust is water.

• You’ll find them in hydrogen cars such as the Toyota Mirai and Hyundai Nexo, in buses and trucks, and in warehouse forklifts.

Solid oxide fuel cells (SOFC)

• The electrolyte is a thin ceramic plate that lets oxygen ions through, but only when it’s very hot.

• They run at several hundred degrees up to around 1,000°C.

• At that heat, they need no platinum, just cheaper nickel and ceramic materials.

• The heat lets them use natural gas directly: the gas is split into hydrogen and carbon monoxide inside the unit, then used as fuel. They can also run on biogas or hydrogen.

• The catch: they are slow to heat up and dislike being switched on and off, so they’re best running steadily, 24 hours a day.

That last point is exactly why solid oxide suits a data center, which also runs steadily, 24 hours a day.

How do Bloom Energy’s fuel cells work?

Bloom Energy is the best-known maker of solid oxide fuel cells. Its product, the Bloom Energy Server, looks like a row of steel cabinets. Inside are stacks of small, thin ceramic cells running at high temperature.

The current model, the Energy Server 6.5, according to Bloom’s February 2026 data sheet:

• makes 325 kW per unit, in a box about 9 m long, 1.3 m wide and 2.5 m tall

• turns 53–65% of the fuel’s energy into electricity (higher when new, lower as the stacks age)

• emits almost no nitrogen oxides (0.003 lb per MWh), because nothing burns

• uses no water in normal operation

• releases exhaust above 350°C, which can be captured for heating, taking total efficiency above 90%

On natural gas it still emits about 679–833 lb of CO₂ per MWh, roughly 308–378 kg. That is similar to the best combined-cycle gas plants Uncle covered on 26 September. It’s cleaner air, not zero carbon, unless it runs on biogas or hydrogen.

Scale check: a 100 MW data center would need about 300 of these boxes lined up outside.

Subscribe now

Why are AI data centers buying fuel cells?

Because the grid is the bottleneck.

As Uncle covered on 15 September, connecting a big new load to the grid can take years in the US. An AI campus that is ready to switch on can sit waiting for power. Fuel cells offer a way around that queue:

• Speed. Bloom says it can deliver on-site power for an entire data center within 90 days of a site being ready. A new grid connection or big gas plant takes years; gas turbine orders are now backed up to around 2031.

• Modular. Need 10 MW more? Add 30 more boxes. There’s no giant single machine to wait for.

• Permits. With almost no NOx and no combustion, fuel cells are easier to permit in areas with strict air-quality rules. Bloom notes that about 46% of announced data center builds sit in such “non-attainment” zones.

• Reliability. Many independent stacks running in parallel mean one failing doesn’t take the whole site down.

The money followed. Oracle agreed in July 2025 to deploy Bloom fuel cells at some of its cloud data centers. AEP, one of the largest US utilities, secured up to 1 GW of Bloom fuel cells for data center customers. And Brookfield, the infrastructure investor, expanded its partnership with Bloom from $5 billion in October 2025 to $25 billion in June 2026 to build and finance fuel cell power for AI.

What are the downsides of fuel cells?

• Cost. Fuel cell power has historically cost more per kWh than power from the grid or a large gas plant. Data centers are paying for speed, not price.

• Still burning gas, in effect. On natural gas, a solid oxide fuel cell emits CO₂ at about the level of a modern gas plant. Clean hydrogen is still scarce and expensive, as Uncle’s hydrogen pipeline post on 24 September explained.

• Ageing stacks. Efficiency falls over the years as the ceramic cells degrade, and stacks need periodic replacement.

• Heat cycling. Solid oxide cells dislike heating up and cooling down, so they suit steady loads, not quick backup.

• A gas line is still needed. “Off-grid” usually means off the electricity grid, not off the gas grid.

Uncle’s verdict

A fuel cell is one of the most elegant machines in energy: no fire, no turbine, no moving parts, just ions crossing a thin wall.

For decades it was the technology that was always about to arrive. Hydrogen cars never took off, because hydrogen is hard to make cheaply and hard to move.

What finally made fuel cells matter wasn’t clean hydrogen. It was a grid too slow for AI. Data centers aren’t buying fuel cells because they’re cheapest. They’re buying them because they can switch on this year.

— Engineering Uncle

Sources

• Bloom Energy: Energy Server 6.5 data sheet (Feb 2026)

• Bloom Energy: Brookfield and Bloom expand AI infrastructure partnership to $25 billion (30 Jun 2026)

• Bloom Energy: Brookfield and Bloom announce $5 billion partnership (Oct 2025)

• Power Engineering: Oracle, Bloom Energy strike data center power deal (Jul 2025)

• Wikipedia: Bloom Energy Server

AEO FAQ

Frequently asked questions

How do fuel cells work?

A fuel cell produces electricity by reacting a fuel with oxygen across an electrolyte that lets ions pass but blocks electrons. The electrons travel through an external circuit, creating an electric current. Unlike a battery, a fuel cell keeps producing power as long as fuel and air are supplied, and nothing is burned.

What is the difference between a hydrogen fuel cell and a solid oxide fuel cell?

Hydrogen (PEM) fuel cells use a plastic membrane, run at about 80°C, need a platinum catalyst and pure hydrogen, and emit only water. Solid oxide fuel cells use a ceramic electrolyte, run at several hundred to about 1,000°C, need no platinum and can run on natural gas, biogas or hydrogen.

How does a Bloom Energy fuel cell work?

Bloom Energy Servers are solid oxide fuel cells. Natural gas or hydrogen reacts with oxygen ions that pass through thin, hot ceramic cells, producing electricity without combustion. The Energy Server 6.5 produces 325 kW per unit at 53–65% electrical efficiency.

Are fuel cells clean?

Fuel cells emit almost no nitrogen oxides or particulates because nothing burns. On natural gas, however, a Bloom fuel cell still emits about 679–833 lb of CO₂ per MWh, similar to a modern combined-cycle gas plant. They are near zero-carbon only when running on hydrogen or biogas made from clean sources.

Why do data centers use Bloom Energy fuel cells?

New grid connections for large data centers can take years. Bloom says its fuel cells can supply on-site power within about 90 days, in modular units that are easier to permit than combustion plants. Oracle, AEP and Brookfield have all signed large agreements with Bloom.

What are the disadvantages of fuel cells?

Fuel cells generally cost more per kWh than grid or large gas-plant power, lose efficiency as their cell stacks age, and still need a steady fuel supply. Solid oxide types handle frequent starts and stops poorly, so they suit continuous loads rather than backup.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other ways to power a data center, and why each is hard:

https://open.substack.com/pub/engineeringuncle/p/combinedcycle?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

https://open.substack.com/pub/engineeringuncle/p/aidatacenterwater?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Friday 16 Oct: Why Can’t California Build? What Hong Kong, Tokyo and Taipei Do Differently. Subscribe so it lands in your inbox.

TL;DR

• Starlink Direct to Cell puts a mobile phone tower on a satellite. More than 650 of these satellites talk to ordinary 4G phones using a slice of partner operators’ spectrum, starting with T-Mobile’s T-Satellite service in the US.

• The hard part is physics: a phone’s signal is weak and built to reach a tower a few kilometres away. A satellite hundreds of kilometres up receives it roughly 10,000 times weaker.

• SpaceX solves it with many smaller satellites. AST SpaceMobile (ASTS) solves it with a few enormous ones: its BlueBird 6 unfolds a 2,400 sq ft (about 223 m²) antenna and is designed for peak speeds of 120 Mbps.

• Today T-Satellite handles texting, location sharing and some app data at about 2 to 4 Mbps per user. SpaceX says its second-generation satellites, expected from mid-2027, will have about 20 times the throughput.

• The real limit is spectrum. In May 2026 the FCC approved SpaceX’s $17 billion purchase of EchoStar’s AWS-4 and H-block frequencies, which gives it spectrum of its own for the next generation.

• Subscribe now

Listen up.

Short answer first: Starlink Direct to Cell works by putting a cell tower in orbit. Special Starlink satellites carry the same kind of radio a phone tower uses, plus a large antenna. Your ordinary phone thinks it has found a normal tower, just one that happens to be moving at 27,000 km/h a few hundred kilometres overhead. AST SpaceMobile does the same job with far fewer, far bigger satellites.

No special phone. No dish. That’s the whole trick, and it’s harder than it sounds.

Why is it so hard for a phone to reach a satellite?

Your phone was designed to talk to a tower a few kilometres away. It transmits with well under one watt, through an antenna smaller than your thumb.

Now move the tower to space. A satellite a few hundred to 500 km up is about a hundred times farther away than a typical tower. Radio signals weaken with the square of distance, so the signal arrives about 10,000 times weaker. That’s before you add the fact that the satellite is racing across the sky, which shifts the frequency (the Doppler effect) and hands you from satellite to satellite every few minutes.

There are only two ways to hear a whisper from very far away:

1. Get closer, with lots of satellites in a lower orbit.

2. Build a bigger ear, with a huge antenna in space.

SpaceX mostly chose the first. AST SpaceMobile chose the second.

How does Starlink Direct to Cell work?

SpaceX flies a special version of its Starlink satellites carrying an eNodeB, the radio that sits in a normal 4G phone tower, and a phased-array antenna that steers its beams electronically.

1. The phone transmits as it always does, on a frequency its mobile operator owns.

2. The satellite picks that signal up with its large antenna, corrects for the Doppler shift and timing delay, and behaves like a tower.

3. Laser links pass the traffic between satellites to a ground station.

4. The ground station hands it to the operator’s normal network, which sends your text on its way.

Because each satellite is only moderately sized, SpaceX needs a lot of them. More than 650 direct-to-cell satellites are now in orbit, covering 22 countries.

In the US the service is T-Mobile’s T-Satellite, launched commercially in July 2025 for about $10 a month on top of a plan. It supports texts, picture messages and location sharing, and since October 2025 limited data for apps like WhatsApp and Google Maps, at roughly 2 to 4 Mbps per user. Voice calls are still being tested. Around 60 recent phone models work.

SpaceX calls today’s version a first generation. Its second-generation satellites, expected from mid-2027 and launched on Starship, will carry custom SpaceX chips, thousands of beams and about 20 times the throughput, aimed at something that feels like normal 5G.

How does AST SpaceMobile work differently?

AST SpaceMobile (ASTS) went for the biggest ear it could build.

Its first five commercial BlueBird satellites, launched in September 2024, each unfold an antenna of about 693 sq ft. Its next-generation BlueBird 6, launched in December 2025, unfolded an array of about 2,400 square feet (about 223 m²), the largest commercial communications antenna ever deployed in low Earth orbit. That’s roughly the floor area of a big family house, folded up for launch.

A bigger antenna hears weaker signals and can focus tighter beams. So each BlueBird can deliver far more per user: AST says BlueBird 6 is designed for peak speeds of up to 120 Mbps, enough for video, straight to an ordinary phone.

The trade-off is numbers and launches. AST aimed to have 45 to 60 satellites up by the end of 2026, saying about 45 would give basic continuous service in key markets. Industry analysts have doubted that pace, because launch slots on SpaceX and Blue Origin rockets are scarce. Its partners include AT&T, Verizon and Vodafone, among more than 50 operators. AT&T planned a beta service for consumers and FirstNet emergency responders in 2026.

Subscribe now

Which approach is better?

Neither, yet. They are betting on different constraints.

• Starlink has scale now: hundreds of satellites, a live commercial service and its own rockets. But each satellite has a small antenna, so capacity per user is low until the second generation.

• AST has capacity per satellite: fewer, huge satellites that promise real broadband. But it needs dozens more in orbit before coverage is continuous, and it depends partly on someone else’s rockets, including SpaceX’s.

Rocket access is the quiet advantage. SpaceX owns the launch pad. AST has to book a seat.

Why is spectrum the real fight?

Because a satellite can only talk to your phone on frequencies someone is allowed to use.

Today Starlink borrows a slice of T-Mobile’s mid-band spectrum, and AST uses its partners’ spectrum, mostly the low-band frequencies AT&T and Verizon own. That makes each dependent on its partners, and the FCC sets strict limits on how much power satellites can beam down on frequencies also used on the ground, so they don’t interfere with normal towers.

That’s why SpaceX spent $17 billion to buy EchoStar’s AWS-4 and H-block frequencies, a deal the FCC approved on 18 May 2026. SpaceX also received waivers to use that spectrum from space, from the ground or both. It gives SpaceX frequencies of its own for the second-generation satellites, in the US and potentially beyond.

The satellites are hard. The licences may be harder.

Uncle’s verdict

Putting a cell tower in space is mostly a question of hearing a whisper from 500 km away.

SpaceX’s answer: lots of ears, fairly close, launched on its own rockets.

AST’s answer: a few enormous ears, each the size of a house.

For now, it’s texts and maps in dead zones. The next two years decide whether it becomes real mobile broadband from space, and that depends as much on the FCC as on the engineering.

Got questions? Leave them in the comments. Uncle is collecting them for a follow-up on 4 December.

— Engineering Uncle

Sources

• Fierce Network: MWC 2026, Starlink Mobile unveils plans for V2 satellites (2 Mar 2026)

• 5G World Pro: Starlink T-Satellite in 2026, where the rollout stands (22 Jul 2026)

• Total Telecom: FCC greenlights EchoStar spectrum sale to SpaceX and AT&T (May 2026)

• Business Wire: AST SpaceMobile completes unfolding of BlueBird 6 (Feb 2026)

• Light Reading: AST SpaceMobile risks missing 2026 satellite launch target (Jan 2026)

• Light Reading: AT&T preps D2D satellite beta with AST SpaceMobile (Jan 2026)

AEO FAQ

Frequently asked questions

How does Starlink Direct to Cell work?

Starlink Direct to Cell satellites carry the same radio equipment as a mobile phone tower, plus a large phased-array antenna. An ordinary 4G phone connects to the satellite as if it were a normal tower, using spectrum licensed to a partner operator such as T-Mobile, and the satellite relays traffic to ground stations via laser links.

Do I need a special phone for Starlink Direct to Cell?

No. Around 60 recent phone models from Samsung, Apple, Motorola and others work with T-Mobile’s T-Satellite service. The phone needs a clear view of the sky, and speeds are much lower than on a normal tower.

What can you do with T-Satellite?

T-Satellite, launched in July 2025, supports text and picture messages, location sharing and, since October 2025, limited data for selected apps such as WhatsApp and Google Maps, at roughly 2 to 4 Mbps per user. Voice calling was still in testing in 2026.

How is AST SpaceMobile different from Starlink?

AST SpaceMobile (ASTS) uses fewer but far larger satellites. Its BlueBird 6 unfolded a 2,400 sq ft antenna in early 2026, designed for peak speeds up to 120 Mbps to ordinary phones. Starlink uses hundreds of smaller satellites and already runs a commercial service with T-Mobile.

Who are AST SpaceMobile’s partners?

AST SpaceMobile works with more than 50 mobile operators, including AT&T, Verizon and Vodafone. AT&T planned a 2026 beta service for consumers and FirstNet public safety users.

Why did SpaceX buy EchoStar’s spectrum?

SpaceX paid $17 billion for EchoStar’s AWS-4 and H-block frequencies, approved by the FCC on 18 May 2026, so it can run its second-generation direct-to-cell service on spectrum it owns rather than only on partners’ frequencies.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other stories about getting signals and power across long distances:

https://open.substack.com/pub/engineeringuncle/p/tsmcfoundry?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Wednesday 14 Oct: How Do Fuel Cells Work? Bloom Energy’s Solid-Oxide Boxes Explained. Subscribe so it lands in your inbox.

TL;DR

• TSMC (Taiwan Semiconductor Manufacturing Company) makes chips by building up to about 100 patterned layers on 300 mm silicon wafers, using a cycle of deposition, lithography, etching, doping and polishing repeated hundreds of times.

• An advanced chip can take 600 to 1,000 or more process steps and roughly three to four months to go from bare wafer to finished chip.

• TSMC is a foundry: it doesn’t design chips, it manufactures them for Apple, Nvidia, AMD, Qualcomm and others. In Q2 2026 it held a record 72.5% of global foundry revenue, according to TrendForce.

• Its newest process, 2 nanometre (N2), entered volume production in late 2025 and first contributed revenue in Q2 2026.

• TSMC is spending $165 billion on six fabs in Arizona, but chips made there have been reported to cost at least 50% more than chips made in Taiwan. A fab is people and suppliers as much as buildings.

Listen up.

Short answer first: TSMC makes chips by printing and carving patterns onto silicon wafers, layer by layer, up to about a hundred times. Each layer goes through the same loop: coat, print with light, etch, dope, polish, check. An advanced chip needs hundreds to over a thousand of these steps and about three to four months in the fab. TSMC doesn’t design the chips. It makes them for almost everyone else, better than anyone else.

Uncle writes this from Taipei, where TSMC is less a company than a national institution. Here’s what actually happens inside.

Subscribe now

What is a foundry?

Most famous chip companies don’t make their own chips. Apple, Nvidia, AMD and Qualcomm design them, then send the designs to a foundry to manufacture.

TSMC invented this model in 1987. It only makes chips; it never competes with its customers by selling its own. That trust is a big part of why nearly every leading-edge chip in the world now comes out of its fabs. In Q2 2026, TrendForce put TSMC at 72.5% of global foundry revenue. Samsung was second with 5.9%, and China’s SMIC third with 5.4%.

What does a chip actually start as?

Sand, eventually. More precisely, a wafer: an ultra-pure slice of crystalline silicon, 300 mm across (about the size of a dinner plate) and less than a millimetre thick. Wafer makers grow single silicon crystals, slice them and polish them flat to near-atomic smoothness before TSMC ever touches them.

One wafer holds dozens to hundreds of chips, depending on how big each chip is. A big AI GPU takes up so much space that only a few dozen fit.

How does TSMC make a chip, step by step?

The transistors are built first, at the bottom. Then layer after layer of copper wiring is built on top to connect billions of them. Every layer goes through roughly the same loop:

1. Deposit. Lay down an ultra-thin film of material (an insulator, a metal or silicon) across the wafer, sometimes just a few atoms thick.

2. Coat. Spin a light-sensitive chemical called photoresist over it, like a thin layer of film in a camera.

3. Print with light (lithography). Shine light through a patterned template called a mask, shrunk down by lenses or mirrors onto the wafer. For the finest layers TSMC uses extreme ultraviolet (EUV) machines from ASML, the $200-million-plus machines Uncle covered on 10 September. Less critical layers use older deep ultraviolet (DUV) machines.

4. Develop. Wash away the exposed (or unexposed) photoresist, leaving the pattern behind.

5. Etch. Use plasma or chemicals to cut the pattern into the material underneath, where the photoresist isn’t protecting it.

6. Dope. Fire charged atoms into the silicon (ion implantation) to change how it conducts electricity. This is how transistors get their “on” and “off” behaviour.

7. Polish. Grind the surface perfectly flat (chemical mechanical polishing) so the next layer sits level.

8. Clean and check. Wash off every trace of residue, then inspect for defects. A single speck of dust can kill a chip.

Then do it again. And again. Modern chips have up to about 100 layers, and each has to line up with the one below to within a few nanometres. An advanced process can involve 600 to 1,000 or more steps. At about a day per layer, a wafer spends roughly three to four months in the fab.

At the end, the wafer is tested, cut into individual chips and sent off for packaging: mounting on a base and wiring to the outside world. For AI chips, TSMC’s own advanced packaging (CoWoS) joins the processor to its memory, which gets its own Uncle deep dive in December.

Why does a fab need to be so clean?

Because the features are smaller than a virus. A leading-edge fab’s cleanroom meets ISO Class 1: no more than about 10 particles of 0.1–0.2 microns per cubic metre of air, which ASML describes as roughly 10,000 times cleaner than outside air.

Workers wear full “bunny suits”. Robots carry wafers between machines in sealed pods, so the wafers barely touch the cleanroom air at all. The air is filtered and pushed downward constantly. The floors sit on isolated foundations, because vibration from a truck outside can blur a pattern measured in nanometres.

What does “2 nanometre” actually mean?

Less than it sounds. Years ago the number matched a real transistor dimension. Today “3 nm” and “2 nm” are marketing names for a generation of process. Each new one packs more transistors into the same space and uses less power, but no feature on a 2 nm chip is literally 2 nm wide.

TSMC’s N2 (2 nm) process entered volume production at the end of 2025 and first contributed revenue in Q2 2026, according to TrendForce. It is the first TSMC node to use gate-all-around (GAA) nanosheet transistors, where the gate wraps around the channel on all sides for better control. That is the biggest change in transistor shape in over a decade.

Why is TSMC so far ahead?

Three reasons, and none of them is a secret machine. Everyone can buy the same ASML tools.

• Yield. The share of good chips per wafer is what decides profit. TSMC is famously good at getting new processes to high yield fast, and that know-how lives in thousands of engineers’ heads and decades of process data.

• Scale and customers. Apple, Nvidia and AMD all use TSMC. Every new customer’s chip teaches it something, and the volume pays for the next generation. Leading-edge fabs are widely reported to cost around $20 billion or more each.

• The cluster. Taiwan’s science parks in Hsinchu, Taichung, Tainan and Kaohsiung are packed with suppliers, chemical makers, equipment engineers and universities, all within a short drive. When a machine breaks at 3 a.m., the fix is nearby.

Why can’t TSMC just build the same fab in Arizona?

It is trying. TSMC has committed $165 billion to Arizona: six fabs, two advanced packaging plants and an R&D centre. The first fab started 4 nm production in early 2025. The second, for 3 nm, is aimed at the second half of 2027, and the third, for 2 nm and beyond, at around 2029. The US gave it $6.6 billion in CHIPS Act grants.

But copying a fab isn’t like copying a building:

• People. TSMC struggled to find enough experienced staff in Arizona and sent US hires to Taiwan for 12 to 18 months of training.

• Construction. US construction costs have been reported at four to five times Taiwan’s, with slower permits and more expensive labour.

• Suppliers. The chemical, gas and equipment ecosystem that surrounds Hsinchu has to be rebuilt nearby, or shipped in.

• Cost. Chips made in Arizona have been reported to cost at least 50% more than the same chips made in Taiwan.

That’s the policy lesson. The CHIPS Act can pay for the building. It can’t pay for 30 years of experience overnight.

Uncle’s verdict

A TSMC fab is a thousand steps, repeated with near-perfect accuracy, millions of times a year.

The machines matter. ASML’s EUV scanners matter. But everyone can buy those. What TSMC really sells is the ability to run that loop at high yield, at scale, before anyone else.

That’s why the world’s most advanced chips still come from a small island, and why moving even part of that to Arizona costs $165 billion and a decade.

— Engineering Uncle

Sources

• TrendForce: Global foundry revenue in Q2 2026 (9 Sep 2026)

• Wikipedia: TSMC Arizona

• Semiconductor Engineering: Battling fab cycle times

• ASML: How microchips are made

AEO FAQ

Frequently asked questions

How does TSMC make chips?

TSMC builds chips on 300 mm silicon wafers by repeating a cycle of deposition, photoresist coating, lithography, development, etching, ion implantation, polishing and inspection for each of up to about 100 layers. An advanced chip takes 600 to 1,000 or more steps and about three to four months in the fab.

What is a semiconductor foundry?

A foundry manufactures chips designed by other companies. TSMC, founded in 1987, pioneered the model and makes chips for Apple, Nvidia, AMD, Qualcomm and others. It held 72.5% of global foundry revenue in Q2 2026, according to TrendForce.

How long does it take TSMC to make a chip?

A leading-edge wafer typically spends about three to four months in the fab, because each of its roughly 80 to 100 layers takes around a day to process. Packaging and testing add more time afterwards.

What does 2 nanometre mean?

“2 nm” is a name for a process generation, not a literal feature size. TSMC’s N2 process, which entered volume production in late 2025, packs more transistors into the same area than 3 nm and is its first node to use gate-all-around nanosheet transistors.

Why is TSMC so far ahead of Samsung and Intel?

TSMC’s lead comes mainly from yield (getting a high share of working chips quickly on new processes), scale across many major customers, and a dense supplier and talent cluster in Taiwan. All leading chipmakers use similar ASML lithography machines.

Why are TSMC’s Arizona chips more expensive?

TSMC Arizona has faced higher construction costs, a shortage of experienced staff and a thinner local supplier network than in Taiwan. Chips made there have been reported to cost at least 50% more. TSMC plans $165 billion of investment across six Arizona fabs.

Subscribe now

Read next

Read next on EngineeringUncle

If this one clicked, these are the other pieces of the chip story:

https://open.substack.com/pub/engineeringuncle/p/aidatacenterwater?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Monday 12 Oct: How Does Starlink Direct to Cell Work? (And How AST SpaceMobile Differs). Subscribe so it lands in your inbox.

TL;DR

• Tokyo’s biggest flood defence is the Metropolitan Area Outer Underground Discharge Channel, nicknamed G-Cans: a 6.3 km tunnel about 10 metres wide and roughly 50 metres underground, in Kasukabe, Saitama, just north of Tokyo.

• When small rivers in a low, saucer-shaped basin overflow, water spills into five giant shafts, flows through the tunnel into a 177-metre-long “underground temple” held up by 59 pillars, and is pumped out into the much larger Edo River.

• Four pumps driven by jet-engine-derived gas turbines move up to 200 cubic metres of water per second, enough to empty a 25-metre school pool in about a second.

• It cost about 230 billion yen and took from 1993 to 2006 to build. It now runs about seven or eight times a year, and during Typhoon Hagibis in October 2019 it pumped out about 17 million cubic metres in roughly two days.

• Japan built it underground because there was no room above ground. Tokyo is now adding more: a 13.1 km tunnel reservoir under Loop Road 7 designed for rainfall of 75 mm an hour.

Listen up.

Short answer first: Tokyo stops floods by giving the water somewhere else to go, underground. The biggest example is G-Cans, officially the Metropolitan Area Outer Underground Discharge Channel. When rivers north of the city overflow, the excess water drops into five huge shafts, runs through a 6.3 km tunnel about 50 metres below the ground, and is pumped out into a larger river that can take it. The pumps can move 200 tonnes of water every second.

It looks like a temple. It works like a giant storm drain.

Subscribe now

Why does Tokyo flood in the first place?

Greater Tokyo is built on a flat river plain, and typhoons and summer downpours dump huge amounts of rain on it in a few hours.

The area north of the city, around Kasukabe in Saitama, is the worst. It sits in a low, saucer-shaped basin crossed by small rivers such as the Naka and Kuramatsu. Water drains in from all sides and has nowhere fast to go. As the farmland turned into suburbs, concrete and roofs replaced the paddy fields that used to soak up rain, so more water reached the rivers, faster.

The big rivers nearby, especially the Edo River, can carry far more water. The problem was getting the flood from the small rivers to the big one.

How does G-Cans work?

Think of it as five giant bathtub drains connected by one very large pipe.

1. The shafts. Five concrete shafts sit next to the small rivers. Each is about 65 metres deep and 32 metres wide, deep enough to stand the Statue of Liberty inside, without its pedestal. When a river rises past a set level, water spills over a weir into the shaft.

2. The tunnel. At the bottom, a tunnel about 10 metres across runs 6.3 km, roughly 50 metres below ground, linking the shafts together. The water flows along it towards the pump station.

3. The “underground temple”. Before the pumps is a huge chamber called the pressure-control tank: 177 metres long, 78 metres wide and 25.4 metres tall. Its roof is held up by 59 concrete pillars, each about 18 metres tall and weighing about 500 tonnes. The tank slows the rushing water and evens out the flow, so the pumps are not hammered by surges.

4. The pumps. Four giant pumps, each moving 50 cubic metres per second, lift the water out and into the Edo River. Together that is 200 cubic metres per second, enough to empty a 25-metre school swimming pool in about a second.

5. The engines. The pumps are driven by gas turbines derived from jet engines, the same family of machine as the combined-cycle plants Uncle covered on 26 September. Jet-type turbines start fast and pack huge power into a small space, which is what you want when a typhoon is arriving.

The whole system can hold about 670,000 cubic metres of water at once. At full pumping speed, it could empty in under an hour.

How well does it work?

It runs about seven or eight times a year on average. By mid-2021 it had been used for flood control 131 times since it first partly opened in 2002.

Its biggest test was Typhoon Hagibis in October 2019, one of the most destructive storms to hit eastern Japan in decades. G-Cans filled to about 90% of its capacity and pumped out about 17 million cubic metres of water in roughly two days, about 14 Tokyo Domes’ worth.

Flood damage in the basin, measured by homes and area flooded, has been reported to have fallen by about two-thirds since it opened. It doesn’t stop every flood. It takes the peak off, which is where most of the damage comes from.

Why did Japan build it underground?

Because there was no room above ground.

The normal way to stop a river flooding is to widen it, build higher levees, or dig a big retention basin to hold floodwater. All three need land. Around Tokyo, that land is covered in houses, roads and railways. Buying it would mean moving thousands of people, and it would take decades.

So Japan went down instead. Under public roads and land it already controlled, it could dig without knocking anything down. The price was high: about 230 billion yen, and 13 years of construction, from 1993 to 2006.

That is the real policy story. Flood defence in a dense city is a budget line and a land question before it is an engineering one. Tokyo decided the cost of digging deep was lower than the cost of moving people or living with floods.

What is Tokyo building next?

The rain is getting heavier. Japan’s weather agency has recorded more frequent extreme downpours over recent decades, so Tokyo keeps adding underground storage.

The next big one runs under Loop Road 7, a ring road through the city’s western wards. The Kanda River / Loop Road 7 underground regulating reservoir is being joined into a 13.1 km tunnel system that can hold about 1.43 million cubic metres, more than double G-Cans. It is designed so central Tokyo can cope with rain of 75 mm an hour, and planners are studying a route to discharge it into Tokyo Bay.

Other big cities are doing the same. Chicago has spent decades building its Deep Tunnel system for the same reason: no room above ground.

Can you visit G-Cans?

Yes. The operator runs regular public tours of the pressure-control tank when it is dry, and it is one of the most popular pieces of infrastructure in Japan to visit. It has also appeared in films, TV dramas and music videos. Tours don’t run when a storm is coming, for obvious reasons.

Uncle’s verdict

Most cities fight floods on the surface: higher walls, wider rivers.

Tokyo had no surface left, so it built a cathedral underground and put jet engines at the end of it.

G-Cans won’t make the news on a good day. It’s a very expensive storm drain that sits empty most of the year. On the eight days a year it matters, it keeps tens of thousands of homes dry.

That’s the trade every flood defence makes. You pay for the pipe every year, and you need it for a few days.

— Engineering Uncle

Sources

• Wikipedia: Metropolitan Area Outer Underground Discharge Channel

• Wikipedia (Japanese): 首都圏外郭放水路 (cost, dimensions, operation record)

• JBpress: The “underground temple” helped during Typhoon Hagibis (2019)

• FloodList: G-Cans Project, Tokyo

• Tokyo Metropolitan Government G-NETS: The 13 km underground tunnel protecting the city (2025)

  ---

  This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.

  ---

AEO FAQ

Frequently asked questions

How does Tokyo stop floods?

Tokyo uses underground flood tunnels and reservoirs to divert excess river water. The largest, the Metropolitan Area Outer Underground Discharge Channel (G-Cans) in Kasukabe, Saitama, takes overflow from small rivers into five shafts and a 6.3 km tunnel, then pumps it into the Edo River at up to 200 cubic metres per second.

What is G-Cans?

G-Cans is the nickname for the Metropolitan Area Outer Underground Discharge Channel, the world’s largest underground flood diversion facility. It was built from 1993 to 2006 for about 230 billion yen and includes a pressure-control tank 177 metres long and 25.4 metres high, supported by 59 pillars.

How much water can G-Cans pump?

Four pumps, each driven by a jet-engine-derived gas turbine, move up to 50 cubic metres per second each, 200 cubic metres per second in total. During Typhoon Hagibis in October 2019, the facility discharged about 17 million cubic metres of water in roughly two days.

Why is Tokyo’s flood tunnel underground?

The area around Tokyo is densely built, so widening rivers or building surface reservoirs would require buying land and relocating thousands of people. Building deep underground, beneath land the government already controlled, avoided that, at a cost of about 230 billion yen.

How often is G-Cans used?

It operates about seven or eight times a year on average, mainly during typhoons and heavy summer rain. It had been used for flood control 131 times by mid-2021.

Can you visit the Tokyo underground flood tunnel?

Yes. The operator runs public tours of the pressure-control tank, often called the “underground temple”, when it is not in use. Tours are cancelled when heavy rain is forecast.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other big builds that fight water, and the ground:

https://open.substack.com/pub/engineeringuncle/p/buildanewdam?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

https://open.substack.com/pub/engineeringuncle/p/combinedcycle?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Friday 9 Oct: How Does TSMC Make Chips? Inside a Leading-Edge Foundry. Subscribe so it lands in your inbox.

TL;DR

• A circuit breaker has to kill the electric arc that forms when its contacts open. Alternating current (AC) passes through zero 100 or 120 times a second, and the arc dies at a zero. Direct current (DC) never passes through zero, so the arc keeps burning.

• An HVDC circuit breaker has to create its own current zero, absorb the energy stored in the line, and do it within about 3 to 5 milliseconds, before the fault current grows too large.

• The design that made it practical is the hybrid breaker, first shown by ABB in 2012. Normal current flows through a simple metal path; in a fault, power electronics take over and cut the current in milliseconds.

• China’s Zhangbei DC grid near Beijing, a four-terminal ±500 kV network in service since 2020, uses 16 hybrid breakers that can each cut 25,000 amps in 3 milliseconds.

• Germany’s four grid operators plan multi-terminal DC hubs at 525 kV, with DC breakers, for North Sea wind, aimed at commercial use from 2029.

Listen up.

Short answer first: an HVDC circuit breaker is a switch that can cut high-voltage direct current during a fault. Ordinary breakers rely on the current passing through zero to put out the spark. Direct current never does. So an HVDC breaker has to force the current to zero itself, soak up the energy, and do it in a few thousandths of a second. That is why it is a special machine, and why real DC grids took so long to arrive.

What does a circuit breaker actually do?

Your house has breakers. Something shorts, one clicks, power dies in that room, and the rest of the house stays on.

A grid breaker does the same job for a transmission line carrying hundreds of megawatts. It must spot the fault, open, and isolate just the broken piece so everything else keeps running.

The hard part is not opening the contacts. It is what happens between them.

Why is AC easy to switch off?

When you pull two contacts apart while current is flowing, the current does not simply stop. It jumps the gap as an electric arc, a very hot channel of glowing gas that conducts electricity.

Alternating current (AC), what homes and most grids use, swings back and forth 50 or 60 times a second. So it passes through zero 100 or 120 times a second. At each zero, the arc briefly has no current to feed it. A good AC breaker blasts the gap with gas at that moment, the arc goes out, and the line stays off.

Nature gives you that zero for free. An AC breaker just waits for it.

Why is DC so hard to switch off?

Direct current (DC) flows one way only, like a battery. At very high voltage, hundreds of thousands of volts, it is called HVDC. It is used for long undersea cables and very long land lines, because it loses less power over distance.

DC never passes through zero. Open a normal breaker on it and the arc just keeps burning. It can destroy the breaker.

It also has to happen fast. On a DC network, fault current climbs very quickly, because nothing in the circuit slows it much. In ABB’s early design work, the fault current rose at about 3.5 kiloamps per millisecond, even with a large reactor added to slow it. Wait 5 milliseconds and the current has already risen by more than 17,000 amps. Wait much longer and the converter stations, the big power electronics at each end of the line, can be damaged.

If you cannot kill the current in a few milliseconds, a fault on one cable can collapse the voltage across the whole DC network, taking down every wind farm and converter on it, not just the sick wire.

How does an HVDC circuit breaker work?

It has to do four jobs, in about 3 to 5 milliseconds:

1. Detect the fault.

2. Force the current to zero, because nature will not.

3. Absorb the energy stored in the line and its reactors, so nothing explodes. This job goes to banks of metal-oxide surge arresters, which turn the energy into heat.

4. Leave the healthy cables running.

There are three main designs.

Mechanical breakers use big contacts plus an extra circuit, a capacitor and inductor, that injects an opposite current to create an artificial zero. They are cheaper and waste almost no power in normal running, but they are slower.

Solid-state breakers do the whole job with power semiconductors. They are very fast. But the full current flows through those chips all the time, and they waste power as heat constantly. ABB estimated that a pure semiconductor breaker could lose about 30% as much power as the converter station it protects.

Hybrid breakers combine both. This is the design that made DC grids practical:

• In normal running, current flows through a simple mechanical path with only a tiny electronic switch in it, so losses are negligible.

• When a fault hits, that small switch pushes the current sideways into a big stack of semiconductor switches. An ultrafast disconnector then opens the mechanical path while it carries no current, which it can do in a couple of milliseconds.

• The semiconductor stack turns off, forcing the current down, and the surge arresters absorb the energy.

ABB announced the first hybrid HVDC breaker in November 2012. Its prototype cut more than 9,000 amps at 320 kV within 5 milliseconds.

Where are HVDC breakers actually in use?

China got there first. The Zhangbei DC grid, which links wind and solar farms in Hebei to Beijing, went into service in 2020. It is a four-terminal, ±500 kV network, described as the world’s first true DC power grid. It uses 16 hybrid DC breakers, four at each station. Each can cut up to 25,000 amps in 3 milliseconds and reclose within 300 milliseconds. China also ran earlier, smaller multi-terminal DC projects at Nan’ao and Zhoushan.

Europe is next. In July 2024, Germany’s four transmission operators (50Hertz, Amprion, TenneT and TransnetBW) and three manufacturers (GE Vernova, Hitachi Energy and Siemens Energy) agreed to develop multi-terminal DC hubs at 525 kV, with DC breakers built in. Implementation starts in 2026, aiming for commercial use in 2029. The goal is to connect large amounts of North Sea offshore wind into one meshed DC grid.

Why do DC grids need breakers at all?

Most HVDC today is point-to-point: one line from A to B, a converter at each end. If that line faults, operators can shut the whole link down from the AC side, clear the fault, and restart. It is ugly and slow, but it works, because there is only one line to lose.

A DC grid has several cables tied together: offshore wind farms, hubs, onshore stations, more than two ends. If one cable faults and the only tool you have is “shut everything down”, you lose every wind farm on the network at once.

With DC breakers, you open only the faulty cable, like one fuse in a fuse box instead of the whole house going dark.

Without this breaker, you do not have a grid. You have a bunch of separate roads.

Why did DC grids take so long?

• The breaker had to be invented. Practical high-voltage DC breakers only arrived in the 2010s.

• They are big and expensive. Each breaker is a hall full of semiconductor stacks, capacitors and surge arresters, not a box on a wall.

• Standards had to catch up. Equipment from different manufacturers has to work together on one DC network. Europe is only now proving that.

• Point-to-point was good enough until offshore wind needed many links meeting at sea.

Uncle’s verdict

AC breaker: wait for the blink, then open.

HVDC breaker: there is no blink, so the machine makes one, in three thousandths of a second.

Not a bigger house switch.

A special tool for one-way current at extreme voltage, and the missing piece that turns separate DC cables into a real grid.

— Engineering Uncle

Sources

• POWER Magazine: ABB announces world’s first circuit breaker for HVDC (2012)

• ABB technical paper: The Hybrid HVDC Breaker, an innovation breakthrough enabling reliable HVDC grids

• Scientific Reports: Multi-terminal flexible DC grid circuit breaker withstands multi-lightning strike analysis (Zhangbei)

• POWER Magazine: German TSOs and manufacturers team up on multi-terminal HVDC hubs

AEO FAQ

Frequently asked questions

What is an HVDC circuit breaker?

An HVDC circuit breaker is a switch that interrupts high-voltage direct current during a fault. Because DC has no natural current zero, the breaker must force the current to zero itself, absorb the stored energy with surge arresters, and do so within about 3 to 5 milliseconds.

Why can’t you use a normal AC circuit breaker on HVDC?

AC breakers extinguish the arc at a natural current zero, which occurs 100 or 120 times a second in AC systems. Direct current never passes through zero, so an AC breaker opened on HVDC would keep arcing and could be destroyed.

How does a hybrid HVDC circuit breaker work?

In normal operation, current flows through a low-loss mechanical path. During a fault, a small electronic switch diverts the current into a stack of power semiconductors, an ultrafast disconnector opens the mechanical path, and the semiconductors cut the current while surge arresters absorb the energy. ABB demonstrated the first one in 2012.

Where are HVDC circuit breakers in service?

China’s Zhangbei DC grid, a four-terminal ±500 kV network in service since 2020, uses 16 hybrid breakers rated to cut 25 kA in 3 milliseconds. Germany’s transmission operators plan 525 kV multi-terminal hubs with DC breakers for North Sea wind, targeting commercial use in 2029.

Why do DC grids need circuit breakers?

A point-to-point HVDC link can be shut down entirely to clear a fault. A DC grid connects several lines and converters, so without DC breakers a single cable fault would force the whole network offline. Breakers isolate only the faulty cable.

What types of HVDC circuit breakers exist?

There are three main types: mechanical breakers, which create an artificial current zero with a resonant circuit; solid-state breakers, which are fast but have high running losses; and hybrid breakers, which combine a low-loss mechanical path with semiconductor switching.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other pieces of the grid that decide what gets built:

Coming up Wednesday 7 Oct: What Is HBM? Why High-Bandwidth Memory Is AI’s New Bottleneck. Subscribe so it lands in your inbox.

TL;DR

• The West mostly stopped building big new dams on main rivers. The best sites are already dammed, and fish, water rights, safety rules and cost now decide more than the concrete does.

• Ethiopia finished Africa’s largest hydropower dam, the 5,150 MW Grand Ethiopian Renaissance Dam, in September 2025 for about $5 billion. There is still no binding deal with Egypt and Sudan on how it runs.

• China is building the biggest hydropower project ever: a planned 60 GW cascade on the Yarlung Tsangpo in Tibet, started in July 2025, costing about 1.2 trillion yuan, in earthquake country upstream of India and Bangladesh.

• California’s Sites Reservoir, first proposed in the 1950s, would be the state’s first major new reservoir since 1979. Its water right is still up for a vote on 4 November 2026.

• Just rebuilding an old dam for earthquakes now costs billions. Anderson Dam in California is about $2.8 billion and runs to around 2034.

Listen up.

Short answer first: in the West we mostly don’t, because the good sites are taken and the river already has owners. Fish, farms, cities, tribes, downstream neighbours and safety regulators all have a say, and the permits can take longer than the pour. China and Ethiopia still build giant dams. They hit the same walls. They just decided to push through them.

People talk about a new dam like it is a swimming pool with a plug. It is a wall across a living river, or a new lake off to the side, plus tunnels, turbines, roads and a fight about who owns the water.

Engineers did not forget how to pour concrete. The river already has a queue.

What does a dam actually have to do?

A dam has to:

• Sit on rock that will not slide. The foundation and the valley walls (the abutments) carry the whole push of the lake.

• Hold a lake that will not leak through cracks in the rock around it.

• Survive the worst flood it could ever see, and in many places a serious earthquake.

• Let enough water through that people and farms below it do not go dry.

• Last a century while silt from upstream slowly fills the lake.

Miss one and you do not have a power station. You have a liability.

Ethiopia: the dam got built, the fight did not end

The Grand Ethiopian Renaissance Dam (GERD) on the Blue Nile is Africa’s largest hydropower plant: 5,150 MW, with a reservoir of about 74 billion cubic metres. It cost about $5 billion. Ethiopia paid for most of it itself, from government funds and bonds sold to its own citizens, with some Chinese loans. It was inaugurated on 9 September 2025, about 14 years after construction began.

That is the engineering win. A poor country finished a giant concrete job.

The hard part in 2026 is not the pour. It is the river below. Egypt gets almost all its fresh water from the Nile and calls the dam an existential threat. Sudan sits in between, with its own dams and its capital downstream. A 2015 Declaration of Principles was signed, but there is still no binding agreement on how Ethiopia fills and runs the reservoir in a drought. Egypt has taken the argument to the UN Security Council.

Lesson: you can finish the wall and still not be “done”. Downstream countries can fight you for a decade after the ribbon.

China: still building giants, in harder places

China already runs the biggest dams on Earth. Three Gorges on the Yangtze, 22.5 GW, is the famous one. Baihetan on the Jinsha, 16 GW, is the world’s second-largest hydropower station. It has been fully running since 2022, sending power east on ultra-high-voltage lines.

Now China has started the next leap. On 19 July 2025, work began on the Medog (Motuo) hydropower project on the lower Yarlung Tsangpo in Tibet, the river that becomes the Brahmaputra in India. The plan:

• Five power stations in a cascade, about 60 GW in total, nearly three Three Gorges

• About 1.2 trillion yuan, roughly $167 billion

• The river drops about 2,000 metres in around 50 km through the “Great Bend”, so most of the water goes through long tunnels bored through the mountains rather than behind one giant wall

• First power aimed at the early 2030s

It sits in one of the steepest, most earthquake-prone gorges on Earth, where the Indian and Eurasian plates meet. India and Bangladesh, downstream, have both raised water-security concerns. India is planning its own large dam on the same river, partly as a buffer.

China can still build at a scale the West will not match. That does not delete geology, resettlement or neighbours. It means China accepted those fights and kept the batching plant running.

The West: new dams are rare, and paperwork is the wall

Europe and the US largely stopped building big new dams on main rivers decades ago. The best sites were already taken, and fish, tribes, farms and “who pays when it goes wrong” won the argument. Much of the Western money now goes into fixing old dams, or removing them. The four Klamath River dams on the California–Oregon border came down in 2023–24, the largest dam removal in history.

California shows the honest 2026 picture.

Sites Reservoir is a new lake off to the side of the Sacramento River, not a plug in it. In wet years, water would be pumped from the river into a valley about 14 miles away and stored there: about 1.5 million acre-feet, roughly 1.85 billion cubic metres. Cost: about $6 billion. It would be California’s first major new reservoir since New Melones in 1979.

It was first proposed in the 1950s. In 2026 the fight is still the water right, meaning how much it may take from the river and when. The State Water Board issued a draft permit in March 2026. The project’s backers said the limits were too tight to justify the cost. A revised draft in September allows up to 986,000 acre-feet a year in high flows, and the board votes on 4 November 2026. If it passes, groundbreaking could come in 2027, with water around 2034.

Anderson Dam, south of San Jose, is not even a new dam. It is an existing earth dam being rebuilt because it sits close to active faults and could fail in a big earthquake. The reservoir has been kept low for years. A $1.88 billion construction contract was awarded in July 2026, the total project is about $2.8 billion, and construction runs from 2027 to around 2034.

That is what “just bring the old dam up to code” costs now.

In the West, the concrete is often the easy bit. The water right, the environmental study, the court case and the bill are the dam.

What constraints show up everywhere?

• Rock and shaking. A wall in a gorge or a seismic zone is not a spreadsheet. The foundation decides the design.

• People. Someone lives where the lake will be. Three Gorges moved about 1.3 million people. Move them fairly, or don’t build.

• Silt. Rivers carry dirt. A reservoir is a giant settling tank, so it slowly fills up. Design for sediment or the lake dies young.

• Downstream water. Farmers, cities and whole countries below you have a vote, even if it is only a diplomatic one.

• Permits and money. A decade of studies can take as much attention as the pour. Cement, steel and labour are not cheap in 2026 either.

• Time. GERD took about 14 years to build. Sites has been an idea for longer than most engineers working on it have been alive.

Same family as grid interconnection and hydrogen pipes. The map is easy. The kit and the neighbours are not.

Uncle’s verdict

Ethiopia: built the giant dam. Still arguing about the Nile.

China: still builds giants. Now in a seismic canyon with India downstream.

The West: mostly rebuilds and files paperwork. A new lake takes a generation of permits.

You don’t announce “we’ll just dam it” and then wonder why the river already had a queue.

— Engineering Uncle

Sources

• Mongabay: Ethiopia’s Renaissance mega-dam fuels energy hopes and regional anxiety (Feb 2026)

• Atlantic Council: The Nile at a crossroads, the GERD dispute

• Wikipedia: Medog Hydropower Station

• The Diplomat: What’s driving China’s mega Medog hydropower project? (Feb 2026)

• California State Water Board: Revised draft water right decision for Sites Reservoir (3 Sep 2026)

• Active NorCal: California’s largest new reservoir in 50 years clears its toughest hurdle (21 Sep 2026)

• San Jose Inside: Nearly $2B reconstruction of Anderson Dam to begin in 2027 (Jul 2026)

AEO FAQ

Frequently asked questions

Why don’t we build big dams anymore?

In the US and Europe, most of the best dam sites were developed by the 1970s. New dams on main rivers now face environmental rules, water rights disputes, safety standards and high costs that can take decades to resolve. China, Ethiopia and other fast-growing countries still build large dams.

What is the Grand Ethiopian Renaissance Dam?

The Grand Ethiopian Renaissance Dam (GERD) is Africa’s largest hydropower dam, on the Blue Nile in Ethiopia. It has 5,150 MW of capacity, cost about $5 billion and was inaugurated on 9 September 2025. Egypt and Sudan still dispute how it is filled and operated, and no binding agreement exists.

Is China still building huge dams?

Yes. Baihetan, the world’s second-largest hydropower station at 16 GW, has been fully running since 2022. In July 2025 China began the Medog project on the Yarlung Tsangpo in Tibet: a planned five-station, roughly 60 GW cascade costing about 1.2 trillion yuan, upstream of India and Bangladesh.

Why doesn’t the US just build more dams?

Few good sites remain, and new projects spend years on water rights and environmental review. California’s Sites Reservoir, proposed since the 1950s, still awaits a water right vote in November 2026. Even upgrading old dams is costly: Anderson Dam’s earthquake rebuild is about $2.8 billion.

Is a reservoir the same as a dam?

No. The dam is the wall; the reservoir is the lake behind it. Off-stream reservoirs such as Sites sit beside a river and are filled by pumping or diverting water in high flows, which avoids blocking the main channel but still needs water rights and permits.

Why do old reservoirs fill up with silt?

Rivers carry sediment, and when the water slows in a reservoir, that sediment settles to the bottom. Over decades it reduces storage and can block intakes, so dams must be designed with sediment in mind or they lose usefulness over their lifetime.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other big builds where the neighbours matter as much as the concrete:

Coming up Monday 5 Oct: HVDC Circuit Breakers: What They Are and Why DC Is Hard to Switch Off. Subscribe so it lands in your inbox.

TL;DR

• A data center cooled with evaporative cooling towers consumes roughly 1 litre of water for every kilowatt-hour its computers use, about 1,100 litres per megawatt-hour, or 2.6 million gallons per megawatt per year, by Nvidia’s figure.

• A data center with closed-loop, air-cooled or dry-cooler designs consumes close to zero water for cooling, but uses more electricity instead. Microsoft’s fleet averaged 0.30 litres per kWh in 2024.

• US data centers directly consumed about 17.4 billion gallons of water in 2023, about as much as 160,000 American households, according to Lawrence Berkeley National Laboratory. That could reach 38 to 73 billion gallons by 2028.

• The power plants that supply data centers usually consume more water than the data centers themselves.

• A single AI chatbot prompt uses very little. Google puts the median Gemini text prompt at 0.26 millilitres, about five drops. The local problem is a whole campus drawing from one town’s supply.

Listen up.

Short answer first: it depends almost entirely on how the data center gets rid of its heat. One that cools by evaporating water consumes roughly 1 litre per kilowatt-hour of computing, which for a big campus can mean millions of gallons a day. One that uses closed-loop or air-based cooling consumes close to none, and pays for it with a bigger electricity bill instead. Same computers, very different water bill.

That choice is now fought over at town council meetings, not just in engineering offices.

Why does a data center need water at all?

Every watt of electricity a computer uses turns into heat. A 100 MW AI campus is, thermally, a 100 MW heater that never switches off. That heat has to leave the building.

Inside the hall, liquid cooling carries the heat off the chips. Uncle covered that in the liquid cooling post on 29 September. But that only moves the heat to the building’s water loop. The real question is what happens at the edge of the site, where the heat has to go into the outside world.

There are two main ways to do it:

1. Evaporate water. Spray the warm water through a cooling tower. Some of it evaporates, and evaporation carries away a lot of heat. This is how most big power plants and many data centers work. It is cheap on electricity and hungry for water.

2. Blow air over radiators. Dry coolers work like a car radiator: fans push outside air over fins full of warm water. No water is lost, but you need a lot more fan power, and on hot days you may need chillers, which are refrigeration machines that use even more electricity.

Water or electricity. Every data center picks its trade-off.

How much water does evaporative cooling use?

Here is the physics.

Evaporating one litre of water absorbs about 2.26 megajoules of heat. One kilowatt-hour of heat is 3.6 megajoules. So if you got rid of all the heat by evaporation, you would need about 1.6 litres per kilowatt-hour.

Real cooling towers do a bit better than that ceiling, because some heat also leaves as warm air. Nvidia says conventional cooling-tower systems consume about 2.6 million gallons per megawatt per year. That works out to about 1.1 litres per kilowatt-hour, or about 1,100 litres per megawatt-hour.

Scale it up. A 100 MW AI campus running flat out on evaporative cooling would consume roughly 700,000 gallons a day. Google’s largest US data center campus, in Council Bluffs, Iowa, has consumed about 2.8 million gallons a day on average. That is roughly what 9,000 American homes use.

And not all of the water taken in is lost. Some goes back to the sewer or river as “blowdown”, flushed out to stop minerals building up, and it comes back warmer and saltier. Up to about 85% of the water a data center uses evaporates.

How much water do US data centers use in total?

Lawrence Berkeley National Laboratory’s 2024 report for the US Department of Energy estimated that US data centers directly consumed about 17.4 billion gallons of water in 2023. That is about 48 million gallons a day, roughly what 160,000 American households use.

By 2028, the same work projects 38 to 73 billion gallons a year as AI campuses multiply, with hyperscale sites making up about half.

In national terms that is still small next to farming, power plants or, in Arizona, golf courses. But water is always local. Seventeen billion gallons spread across the US is one thing. Two million gallons a day from one aquifer under one desert town is another.

Do power plants use more water than data centers?

Often, yes. This is the part most headlines skip.

Thermal power plants (gas, coal, nuclear) also cool themselves by evaporating water. On average, US thermoelectric plants consume roughly 0.47 gallons, about 1.8 litres, per kilowatt-hour they generate. So a data center running on grid power from those plants can be responsible for more water loss at the power station than at its own cooling towers.

That is why switching a data center to air cooling does not remove its water footprint entirely: it uses more electricity, and that electricity has its own water cost, unless it comes from wind or solar.

How much water does one AI prompt use?

Very little. In August 2025 Google reported that its median Gemini text prompt used 0.24 watt-hours of energy and 0.26 millilitres of water, about five drops. Critics noted that the figure leaves out the water used at power plants.

The point is not one prompt. It is billions of prompts a day, concentrated in a handful of campuses, each drawing on one local water system.

Can data centers use zero water?

For cooling, nearly. It costs electricity.

• Microsoft says all its new data center designs since August 2024 use closed-loop, zero-evaporation cooling. Pilots in Phoenix, Arizona and Mount Pleasant, Wisconsin come online in 2026, and each site avoids more than 125 million litres of water a year. Microsoft admits the trade-off: “a nominal increase” in energy use. Its existing fleet averaged 0.30 litres per kWh in 2024, down from 0.49 in 2021.

• Nvidia says its Rubin systems can take coolant at up to 45°C. Water that warm can be cooled by outside air on most days, so dry coolers replace cooling towers. In favourable climates, Nvidia says, facility cooling water use can drop to near zero.

• Google pledged in June 2026 to replenish 120% of the water its data centers consume by 2030. Its 2024 consumption was about 9.5 billion gallons worldwide.

Note the difference. Microsoft and Nvidia change the engineering to use less. Google offsets what it uses by funding water projects elsewhere, which does not change the draw on the local aquifer.

Why is water the fight at every data center hearing?

Because water is local, visible and already scarce in many of the places data centers want to go. Sunbelt states have cheap land and power, and they are dry.

Tucson, Arizona, is the textbook case. In August 2025 the city council unanimously rejected Project Blue, a large data center proposal, including any use of city water. The developer switched to a closed-loop, air-cooled design, bought land outside city limits, and in May 2026 won state approval for groundwater wells of about 31 million gallons a year, mostly for offices and the initial fill. Earlier that month, the city had cut off a temporary water meter the contractor was using for construction dust control.

That is the pattern in 2026:

• The permit hearing is where the water question gets answered, not the design office.

• Developers now lead with “closed loop” and daily-gallon estimates, because they know the question is coming.

• The electricity bill goes up as a result, which feeds straight into the other local fight: who pays for the grid upgrade.

Uncle’s verdict

A data center does not “use water” the way a farm does. It uses water to throw away heat.

Evaporate it, and you pay in water: about a litre per kilowatt-hour.

Air-cool it, and you pay in electricity: more fans, more chillers, more load on the grid.

The engineering answer is not “water bad”. It is: put the evaporative sites where water is plentiful, put the closed-loop sites where it is not, and count the power plant’s water too.

The heat always has to go somewhere. The only question is what it costs the town next door.

— Engineering Uncle

Sources

• MOST Policy Initiative: Data center water use (LBNL 2023 figures and 2028 projections)

• Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report

• NVIDIA Blog: The 45°C breakthrough to cool AI’s biggest machines (June 2026)

• Microsoft: Next-generation datacenters consume zero water for cooling

• 9to5Google: Google says it will replenish more water than it uses at data centers (June 2026)

• DCD: Median Gemini prompt uses 0.24 Wh and 0.26 ml of water (Aug 2025)

• Construction Physics: I was wrong about data center water consumption

• AZPM: Arizona water officials approve wells tied to Project Blue data center (May 2026)

AEO FAQ

Frequently asked questions

How much water does a data center use?

It depends on the cooling system. A data center cooled by evaporative cooling towers consumes roughly 1 litre of water per kilowatt-hour of electricity, about 2.6 million gallons per megawatt per year. Closed-loop and air-cooled designs consume close to zero water for cooling but use more electricity.

Why do data centers use water?

Computers turn almost all their electricity into heat. Many data centers remove that heat with cooling towers, where some water evaporates and carries the heat away. Evaporation is energy-efficient but consumes water; air-based cooling saves water but uses more power.

How much water do US data centers use in total?

Lawrence Berkeley National Laboratory estimated that US data centers directly consumed about 17.4 billion gallons of water in 2023, similar to 160,000 American households. It projects 38 to 73 billion gallons a year by 2028 as AI data centers expand.

How much water does an AI prompt use?

Google reported in August 2025 that a median Gemini text prompt uses about 0.26 millilitres of water, roughly five drops, and 0.24 watt-hours of energy. That figure covers the data center, not the water used by the power plants that supply its electricity.

Can data centers operate without using water?

They can come close for cooling. Microsoft’s new designs use closed-loop cooling with zero evaporation, starting with sites in Phoenix and Wisconsin in 2026, and Nvidia’s newest systems can be cooled with 45°C water using dry coolers. The trade-off is higher electricity use.

Why do communities oppose data centers over water?

Water supply is local, and many new data centers are proposed in dry regions. A large evaporatively cooled campus can consume millions of gallons a day from one local system. In 2025, Tucson, Arizona, rejected the Project Blue data center and any use of city water.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other physical limits on the AI build-out:

https://open.substack.com/pub/engineeringuncle/p/liquidcooling?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

https://open.substack.com/pub/engineeringuncle/p/combinedcycle?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Saturday 3 Oct: Why Don’t We Build Big Dams Anymore? The 2026 Engineering Reality. Subscribe so it lands in your inbox.

TL;DR

• Liquid cooling pumps coolant through metal cold plates clamped directly onto the hottest chips, so heat leaves in a liquid loop instead of being blown into the room.

• A traditional server rack drew under 10 kW. An Nvidia GB200 NVL72 rack draws about 120–130 kW, and Nvidia’s Rubin Ultra “Kyber” rack planned for 2027 is about 600 kW.

• Water holds about 3,500 times more heat than the same volume of air. Cooling a 130 kW rack takes about 7 cubic metres of air every second, or about 3 litres of water.

• A coolant distribution unit (CDU) pumps the clean chip loop and passes its heat to the building’s water through a heat exchanger. The two loops never mix.

• Nvidia says its Rubin systems can take coolant at up to 45°C. In many climates that means no chillers at all, just dry coolers outside.

Listen up.

Short answer first: data center liquid cooling works by bringing the water to the chip. A metal cold plate sits on each processor. A clean liquid runs through the plate, picks up the heat and carries it to a pump box called a coolant distribution unit (CDU). The CDU hands that heat to the building’s water, which takes it outside. Air lost because AI racks got too hot for a breeze to carry the heat away.

How did data centers cool computers before?

With air. Cold air in the front of the rack, hot air out the back, big air-conditioning units around the room.

That worked for decades because each rack was only a little warm. Uptime Institute’s surveys have found that most racks draw under 10 kW, about as much as a few electric kettles running nonstop. Uptime reckons that above about 20–25 kW per rack, liquid cooling starts to make more sense than air.

Then AI arrived.

Why did air lose?

Because the heat got too big for a breeze.

An Nvidia GB200 NVL72 rack, 72 GPUs in one cabinet, draws about 120–130 kW. Nearly all of that becomes heat. That is more like a small furnace in a cupboard than a computer.

Now the physics. Water holds about 3,500 times more heat than the same volume of air. Uncle’s numbers for carrying away 130 kW:

• With air, warming it by 15°C: about 7 cubic metres every second, roughly 15,000 cubic feet per minute, through one cabinet. That is a gale in the aisle, and the fans would burn a pile of power.

• With water, warming it by 10°C: about 3 litres every second, through hoses about the size of a garden hose.

Air did not lose because fans are “old”. It lost because the maths stopped working.

It gets worse. Nvidia’s Vera Rubin NVL144 racks arriving from late 2026 stay around the same 120–130 kW. The Rubin Ultra “Kyber” racks planned for 2027 are about 600 kW each. No air system comes close.

Subscribe now

What is a cold plate?

A metal block, usually copper, clamped onto the chip. Inside are hundreds of tiny channels, like fins turned inside out. Coolant flows through the channels, picks up the heat, and leaves through a hose.

Same idea as a car engine and its radiator, except the first radiator is a plate sitting right on the silicon.

The liquid stays inside the plate and the hoses. It never pours over the circuit board. People worry about “water next to computers”. Fair. The answer is sealed plates, dripless quick-disconnect fittings, leak sensors and proper installation, not a bucket.

Cold plates take most of the heat but not all of it. Memory, network cards and power supplies in the same rack still give off some heat into the air, so most liquid-cooled halls keep some fans and air handling too.

What is a CDU?

CDU stands for coolant distribution unit. It is a box with pumps, a heat exchanger, filters and controls.

Most liquid-cooled halls run two separate loops:

1. The chip loop, or technology cooling system. Clean, treated coolant, often water mixed with propylene glycol, flowing to the cold plates and back.

2. The building loop, or facility water. The building’s water, which carries the heat outside to cooling towers, dry coolers or chillers.

The CDU sits between them. It pumps the chip loop and passes the heat across a heat exchanger to the building water. The two liquids never mix, so building water never goes into the servers.

CDUs come in two main sizes: small ones that sit inside a rack, and big “row” units that serve many racks at once.

What about rear-door heat exchangers and immersion?

Two other options you will hear about:

• Rear-door heat exchangers. A radiator on the back door of the rack cools the hot air as it leaves. It helps, and it is a good half step for older halls, but it still relies on air to pull heat off the chip. It runs out of steam well below AI rack densities.

• Immersion cooling. The whole server is dunked in a tank of non-conductive fluid. It works and handles very high densities, but it changes how you service every machine. For now, the big AI builds have gone with cold plates.

Why does Nvidia cool its chips with 45°C water?

This is the part that surprises people.

Nvidia says its Rubin systems can take coolant at up to 45°C, hotter than a hot tub, and send it out at about 55°C. That sounds backwards. It isn’t.

The chip runs much hotter than 45°C, so even warm water can carry heat away from it. And water that only needs to be 45°C can be cooled by outdoor air on most days of the year. No refrigeration needed, just dry coolers: big radiators with fans outside.

Nvidia says that in favourable climates this can take cooling water use from about 2.6 million gallons per megawatt per year with conventional cooling towers to near zero, and remove chillers entirely. In hot places, chillers may still run a few days a year.

The liquid does not make the heat disappear. It just makes it much cheaper to get rid of.

What still goes wrong?

• Leaks. Rare with good kit and good installation, ugly without them. Plan for drips the way a plant room plans for spills: sensors, drip trays, automatic shut-offs.

• Dirty coolant. Particles and corrosion clog the tiny cold plate channels and wear out pumps. Treat the chip loop like boiler water, not tap water: filtered, chemically controlled, tested.

• The heat still has to leave the site. Liquid cooling moves heat better inside the hall. Outside, you still need cooling towers, dry coolers or chillers big enough for the whole load.

• Old halls need surgery. An air-cooled data center has no water pipes to the racks, often weak floors for heavy liquid-cooled racks, and not enough power. Converting one means new pipes, new CDUs and often new electrical kit.

Uncle’s verdict

Air cooling: blow on the oven.

Liquid cooling: strap a small radiator to the oven and pump the heat outside.

Cold plate = the strap-on radiator.

CDU = the pump box in the middle.

Air lost because the oven became a furnace, and the next furnace is five times bigger.

— Engineering Uncle

Sources

• NVIDIA Blog: Hotter Than a Hot Tub, the 45°C breakthrough to cool AI’s biggest machines

• Introl: NVIDIA Vera Rubin and 600 kW racks by 2027

• Introl: GB200 NVL72 deployment

• Uptime Institute: Rack density is rising

• Uptime Institute: AI and cooling, methods and capacities

AEO FAQ

Frequently asked questions

How does data center liquid cooling work?

Direct-to-chip liquid cooling pumps coolant through metal cold plates attached to processors. The warmed coolant flows to a coolant distribution unit (CDU), which transfers the heat to the building’s water loop through a heat exchanger. That water then carries the heat outside to cooling towers, dry coolers or chillers.

Why can’t air cool AI data centers?

AI racks such as the Nvidia GB200 NVL72 draw about 120–130 kW, compared with under 10 kW for a typical traditional rack. Water holds about 3,500 times more heat than the same volume of air, so removing that much heat with air would take roughly 7 cubic metres of airflow per second per rack.

What is a cold plate?

A cold plate is a metal block, usually copper, clamped onto a chip. Coolant flows through small internal channels, absorbs the chip’s heat and carries it away through hoses. The coolant stays sealed inside the plate and never touches the circuit board.

What is a CDU in a data center?

A coolant distribution unit (CDU) contains pumps, a heat exchanger, filters and controls. It circulates clean coolant through the servers’ cold plates and passes the heat to the separate facility water loop, so the two liquids never mix.

Why does Nvidia use 45°C water for cooling?

Chips run much hotter than 45°C, so warm coolant can still remove their heat. Water at that temperature can be cooled by outside air using dry coolers on most days, so chillers can be reduced or eliminated. Nvidia says this can cut cooling water use to near zero in favourable climates.

Does liquid cooling remove the need for cooling towers?

No. Liquid cooling moves heat out of the servers more efficiently, but the heat still has to leave the site. Facilities still need cooling towers, dry coolers or chillers sized for the full heat load.

Read next

Read next on EngineeringUncle

If this one clicked, these are the other bottlenecks inside the AI build-out:

https://open.substack.com/pub/engineeringuncle/p/combinedcycle?r=8suf21&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Coming up Wednesday 7 Oct: What Is HBM? Why High-Bandwidth Memory Is AI’s New Bottleneck. Subscribe so it lands in your inbox.

TL;DR

• A combined-cycle plant burns gas in a gas turbine, then uses the turbine’s 600°C-plus exhaust to boil water for a second, steam turbine. Two machines, one fuel bill.

• A gas turbine on its own turns about 35–43% of the fuel’s energy into electricity. The steam half recovers about a third of the heat that is left, which lifts the plant past 60%.

• The record is 63.08%, set at Chubu Electric’s Nishi-Nagoya plant in Japan in 2017–18. The newest GE and Mitsubishi turbines are sold as capable of more than 64%.

• Brochure numbers are best-case: new plant, full power, cool day. The average US combined-cycle plant runs at about 51% on the same basis.

• Demand for these plants from AI data centers has outrun supply. GE Vernova’s gas turbine backlog reached 116 GW in mid-2026, and a new heavy-duty order placed now arrives around 2031.

Listen up.

Short answer first: a combined-cycle plant reaches 60% by using the same fire twice. A gas turbine makes electricity and throws out very hot exhaust. A steam plant bolted onto that exhaust makes more electricity from heat that would otherwise go up the chimney. Neither machine can do 60% alone. Stacked, they can.

What is a simple-cycle gas turbine?

A gas turbine is a jet engine bolted to the ground. Suck air in, squeeze it, burn gas in it, let the hot gas spin a turbine, and the turbine spins a generator. Simple and fast.

Also wasteful. The exhaust leaves at around 600 to 660°C. Mitsubishi lists 646–663°C for its biggest J-class turbine. That is hot enough to glow. In a “simple-cycle” plant, that heat just goes up the stack.

The best big turbines on their own reach roughly 40–43% efficiency: about 40 units of electricity from every 100 units of energy in the fuel. The other 60 leave as hot exhaust.

How does the second machine work?

A combined-cycle plant refuses to throw that heat away.

The hot exhaust goes through a heat-recovery steam generator (HRSG). Fancy name. It is a kettle the size of an apartment block: walls of tubes full of water, with the turbine exhaust blowing through. The exhaust boils the water. The steam spins a second turbine. More electricity. Same original gas.

Two textbook cycles, stacked:

• Gas side: the Brayton cycle. Squeeze, burn, expand.

• Steam side: the Rankine cycle. Boil, expand, condense.

Together: a combined cycle.

Why does stacking get you past 60%?

Here is the maths, and it is simpler than it looks.

Say the gas turbine turns 40% of the fuel into electricity. That leaves 60% as heat, mostly in the exhaust. The steam half can turn about a third of that leftover heat into electricity.

• 40% from the gas turbine

• plus one-third of the remaining 60%, which is another 20%

• equals about 60%

That extra 20 points is not a miracle chip. It is the second machine. Push the gas turbine to 43% and the steam half a little higher, and you reach the 63–64% the newest plants claim.

What is the record?

The verified record is 63.08%, set in 2018 at Chubu Electric’s Nishi-Nagoya Block-1 in Japan, which runs three GE 7HA.01 turbines. Before that it was 62.22% at EDF’s Bouchain plant in France in 2016.

Today GE Vernova (H-class) and Mitsubishi Power (J-class) both sell turbines they say can pass 64% in combined cycle.

What actually gets you over 60%?

Not a sticker on the shed. Four things.

1. A hotter fire, carefully. The hotter the gas going into the turbine, the more work it gives. In the latest turbines that gas is around 1,600°C, hotter than the melting point of the nickel alloys the blades are made of. The blades survive because cooling air is pumped through tiny internal channels and out of holes in their surface, and a ceramic thermal barrier coating takes the direct heat. Hotter inlet gas means more work from the first machine, and exhaust still hot enough to feed the second.

2. A proper kettle. A good HRSG does not boil water in one crude pass. It has three pressure levels, high, intermediate and low, plus a reheat stage, so it collects heat in steps as the exhaust cools. Get that wrong and you leave steam on the table.

3. A steam turbine and condenser that fit. The steam turbine has to be sized for the exhaust it gets. After the steam has done its work, a condenser dumps the last low-temperature heat into a river, the sea or cooling towers. If the cooling is poor, as on a hot day with warm river water, the steam half sulks and efficiency drops.

4. Clean air and a steady load. Gas turbines lose output on hot days, because hot air is thinner. Dirty inlet filters and worn blades cost more. And a plant running at half power because the grid does not need it will never hit the brochure figure.

Why don’t real plants hit the brochure number?

Because 60% is a design point, not a law of nature that holds at 2 a.m. in August.

There is also some small print. Headline figures are usually gross (before the plant’s own pumps and fans) and measured on the fuel’s lower heating value, which ignores the energy in the water vapour in the exhaust. The US Energy Information Administration measures on the higher heating value, and it found the average US combined-cycle plant ran at a heat rate of 7,340 Btu/kWh. That is about 46.5% on the US basis, or roughly 51% on the brochure basis.

That fleet includes older plants, part-load running and hot summers. It is still the most efficient way to burn fuel for electricity at scale.

Why does 60% matter?

Most of the world’s thermal power plants run in the thirties and forties. On EIA’s figures, the average US coal plant used about 25% more fuel energy per kilowatt-hour than the average combined-cycle gas plant.

Higher efficiency means less fuel for the same lights, and less CO₂. Uncle’s rough numbers for burning natural gas:

• Simple-cycle at 40%: about 450 grams of CO₂ per kWh

• Combined-cycle at 60%: about 300 grams per kWh

• For comparison, an average coal plant: close to 900 grams per kWh

It is still gas. It is just much less sloppy gas.

Why do AI data center stories keep mentioning these plants?

Because if you need a lot of power, all day, soon, and the grid connection is years away, a combined-cycle plant is the most efficient fossil option you can build.

The problem is that everyone had the same idea. GE Vernova’s gas turbine backlog reached 116 GW by the end of June 2026, counting firm orders and paid slot reservations. Order a large heavy-duty turbine now and it arrives around 2031. Turbine prices are forecast to reach about $600 per kW by the end of 2027, nearly three times 2019 levels. GE Vernova, Siemens Energy and Mitsubishi Power are all expanding their factories.

The gas turbine has joined transformers, grid connections and memory chips on the list of AI build-out bottlenecks.

What a combined-cycle plant is not

• Not free energy. You still buy the gas, and 36–40% of it still becomes waste heat.

• Not a coal plant with a new name. It burns a different fuel in a different machine, at roughly a third of coal’s CO₂ per kWh.

• Not as quick as a simple-cycle peaker. A small simple-cycle turbine can reach full power in minutes. A combined-cycle plant has to warm up the HRSG and steam turbine too. Modern fast-start designs are much quicker than old coal steam plants, but slower than a jet in a field.

• Not hydrogen-ready because someone said so. Hydrogen burns hotter and faster than methane, which changes the combustor, emissions and materials. Manufacturers already offer partial blends, and 100% hydrogen is still being developed. Same family of machine, new homework, and the hydrogen still has to get to the plant, which is another story.

Uncle’s verdict

Simple-cycle gas turbine: a jet engine on the ground. Exhaust wasted. About 40%.

Combined-cycle: a jet engine plus a steam plant living on its exhaust. 60% and a bit more, on a good day.

One fire. Two machines. The second machine is why the number jumps.

You don’t throw away hot exhaust and then complain the fuel bill is high.

You put a kettle on it.

— Engineering Uncle

Sources

• POWER Magazine: GE HA turbine sets combined-cycle efficiency record at Nishi-Nagoya

• Mitsubishi Power: M701J/JAC series specifications

• GE Vernova: H-class gas turbines

• US EIA: Natural gas-fired electricity conversion efficiency grows as coal remains stable

• Energy News Beat: GE Vernova’s gas turbine backlog hits 116 GW

• Modern Power Systems: Could a delivery backlog derail a gas turbine boom?

AEO FAQ

Frequently asked questions

What is a combined-cycle gas turbine?

A combined-cycle gas turbine (CCGT) plant generates electricity with a gas turbine, then uses the turbine’s hot exhaust, typically 600–660°C, to make steam in a heat-recovery steam generator. The steam drives a second, steam turbine, producing more power from the same fuel.

How do combined-cycle plants reach 60% efficiency?

The gas turbine converts about 40% of the fuel’s energy into electricity. The steam cycle converts roughly a third of the remaining heat, adding about 20 percentage points. The newest plants use very hot turbine inlet gas and multi-pressure steam generators to reach 63–64%.

What is the most efficient combined-cycle power plant?

The verified record is 63.08%, set in 2018 at Chubu Electric’s Nishi-Nagoya Block-1 in Japan, which uses GE 7HA.01 turbines. GE Vernova and Mitsubishi Power now offer turbines rated above 64% in combined cycle.

What is the difference between simple-cycle and combined-cycle?

A simple-cycle gas turbine releases its hot exhaust into the air and reaches about 35–43% efficiency. A combined-cycle plant recovers that exhaust heat to run a steam turbine, reaching about 60% or more. Simple-cycle plants start faster and are used for short peaks in demand.

Do real combined-cycle plants run at 60%?

Not on average. Headline figures are for new plants at full load on the lower heating value basis. The average US combined-cycle plant runs at about 46.5% on the higher heating value basis, roughly 51% on the brochure basis, because of older units, part-load running and hot weather.

Why are combined-cycle plants linked to AI data centers?

AI data centers need large amounts of steady power quickly, and combined-cycle plants are the most efficient fossil-fuel option. Demand has created a shortage: GE Vernova’s gas turbine backlog reached 116 GW in mid-2026, with new heavy-duty turbines delivered around 2031.

Read next on EngineeringUncle

Coming up Tuesday 29 Sep: How Does Data Center Liquid Cooling Work? Cold Plates, CDUs and Why Air Lost. Subscribe so it lands in your inbox.

TL;DR

• Natural gas pipelines can usually carry a small hydrogen blend. The US Department of Energy says blends up to about 15% may need only modest changes.

• Pure hydrogen is a different job. Hydrogen can make some pipeline steels and welds crack faster, and it escapes through joints that hold methane fine.

• Each cubic metre of hydrogen carries about a third of the energy of natural gas. It flows faster, so a converted pipe still moves roughly 85% of the energy, but the compressors have to push about three times the volume.

• A 20% hydrogen blend by volume replaces only about 7% of the energy in the gas.

• Real hydrogen networks are being built for factories, not kitchens. Germany’s 9,040 km core network reuses about 60% old gas pipe, but only after testing and rebuilding.

Short answer first: partly. An old gas pipe can often take a little hydrogen mixed into the gas. Pure hydrogen usually needs new pipe, or old pipe tested and rebuilt on purpose. And the customer is a factory, not your kitchen.

Now the long answer. Two questions first, because they are the ones that get people lost.

Is a hydrogen pipeline a water pipe?

No. Water is H₂O: hydrogen stuck to oxygen. Taps and toilets.

A hydrogen pipeline carries hydrogen gas: just H₂. A fuel. Invisible. Burns.

Is a hydrogen pipeline just a gas pipe?

Same family as the pipe that brings cooking gas to your street: steel tube, compressors, valves, moving a gas.

Not the same job. You cannot treat the town gas main as a hydrogen main with a new coat of paint. The rest of this post is why.

Why build hydrogen pipelines at all?

Because factories already use a lot of hydrogen.

The world used just over 100 million tonnes of hydrogen in 2025, according to the International Energy Agency (IEA). Almost all of it goes to making fertiliser (ammonia), refining oil, making methanol and treating steel. Nearly all of it is made from natural gas or coal, usually on the same site that uses it. Low-emissions hydrogen was close to 1 million tonnes, about 1% of the total.

Climate plans want more of that hydrogen made with clean electricity somewhere else, then sent to the factory. A pipe is the boring way to move a lot of gas every hour, all day, cheaper than a never-ending line of trucks.

This is not new. The United States already has about 1,600 miles (2,600 km) of hydrogen pipeline, mostly along the Gulf Coast, owned by industrial gas companies and feeding refineries and chemical plants. It works. It is also short, private and built for a handful of big customers.

Then the brochure jumps to: “Just use the gas pipes already under the street.” That is where the engineering says wait.

How is hydrogen different from natural gas?

Natural gas is mostly methane, CH₄. We have piped it for a century.

Hydrogen, H₂, is the smallest and lightest molecule there is:

• About one-eighth the density of methane.

• About one-third the energy per cubic metre: roughly 10.8 megajoules versus 35.8 for methane.

• It burns with a flame that is nearly invisible in daylight.

Three real constraints follow.

Constraint 1: Hydrogen leaks more easily

Small molecules find small gaps. Through the same hole, hydrogen escapes at roughly 2.8 times the volume of methane, because it is so light. Threaded joints, old seals and valve stems that hold methane fine can weep hydrogen.

Like natural gas, hydrogen has no smell of its own. Town gas smells because the network adds an odorant, and that odorant has to be proven safe for hydrogen equipment too.

Leaks also cost you twice. You lose product, and hydrogen in the air slows the breakdown of methane, which makes it an indirect greenhouse gas. A leaky hydrogen network gives back part of the climate benefit it was built for.

So hydrogen lines need tighter joints, more welded connections, and better leak detection than an ordinary street main.

Constraint 2: Hydrogen can weaken steel

This is the big one, and it has a name: hydrogen embrittlement.

Under pressure, hydrogen atoms work their way into the steel, especially at tiny cracks, hard spots and welds. There they make the metal less tough. Every time the pipeline pressure goes up and down, an existing crack grows a little further, and with hydrogen present it grows faster than it would with methane.

The steels most at risk are the high-strength grades used for modern high-pressure lines, and hard weld zones. A pipe that lived happily on methane for 40 years might be fine on hydrogen, fine at a lower pressure, or not fine at all. You only know after testing that specific steel and those specific welds.

The US Department of Energy lists embrittlement of steel and welds as the first research challenge for hydrogen pipelines. The US design code for hydrogen piping, ASME B31.12, adds its own rules on top of the normal gas pipeline code.

“We already have pipes” is true. “Those pipes can take pure hydrogen at full pressure” is a different sentence.

Constraint 3: Same pipe, a bit less energy, much bigger compressors

Here is where the maths gets interesting.

Each cubic metre of hydrogen carries about a third of the energy of methane. That sounds like the pipe loses two-thirds of its job. It doesn’t, because hydrogen is so light that it flows about 2.8 times faster through the same pipe for the same pressure drop.

Uncle’s back-of-envelope: one-third of the energy, times 2.8 times the flow, gives roughly 85% of the energy a natural gas pipe of the same size would move. Detailed studies land in a similar range, depending on pressure and pipe.

The catch is the compressors. To move that energy, they must push about three times the volume of gas. And hydrogen is awkward to compress: the light molecules slip back in the big centrifugal machines that gas pipelines use, so hydrogen needs more stages or piston compressors. Existing compressor stations usually cannot just be switched over.

The energy those compressors burn is small, around 1.5 to 2% of the hydrogen’s energy per compression step in manufacturer examples. The real cost is the new machines themselves.

That is why hydrogen is happiest close to the plant that uses it. A pipe across a whole country is possible. It is expensive, and it is not “the old grid, new sticker.”

Can you just mix a little hydrogen into the gas grid?

Yes, a little. This is called blending.

The US Department of Energy says blends up to about 15% hydrogen may need only modest changes to a natural gas pipeline. In the UK, the HyDeploy trials ran a 20% blend to homes at Keele University and then to about 670 homes in Winlaton, Gateshead, without customers changing their boilers or cookers.

But look at what 20% means. Those percentages are by volume. Because hydrogen carries a third of the energy per cubic metre, a 20% blend replaces only about 7% of the energy in the gas, and cuts carbon dioxide from burning it by about the same. A 10% blend replaces about 3%.

So blending is a real, small step. It is not a whole city running on hydrogen.

What hydrogen pipelines are actually being built?

Two European projects show what the real thing looks like.

Germany’s hydrogen core network. Approved by the Federal Network Agency in October 2024: 9,040 km of pipeline by 2032, costing about €18.9 billion. About 60% is converted natural gas pipe. The first converted sections carry hydrogen, and companies could book capacity from March 2026.

Notice that 60%. Reusing old pipe does happen, but these are big, thick transmission lines with good records, chosen after testing, fitted with new valves and compressors. That is “rebuilt on purpose,” not “new sticker.”

The Dutch national hydrogen network. Gasunie opened the first 32 km section, Maasvlakte to Pernis in the Port of Rotterdam, on 21 May 2026, much of it repurposed gas pipe. Hydrogen is not expected to flow until late 2026, when Shell’s electrolysis plant starts up to supply its refinery. The full plan is about 1,200 km. The Dutch Court of Audit has warned that large parts may never be finished and that loss compensation could reach about €2.5 billion, because nobody wants to build supply before customers exist, or sign up as a customer before the pipe exists.

That chicken-and-egg problem is the real constraint behind all the engineering.

What about the gas main under your street?

This is the part people get backwards.

Many town gas mains are now polyethylene plastic, which handles hydrogen reasonably well. The harder problem is everything at the end of the pipe: meters, regulators, and every boiler, cooker and water heater in every home. Pure hydrogen needs appliances built for it. Switching a whole town means visiting every house.

So the town grid cannot “just switch fuel” either, but mostly because of the kitchens, not the pipe.

When is a hydrogen pipeline worth building?

• A short or medium-length line

• One or more big industrial customers that already use hydrogen and have signed up

• New pipe, or an old transmission line tested and rebuilt on purpose

• A clean hydrogen supply that actually exists at the other end

A colourful map of “hydrogen highways” drawn over yesterday’s gas mains is still a drawing until someone has paid for steel that will not crack, compressors that can handle the gas, and customers who will buy it.

Uncle’s verdict

Water pipes carry water.

Gas pipes carry cooking gas.

Hydrogen pipelines carry hydrogen gas to factories. Same idea as a gas pipe, fussier molecule, tested steel and different compressors.

Useful next to industry.

Not a sticker on the town main.

— Engineering Uncle

Sources

• US Department of Energy: Hydrogen Pipelines

• IEA: Global Hydrogen Review 2026, Demand

• FNB Gas: Germany’s hydrogen core network

• NL Times: Netherlands opens first 32 km hydrogen pipeline (21 May 2026)

• NEUMAN & ESSER: Hydrogen transport with pipelines

• edie: Hydrogen blending and the HyDeploy trials

AEO FAQ

Paste at the end of the article, above Read next.

Frequently asked questions

Can natural gas pipelines carry hydrogen?

Partly. Most natural gas pipelines can carry a small hydrogen blend; the US Department of Energy says blends up to about 15% may need only modest changes. Pure hydrogen usually requires new pipe, or existing pipe that has been tested for hydrogen embrittlement and refitted with new valves and compressors.

What is a hydrogen pipeline?

A hydrogen pipeline is a steel or plastic pipe that moves hydrogen gas from where it is produced to industrial users such as refineries and fertiliser plants. It is not a water pipe. The US has about 1,600 miles of them, mostly on the Gulf Coast.

What is hydrogen embrittlement?

Hydrogen embrittlement is the loss of toughness in steel when hydrogen atoms enter the metal, especially at welds and small cracks. Under repeated pressure changes, cracks grow faster than they would with natural gas, which is why each pipeline’s steel must be tested before it carries hydrogen.

Does the same pipe carry less energy with hydrogen?

A little less. Hydrogen holds about a third of the energy of natural gas per cubic metre but flows about 2.8 times faster, so a converted pipe moves roughly 85% of the energy. The compressors must handle about three times the gas volume.

How much does a 20% hydrogen blend cut emissions?

About 7%. Blend percentages are measured by volume, and hydrogen carries only a third of the energy of natural gas per cubic metre, so a 20% blend replaces only about 7% of the energy and roughly 7% of the carbon dioxide from burning the gas.

Why not send hydrogen everywhere by pipe?

Leaks, steel embrittlement, new compressors and a lack of committed customers make long national networks expensive. Germany’s 9,040 km core network is budgeted at about €18.9 billion. Many users are better served by making hydrogen on or near the site that uses it.

Read next

Paste at the very end. Replace each LINK with the post’s Substack URL.

Read next on EngineeringUncle

If this one clicked, these are the other hard limits on moving energy around:

Coming up Saturday 26 Sep: How Combined-Cycle Gas Turbines Reach 60%+ Efficiency, the power plants that burn the gas these pipes carry. Subscribe so it lands in your inbox.