AI demand grows 10x a year; supply is booked through 2028

1. Bottleneck south of the model — ▶ 1:12

   Every dramatic new way of computing needs new infrastructure, and this one needs all of it: new chips, new system software, new ways of doing power, and a replacement for copper. Infrastructure now runs "all the way down to the mines." Over the last 3 years the models stopped being the bottleneck and now improve faster using AI itself, so the constraint sits "south of the model." Founders have noticed. Hardware was roughly 3-5% of deals from top founders and is now north of 20-30%.

2. Demand is not hype — ▶ 4:23

   Hyperscaler capex is about $700 billion this year and heads toward $1 trillion next year, from the buyers who see labs, AI natives, enterprises and foreign demand at once. Chip prices are rising, which has never happened before. Supply is booked out to 2028, there are multi-day auctions for a few thousand GPUs, and GPUs resell at 4x purchase price. The dot-com comparison fails: dark fiber was laid on speculation with too few users to fill it, whereas here every GPU is already pre-sold.

3. Nobody could have built it — ▶ 8:59

   A CFO found the memory in his company's servers had appreciated enough to fund the entire cloud migration he had resisted. A leading memory vendor said at Hot Chips that today's demand alone, not future demand, takes 3 years of capacity to supply. Even with a perfect oracle, chip cycles run 3-4 years and data centers 4-5 years plus power, and the boom is only 4 years old. An industry used to 20-30% growth is now tied to AI software growing at triple digits.

4. Tokens grow 10x a year — ▶ 11:57

   Scaling now runs through inference: reinforcement learning, chain of thought and long-running agents all burn tokens. A chat takes about 100 tokens; an agent takes thousands. AI is also used to build AI, as with GPU kernels. The audience widens in steps: casual chat with a billion weekly users, then 30 million developers, then over 1 billion knowledge workers, then back-office agents and robots. Token demand is put at "close to 1,000% a year", later "10x a year", and supply cannot grow that fast.

5. Money replaces engineers — ▶ 14:32

   In software, a 2-year lead was safe because a rival hiring 1,000 engineers would wreck its own company. That is the mythical man-month. Now "taking $3 billion and lighting up a magnificent cluster" works, and Grok or Kimi can arrive from nowhere. Any problem can be solved with enough GPUs, power and money. There is no natural regulator between money in and capability out, so "until we run out of problems, we're not going to run out of demand." Supply is the only limit.

6. Agents as employees — ▶ 20:55

   The product idea moved from AI as a search bar, to chat, to an extension of you holding your keys, to an employee with its own computer and browser. It triages email and checks before acting. Horowitz is blunt about the downsides: agents burn tokens for nothing, forget, make things up and create security problems, but can also be super productive. The goal is making humans superhuman, not automating the firm. Treating agents as people has been "the most durable way" to bring them in.

7. Per-model ASIC math — ▶ 26:44

   A frontier model costs $3-5 billion to train. Inference must pay back at least that, say 2x, so $10 billion. A 20% efficiency gain is worth $2 billion, and an ASIC costs about $2 billion. Model weights are fixed, unlike stateful software, so a chip per model could make sense. They admit they do not know if the world goes there. No digital artifact has ever absorbed $5 billion directly. Founders are rebuilding memory, compute, interconnect, power and cooling from fundamentals.

8. Data centers go obsolete — ▶ 28:08

   Rack power is moving from 5-10 kilowatts toward 100 and beyond, with compute density up about 70x. At that level AC stops working, so data centers shift to DC at 800 volts, which needs its own cooling and is dangerous. Only 2% of US electricians are certified on DC power, and Meta trains people for free. Liquid cooling is already standard. Floors must hold more weight, walls must be thicker against noise, and reinforced concrete is among the fastest-rising prices. About 10% of sites waste water or take power without giving back.

9. Power gap and politics — ▶ 34:04

   New data centers need roughly 44 gigawatts by 2028 against about 25 gigawatts of expected grid additions. A gigawatt is about 50,000 homes. Builders need permits, grid access and their own generation, and transformers and turbines are short. "This is not a software problem", and lead times will not compress. New GPU buyers are going to Mexico or Australia, so banning data centers in the US hands jobs abroad. Their proposed standard is no noise, no water issue, more jobs, and power given back. Some sites already do it.

10. Incumbents leave silver bricks — ▶ 39:45

    Any 10x on tokens per watt or per rack needs new innovation, which traditionally comes from founders working from first principles. Incumbents worth multiple trillions are busy with the 90%, and 5% of that value is a massive private company. Nvidia has "so many gold bricks" it will not chase silver ones. Markets fragment as they expand and consolidate when growth slows, as Ford's Rouge River plant gave way to supplier tiers. Margins matter too: unlike software, AI may make hardware optimization central to business upside.

11. Systems founders, often older — ▶ 45:53

    Hardware startups take hundreds of millions in first rounds before any product. Founders must be systems thinkers who architect the chip, plan manufacturing and think about suppliers before design. Jensen is "the Michael Jordan of this." Labs are desperate enough to sign with startups before hardware exists, which was unthinkable 5 years ago. Founders skew older because hardware was defocused by industry and academia for 20 years and "you don't go intern and like build a chip." They expect a SpaceX-style alumni wave, and one partner closes on keeping America ahead.

The limiting factor on AI is no longer intelligence. It is concrete, copper, transformers, turbines and time. The partners say the models stopped being the bottleneck about 3 years ago, and everything south of them now sets the pace. That is why they think the constraint outlasts any single product cycle.

Start with why they believe demand is real. The buyers with the best view are spending the most. Hyperscaler capex is about $700 billion this year and heads toward $1 trillion next year. Those companies see frontier labs, AI-native startups, enterprises and foreign buyers at once. Chip prices are rising, which the partners say has never happened. Supply is booked out to 2028. There have been multi-day auctions for a few thousand GPUs, and GPUs resell at 4x what was paid. They reject the dot-com comparison. Dark fiber was laid on speculation, with too few users to fill it. Here, every GPU is pre-sold.

Supply cannot catch up quickly, because build cycles are long and demand does not wait. Chips take 3-4 years. Data centers take 4-5 years, plus power. The boom is only 4 years old, so even a perfect forecast could not have produced the capacity. A leading memory vendor said at Hot Chips that today's orders alone take 3 years of capacity to fill. Demand is put at 10x a year. An industry used to 20-30% growth is now tied to software growing at triple digits.

Demand compounds because AI improves by consuming more AI. Reinforcement learning, chain of thought and long-running agents are all inference-heavy. A chat takes about 100 tokens. An agent takes thousands. AI also builds AI, as with GPU kernels. The user base widens in steps: casual chat, then coding, then more than 1 billion knowledge workers, then back-office agents, then robots. Agents are being treated as a new kind of employee. They can burn tokens and produce nothing, forget, make things up and create security problems. They can also be extremely productive. Horowitz's firm found that treating them as people was the most durable way to bring them in.

Underneath is a change in what scaling means. In software, a 2-year lead was safe, because a rival hiring 1,000 engineers wrecked its own company. Now a rival can spend $3 billion on a cluster and arrive from nowhere, as Grok and Kimi did. One partner says any problem "can be solved with enough infrastructure." There is no natural regulator between money in and capability out. So the only brake is supply, and supply is physical.

That physical layer is harder than it looks. Racks are moving from 5-10 kilowatts toward 100 and beyond, with density up about 70x. At that level AC power stops working, so data centers go to DC, at 800 volts. Only 2% of US electricians are certified on DC power. Floors must carry more weight. Walls must be thicker because the racks are loud. New data centers need roughly 44 gigawatts by 2028 against about 25 gigawatts of expected grid additions. Builders need permits, grid access and their own generation, and they face shortages of transformers and turbines. Working weekends does not compress those lead times.

The politics are part of the engineering. About 10% of data centers waste water or lean on the grid without giving power back, and the partners expect that to end. Their standard is simple: power gets better, no noise, no water problem, more jobs. A data center runs flat day and night while a city does not, so a site with its own generation can feed the grid by day and draw at night. Meanwhile new GPU buyers are going to Mexico and Australia, because building in the US is so hard.

This is where the fund places its bet. Every layer was designed for a different era and is reaching its physical limits. The arithmetic of a frontier model shows why bespoke hardware now pays. A model costs $3-5 billion to train. Inference must earn perhaps 2x that, or $10 billion. A 20% efficiency gain is worth $2 billion, which roughly buys an ASIC. Weights are fixed, so a chip per model could make sense. On incumbents, the logic is that Nvidia has "so many gold bricks" it cannot pick them all up, so it will not chase silver ones. Markets fragment as they expand and consolidate when growth slows.

The founders have to be systems founders. They architect the chip, plan manufacturing and think about suppliers before they design. They skew older, because complex supply chains reward experience and because hardware was defocused for 20 years. "You don't go intern and like build a chip." First rounds run to hundreds of millions, because money goes in long before there is a product. Labs, desperate for supply, now sign with startups before hardware exists.

The obvious objection is that the evidence is circular. Nearly every signal cited is someone spending money: capex, pre-sold GPUs, resale prices. That shows buyers believe in demand. It does not show that revenue will meet it. The 10x a year figure is stated, not sourced. The ASIC arithmetic opens with "let's say": the 2x payback is assumed, and the partners concede they do not know if the world goes that way. They did not discuss what happens to pre-sold orders if their buyers' growth slows. They also note that fragmentation ends in consolidation once growth slows. Nothing in the conversation says when that is.

What stays unresolved is whether the supply gets built at home. One partner ends by saying that if America loses its lead in technology, the next era belongs to another country with different values. The fund depends on permits, turbines and electricians as much as on founders.