The Economics of AI Compute

These are distilled notes from Dylan Patel of SemiAnalysis, in conversation with Dwarkesh Patel — go to the original for the full three hours. Everything below is his argument, restated tightly; where a number is his own extrapolation rather than a filed fact, it says so. The reason to read it here is that almost every claim maps onto a name DeadRisk already covers.

The frame is a relay of bottlenecks. The big four — Amazon, Meta, Google, Microsoft — carry roughly $600 billion of forecast capex this year; across the supply chain the figure approaches a trillion. At current rental prices that money buys on the order of 50 gigawatts of compute, arriving over several years, not this one. The constraint that binds keeps moving: first cooling, then data centers, then power, and now — the part almost nobody had priced — chips and memory.

What matters is that each prior bottleneck could flex. Cooling and construction could borrow labor and capacity from adjacent industries. The last two cannot. There is no idle wafer capacity to slide over from phones, and no spare memory fab waiting to be switched on. That is the whole story of the AI-capex loop seen from the supply side: the money is not the scarce input anymore.

The number that governs everything is roughly $10–13 billion to rent a gigawatt of AI compute for a year. Once you hold it, every capex headline becomes checkable: an announcement is just a claim about gigawatts, and a gigawatt has a price.

Run it on Anthropic. At ~$20 billion of annualized revenue and sub-50% gross margin, the implied compute bill is ~$13–14 billion a year — the output of roughly $50 billion of someone else’s capex. To keep adding revenue at the pace it has been, Patel estimates it needs about four more gigawatts of inference capacity this year alone, with its training fleet held flat.

So the labs’ giant raises stop looking strange. A single ~$30–110 billion round is not paying for a model; it is pre-buying gigawatts — locking in compute and price years ahead. The value of holding the best model for a few months is that it lets you sign those deals before the crunch prices them. It is the same reasoning CoreWeave and IREN sit inside from the other side: whoever contracts the power early owns the cheap position.

The bear case, Michael Burry’s, is an accounting one: if a GPU’s useful life is really two or three years, not the five the hyperscalers depreciate over, then true amortized capex is far higher than reported, and the whole cloud economics look worse.

Patel’s counter is the interesting part. In a supply-constrained world a chip’s value is not set by “what could I buy today at the same price” — which would make older chips decay as faster ones ship. It is set by what the best model that can run on it is worth. Because GPT-5.4 serves more tokens per H100, of higher value, than GPT-4 ever did, he argues an H100 is worth more today than it was three years ago.

That is a falsifiable claim, and the honest way to hold it is backward: it predicts filings should be extending useful-life assumptions, not shortening them, and secondary-market Hopper prices should hold or rise. Watch the record, not the forecast — it is the quality-of-earnings question underneath Nvidia, and exactly the split between price and worth that Graham’s method was built to force.

A smaller wrinkle rides along. As the fixed cost of compute climbs, the relative gap between a frontier model and a cheaper one narrows — so buyers rationally trade up. Patel notes the price differential between the top tier and the mid tier has already compressed as the underlying GPU-hour rose.

Two labs, two strategies. Anthropic stayed conservative — its own founder said he would not go crazy on compute for fear of a revenue inflection at the wrong moment. OpenAI signed aggressively, and by year-end has materially more compute secured.

The cost of being caught short is concrete. To acquire compute in a pinch you go to lower-quality providers you would have skipped, or you take it through a hyperscaler — served on Bedrock, Vertex, or Foundry with a revenue share — paying a markup you would not have owed had you bought early.

The mirror image is a margin locked in by whoever moved first. Spot Hopper has inflected toward $2.40 an hour against a ~$1.40 amortized cost; a five-year deal signed early now sits on a large embedded spread. This is why CoreWeave’s disclosed >90%-on-3-year-plus contract book is a feature, not a risk — it cannot reprice up, but it cannot be crowded out either.

Follow the margin up the stack. The labs cannot hold pricing power — they are capacity-constrained and competing. The clouds cannot flex price much either — most of the book is multi-year. The pricing power sits with the chip and memory vendors, and under them with the one node nobody out-negotiates. Nvidia is carrying roughly $90 billion of long-term contracts and negotiating three-year memory deals; that is a cornered resource in the textbook sense.

There is a deliberate move inside it. Patel argues Nvidia fractures its allocation across many neoclouds on purpose, so no single customer accumulates enough volume to negotiate against it — the same tactic the labs run on their own data-labeling suppliers. The equity stakes read less as financial engineering than as a strategy to keep the customer base too split to push back.

Here is the bottleneck that cannot flex. Nvidia is already the largest customer at both TSMC and SK Hynix, so there is no adjacent capacity to borrow from PCs and phones. The chip supply is the wall the money runs into.

The arithmetic is stark. A gigawatt of Rubin-class compute needs, on Patel’s math, on the order of two million EUV lithography passes — about 3.5 EUV tools’ worth. ASML makes ~70 tools this year, perhaps ~100 a year by 2030 even under aggressive expansion, stacking to ~700 tools by decade’s end. That caps total AI-chip capacity near 200 gigawatts a year — against a stated 50-gigawatt-a-year ambition from a single lab. (The ceiling is his extrapolation of ASML’s own disclosed trajectory, not a fixed fact.)

And ASML cannot simply build more. The tool rests on a ~10,000-person, single-source supply chain — Zeiss optics, Cymer’s tin-droplet light source, overlay tolerances below a nanometer — that nobody was “AGI-pilled” enough to pre-fund. The monopoly is itself bottlenecked by component makers who have not tried to expand. It is the strongest structural argument in the whole conversation, and it sits on top of TSMC.

A third of 2026 big-tech capex now goes to memory alone. The mechanism is demand destruction: AI’s appetite for HBM and KV-cache is pulling DRAM and NAND away from consumer electronics, and because memory fabs take two years and none were started in 2023–25 (the vendors were losing money then), the crunch cannot be supply-fixed before 2027–28.

The consumer eats it. Patel walks the phone math: an iPhone’s DRAM cost roughly triples, from ~$50 to ~$150, adding perhaps $250 to the sticker once NAND and margin pass through — and low-to-mid-range smartphone volumes could fall from ~1.4 billion toward 500–600 million within two years, because that segment has the thinnest margin to absorb it.

Cheap memory cannot substitute. An HBM stack moves ~2.5 terabytes a second against 64–128 gigabytes for equivalent commodity DDR — the reason accelerators cannot fall back to the cheap part even where a workload would tolerate the latency. This is the gap DeadRisk covers least today, and the one the record is about to fill: Micron, SK Hynix, Samsung.

The power wall shows up here too. There are only three makers of combined-cycle gas turbines — GE Vernova, Siemens Energy, Mitsubishi Heavy — and their order books run years out. So builders reach past them: reciprocating engines from the likes of Cummins, more than a dozen makers of gas power-generation gear, anything that can be stood up behind a fence line faster than the queue.

The scramble is visible in the record. Google, once it saw its own Gemini demand inflect, went — in Patel’s words — absurdly aggressive: buying an energy company, putting deposits on turbines, buying up powered land. Nebius runs cargo-ship engines for a Microsoft build; SoftBank Energy, which had never built a data center, is building them for OpenAI. It is the same move The Power Wall tracks for Bloom Energy: when the grid will not come in time, you make your own power and dare the county to say no.

The financing scare of last year now reads as a dress rehearsal. OpenAI signed deals the market said it could not pay for; Oracle and CoreWeave stocks wobbled, credit markets tensed on the fear that the end buyer could not fund the compute — and then the labs raised, and the fear reset. The circularity is the point: chips fund the clouds that buy the chips that train the models that raise the money.

The lens that fits is Dalio’s debt cycle. A buildout funded out of cash flow and a buildout funded on debt are different objects with different failure modes, and his template asks the one question that separates them: when the revenue inflection comes late, who is forced to sell? For now the answer is nobody — the money keeps arriving. The record’s job is to notice the quarter that stops being true.