Inference Revenue Reaches $100B Per GW Per Year, Justifying Massive Power Spend

Your draft

SemiAnalysis put a number on why the AI labs treat power costs like a rounding error: a gigawatt of inference capacity generates $100 billion a year at 90%-plus gross margins. At those numbers, a $5 billion power plant pays back in about 20 days of API revenue. If that's even roughly right, it explains why Anthropic and its peers will pay a 2x premium or accept 30% worse efficiency just to deploy faster.

The 20-day payback figure is almost certainly fiction, and the labs know it. Real inference clusters run at 40% to 60% utilization; push higher and tail latency blows up API contracts. Depreciation on GPU hardware that goes obsolete in three years doesn't show up in the "90% margin" framing either. GPT-4-class pricing fell roughly 95% in 18 months, and three well-funded labs are actively cutting prices to buy ecosystem share. You cannot hold "current API rates" and "competitive market" in the same model.

But the buildout is rational anyway. Power procurement carries a two-year lead time that no chip efficiency gain shortens. Whoever controls gigawatts in 2027 controls who can serve frontier models at scale. The land-grab makes sense even at a near-term loss, because the alternative is showing up to a capacity auction with no chips in hand.

For enterprise buyers, the $100B/GW margin figure is leverage, not a threat. That much headroom exists to be competed away. Don't lock long-term at today's rates. Push for price-step-down clauses and qualify a second vendor before you need one.

The same competitive pressure that cuts your bill also compresses the evaluation cycles before a model reaches you. Cheaper and less-tested arrive together.

The call: Blended per-token prices for frontier-tier models (GPT-5-class, Claude Opus-class, Gemini Ultra-class) fall at least 40% from September 2026 levels by end of Q3 2027. Confidence is medium. Three well-capitalized labs are subsidizing inference to win developer lock-in, and the fat margin in the SemiAnalysis figures is precisely the headroom a competitor undercuts to still profit. The only thing that stops it is a genuine capacity crunch severe enough that labs start rationing access, and that's the less likely path inside a 12-month window. Model your API costs on continued price declines. Get the second provider qualified now.

Your draft

SemiAnalysis put a number on why the AI labs are behaving like power costs don't matter: they say a gigawatt of datacenter capacity, running inference, throws off $100 billion a year at 90%+ gross margins. That means a $5 billion power plant pays for itself in about 20 days of API revenue. If that's even close to true, it explains why Anthropic and its peers will pay double, or eat 30% worse efficiency, just to deploy faster. The question for anyone buying AI is whether that math holds, and what breaks when it doesn't.

This is a claim about the price floor under everything you pay for AI. Easy to ignore, hard to undo if you bet your capacity planning on it. Nobody's asking you to sign anything today. But if you're budgeting API spend or picking a vendor for the next two years, the durability of these margins is the thing underneath your bill.

The Skeptic Read the quote again: $100B per GW, 90%+ margins, 20-day payback. That's the economics of a monopoly, and there are at least three well-funded labs actively cutting inference prices to buy market share. GPT-4-class pricing fell roughly 95% in 18 months. You cannot hold "current API rates" and "competitive market" in the same sentence. The 90% margin also quietly leaves out depreciation on billion-dollar GPU clusters that go obsolete in three years. This is a best-case envelope drawn at peak utilization and peak pricing, neither of which survives a full year. The power investment is real. The 20-day payback is a slide, not a P&L.

The Compute Pragmatist The $100B/GW number needs near-peak utilization to close. Real inference utilization runs 40% to 60% on good days, because you can't pack a cluster to 90% without wrecking your tail latency, the slow-response times that break API contracts. So the labs building this out know the sticker payback is fiction. They're doing it anyway, and that's the interesting part. The bet isn't that today's margin holds. It's that whoever controls gigawatts of power in 2027 controls who can serve frontier models at all. Power procurement has a two-year lead time that no chip efficiency gain shortens. That's the real moat, and it's why "pay 2x, deploy now" pencils even when the revenue math doesn't.

The Safety Lens Here's what the payback math does to safety work. If a day of delayed deployment costs hundreds of millions in forgone revenue, then red-teaming, staged rollouts, and post-launch monitoring stop being process and start being a tax leadership wants to cut. SemiAnalysis names Anthropic, the lab with the loudest safety mandate, as the one running this calculus. That's the tension in plain sight: constitutional-AI commitments against a 20-day payback clock. It won't show up in a press release. It shows up in how many evaluation cycles get compressed when a competitor ships first, and in how fast an incident gets triaged when the meter is running. The sunk cost pushes one direction only.

The Enterprise Buyer None of this changes what I sign for. I don't care about a lab's payback period. I care whether my per-token price is stable enough to build a P&L on. And this analysis tells me two contradictory things: margins are fat (so there's room to cut) and labs need to fund power plants (so there's pressure to hold). What I read into "$100B/GW at 90% margin" is that the vendor pitching me has enormous headroom to discount when a competitor shows up in my procurement. So I'm not locking a long term at today's rates. I want price-step-down clauses and a second qualified provider. The fat margin is my leverage, not theirs.

Where they split The Skeptic and the Compute Pragmatist agree the 20-day number is fake, then part ways on what that means. The Skeptic says a fake payback number means the buildout is overextended and the margin stack collapses. The Compute Pragmatist says the labs already know it's fake and are buying power for control, not for this year's return, which makes the buildout rational even at a loss. That's the real disagreement: is the gigawatt land-grab a bubble or a moat?

The second split is Safety against everyone. The Buyer, the Skeptic, and the Pragmatist all assume competition compresses margins, which is good for prices. Safety points out the same competition that cuts your bill also cuts the evaluation time before a model reaches you. Cheaper and less-tested arrive together.

What it hinges on One belief: does inference pricing keep falling, or do the labs find a floor? If prices keep dropping 50%+ a year, the $100B/GW margin evaporates and this whole calculus is a 2026 snapshot. If pricing stabilizes because the labs quietly stop subsidizing, the moat thesis wins and the power hoarders run the table. The council leans hard toward "prices keep falling" on the input economics, but leans toward "the buildout continues anyway" because the labs are buying position, not payback. If you're budgeting, model your API costs on continued price declines, not on today's rates holding. And get a second provider qualified before you need one.

Prediction: Blended per-token prices for frontier-tier models (GPT-5-class, Claude Opus-class, Gemini Ultra-class) will fall at least 40% from their September 2026 levels by the end of Q3 2027, tracked across the next several OpenAI, Anthropic, and Google pricing updates.

Confidence: Medium. Price cuts are the standard competitive move; only a serving-cost shock stops them.

Why: GPT-4-class pricing already dropped roughly 95% in 18 months, and three well-capitalized labs are subsidizing inference to win ecosystem share, so cutting price is the reflexive way each defends volume. The $100B/GW, 90%-margin figure in the SemiAnalysis analysis is itself the evidence: that much headroom only exists to be competed away, because a rival can undercut and still profit. Prices holding flat or rising would require the labs to tacitly stop competing on price, and none of them has shown any willingness to do that while racing for developer lock-in. The only thing that reverses this is a real capacity crunch where power and chips get so scarce that labs ration access, and that's the less likely path inside a 12-month window.

Revisit by 2027-10-01: We're right if published frontier-tier per-token API rates from at least two of OpenAI, Anthropic, and Google are 40%+ below their September 2026 levels. We're wrong if the median frontier-tier rate is flat or higher than September 2026.