Two buyers, one price: Patel says 2027's new compute is already half spoken for

Two of Patel's per-megawatt figures do not sit comfortably together. He puts Anthropic's peak at $50M per megawatt against a deployment cost of $10-15M ^[5][4], a multiple of roughly three to five ^[1], where serving GPT-4 on Hopper had run at negative gross margin ^[6]. He then gives an end-2027 base case of $70-80M per megawatt blended across the whole company ^[18]. But only about 40% of a lab's compute budget serves inference, with the rest in training and research ^[14]. Divide the blended figure by that share and the serving megawatts have to be clearing $175-200M each ^[2]. Either the mix swings hard back toward inference, or those two numbers are not measured on the same base.

The floor is easier to pin down. Anyone can stand up open weights, Kimi on vLLM or SGLang, at $10-15M per megawatt and clear positive revenue through OpenRouter ^[10], so a megawatt has a clearing price that owes nothing to what a frontier lab earns on it. Patel's own arithmetic needs that price to move to $25-50M for the two labs to hold 100 GW in 2028 ^[9][11], which is 67% to 233% above the top of today's base ^[4]. Price then tracks who can pay most at the margin rather than what the capacity cost to build.

What breaks the story is not demand. From the middle of 2026, marginal revenue per new megawatt flattened even as absolute capacity kept climbing ^[16], and Patel reads that as marginal megawatts going into research rather than serving ^[15]. For a buyer of tokens that is the worse reading: compute absorbed into internal R&D does not appear as available capacity at any price. The shelving of Astra and of Anthropic's Model 2 works the same way ^[17]. The models that would justify $100M per megawatt are the ones not being shipped.

Supply will not rescue the residual buyer either. New York State has banned data center construction and Texas has paused approvals, with Ohio making data centers pay property tax on surrounding land ^[20], and the tooling chain upstream cannot be widened on a two-year horizon even with capital thrown at Carl Zeiss ^[22]. The silicon going in during 2026 runs three to five times the performance per watt of the last generation ^[12], so half the physical megawatts is more than half the usable FLOPs. Meta and SpaceX can build with no end customer and then choose between using the capacity and reselling it at $25-50M per megawatt ^[19].

Every number here is Patel's, relayed second hand in a summary of the conversation ^[24]. The structural claim survives that. In 2024 and 2025 the hardware chain took nearly all the gross margin while the model layer ran negative ^[21], and compute was at least sold to whoever turned up with money. On these figures, within one to two years most of the world's compute is owned by or serving two labs ^[3], and everyone else trades in what they leave.