The most useful number a founder will read this month came out of a securities filing, not a launch. SpaceX's S-1, filed May 20, disclosed that Anthropic is paying $1.25 billion every month through May 2029 for exclusive access to one data center — Colossus 1 in Memphis, 220,000-plus Nvidia GPUs and 300 megawatts of power — for a total north of $40 billion. Google has its own SpaceX compute deal at roughly $920 million a month. These are the pipes your product runs through, and the S-1 finally put a price tag on them.
Here's why that tag matters more than any token-price headline. The number to fixate on isn't the $40 billion total. It's the shape: a fixed commitment, for three years, that does not move when a more efficient model ships next quarter.
The back-of-envelope#
Take the lease at face value. $1.25 billion a month across 220,000 GPUs is about $5,700 per GPU per month — roughly $7.80 per GPU-hour, all in. That number looks high next to a spot H100 because it isn't just silicon; it bundles the 300 MW of power and the whole sited, interconnected facility. That bundling is the point. What a lab is buying — and what a fixed lease through 2029 actually reserves — is capacity, not chips.
The chip was never the scarce thing. Power, and GPUs that sit close enough together to train as one machine, are the scarce thing — and Anthropic just locked a building's worth of it up at a fixed monthly price until 2029.
What it does to the prices you budget against#
Now connect it to the thing you actually see: the token price war. This week Gemini 3.6 Flash undercut token prices again. The intuitive reading is "compute is getting cheaper, and the savings are being passed to me." The lease says otherwise. If the underlying compute is committed at a fixed $1.25 billion a month regardless of model efficiency, then a price cut isn't a cost reduction flowing downstream — it's demand acquisition funded out of already-committed capacity and investor cash. The bill is sunk. The discount is a land-grab.
That's not a cynical flourish; it's the difference between a cost curve and a pricing strategy. Costs that are contractually fixed for three years don't fall on the schedule that headline prices do. We've made the narrower version of this point before — the price fell and the bill rose once usage climbed — but the S-1 adds the supply-side half: even the provider's own cost isn't dropping the way the sticker suggests.
What a solo founder should actually do#
Three moves, none of them exotic:
- Price your unit economics at today's rates, not projected discounts. If your margin only works after tokens get 10x cheaper "on schedule," you don't have a margin — you have a bet on someone else's capex. The demand-side price war is real, but it's a subsidy, and subsidies are policy, not physics.
- Keep your prompts and routing provider-portable. The lab whose economics look best today is not guaranteed to look best in 2027, and a fixed $40B commitment is exactly the kind of pressure that eventually reprices. A thin routing layer is your insurance; our model routing map is a starting point.
- Treat every free or below-cost tier as temporary. It exists to acquire you during the land-grab. Build so that its withdrawal is a line-item change, not a business-model crisis.
The bullish story about AI costs is that efficiency compounds and prices fall forever. The S-1 doesn't refute that — models really are getting more efficient — but it shows you the counterweight in black and white. Somewhere under the token price you're quoted is a Memphis warehouse drawing 300 megawatts, invoiced at $1.25 billion a month until 2029, whether or not the next model is twice as efficient. Budget like you know that floor is there. It is.
