The most quietly absurd disclosure of the month was buried in SpaceX’s S-1 filing last week. Anthropic, the AI lab whose stated mission is to build AI safely, has signed a $45 billion compute contract through May 2029 with Colossus and Colossus II, the data centers built originally to power Elon Musk’s xAI and its Grok chatbot. The arrangement comes out to roughly $1.25 billion a month for access to more than 220,000 Nvidia GPUs and over 300 megawatts of power. Either party can walk with 90 days notice, but for now, Anthropic is paying its largest competitor’s data center landlord more per month than most public software companies generate in annual revenue.
Stack this against the other deals Anthropic announced in the last month and you start to see a pattern that is much weirder than any single line item.
Google just put another $40 billion into Anthropic and committed to a long-term Google Cloud spend. Amazon put in another $5 billion, packaged with a roughly $100 billion AWS compute commitment over the deal life. Broadcom signed on to co-design custom AI silicon, sharing a chip pipeline that Anthropic now runs alongside its rented Nvidia capacity. Add the existing capex commitments on top, and the company is sourcing compute, in 2026 dollars, from Google, Amazon, Broadcom, Nvidia direct, and the Musk data center stack simultaneously. The total compute commitment, by the disclosed numbers, lands somewhere north of $85 billion through 2029.
Two years ago, the dominant question in any AI lab’s board meeting was a clean binary: are you a Google company, an Amazon company, or a Microsoft company. The hyperscaler you picked determined your cost structure, your geopolitical exposure, your enterprise sales motion, and roughly half of your strategic direction. OpenAI picked Microsoft, Anthropic picked Google plus Amazon, Mistral picked Microsoft and then half of France. The decision was load-bearing.
That question has now been replaced by something much more boring and much more accurate: where can you get the watts. Not who do you trust, not who do you owe favors to, not who do you ideologically want to be associated with. Where, physically, on the surface of the planet, are 300 contiguous megawatts of data center capacity available in 2026, and what is the dollar-per-token cost to use them. The answer this week was “in two buildings Elon Musk built for Grok,” and so Anthropic wrote the check.
The infrastructure shortage decides everything
The reason this happened is genuinely simple. There is not enough advanced data center capacity in the United States to satisfy the compute demand from frontier AI labs in 2026. Not within a year. Not within two. The grid interconnect queues are years long. The transformers are backordered. TSMC’s CoWoS packaging is the gating constraint on every GPU shipped by every vendor, which is why Jensen Huang spent the holiday weekend in Taiwan, which is the subject of a different brief on this site. Capacity is the bottleneck, capacity is the bottleneck, capacity is the bottleneck, and once you accept that as the operating reality, the bizarre alliance map starts to make sense.
xAI sunk $45 billion into Colossus and Colossus II to train Grok. Grok is good. Grok is also not eating 1 GW of power around the clock. That leaves a lot of fungible watts in the building. Musk’s choice is between letting them sit idle and renting them to someone, and “someone” in this market means one of three or four companies, and one of those three or four companies is Anthropic. The deal is dispassionate in a way that is structurally embarrassing for everyone involved but very easy to write a CFO memo about. The compute is there, the demand is there, the contract has a 90-day exit clause, and the alternative is that Anthropic does not have enough capacity to ship Claude 5.
The same logic explains why Anthropic is on Google Cloud and AWS simultaneously, despite Google and Amazon being direct competitors in nearly every market that buys their cloud services. Each hyperscaler can supply some of what Anthropic needs. Neither can supply all of it. So you stack them.
The implications, calibrated
The first implication is that the next round of AI ecosystem reporting needs to stop treating “X is an Anthropic partner” as if it means anything. Every major infrastructure player is an Anthropic partner. Every major infrastructure player is also an OpenAI partner. The interesting question is the share-of-spend, not the existence of the contract.
The second implication is for the equity story. Anthropic just told investors it will post its first operating profit this quarter on $10.9 billion in revenue. The $85 billion compute commitment is the line that has to be subtracted from that revenue line going forward. Either the revenue keeps doubling, or the unit economics get visibly ugly, or the model gets ten times more compute-efficient. Probably some combination of all three. The IPO speculation is going to live or die on which combination it is.
The third implication is for the geopolitics nerds. The story for the last three years has been “Microsoft and Amazon and Google are jockeying for which AI lab wins, by sponsoring different labs.” The story for the next three years is closer to “all three are landlords, and the labs are tenants who can change leases when the power bill comes due.” That is a structurally different industry than the one the analyst notes are still pricing.
The last implication, and the funniest one, is that Sam Altman’s recent comments about Musk being “an unhealthy person” and Musk’s recent comments about Anthropic’s safety theater both get a little more interesting once you remember that Anthropic is now writing Musk’s company a $15 billion annual check. Tech-feud kayfabe is harder to maintain when one party is the other party’s customer. The next round of subtweets is going to have to step around a wire transfer.
The map of AI alliances was always going to look weird once the infrastructure layer became the binding constraint. It just turns out “weird” looks like Anthropic and xAI sharing a parking lot.