Anthropic Bought Compute Where the Chair Was

RUM Group disclosed the deal on August 24 without naming the buyer. The Information named Anthropic on September 13. Roughly three weeks of public silence while the stock moved and the tech press wrote about a 'mystery customer.' That is not how you run an optics campaign. That is how you close a deal under time pressure and hope the name stays buried a little longer.

The consensus read is that this is an awkward political story: Anthropic, the AI safety company whose CEO called for limits on training compute, just cut a nine-figure-per-year contract with the platform that hosts Truth Social. The contradiction writes itself. But the most likely explanation is not ideological compromise. It is capital structure arithmetic.

What the SEC Filing Actually Says

RUM Group said in regulatory filings that it currently lacks the financing required to fulfill the contract. Translation: we signed a $13.7 billion deal without the money to build the thing.

The warrant structure confirms it. In tandem with the commercial agreement, the parties signed a binding term sheet for a warrant allowing the customer to buy up to 50,808,408 Class A shares at $0.01 per share, vesting based on purchase milestones. At RUM's September 11 close of $7.17, that warrant equals about 18% of RUM Group's Class A stock outstanding. You do not price equity at one cent per share when traditional financing is available.

RUM Group had $203.3 million of cash at June 30 with no committed financing for the buildout. The Maysville, Georgia facility was projected at 120MW capacity, with potential expansion to 180MW, expected operational in Q1 2027. A 120MW campus costs well north of a billion dollars before a single GPU rack goes in. RUM Group's cash covers permitting and early concrete.

The Scarcity Math Anthropic Is Running

The most likely explanation for why Anthropic accepted this structure is not that RUM Group was the best partner available. It is that RUM Group was the available partner. Anthropic has signed agreements with several cloud service providers, including Google, SpaceX, and Nscale, with a cumulative computing capacity of at least 14.8 gigawatts and a total cost potentially reaching $517 billion over the next decade. The company is building a compute portfolio the way a CFO builds a bond ladder: different maturities, different counterparties, whatever clears the market.

RUM Group brought something real to the table. The addition of Northern Data brought 22,000 NVIDIA Hopper architecture GPUs to RUM Group, most of which are currently being leased out in the European and American markets. The GPU inventory is actual. The Maysville expansion is the bet.

The third tranche carries a kill switch: before obligations relating to that tranche become effective, the customer must determine the proposed delivery date is acceptable in its reasonable discretion. Anthropic negotiated an exit ramp on the final third of the commitment. That is not the clause of a company with unlimited options.

If you are a CTO running Claude in production at scale, the counterparty risk is now layered. Anthropic's inference capacity is partly contingent on a company that has no committed financing, an unbuilt facility, and a Q1 2027 operational target seventeen months away. The warrant aligns Anthropic with RUM's equity upside. It does not solve the construction timeline.

This deal will not be the last of its kind. When the largest AI safety company on the planet holds warrants in the platform that livestreams White House press briefings, the polite fiction that compute procurement is governed by values alignment dies. GPU utilization beats affiliation. Every time.

What to Watch

First: does Anthropic's next funding round require warrant disclosure, and how does the Trump-adjacent equity position read to institutional LPs eyeing an eventual IPO? Second: does RUM Group close committed financing before Q1 2027, or does the Maysville timeline slip and trigger Anthropic's approval clause? Third: watch whether any other top-tier AI lab files a similar structure with a non-institutional infrastructure partner in the next six months. If they do, the inference becomes a pattern.