In the space of a few days, the financial architecture of frontier AI became legible. Two confidential S-1 filings, a $35 billion private credit deal, and a $920 million-per-month GPU lease settled into the public record at roughly the same moment. Add a US government equity proposal still taking shape, and the capital stack underpinning large-scale AI is, for the first time, measurable in numbers that institutional investors, operators, and policymakers can argue over.
OpenAI files a confidential S-1 at an $852 billion valuation
OpenAI submitted a draft S-1 to the US Securities and Exchange Commission on 8 June, with Goldman Sachs and Morgan Stanley as lead underwriters. The company announced the filing proactively, noting that "there are things we want to do that are likely easier as a private company" while preserving the option to list sooner if conditions allow. No IPO date has been set; reports point to a possible autumn 2026 window, which OpenAI has neither confirmed nor denied.
At $852 billion, OpenAI would rank among the most valuable companies ever to debut on public markets. The filing came one week after Anthropic's own confidential S-1 at a $965 billion valuation. Both companies have signalled that they are not in a hurry to list; the S-1s buy time, confer institutional credibility, and give public-market investors a first structured view of the economics underlying frontier AI at scale.
For an operator, the near-term consequence is pricing transparency. When the S-1s are eventually made public, the true cost structure of running Claude and GPT at scale, and the revenues required to justify these valuations, will be visible for the first time. Customers who treat AI as a commodity line item should prepare for that data to reset their assumptions.
A $35 billion private credit deal secures Anthropic's compute supply
Apollo Global Management and Blackstone finalised a $35 billion debt financing package for Anthropic, structured across three tranches against a portfolio of Google tensor processing units. Morgan Stanley arranged the transaction; roughly half of the debt was syndicated to third-party investors.
The structure is worth understanding. A special-purpose vehicle borrows the funds, acquires the TPUs, and leases them back to Anthropic, keeping the hardware off Anthropic's balance sheet while giving creditors a secured claim on physical assets. Broadcom provided a residual-value guarantee covering the two senior tranches: $6 billion of A1 notes and $24 billion of A2 notes. If Anthropic defaults and the chips are sold below cost, Broadcom absorbs the shortfall for holders of those two tranches. The remaining $4.5 billion B tranche, unsupported by that guarantee, priced at an 8.5% coupon to compensate for the additional risk.
The mechanism solves a problem that every frontier AI company now faces. Training and inference require hardware that depreciates rapidly, but financing it with equity dilutes investors in companies whose valuations are already approaching a trillion dollars. Any other lab that needs to scale compute without issuing more shares will study this deal as a template.
Google rents 110,000 NVIDIA GPUs from SpaceX for $920 million a month
SpaceX disclosed in an SEC filing that Google has agreed to pay $920 million per month from October 2026 through June 2029 for access to approximately 110,000 NVIDIA GPUs at the Colossus campus in Memphis, the data centre originally built for xAI. A reduced ramp fee applies through September 2026; if SpaceX cannot deliver the full committed capacity by that date, Google retains the right to terminate.
Google described the arrangement as bridge capacity for its Gemini Enterprise agent platform. That framing is notable: Alphabet raised $84.75 billion in its own equity offering earlier this month and is nevertheless renting surplus capacity from a rival's campus to serve near-term agent demand. Supply is that constrained.
The monthly rate, annualised to roughly $11 billion, establishes a pricing benchmark for GPU-class infrastructure at scale. Operators building on managed inference APIs should understand that the capacity they consume is priced at these levels upstream, and that the compute required for sustained agent workloads is scarce enough to command that premium from one of the world's largest technology companies.
The Trump administration opens equity talks with OpenAI
Senior US government officials have held preliminary discussions with OpenAI about the federal government acquiring a stake in the company. Under the structure being discussed, OpenAI would donate equity rather than sell it, seeding something akin to a public wealth fund that would distribute AI-derived returns to American households. Sam Altman has outlined a similar vehicle in policy documents since April 2026, and has discussed the concept periodically with senior administration officials since the start of Trump's second term.
President Trump raised the idea publicly on 5 June. No terms have been agreed. Anthropic is notably absent from these discussions; a February 2026 dispute in which Anthropic declined Pentagon deployments without safety guardrails created lasting friction with the administration. NOTUS separately reported that officials are reviewing whether other AI companies beyond OpenAI might eventually be approached.
The structural implication is direct: if the US government holds equity in frontier AI companies, its regulatory calculus shifts. An administration with an ownership stake has a financial incentive to see those companies succeed, which will shape how any future AI legislation is framed and enforced. For operators assessing long-term platform risk, that alignment between government and the two largest AI providers is a variable that did not exist a year ago.
The pattern across all four stories is the same: capital is converging on frontier AI at a scale now visible in institutional debt markets, sovereign wealth fund proposals, and nine-figure monthly compute leases. The question for any operator is no longer whether frontier AI is expensive. It is. The question is whether the infrastructure you are building on has a durable claim on the compute required to remain at the frontier as demand continues to outpace supply.