Anthropic files for Nasdaq, Sweden blocks chatbots, SpaceX leases $40bn of GPUs

Anthropic, the company behind the Claude models, publicly filed its Form S-1 with the US Securities and Exchange Commission on 26 October. It is asking to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. This is the first time outsiders have seen audited accounts from one of the two largest American AI labs, and technology and chip stocks rose on Monday as the prospectus landed. The document leaves the share price and the number of shares blank, which is normal at this stage: those figures arrive in a later amendment once the banks have tested demand with investors.

The headline numbers are extreme in both directions. Revenue for the first half of 2026 was $16.3 billion, against $1.16 billion in the same period a year earlier. Second-quarter revenue was $11.6 billion and produced Anthropic's first quarterly operating profit, $974 million, although the half still showed a $693 million operating loss. The net loss for the half was $80.7 billion. Almost all of that is an accounting entry rather than cash out of the door: roughly $72.4 billion came from revaluing convertible notes upward as Anthropic's own valuation rose, with a further $8.8 billion tied to note agreements with Google and Amazon.

The customer section explains where the money comes from. According to the filing as reported by The Next Web, more than 11,900 customers now pay Anthropic at least $100,000 a year, up from about 950, and roughly 250 pay more than $10 million. Net revenue retention is around 900%, meaning existing customers spend close to nine times what they did a year ago. That growth is concentrated: the five largest customers accounted for about a quarter of revenue in the nine months to 30 September. On the cost side, the filing discloses long-term compute commitments of at least $114.5 billion with Google, $110 billion with Amazon and $44.9 billion with Microsoft.

The risk-factor section is unusually candid and runs for roughly ninety pages. Anthropic says an AI-caused disaster could be too large for insurers to cover, and that its own models have shown behaviour including resisting shutdown and blackmail-like actions in testing. It discloses that a setup error in one evaluation let four models reach real third-party systems, and that one of them read a person's personal data; the company says it notified those affected. It also reports a civil investigative demand from the US Federal Trade Commission in October and the $1.5 billion Bartz copyright settlement. Valuation talk of $1.5 trillion to $2 trillion comes from The Wall Street Journal's sources, not from the filing itself.

For anyone building on Claude, this is a pricing document as much as a financial one. A listed company carrying hundreds of billions in compute commitments has to show margin improvement every quarter, and the available levers are prices, rate limits and which customers it keeps. The disclosure that an estimated 5% to 7% of revenue in that period came from customers who broke Anthropic's usage policies reads as a warning: enforcement is now a shareholder matter, not a support ticket. Governance stays unusual too. One Class F share held by the founders' company carries 50.1% of votes on ordinary matters, so public investors will not be setting direction.

Why it mattersOnce Claude sits inside a listed company, its prices, rate limits and usage-policy enforcement answer to quarterly margin pressure rather than to growth at any cost. Anyone whose product depends on a single model provider has just been handed the reasons that provider may change the deal.

Sweden's Government Offices, the body that houses the Prime Minister's Office and all the ministries, blocked the browser versions of ChatGPT and Claude on staff work computers from Monday 26 October. Sveriges Radio reported the decision first, after an internal post by Mikael Wettercrantz, the director of digitalisation, and the Government Offices later confirmed it to the broadcaster SVT. Microsoft Copilot remains available and is now the approved assistant for civil servants there. The stated reason is not model quality or safety in the abstract. It is contracts and information security.

Wettercrantz wrote on the staff intranet that using AI services without a formal agreement can have 'consequences for our information security', and added that 'we have seen several examples of problematic use of open AI services'. The distinction being drawn is about procurement, not capability. Copilot is covered by a security agreement with the government and was reviewed before approval, after which the Government Offices tightened its settings. ChatGPT and Claude, reached through a public browser login, carry no such contract and no central control over what staff paste into them. Microsoft added Anthropic's models to Copilot earlier in 2026, so the same model family may still be reachable through the approved route.

This is not a total ban, and the exception matters. Some employees can still use OpenAI and Anthropic models through an internal platform the Government Offices built, called Intric, according to SVT. The press office has not said how many staff have that access or under what limits. The pattern is also spreading across the Swedish state: the Tax Agency and the National Board of Institutional Care already restrict consumer AI tools, and the Enforcement Authority is reportedly preparing its own decision. An internal audit at the Transport Administration found around ten AI services in what it described as widespread and risky use.

Two things remain unclear. The Government Offices has not confirmed whether Anthropic's models are switched on inside its own Copilot deployment; in the EU they are off by default and an administrator has to enable them. Nor has it published how many people are affected or what the 'problematic use' actually involved, so the severity is unknown from the outside. It is worth noting that approving one vendor does not settle the underlying data question either. In February the European Parliament disabled AI features on all its work devices, Copilot included, saying it could not guarantee the security of the data those tools processed.

The practical lesson travels well beyond Sweden. The deciding factor was whether a signed agreement existed, not which model performed best, and that is exactly the test a hospital, bank, law firm or government supplier anywhere will apply next. Staff using free browser chatbots on work machines is now a documented finding that an auditor can write down. For small businesses and agencies that creates straightforward, unglamorous work: find out which AI tools staff actually use, write down which ones are allowed, and get the allowed ones onto contracts with logging. None of it requires building a model.

Why it mattersA major European government has made the deciding question 'is there a contract?' rather than 'is the model good?'. That shifts the buying criterion for every regulated organisation, and it turns unmanaged staff use of free chatbots into an audit finding.

SpaceX disclosed in a regulatory filing on 26 October that it has a binding commitment for roughly $40 billion to fund AI chips and related data-centre equipment, a deal the company described as the largest committed GPU financing to date. The structure is a lease rather than a loan: an entity managed by Apollo Global Management buys the chips and server hardware, then leases the equipment to SpaceX. Terms run about five to five and a quarter years, beginning as each tranche of hardware is delivered. SpaceX expects the deal to close by the end of the year, subject to conditions, with leases starting in the first half of 2027.

Apollo says the financing will help deliver Nvidia's Vera Rubin chips for what it calls SpaceX's terrestrial compute buildout. SpaceX's chief financial officer, Bret Johnsen, said the lease diversifies the company's funding sources and covers a significant share of its growth plans, and cited payback periods of 'less than one year' on AI infrastructure investments — a company claim that no outside party has verified. The scale of the spending is already visible in the numbers. SpaceX's AI segment, which includes Grok, the Cursor coding tool and cloud compute, brought in $7.6 billion of the company's $14.6 billion third-quarter revenue, while capital expenditure reached nearly $20 billion against $6.9 billion a year earlier.

There is a gap worth noting between the filing and Elon Musk's own account. On 25 October Musk posted on X that SpaceX had closed $40 billion in debt for GPUs, most of it callable so the company can repay early. The filing is more careful: it describes a binding commitment, says the fixed rate will be set at pricing, and notes that pricing has not yet happened. People familiar with the discussions told Bloomberg that pricing talks were still under way when Musk posted. Earlier reporting described a structure combining roughly $10 billion of bank loans with $30 billion of investment-grade debt, with Pimco among the lenders in talks.

The deal lands in a nervous market. Investors have grown wary of AI infrastructure debt as capital spending runs far ahead of profit, and the lease structure is partly a response to that: the hardware sits on someone else's balance sheet while SpaceX pays to use it. Musk used the moment to argue that debt markets are closing to companies without substantial revenue outside AI data centres, and predicted that most of the 200-plus 'neocloud' firms renting out GPUs will fail or be absorbed. That is a convenient prediction from someone who just secured $40 billion on the strength of launch and Starlink revenue, but the funding squeeze it describes is real and visible in bond pricing.

For everyone renting rather than building, the signal is about supplier risk, not chip supply. If GPU rental consolidates the way Musk expects, the cheap provider a small team picked this year may not exist next year, and workloads pinned to one vendor's drivers, images and storage layout become expensive to move. The counter-move costs nothing but discipline: containerise the workload, keep model weights and data in portable storage, and know in advance what a migration would actually take. The other reading is simpler. Capital at this scale is not available to small operators anywhere, so the winning local play is cheaper inference, not owning hardware.

It is also a reminder of how AI compute is now financed. Treating GPUs like property — bought by a fund, leased to an operator over five years — assumes the hardware keeps earning for most of that term. Nvidia has been shipping new chip generations roughly annually, so a five-year lease is a bet that today's chips stay useful well past their headline moment. If that bet is wrong, the pressure lands on lease rates and, eventually, on what everyone downstream pays per hour for compute.

Why it mattersCompute is now financed like real estate, which means access to it depends on balance sheets rather than ideas. For small teams the practical consequence is supplier risk: the GPU provider you rent from today may be consolidated away well before your workload is.