Anthropic's 900B compute bill arrives before the revenue

The big AI numbers are promises, not money spent.

Anthropic's prospectus landed and I went looking for the valuation, like everybody else. The number that stopped me sits further down, in the note listing what the company has already agreed to buy. Then Databricks put two European figures out the same day, and those are years away too.

Advertisement

1. Anthropic has promised $895.4 billion for compute

The terms: 825 million Class A shares at $70 to $78, up to $64.4 billion raised, a target valuation near $1.33 trillion. Alongside it, Google, Nvidia and AMD are putting $26 billion in through a private placement.

Further down sits the commitments note: $895.4 billion of purchase and lease obligations, with $447.5 billion of it falling before the end of 2030. Four fifths of the company's revenue comes from pay-as-you-go API usage rather than subscriptions.

$69.8 billion is due inside the next twelve months.

We ran the one-line version on 8 November, when the listing was still something Anthropic was aiming at.

The week around it is worth reading next to the filing. On 9 November Anthropic shipped Claude Fable 5.2 and Mythos 5.2, reported as a 20% cut to token prices ahead of the listing. The S-1 itself went in on 26 October and named Accenture, METR and SecureBio as outside evaluators for its safety assessments.

I do not read $895.4 billion as confidence. It is a cost base already fixed in contracts, carried by a seller that just cut the price of the line generating most of its money, and that line is metered rather than committed. If I were renewing an enterprise agreement with a lab in this shape, I would ask for the rate card in writing before the listing rather than after it. A listed company answers to a different set of people about margin than a private one does.

Read the source →

Our research · Two in three launches, partnerships and executive changes we read are not on Crunchbase's cards

Of 50 real launches, partnerships and executive changes, 33 were not on the company's Crunchbase event cards, and a news search on the name missed almost half.

Read the research →

2. Databricks named $850 million for Britain and €500 million for the Netherlands

Three commitments in one announcement:

• More than $850 million in the United Kingdom, over the next three years.
• €500 million in the Netherlands, with a plan to double its headcount there.
• With the Danish consultancy Valcon, an on-premises, air-gapped platform called the Disconnected Data Platform.

The third one carries no date, and it is the only one a buyer can test.

On the same day a consultancy walked through the deepened Microsoft arrangement, with Azure Databricks reaching further into Power BI, OneLake, Purview and Copilot, and Commvault named Databricks alongside Nvidia and Zscaler in its recovery work for AI systems.

The €500 million will be read in the Netherlands as a win, and the air-gapped platform is the part I would put in a contract. A three-year country commitment gets revised when a quarter goes badly and nobody publishes the revision, whereas a platform that runs with the outbound connection cut is a property you own and can check on a Tuesday afternoon. My guess is it sells mostly to the people who were told they could not put Databricks anywhere, which is a narrower market than the announcement implies and a far stickier one.

Read the source →

Advertisement

3. Alphabet paid more to borrow in August than in February

Bessemer Trust published its read of how the buildout is being financed, and one detail is worth keeping. Alphabet's $20 billion bond issue in February 2026 needed spread concessions of 7 to 13 basis points to clear. Its $25 billion issue in August needed 25 to 27.

Same issuer, same year, roughly double the concession.

A spread concession is the extra yield a borrower hands over to get the paper away. It is what the bond market charges for doubt, priced by people who have to settle the trade.

Spreads are the dullest figure in today's letter and the only one with a counterparty standing behind it. A survey can be wrong about sentiment, but a buyer who holds out for a wider spread has put money where the opinion is, which is why I track the coupon rather than the capital expenditure announcements. Announcements get reannounced. If the next large hyperscaler issue clears tighter than August's, the funding worry was seasonal, and if it clears wider, every multi-year commitment above needs rereading with the dates in front of you.

Read the source →

Advertisement

Today's context on 7wData

• Has the Cohere and Aleph Alpha merger closed? Cohere and Aleph Alpha signed a definitive merger agreement on 16 September 2026, but no source reports the close: the deal still needs regulatory approval, and the Series E led by Schwarz Group is mostly secured, not closed.
• Is Cyera raising $800 million at a $20 billion valuation? On 9 November Calcalist reported that Cyera, the data security company, is in talks to raise about $800 million at a valuation of about $20 billion.

Also today

• IBM and Microsoft launched a joint AI agent platform. (1 min read)
• Adobe is wiring its enterprise agent into ChatGPT, Gemini, Microsoft Copilot, Claude and Databricks Genie. (1 min read)
• Dreeze came out of stealth with $33 million for recovering cloud applications after an outage. (4 min read)

Reality CheckThe speedup nobody measured

Your coding agent tells you it is done. In most teams that line is the closest thing to a productivity measurement anybody holds, and it is a self-report from the party being assessed. The one controlled trial I can find points the other way: sixteen experienced developers took longer with the tools than without, while believing they had gone faster. The caveat is as important as the finding, and it is in the piece below.

What was your cycle time in the month before you switched the agents on?

My call: by 30 June 2027, at least one listed software company reports a measured change in developer productivity from AI coding tools, with the baseline and the method, in a quarterly report or an annual filing.

Read the deep dive →