Monday, October 26, was the day the AI boom got an income statement. Anthropic published the first detailed public financials from a frontier AI lab, OpenAI answered with a cheaper agent model, and Broadcom, SpaceX and Meta announced three very different ways of paying for chips and data centers.
Key points
- Anthropic publicly filed an S-1 to list on Nasdaq under the ticker “ANT,” with no share count or price range set yet.
- The filing shows $16.3 billion in first-half 2026 revenue and an $80.7 billion net loss that includes $72.4 billion in fair-value changes, mainly on convertible notes.
- OpenAI says GPT-6.1 Terra’s standard API prices are 50% lower than GPT-5.6 Terra’s, and all of its benchmark figures are vendor-reported.
- The Broadcom-Anthropic $63 billion SPV financing, SpaceX’s roughly $40 billion GPU lease and Meta’s up-to-$95 billion equity raise are different kinds of money, so we don’t add them up.
Here’s what happened, what the numbers actually say, and what we’re watching.
Anthropic files to go public as “ANT”
After weeks of leaked drafts, Anthropic announced that it had publicly filed a Form S-1, the registration statement a company files with the US Securities and Exchange Commission before an initial public offering. It plans to list its Class A stock on the Nasdaq Global Select Market under the ticker “ANT.” The share count and price range haven’t been set.
Goldman Sachs and Morgan Stanley are named as the contacts for the preliminary prospectus.
The Associated Press, in a syndicated summary, reports that the filing sets up a Wall Street debut in November. The AP says the offering could raise as much as $100 billion, which would top SpaceX’s record $86 billion IPO in June, and that Anthropic could be valued at as much as $2 trillion. Those are estimates. Nothing about the price is final.
What the S-1 actually shows
The headline numbers, per TechCrunch’s read of the filing:
- Revenue: $16.3 billion in the first half of 2026, up from $1.2 billion in the same period of 2025.
- Operating loss: $693 million, down from $2.1 billion a year earlier.
- Net loss: $80.7 billion.
That last number needs context. TechCrunch says it was largely driven by non-cash accounting adjustments tied to convertible debt. The S-1 says the loss includes $72.4 billion from changes in the fair value of convertible notes, derivative liabilities and SAFEs, mainly the convertible notes, as the company’s valuation rose. In plain terms: when your stock gets more valuable, debt that converts into that stock costs more on paper.
Anthropic still warns that it “may not achieve or sustain profitability in the future.”
The enterprise engine
The customer numbers are where the filing gets interesting. As of September 30, Anthropic counted more than 11,900 customers spending at least $100,000 a year, more than 1,900 spending $1 million or more, and about 250 spending $10 million or more.
A second TechCrunch piece digs into how those customers spread across products. More than half of the $100,000-plus customers now use both the Claude platform and a first-party app like Claude Code or Cowork, up from about 30% at the end of last year. The filing cites net revenue retention of about 900% at the end of the third quarter. Net revenue retention measures how much existing customers spend now compared with a year ago; 900% means, on average, roughly nine times as much.
The filing also says that in the third quarter, Claude Enterprise users who didn’t use Claude Code spent less than $75 a month on average, while the top 5% of users who did spent about 20 times that.
Our read: this is the classic land-and-expand playbook, and on these numbers it’s working. Start with the API, add Claude Code for engineers, then Cowork for everyone else.
The catch: the same filing lists the risks. The cloud companies that resell Claude, namely Amazon, Google and Microsoft, increasingly see Anthropic as a competitor and may prioritize their own needs or rivals they’ve invested in when allocating compute. Anthropic also flags customer concentration: losing one or more of its largest customers could hurt results.
Who controls the company
Governance is unusual, and contested. Anthropic stays a public benefit corporation, a legal form that lets a company weigh a stated public mission alongside shareholder returns.
According to TechCrunch, the seven co-founders (Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah) will hold a “Founder LLC” with 50.1% of shareholder voting power. The Long-Term Benefit Trust, set up in 2023, will appoint most of the board.
TechCrunch notes why that’s controversial: voting power isn’t proportional to economic ownership, so public shareholders will have little practical control over a company that describes its own technology as posing “catastrophic or existential risks to humanity.”
The AP also highlighted the risk factors, which range from the mundane to cyberattacks, misinformation and the possibility that advanced AI could threaten humanity itself. Anthropic presents the IPO as a step toward mitigating those risks.
Why it matters: investors buying ANT will be buying economic exposure, not control. That’s a familiar trade at founder-led tech companies. It’s a less familiar one when the company’s own prospectus warns about existential risk.
OpenAI’s GPT-6.1 Terra halves the price of agents
On the same day, OpenAI launched GPT-6.1 Terra, a fast mid-tier model built for computer use (operating apps on screen), visual reasoning and everyday writing.
Pricing is the headline. OpenAI says Terra’s standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol: $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. Tokens are the chunks of text a model reads and writes, roughly three-quarters of a word each.
These are OpenAI’s own benchmark numbers:
- Terminal-Bench 4.0 (solving problems from the command line): 50.3% at maximum reasoning effort, more than double GPT-5.6 Terra’s 21.5%.
- Agents’ Last Exam (professional computer-use tasks): 15 percentage points above GPT-5.6 Terra at medium effort, at about one-third the cost, and on par with GPT-6 Astra at one-tenth the price per token.
- Error-inducing prompts (an internal test): the error rate at low effort fell from 53.6% to 16.7%.
OpenAI still recommends GPT-6.1 Sol or GPT-6 Astra for the hardest coding work. It also says Terra meets the “Critical” cybersecurity threshold under its Preparedness Framework, OpenAI’s internal risk-rating system, and carries over the safeguards built for GPT-6 Astra.
OrigoTech reports that GitHub made Terra available in Copilot the same day.
Our read: agents that click through real software are expensive to run at scale. Halving the price makes high-volume automation easier to justify, like reading invoices and keying them into accounting software, the example OpenAI uses.
The catch: every benchmark here is vendor-reported, and several are internal. Independent results will tell us more than the launch post.
The money machine: $63B, $40B and up to $95B
If the S-1 showed what AI earns, the day’s financing deals showed what it costs.
Broadcom and Anthropic, $63 billion. Bank of America, Blackstone, Citi and Morgan Stanley announced financing for a newly formed special purpose vehicle (SPV), a standalone company created to hold specific assets and debt. The SPV will buy Broadcom XPU AI accelerators and networking gear and lease them to Anthropic, with deliveries expected to start in 2027. Broadcom is providing residual value support, a promise that backs what the hardware will be worth, for an amount equal to the first-lien portion of the financing. Broadcom’s law firm, Sullivan & Cromwell, describes it as part of the compute capacity agreement Broadcom previously announced. OrigoTech, citing Broadcom’s transaction presentation, puts peak residual-value exposure at $43 billion, declining to zero over five years of lease payments.
SpaceX, $40 billion. SpaceX reached a binding commitment to an approximately $40 billion master finance lease with an Apollo-managed entity for Nvidia Vera Rubin GPUs and related equipment. SpaceX calls it the largest committed GPU financing. Apollo and Goldman Sachs lead the financing, with anchor commitments from funds managed by BlackRock, Goldman Sachs Asset Management and KKR-advised accounts. TechCrunch reports leases of about five to five and a quarter years, starting as hardware arrives in the first half of 2027. SpaceX CFO Bret Johnsen cited “less than one year payback periods” on its AI infrastructure.
Meta, up to $95 billion. Meta plans to raise $35 billion through underwritten equity offerings, then up to $60 billion more through an at-the-market program, which sells shares gradually into the market. CNBC reports the money will support AI infrastructure and superintelligence. On Tuesday, Nvidia and AMD said they would invest $3 billion and $1 billion respectively in the offering.
Those three numbers are big, but they aren’t the same kind of money: debt financing for a leasing vehicle, a lease commitment, and an equity raise whose largest piece is a ceiling that may never be fully used. Adding them into a single total would overstate what was actually committed, so we don’t.
Our read on the financing
The pattern is the story. Buying compute now runs through leases, SPVs and equity raises funded by banks and private credit, with chipmakers sometimes investing in their own customers. Nvidia buying Meta stock while Meta buys Nvidia chips is the clearest example.
The catch: these structures spread risk around rather than removing it. If AI demand slows before the leases run out, someone is left holding expensive, aging hardware. In Broadcom’s case, part of that someone is Broadcom.
TechCrunch also notes the backdrop: a rough stretch in debt markets for AI infrastructure companies as investors grow wary of runaway spending and limited profits. Elon Musk argued that debt markets have largely closed to GPU-rental companies without other revenue, predicting “death and consolidation” among the more than 200 so-called neoclouds. That’s a competitor’s claim, not a market finding.
The human and local side of the build-out
Microsoft offered a voluntary exit package to eligible employees in the US, UK, Canada and India. According to TechCrunch and the AP, about 60,000 people, roughly 27% of Microsoft’s approximately 223,000 full-time employees, are eligible. The AP ties the move to the company’s heavy AI spending. TechCrunch reports terms similar to Microsoft’s standard severance, with some groups, including data center staff, excluded. In the note to employees quoted by TechCrunch, Microsoft calls the offer “new for us.” Chief People Officer Amy Coleman announced it in a blog post, according to the AP.
Meanwhile, the AP reports that towns across Connecticut are pushing back on data centers. East Lyme has imposed an indefinite ban, other towns have passed temporary moratoriums, and state leaders are considering statewide moratoriums. Residents cite pollution, energy use and chemical runoff.
Why it matters: the money is national, but the costs land locally, in power bills, water use and jobs. That’s where the political fight over AI infrastructure is heading.
The takeaway
The takeaway: for years, the AI race was judged on benchmarks and demos. This week it got numbers filed with a regulator. Anthropic’s filing shows real, fast-growing revenue next to enormous paper losses and heavy compute commitments. OpenAI is pushing the price of capable agents down. And the infrastructure bill is being financed in ever more creative ways.
What to watch:
- Anthropic’s price range and share count, and whether the November debut holds.
- Independent benchmarks for GPT-6.1 Terra, and whether rivals cut prices to match.
- Whether more AI hardware deals move into SPVs and leases, and who carries the residual risk.
- How many Microsoft employees take the buyout, and whether Connecticut moves to a statewide moratorium.
Update (Oct 29, 2026): this story’s headline and introduction earlier added the Broadcom, SpaceX and Meta deals into a “nearly $200 billion” total. Those are different kinds of financing, and part of Meta’s figure is a maximum, so we removed the total. We also corrected the S-1’s Claude Code comparison, which is about individual users, not customers, and now cite the filing directly for the $72.4 billion fair-value charge.
Image: ajay_suresh / Wikimedia Commons, CC BY 2.0
.jpg){kind=link}