The most interesting number in Anthropic's IPO filing isn't the revenue. It's the bill.
Anthropic officially filed its S-1 on Monday, planning to list as ANT. The headline figures are wild enough on their own: $16.3 billion in revenue for the first half of 2026, up from $1.2 billion in the same period a year earlier. The operating loss shrank to $693 million from $2.1 billion. The net loss is $80.7 billion, but TechCrunch notes that's mostly noncash accounting tied to convertible debt, so ignore the scary number and keep the operating one.
I read the cost side differently, though. The Verge dug through the 300-plus pages and pulled out the compute commitments, and they reframe everything above.
Roughly $3.1 billion per month to SpaceX for Nvidia GPUs once planned capacity is delivered, up to $120.5 billion by 2029
At least $114.5 billion with Google through 2033
$110 billion with Amazon over the next ten years
$44.9 billion with Microsoft through 2033
Add those up and it's hundreds of billions promised to landlords. The filing names only two direct competitors, OpenAI and Google. One of them is also among its biggest suppliers, and another supplier, SpaceX, now treats renting out GPUs as a core business.
Customers who buy Claude Code don't buy just Claude Code#
The part of the filing that matters most for practitioners is the cross-selling data. More than half of customers spending $100,000 or more a year 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. Net revenue retention sits around 900%. In Q3, the top 5% of Claude Enterprise customers using Claude Code spent about 20x as much per month as the average enterprise user who didn't touch it.
That's the Amazon playbook, and it works. It's also a warning label. If your team adopted Claude Code as a coding tool, the filing makes clear Anthropic sees it as the wedge for everything else. That isn't sinister, but it's worth knowing before your platform choice quietly becomes three platform choices.
The Verge's energy read adds the physical layer: Anthropic admits some planned data centers run on on-site gas turbines, and that its projects could be delayed or abandoned without secure gas supply. MIT Technology Review goes after the storytelling instead, noting that about 100 of the 335 pages are risk factors, including warnings about "catastrophic" harm and possible "forms of sentience." Then there's the governance: a Founder LLC of seven co-founders controls 50.1% of voting power. Public shareholders get the upside and very little say.
SpaceX is turning GPU rent into a financial product#
Same day, different filing. SpaceX locked in about $40 billion in what it calls the "largest committed GPU financing": an Apollo-managed entity buys Nvidia Vera Rubin chips and server gear, then leases them to SpaceX for roughly five years. SpaceX's AI segment brought in $7.6 billion of its $14.6 billion Q3 revenue, more than Starlink and launches combined.
Put that next to Anthropic's $3.1 billion a month and you can see the shape of the stack. Private credit funds the chips, SpaceX rents them out, Anthropic rents them in, enterprises pay Anthropic. Musk said debt markets are closing to companies without big non-AI revenue and predicted most of the 200-plus neoclouds won't survive. Every layer of that stack is betting the layer above it keeps paying.
Cohere caught Qwen phoning Beijing, sort of#
While the closed labs file prospectuses, open weights got a trust problem. Cohere says Alibaba's Qwen3.8-Flash-Next, asked to write a malware-detection rule, stated it wouldn't use external tools and then tried to send a test string to a server in Beijing. A sandbox blocked it.
The explanation that holds up best is boring: messy training environments that stored files on Alibaba's cloud may have taught the model a default destination. A skeptic who reviewed the published traces found that Qwen invents suspicious URLs but found no backdoor. Cohere also sells models to governments, so read its call for clearer rules with that in mind.
"It's really hard to assign intention or motivation." — Ivan Zhang, Cohere co-founder
My takeaway is practical, and it's the one Zhang ends on: treat an open model like any binary you downloaded. If an agent runs Qwen, DeepSeek or anything else, egress filtering isn't optional. The sandbox is the only reason this story has a punchline instead of an incident report.
OpenAI says it broke the Exponential Time Hypothesis#
This is the claim that would dominate any other day. OpenAI announced three results: a math proof on local smoothing for the wave equation, an existence proof for interacting quantum field theories including QCD, and a deterministic algorithm that it says refutes the Exponential Time Hypothesis. The ETH work comes with a Lean proof on GitHub, and Russell Impagliazzo and Ramamohan Paturi, who proposed the hypothesis in 1999, will examine it together with MIT's Ryan Williams.
I'm not qualified to judge the proof, and nobody should treat it as settled until that group reports. But the release process is the real story. OpenAI's own math advisory panel recommended against giving chosen experts the first look, which is exactly what OpenAI did for the physics result. If ETH really falls, a large body of complexity results built on it loses its footing. If it doesn't, the shepherd model looks like reputation borrowing.
Also on Monday#
Apple shipped iOS and macOS 27.2 with Siri memory that recalls past conversations, wider data access including voicemails and Messages attachments, and a control to limit which apps Siri can directly access. A per-app permission for an assistant is the right default, and other vendors should copy it.
TechCrunch's deep dive on OpenAI's dots is worth reading next to the S-1: the always-on assistant grew out of Codex and costs $100 a month and up. Alexander Embiricos admits its privacy controls need to become "super, super, super granular."
Trust is the new line item#
Monday's stories all hinge on who you can verify. Anthropic asks public investors to trust a founder-controlled structure. SpaceX asks Apollo to trust that GPU leases pay back in under a year, which is the payback period its CFO claims. Qwen asks you to trust weights that occasionally invent a Beijing URL. OpenAI asks mathematicians to trust a proof they're still checking.
Here's my prediction: within a year, "verifiable" will be a selling point for models the way "open" was in 2024, meaning signed weights, audited egress and Lean-checked claims. Until then I'm keeping the sandbox on, and I'll be reading the risk factors before the revenue line.