Anthropic's IPO filing shows $17 billion in Q3 revenue and a $23.3 billion net loss
Anthropic's amended S-1 reports about $17 billion in third-quarter revenue, a $567 million operating loss and $895.4 billion in purchase and lease commitments. Marvell agrees to buy Lumentum in an all-stock deal valuing it at $115 billion, Google's Mandiant details LLMjacking bills above $100,000 a day, and GitHub will block pull_request_target in public repositories by default from 2 December.
- Anthropic’s amended S-1 puts third-quarter revenue at about $17 billion, roughly 12 times a year earlier, with a $567 million operating loss, a $23.3 billion net loss and $895.4 billion in purchase and lease commitments.
- Marvell agreed to buy optical-components maker Lumentum in an all-stock deal valuing it at $115 billion including debt, one of the largest tech acquisitions ever; it is expected to close in the first half of 2028.
- Google’s Mandiant says attackers using stolen cloud and LLM keys ran up more than $100,000 in a single day at some victims; one actor exploited a LiteLLM privilege-escalation bug, CVE-2026-40217.
- GitHub will block
pull_request_targetworkflows by default in public repositories from 2 December, with two-hour test blackouts on 9 and 16 November. - The New York Times and three Alden-owned papers filed two suits against Brave, alleging its 30-billion-page index, sold to Anthropic, Perplexity, Mistral, xAI and Cohere, misappropriates their journalism.
- Elon Musk says Grok Bot cost “unsustainably high relative to revenue” after subscribers hit weekly limits over the weekend.
Anthropic’s first full quarter in the open
Anthropic’s amended S-1, filed Monday, reports about $17 billion in revenue for the quarter to 30 September, up about 1,127% year on year and about 47% on the second quarter, CNBC reports. Growth was 144% in Q2. The company swung to a $567 million operating loss from $974 million of operating income in Q2, “primarily driven by $5.0 billion of compute under short-term commitments”, according to Business Insider, which also notes about $2 billion of non-cash charges for its donation-matching programme. The net loss was $23.29 billion; CNBC attributes most of the nine-month loss to fair-value changes on convertible notes and derivatives. Nine-month revenue is $33.24 billion. Purchase commitments and leases now total about $895.4 billion, “substantially all” compute and infrastructure. There is still no price range or share count.
Why it matters
The FT reported in October that Anthropic told investors Q3 had an 11% adjusted operating margin, excluding stock-based compensation; on a GAAP basis it lost money. Expect investors to price the IPO on that gap and on sequential growth slowing to 47%. Our S-1 coverage has the first-half figures.
The commitments are the real headline
$895.4 billion is roughly 27 times this year’s revenue so far, and nearly half is expected to fall due by the end of 2030, per Business Insider. Anthropic is betting that revenue keeps compounding fast enough to pay compute contracts written years ahead.
Marvell buys Lumentum for $115 billion
Marvell announced on 2 November that it will acquire Lumentum, which makes the lasers and optical components that link AI data centres, in an all-stock deal with an enterprise value of $115 billion, TechCrunch reports. Lumentum holders get 4.5 Marvell shares each, about $1,132 a share according to the companies’ announcement, and Marvell holders will own about 67% of the combined company. The companies promise more than $1 billion in annual cost synergies. Closing needs both shareholder votes and regulators, and is expected in the first half of 2028. Nvidia invested $2 billion in each company in March.
Why it matters
Moving data between GPUs, increasingly over light rather than copper, is now one of AI’s tightest bottlenecks. One company selling both the custom chips and the optics gives hyperscalers fewer suppliers to play off against each other.
The long close is the risk
Lumentum tells customers to expect 12 to 18 months to close, which points to a long antitrust review; China’s is our guess for the slowest. Broadcom, Marvell’s main rival for custom AI chips, now faces a competitor that also owns the optics.
Stolen AI keys are running up $100,000 days
Google’s Mandiant and Threat Intelligence Group describe “LLMjacking” incidents at organisations in North America, South America and Asia Pacific, with fraudulent charges above $100,000 in a single day and up to $1 million before access was cut. Most attacks started with stolen cloud identity tokens, service account keys or GitHub tokens. Attackers tested harvested keys with max_tokens=1 requests to stay under volume alerts; one used CVE-2026-40217 in LiteLLM’s /key/regenerate endpoint to turn ordinary virtual keys into admin credentials. Stolen access is pooled and resold. One reseller, Jet Night, sold “unlimited” plans for about $60 a month, made more than 500,000 API calls in under eight hours, and passed nearly $1 million in token costs to one victim.
Why it matters
Patch any LiteLLM proxy, cap spend per model, and alert when someone enables an AI service you never turned on. On Google Cloud, the data-access logs that show model calls are off by default.
The key is now worth more than the server
This fits Wiz’s “Black October” count of AI platforms breached for their stored keys, and Sunday’s Trigger.dev disclosure. Spending caps are the control that limits the damage.
GitHub closes the pull_request_target door
GitHub Actions will block pull_request_target events by default in public repositories on GitHub Free, Team and Enterprise Cloud from 2 December, unless an Actions event policy explicitly allows them. Enforcement was previously due 2 November. Until then the rule runs in evaluate mode with warnings, except for two brownouts, 16:00–18:00 UTC on 9 and 16 November, when affected workflows fail. Private and internal repositories and GitHub Enterprise Server are unaffected.
Why it matters
If your public repo labels, comments on or tests fork PRs with pull_request_target, check Insights before 9 November and allow only the workflow files that need it.
A default that should have come years ago
The trigger runs the base repository’s workflow with its secrets, so a workflow that checks out and runs a fork’s code hands those secrets to the fork; that pattern has been behind many CI supply-chain compromises. Opting in per workflow keeps the legitimate uses while making the dangerous one deliberate.
Quick hits
- The New York Times and three Alden papers sue Brave. The suits, filed in Manhattan, target the Brave Search API, which the complaints say is sold to Anthropic, Perplexity, Mistral, xAI and Cohere, plus Ask Brave and Brave’s paywall blocker; Brave hadn’t commented.
- Musk says Grok Bot was too costly to serve. Subscribers found weekly limits maxed out over the weekend; Musk called the cost “unsustainably high relative to revenue” and hasn’t said what the new limits are.
- OpenAI is merging ChatGPT’s Chat and Work tabs. Select users get it first, Codex stays separate, and ChatGPT and Codex will draw on one token allowance.
- Oracle files to sell $10 billion in stock and $10 billion in notes. The stock and notes offerings were filed separately, and neither depends on the other closing.
- SUSE is cutting about 10% of staff. The CEO’s memo says “automation and AI are fundamentally shifting how software companies operate” and that “the requirement for some existing roles has greatly diminished.”
- LiteLLM moves releases to Tuesdays and pilots an LTS line. Version 1.108 gets a one-month long-term-support trial, useful for teams that want fixes without upgrading weekly.
- Vercel Sandbox with Drives starts up to 2.4x faster. Median creation time fell from 1.7s to 0.7s, with no code changes or price change.