Good morning, AI builders. Anthropic just filled in the numbers investors were waiting for: about $17 billion in third-quarter revenue, up 1,127% from a year earlier. Its filing also lists $895.4 billion in purchase and lease commitments, nearly all of it for compute.
That's the clearest public look yet at what frontier AI costs to run, and those costs eventually shape the prices you pay for tokens.
Welcome to the 14,152 builders reading today. Forwarded this? Join here.
The 30-second brief:
Anthropic's numbers: $17B in Q3 revenue, against $895.4B in commitments
Marvell buys Lumentum: a $115B all-stock bet that AI data centers go optical
Goldman pulls back vibe coding: Claude Code and Copilot are now for engineers only
Build with AI: run DeepSeek's open-source coding agent on a real repo
ANTHROPIC
Source: Anthropic's amended S-1, via TipRanks and BigGo Finance. Q1 is derived from the nine-month total.
The scoop: Anthropic's amended IPO filing reports about $17 billion in third-quarter revenue and a $23.29 billion net loss, narrower than $66.3 billion in Q2, per Reuters.
Key details:
Growth is slowing fast: revenue rose 47% from Q2, after 144% the quarter before. Nine-month revenue hit $33.24B.
The loss is mostly accounting: about $94.3B of the $104B nine-month loss was a noncash charge tied to convertible notes and derivatives.
Compute bit in Q3: a $974M operating profit in Q2 became a $567M operating loss, as pricier short-term compute and a $2B noncash donation-matching charge landed.
Still unconfirmed: no price range is set. TipRanks points to a late-November Nasdaq listing under "ANT," raising about $100B at a $1.5T to $2T valuation.
Our take: The number to watch isn't the net loss, it's the $567M operating loss. Anthropic has $146.6B in liquidity against $895.4B in commitments, so its whole story now rests on turning expensive short-term compute into cheaper long-term capacity before growth slows further. If you build on Claude, expect pricing and rate limits to follow that math.
MARVELL × LUMENTUM
Source: Marvell and Lumentum deal announcement, via The Next Web and Yahoo Finance
The scoop: Marvell agreed to buy laser and optical-networking maker Lumentum in an all-stock deal valued at $115B including debt. Bloomberg calls it 2026's largest deal by equity value.
Key details:
Terms: 4.5 Marvell shares per Lumentum share, pitched as $1,132 a share, a 23.9% premium. Marvell holders would own 67%.
The logic: copper is hitting physical limits as AI clusters grow. The companies cite a $550B market and over $1B a year in cost savings.
The wait: closing is expected in the first half of 2028, pending shareholder and regulatory approval.
Our take: Stock deals move with the buyer's shares: at Monday morning's $190.43, 4.5 Marvell shares were worth about $857, not $1,132. Investors seem to like the optical thesis more than the price, and with closing more than a year away, they have plenty of time to reprice it.
BUILD WITH AI
Fix a failing test with DeepSeek's open-source coding agent
What you'll make: a working fix for a real bug in your repo, using DeepSeek Harness (dsh), an MIT-licensed agent you run locally. It's a low-cost way to compare agents while compute prices are in the news.
Step by step:
Install Node.js if you need it, and create an API key in your DeepSeek account.
Run
npx @deepseek-ai/dsh web. The web UI opens at http://127.0.0.1:3080.Connect your key, open a repo folder and paste: "Run the test suite, find the first failing test, explain the root cause, then propose the smallest fix."
Read its explanation before accepting anything, then ask: "Re-run the tests and add one test that would have caught this bug."
Review the diff yourself, then commit it on a new branch.
Pro move: Run big jobs off-peak. One breakdown lists deepseek-flash at $0.30 in and $1.20 out per million tokens at peak, and half that off-peak. Check DeepSeek's current rates first; dsh is a developer preview, so expect breaking changes.
GOLDMAN SACHS
The scoop: Goldman has restricted advanced AI coding tools, including Claude Code and GitHub Copilot, to engineers, according to Bloomberg, which cited internal guidelines from October.
Traders had spent months using them to monitor trading flows and create content, and the change reportedly frustrated some staff.
A spokesperson cited "productivity and risk management." In July, CIO Marco Argenti was praising vibe coding by non-technical staff.
Our take: The first enterprise vibe-coding wave is hitting governance, so if you sell agents to regulated companies, your pitch now needs audit trails and access controls as much as speed.
FROM THE CHANNEL
Anthropic's IPO filing and the GPT-6.1 Terra launch, explained
Watch for the backstory on today's lead, plus a look at GPT-6.1 Terra and the latest Qwen 4 leaks.
⚡ Rapid fire
🏢 SUSE is cutting about 10% of its roughly 1,900 staff, with its CEO's memo citing "automation and AI."
💰 Tutor Intelligence raised $75M in an a16z-led Series B to train warehouse robots on data from their own completed tasks.
🔬 MIT CSAIL researchers had GPT-6 Astra edit model weights directly, lifting a small Qwen model's MATH-500 score from 7.9% to 48.1% by changing 0.016% of them.
🏛️ South Korea's 2027 budget sets aside 21.3T won for chips, data centers, physical AI and AI, plus 4.6T won for a top-tier sovereign model.
👀 Unconfirmed: GPT-6.2 Astra and Claude Mythos 5.5 may be close, based on leaks and developer posts about Codex traffic and Anthropic's site code. Neither company has announced them.
🧰 Tool drop
LongCat-2.5: Meituan's new 1.6T-parameter multimodal model for long agentic coding runs, available by chat and API.
GDPx: SemiAnalysis's benchmark of 11 high-value domains, where even the best model scores just 12%.
DeepSeek Search and Fetch API: web search and page retrieval for agents, listed at $0.75 per 1,000 successful requests per tool.
What Do You Rate Today's Newsletter?
That's the brief. See you tomorrow at 6am.
The World of AI team