AI Briefing · Wednesday, October 14, 2026

The day in brief. Three threads matter today. First, frontier-lab economics are now on the record. Per Reuters, Anthropic's IPO documents show it spent $12.6B to earn $4.6B in 2025, and compute alone came to 159% of revenue. Second, the FT found that OpenAI's $70B ARR counted cloud-partner sales; the net figure was about $50B. Third, Meta's Muse Wonder Full now ranks #2 on Artificial Analysis, which gives app-layer companies another frontier supplier. Diligence ask this week: gross versus net revenue, and gross margin after inference cost.

Top of mind

• Lab losses are now in filings. Per Reuters, Anthropic spent $12.6B against $4.6B of 2025 revenue. Compute alone was $7.33B, and the operating loss was $8B. 1
• ARR needs a gross/net check. The FT found OpenAI's reported $70B ARR included cloud-partner sales. The net figure was about $50B. 2
• Meta is back near the frontier. Artificial Analysis ranks Muse Wonder Full #2. It sits behind Claude Opus 5.5 and ahead of Gemini 4 Argon and GPT-6 Astra. 5 9
• Acqui-hire leftovers end in court. Character.ai filed for Chapter 11 to run a court-supervised sale, two years after Google took its core engineering team. 4

Must read

1. Anthropic's 2025 financials show $8B operating loss

A reader · source · from Tue Sep 29

• What happened: Per Reuters, citing IPO documents: Anthropic spent $12.6B to earn $4.6B in 2025 revenue. $7.33B of that spend was compute, for an $8B operating loss. Zitron says losses are worsening year over year but gives no prior-year figures.
• Agenda: Ed Zitron (economics/financing bear) reporting on Reuters' findings
• Evidence check: Mixed. The loss figures are consistent and sourced. Compute alone is 159% of revenue. 'Dog' and 'worsening' have no trend data.
• Our read: We read compute at 159% of revenue as subsidised pricing that cannot last. Stress-test every app-layer company's gross margin at 2x today's token prices; a 2026 filing with compute below revenue would change our view.
• The AGI-forward view: Revenue of $4.6B with compute as the dominant cost is what scaling looks like. If capability keeps compounding, the $8B loss buys the lead that captures the market.

2. Ed Zitron questions OpenAI's $70B ARR claim

Ed Zitron · source · 2026-10-13

• What happened: Axios reported OpenAI's ARR as $70B and the FT confirmed it. The figure was grossed up to include cloud-partner sales and was really $50B. When the FT asked at the time, OpenAI called the $70B a "net" metric.
• Agenda: Ed Zitron is a finance/economics bear who calibrates for downward slant on AI bubble claims
• Evidence check: Mixed. The FT-sourced gross-versus-net finding holds. The jump from one restated metric to "entire AI bubble" is overstated.
• Our read: Gross-versus-net is the revenue game we expect most often in seed and Series A AI companies that resell. Ask for net revenue recognised under accounting rules alongside any ARR figure, and treat a refusal as the answer.
• The market view: Even net, $50B is enormous. Sales through cloud partners are real customer demand and a distribution strength, so the restatement changes the multiple, not the thesis.

3. Meta CFO reportedly warns AI capex will be 'really bad' if wrong

Ed Zitron · source

• What happened: Meta CFO warns AI capex will be 'really bad' if strategy is wrong (Ed Zitron). Zuckerberg admitted spending capex without knowing ROI on an unnamed podcast, following a Meta tech guy reporting CFO's 'hey this better fucking work' offsite comment (Ed Zitron). Ed Zitron sarcastically mocks an unnamed company making $200B AI bet with no clear purpose while CFO warned of financial consequences (Ed Zitron).
• Evidence check: Meta CFO warning and Zuckerberg admission are reported via Zitron's paraphrase without direct quotes, source links or numbers. The $200B claim is unverified sarcasm.
• Our read: We discount this story heavily: no direct quotes, no numbers, no links. The useful point survives: hyperscaler capex is discretionary, so ask any infra-dependent startup what happens to its costs and customers if Meta-scale spending pauses.
• Sources: Ed Zitron: Meta CFO: AI Capex Will Be Bad If… · Ed Zitron: Zuckerberg admits AI investment uncertainty post-CFO warning · Ed Zitron: Ed Zitron on CFOs and AI Capex Risks

4. Character.ai files for Chapter 11 bankruptcy after Google acquisition

Kashmir Hill · source · 2026-10-13

Developing story · day 2 · first in the briefing Oct 13

• What happened: C.AI (Character.ai) filed for voluntary Chapter 11 bankruptcy to facilitate a court-supervised sale, following Google's $2B+ acquisition of its core engineering team in 2024.
• Agenda: Company announcing its own bankruptcy filing (Character.ai)
• Evidence check: Holds. Character.ai confirms it filed for Chapter 11 to sell itself; the $2B+ figure and "shell" label are Hill's framing.
• Our read: Google paid for the team; the company left behind could not stand alone. Ask any founder what preferred and common holders get if a lab licenses the tech and hires the team. That exit path is now routine.

5. Meta releases Muse Wonder models; Artificial Analysis ranks them near top

The Rundown AI · Simon Willison · Superhuman AI · 3 sources · source

• What happened: Meta released Muse Wonder and Muse Wonder Full (available via Muse Code), with Artificial Analysis ranking Muse Wonder Full #2 on its Intelligence Index, behind Claude Opus 5.5 (53 points) and ahead of Gemini 4 Argon, GPT-6 Astra, and Fable 5.1 (The Rundown AI). Muse Wonder beats Opus/Astra on legal/long coding/memory; Opus leads health/computer use/coding (The Rundown AI). Meta claims Muse Wonder is 'GPT-6.1 equivalent' based on Artificial Analysis's score, but Willison notes progress via pelican drawings (Simon Willison).
• Evidence check: Mixed. #2 ranking confirmed by third-party index; GPT-6.1 parity claim from Artificial Analysis not Meta's own assertion; long-term-memory strength unverified as Meta's internal claim.
• Our read: Another frontier supplier helps app-layer companies that route across models: more pricing leverage, and a new option for legal and long-coding work. Ask portfolio teams to run their own evals; an index rank is not their workload.
• The AI-realist view: One aggregate index is not capability. The 'GPT-6.1 equivalent' framing rests on a single score, and the long-term-memory strength is Meta's own unverified claim.
• Sources: The Rundown AI: Meta's Muse Wonder ranks second behind Claude Opus… · Simon Willison: Meta releases Muse Wonder models with GPT-6.1 equivalent… · Superhuman AI: Meta's Muse agent tops App Store, new models…

Worth knowing

• 6. ASML raises 2026 forecast, says it can't meet 2027–28 AI demand. ASML states it cannot fully meet AI-driven demand for chipmaking tools in 2027 and 2028, with strong interest in orders for 2029-2030 (Reuters). ASML reported Q3 net sales of €11.66B (vs. €11.72B est.), net profit of €4.96B (vs. (reuters.com via Techmeme, cnbc.com via Techmeme)
• 7. Cohere releases dataset on agentic tool usage patterns. Cohere released a dataset (Agentic Task Ecosystem) of 694,411 public MCP tools from 123,969 listings, showing only 2.6% map cleanly to ONET tasks across 178 occupations. (Thu Oct 8)* (Cohere / Cohere Labs)
• 8. OpenAI in early talks to acquire security firm Trail of Bits. OpenAI is in early talks to buy Trail of Bits, a vulnerability-finding and fixing company founded in 2012 with over 40 employees. (theinformation.com via Techmeme)
• 9. Gemini 4 Argon model released at $2/million input. Gemini 4 Argon launched with $2/million input and $10/million output pricing; set to double on Dec 1, costing $2,407 for Artificial Analysis benchmark vs. GPT-6 Astra's higher cost. Developing story · day 2 · first in the briefing Oct 13. (Simon Willison)

Open questions

• Do Anthropic's IPO documents show compute cost per dollar of revenue falling from 2024 to 2025? Shows whether lab losses come from scaling or are structural; this drives future API pricing. Settled by: Prior-year figures in the IPO filing, or Reuters follow-up reporting. 1
• Will OpenAI disclose net revenue, and do other labs also count cloud-partner sales in ARR? Sets the honest revenue baseline for AI comps and valuations. Settled by: Audited or IPO-disclosed net revenue; FT or Reuters follow-ups on other labs. 2 1
• Does Gemini 4 Argon's price doubling on Dec 1 signal a broader rise in frontier pricing? App-layer gross margins assume flat or falling token prices. Settled by: Pricing moves by OpenAI, Anthropic and Meta before year-end. 9 1 5

38 items analyzed from 15 sources · 74 filtered as off-topic · 23 more stories in the dashboard. Analysis: claude_code:default, local:qwen3:30b. Our read, the other views and the day in brief are machine-written. Generated 2026-10-14 06:37 EDT.