AI Briefing · Friday, October 16, 2026

The day in brief. Three threads today. Anthropic's Q3 figures ($16.8B, up from $11.5B) come from an unaudited investor briefing; the 900% NRR claim is unverified, and 5–7% of revenue came from rivals distilling its models. Model companies are buying security capability: Cohere signed an LOI for AISLE. Small-model prices keep falling: The Batch reports Haiku 5.5 at a tenth of Haiku 4.5's price. A question worth asking in diligence this week: what happens to gross margin and pricing if inference costs fall 10x again?

Top of mind

• Anthropic's numbers are big and unaudited. An FT-reported investor briefing puts Q3 revenue at $16.8B, up from $11.5B; the 900% NRR claim has no outside confirmation. 1
• Cohere moves into AI security. Cohere signed an LOI for AISLE, whose self-reported 500+ CVEs can be checked against OpenSSL and curl advisories. 2
• Small-model prices fall again. The Batch reports Haiku 5.5 at $0.10/$0.50 per million tokens, down from $1.00/$5.00; its 72.4% OSWorld score is Anthropic's own number. 6
• Capex is leaning on debt. Noah Smith cites research putting AI investment at 3.63% of GDP versus railroads' 2.24%; Nscale's IPO at up to $35B faces delays and penalties. 7 5

Must read

1. Anthropic reportedly told investors Q3 revenue reached $16.8B

Simon Willison · Exponential View (Azeem Azhar) · 1 report + 1 hands-on test + 1 commentary · source

Developing story · day 2 · first in the briefing Oct 15

• What happened: Anthropic reported Q3 revenue of $16.8B (Q2: $11.5B), with 50% quarterly growth per FT sources citing Anthropic's investor briefing; distillation by rivals contributed 5–7% ($500M+ at Q2 levels) and was curbed in Q3 (Simon Willison). Leaked briefing claims revenue exceeded 2026 forecast, annualized exit at $76B with unverified 900% NRR (Exponential View).
• Evidence check: Holds. Anthropic's unaudited pre-IPO figures via FT are the source of all numbers; distillation math checks out but relies on Anthropic's own data. Unverified claims like 900% NRR lack external confirmation.
• Our read: The distillation disclosure matters more than the headline: if 5–7% of revenue came from rivals harvesting outputs, some growth was competitors, not end customers. We would wait for audited S-1 figures before using the $76B run-rate as a comp.
• The AGI-forward view: Revenue rising from $11.5B to $16.8B in one quarter, while distillation was being curbed, is customers paying for capability; demand compounds as the models improve.
• Sources: Simon Willison: Anthropic reports Q3 revenue at $16.8B with 50%… · ft.com via Simon Willison: Anthropic says model distillation contributed 5-7% of revenue · Exponential View (Azeem Azhar): Anthropic Q3 revenue $16.8B exceeds 2026 forecast

2. Cohere proposes acquisition of AISLE for sovereign AI security

Cohere / Cohere Labs · primary source · source · 2026-10-15

• What happened: Cohere signed a letter of intent to acquire AISLE, citing 500+ CVEs across 100+ codebases, including all 12 in OpenSSL's January release and 18 of 24 in curl's. Vlcek leads Cohere Cyber; Baloo becomes company CISO; Fort, VP AI Security.
• Evidence check: Holds. Cohere's own post confirms a proposed AISLE deal; AISLE's CVE record is self-reported but checkable against OpenSSL and curl advisories.
• Our read: A checkable CVE record is a better acquisition asset than a demo, but an LOI is not a close, and we would want AISLE's false-positive rate before calling it a moat.

3. NVIDIA releases Nemotron 3.5 Super VL and Bolt models for edge deployment

NVIDIA (blog) · primary source · source · 2026-10-15

• What happened: NVIDIA released Nemotron 3.5 Super VL (120B open model) with frontier accuracy at 1/5 size of larger models, plus two edge-optimized Bolt models: Bolt 74B-A9B (24GB memory) and Bolt E4B (4GB). All cover agent deployments from enterprise to edge.
• Agenda: NVIDIA blog announcing its own product; sells chips
• Our read: Open 120B and 4GB-class models from the chip vendor may help vertical agent startups that need on-device or on-prem deployment without frontier API bills. We would want independent evals before accepting "frontier accuracy", a claim from the GPU seller.
• The market view: NVIDIA gives away capable open models because cheaper models mean more deployments on its hardware; commoditising the model layer pressures labs that price by API.

Worth knowing

• 4. Microsoft unveils MAI-Base-2 with 2.8T total and 80B active parameters. Microsoft claims MAI-Base-2 is a MoE model with 2.8 trillion total parameters and 80 billion active parameters, stating it achieves 4x-12x more pretraining FLOPs-efficiency than DeepSeek-V4-Pro-Base. (microsoft.ai via Techmeme)
• 5. Nscale delays IPO amid Norway penalty and Microsoft/ByteDance disputes. Nscale faces a six-month delay in Norway, owes penalties to Microsoft and ByteDance, and is preparing for an IPO at up to $35B valuation. (bloomberg.com via Techmeme)
• 6. Anthropic releases Claude Haiku 5.5 with new pricing and claimed gains. Anthropic released Claude Haiku 5.5 (The Batch), offering prompts up to 100,000 tokens at $0.10/$0.50 per million input/output (vs $1.00/$5.00), claiming it covers ~90% of Haiku 4.5 traffic and achieves 72.4% OSWorld performance (vs 15.7%), ahead of GPT-6 Luna on some tasks. (The Batch (DeepLearning.AI), The Batch (DeepLearning.AI))
• 7. AI data center spending outpaces railroads in GDP share. Noah Smith cites Van Nieuwerburgh (2026) showing AI investment at 3.63% of GDP vs railroad boom's 2.24%, with some sources claiming up to 6% for AI including all infrastructure; hyperscalers now borrowing heavily as cash flow turns negative. (Noah Smith)
• 8. US open-weight models close gap with Chinese counterparts. US open-weight models now have a performance gap with Chinese models that is at least halved compared to February 2026, reducing the time difference by approximately three months. (Interconnects (Nathan Lambert))

Recommended reading

Anthropic reports Q3 revenue at $16.8B with 50% quarterly growth · Simon Willison · hands-on test

Why it matters: Willison works through the distillation arithmetic: how much of Anthropic's revenue came from rivals harvesting outputs, and what curbing it did to Q3. That changes how to read lab revenue as a demand signal for application startups.

In brief

• White House AI task force starts 120-day evaluation period (semafor.com via Techmeme)
• Chamath Palihapitiya: 70-80% of AI company revenues from other AI companies (x.com via All In podcast)
• Facebook releases competitive US open-weight AI model (Mark Riedl)
• Meta's Muse Wonder rivals top models (The Batch (DeepLearning.AI))
• Ramp and Stripe build internal vibe-coding platforms for non-engineers (The Pragmatic Engineer (Gergely Orosz))
• Benchmark says AI erodes venture pact against backing rivals (techcrunch.com via TechCrunch)
• Ed Zitron on AI economic reality and potential bubble (Ed Zitron)
• Joanna Bryson warns against using generative AI in direct operations streams (Joanna Bryson)

Open questions

• Will Anthropic's audited filings confirm $16.8B Q3 revenue and break out distillation revenue and NRR? Lab revenue sets the comps and demand assumptions behind AI startup valuations. Settled by: An S-1 or audited financials with revenue by customer type. 1
• Does Haiku 5.5's 72.4% OSWorld score hold up in independent testing? Cheap, capable computer-use agents would reset unit economics for agent startups. Settled by: Third-party OSWorld runs or independent index results. 6
• Do AISLE's CVE credits match OpenSSL and curl advisories, and does the Cohere LOI close? Verified findings would make AI security a product, not a demo. Settled by: Advisory credits matching AISLE's list; a signed definitive agreement. 2

40 items analyzed from 20 sources · 61 filtered as off-topic · 21 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. Sources are weighed by how they know: a primary source, a report or a hands-on test establishes a fact; analysis, opinion, summaries and posts are commentary, however many there are. Generated 2026-10-16 06:50 EDT.