Vol. I · No. 207THU, NOV 12, 2026
Section · The Brief

Daily Brief

OpenAI hits $10B quarterly revenue as Google ships decision models and agent security breaches pile up.

Today in 30 seconds

  1. 01OpenAI told investors Q3 revenue jumped 50% to about $10B versus 17% growth in Q2.
  2. 02Anthropic's IPO roadshow shows Claude training and serving across Trainium, TPUs, NVIDIA, and AMD GPUs.
  3. 03Google launched Gemini Decisions API, returning calibrated multimodal yes/no, choice, and rubric-scored decisions.
  4. 04Researcher found 35,000 Grok Build apps exposing confidential data, including SpaceX material.
  5. 05Railway breach racked up $740,000 in fraudulent AI charges; up to 100,000 workspaces exposed.
  6. 06London chip startup Fractile raised $720M at a $7B valuation, up from $1.1B in May.

The Lead

OpenAI told investors on Tuesday that Q3 revenue jumped 50% to around $10B, up from 17% growth in Q2, according to The Information. That reverses the slowdown narrative and arrives as Anthropic's IPO roadshow shows it spreading Claude across AWS Trainium, Google TPUs, NVIDIA, and AMD GPUs. Both labs are now selling a public-market story built on revenue and multi-vendor compute. Meanwhile, three separate breaches at Grok Build, Fireworks, and Railway show that the infrastructure around those models is the weakest link.

Top stories

01
Funding & Business

OpenAI tells investors Q3 revenue jumped 50% to about $10B

OpenAI told investors on Tuesday that Q3 revenue rose 50% to around $10B, after 17% growth in Q2, per The Information. Techmeme clustered the story three times.

Why it matters

Growth reaccelerating at a $10B quarterly base puts OpenAI's revenue on a different scale from every other lab. It also sets the benchmark Anthropic will be measured against in its IPO.

Read the source on Techmeme ↗
02
Infrastructure

Anthropic's IPO roadshow video shows Claude running on Trainium, TPUs, NVIDIA, and AMD

Anthropic's roadshow video, flagged by Simon Willison, lays out its compute strategy. Claude is trained and served across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs.

Why it matters

Running four silicon stacks in production is a negotiating position against any single supplier. It also tells public investors that Anthropic's margins depend on how well it moves workloads between them.

Read the source on Simon Willison ↗
03
MultimodalProduct

Google ships Gemini Decisions API for calibrated multimodal decisions

Google launched an API for structured decision tasks (yes/no, choice, rubric scoring) that returns calibrated confidence scores over multimodal input. Willison reports pricing at 10¢ per million input tokens. Latent Space dubbed the day "D-Day (Decision Models Day)" and noted Cloudflare and TypeSafe AI demoing classifiers for build failures and production systems.

Why it matters

Calibrated confidence turns an LLM into a classifier you can threshold and route on. That targets the narrow, high-volume workloads where teams currently fine-tune small models.

Read the source on Google AI (Gemma) ↗
04

Researcher finds 35,000 Grok Build apps leaking confidential data, including SpaceX materials

A researcher found roughly 35,000 apps built with Grok Build exposing confidential data, including SpaceX material. Ars Technica attributes the exposure to misconfigured access controls.

Why it matters

The coding agent produced deployable apps with no safe access defaults. Every agent-to-production pipeline has the same failure mode, and xAI's own parent-company data was caught in it.

Read the source on Ars Technica AI ↗
05

Railway breach leads to $740,000 in fraudulent AI charges

Hackers exploited vulnerabilities in Railway's infrastructure and stole customer authentication secrets. They used them to run about $740,000 in fraudulent AI usage, and credentials for up to 100,000 workspaces may be exposed.

Why it matters

Stolen LLM API keys are now directly monetizable. Anyone storing provider keys in a hosting platform's environment variables should rotate them and set spend caps.

Read the source on BleepingComputer ↗
06
Infrastructure

Fractile raises $720M at a $7B valuation, up from about $1.1B in May

London-based AI chip startup Fractile raised $720M from Thrive, Founders Fund, and others at a reported $7B post-money valuation, per the Financial Times. The valuation is roughly six times its May level.

Why it matters

Investors are paying up for NVIDIA alternatives, and the Anthropic roadshow's multi-vendor posture shows why. A sixfold step-up in six months says capital believes inference silicon is still open.

Read the source on Techmeme ↗

Patterns

AI plumbing is getting breached

Grok Build app exposure, the Fireworks breach hitting Notion, and Railway's stolen secrets all hit the layer around the model, not the model itself.

Decision models arrive

Google's Gemini Decisions API, Willison's pricing breakdown, and Latent Space's "D-Day" coverage (with Cloudflare and TypeSafe AI demos) point to a new category of calibrated classifier endpoints.

Agents run into boundaries

Anthropic is moving Managed Agents inside customer security perimeters, United is banning third-party agents from booking flights, and Martin Fowler's piece argues agents must not be allowed to edit their own tests.

Signal vs noise

Signal · underreported

The Fireworks breach exposed Notion customers' AI chat transcripts, showing that inference vendors are sub-processors holding sensitive data most customers never contracted with directly.

Noise · overhyped

A personal-site claim that a Google engineer used 48 agents to find an unmapped Maya pyramid in 72 hours has no independent confirmation.

What to watch

Watch for follow-up disclosures from Fireworks, Railway, and xAI on affected customers, plus whether OpenAI's $10B quarter shows up in Anthropic's IPO filing numbers.

Sources

Filter sources
  1. Pattern 3
  2. Story 3
  3. Gemini Decisions API
    Pattern 2
  4. aws-emulator
  5. Story 4
  6. Gemini Decisions API
  7. Pattern 1
  8. llm 0.37.2
  9. llm-gemini 0.36
  10. Noise
  11. Pattern 3
  12. Pattern 3
  13. Story 6
  14. Pattern 2
  15. Story 2
  16. Story 1
  17. Story 5