Today’s brief covers OpenAI pushing ChatGPT into consumer health records, voice becoming a control surface at two labs, a compute money chase reshaping chip supply, and a fresh round of US-China model claims, plus new model releases, agent acquisitions, and a case for building your own infrastructure.
Data Nobody Outside the Company Can Verify: Health Records, Voice Agents, and Economy-Wide Tracking
OpenAI, Anthropic, and Google all pushed AI deeper into sensitive personal and economic data this week, each asking users to trust promises only the company itself can confirm.
- OpenAI Opens ChatGPT Health to All US Adults on Its Own Terms. OpenAI is connecting ChatGPT to Apple Health and medical records for US adults and says the data will never train its models or target ads, a promise nobody outside the company can verify.
- OpenAI Wires Full-Duplex Voice Into Codex and ChatGPT Work. Voice becomes a control surface for parallel coding agents rather than a chattier chat window, landing first on paid ChatGPT desktop plans.
- Claude’s Voice Mode Drops Its Haiku-Only Ceiling. Anthropic now lets paid voice conversations run on Sonnet and Opus and read Gmail, Slack, and Calendar, while free accounts stay capped at Haiku and a single app.
- Google’s New AI Usage Data Set Doubles as a Policy Pitch. Google’s ATLAS project tracks 15 million Gemini interactions across 150 countries, but the company measuring AI’s economic footprint also profits from how that footprint gets read.
The Compute Money Chase: Intel’s Fastest Growth in Nearly 15 Years Meets Chips Built to Outrun a Shortage
Money and silicon moved together this week, from Intel’s revenue rebound and multi-year supply contracts to inference chips designed around the shortage rather than around raw transistor counts.
- Intel’s 25 Percent Revenue Jump Comes With a New Tactic: Multi-Year CPU Deals. Ten long-term supply deals and a supply-constrained CFO suggest the AI compute shortage has spread from GPUs to server processors, and Intel just posted its fastest revenue growth in nearly 15 years.
- A GPU-Hour Futures Market Won’t Hedge the Cluster You Actually Need. Vast.ai data shows GPU-hour prices barely move even as the co-located clusters large workloads need simply vanish, and that gap is exactly where a hedge breaks.
- Etched Says New Chip Design Runs Math Blocks at Half Voltage. The AI inference startup claims a voltage cut, not a bigger transistor count, is what unlocks more FLOPs without thermal throttling.
- AMD and Cerebras Split Inference in Two to Cut Latency. The chipmakers pair rack-scale throughput with wafer-scale decode speed, targeting agentic workloads where every round trip compounds.
US-China Model Claims: Huawei Training Numbers, Open Weights as Strategy, and a Model That Thinks Before It Codes
Chinese labs made three separate claims about how they train and reason this week, each drawing scrutiny about what the numbers actually show.
- DeepSeek’s Huawei Training Numbers Arrive, So Do the Doubts. A Huawei-led report puts DeepSeek’s V4 post-training at 34.22 percent chip efficiency on Ascend hardware, but it does not show what trained the model originally.
- DeepSeek’s Founder Calls Open Weights Strategy, Not Charity. Liang Wenfeng framed restraint and open-sourcing as vision on an investor call, but the mechanics point to a deliberate move against better-funded rivals.
- Kimi K3’s Real Edge Is What It Does Inside Its Own Head. Design Arena’s trace analysis finds Moonshot’s model burns over 12 times Opus 4.8’s reasoning tokens by simulating an agent loop before writing final code.
New Models on the Board: A World-Model Bet From Black Forest Labs and Microsoft’s Push Into Its Own Stack
Two labs shipped new model families this week, betting on different ideas about what a model should be built to do next.
- Black Forest Labs Reframes Itself as a World-Model Company. FLUX 3 fuses image, video, and audio generation in one architecture, and Black Forest Labs says the same backbone will eventually drive robots.
- Microsoft Ships MAI-Image-2.5-Pro and a Faster, Cheaper Voice Model. Two public-preview models extend Microsoft’s push to replace OpenAI’s models inside Copilot and Bing with its own.
Buying and Building the Agent Stack: Sierra’s Acquisition and Two Bets on Small Teams
Sierra moved to consolidate the long-horizon agent market while Poolside and Andrew Ng made the opposite case, that small teams and local control can outrun bigger budgets.
- Sierra Acquires TakeOff to Own the Long-Horizon Agent Stack. Bret Taylor announced the deal on X, but Sierra and TakeOff have disclosed no financial terms, valuation, or integration plan.
- Poolside Built Laguna S With Under 70 Researchers. Co-CEO Eiso Kant says the 118B model shipped in eight weeks because of a homemade training pipeline, not a bigger budget or a bigger team.
- Andrew Ng Open-Sources a Desktop AI Agent With Full File Access. OpenWorker runs locally, plugs into dozens of apps, and asks permission before it acts, a bet on control over convenience.
Today’s Quick Hits
- Sakana AI Ships Fugu-Ultra v1.1, Still Betting on Orchestration. The Tokyo lab claims broader coding and reasoning gains by routing tasks across models instead of training one, with no independent benchmarks yet to confirm it.
- Runway Adds a Router That Picks Its Own Models. Runway Media Router chooses image, video, and audio models by cost, speed, and quality, extending model routing into generative media.
- Cognition Buys Poke Maker, Its Second Deal in Two Days. Cognition’s second acquisition in two days pairs Devin’s developer tooling with a consumer texting agent, testing whether it is buying capability or reach.