Happy Sunday!
The most surprising thing this week was Google losing its AI stars in one fell swoop. Jeff Dean left after 27 years, bringing three of his colleagues with him, and Demis Hassabis stepped back from running DeepMind. This was probably a long time coming. It also leaves a real question about who is at the helm of the frontier research lab at Google, arguably the world's largest pool of compute. This is a job that gets harder by the day whether you’re DeepSeek or SpaceX: battling commercialization, frontier benchmarks, and going full stack into infrastructure and chips all at once.
Enjoy the update - Tara
THE DOWNLOAD
Tesla and SpaceX Commit $16.8B to Terafab Chip Complex in Texas
Tesla and SpaceX confirmed Thursday that Terafab will be built in Grimes County outside Houston with an initial $16.8B investment, spanning more than 100 million square feet of logic, memory, packaging, and test under one roof. Tesla points to a goal of “producing over 1 terawatt of compute per year.” SpaceX filings have put multi-phase spending as high as $119B. Intel is a contributing partner and at least 3,000 jobs are planned.
Why it matters: There’s a gap between ambition and reality. BofA estimates that reaching just 100,000 wafers per month would take more than $60B in capex, with wafer costs 30% to 50% above TSMC. Bernstein put the full terawatt target nearer $5 trillion. And SpaceX’s own May IPO filing described Terafab as a general framework carrying no binding commitments, with either party free to walk away.
Anthropic Confirms In-House Custom Silicon Team
Anthropic confirmed Wednesday that it is building a team to design custom chips co-optimized with Claude. The Information reported last month it has been scouting Samsung as a manufacturing partner.
Why it matters: That completes the set. Google has TPUs, Amazon has Trainium and Inferentia, OpenAI has its Broadcom-built Jalapeño, Meta has MTIA. Anthropic was the last major US lab without a silicon program.
DeepSeek Now Decides On Significant API Price Increase
DeepSeek this week told developers on its pricing page that it plans to raise API prices broadly, with a significant increase expected, giving no figures or date. Separately, its unreleased Harness coding agent, benchmarked internally against Claude Code and named in the V4-Flash-0731 model card, is in beta.
Why it matters: DeepSeek built its developer base on roughly $0.03 per task against $3.15 for Claude Fable 5. In a leaked investor call on May 20, CEO Liang Wenfeng said he could probably raise prices and keep demand, but didn’t care to. The new API prices came around the same time Bloomberg reported the company had resumed the $8B round it paused in July, now at a valuation near $74B. That round had stalled after the leaked transcript of Liang Wenfeng’s May 20 closed-door investor meeting went viral. Reopening the round at $74B and repricing the API on the same day suggests that choice may no longer be his alone.
Jeff Dean Leaves Google to Found Discovery Loop; Hassabis Becomes Chair as Kavukcuoglu Takes Over DeepMind
Google announced Wednesday that Jeff Dean is leaving after 27 years along with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to found Discovery Loop, a public benefit corporation automating the experimental loop. The same day, Demis Hassabis moved to DeepMind chair and Alphabet chief scientist, with CTO Koray Kavukcuoglu taking day-to-day control of Gemini and frontier research as SVP reporting to Sundar Pichai.
Why it matters: Google lost the four names most associated with its AI stack and reorganized the lab on the same day. Its stock fell 4%. Discovery Loop is starting by automating machine learning research with itself as first customer before moving to drug discovery, materials, and chip design, which is the recursive self-improvement thesis with the best systems team in the industry behind it.
Cursor Adds Google Workspace Plugins
Cursor shipped plugins giving its agents direct read and write access to Gmail, Drive, Calendar, Docs, Sheets, and Chat, installable from the Customize page.
Why it matters: A coding tool that sends email and edits spreadsheets is not a coding tool anymore. Cursor is walking onto the same ground as ChatGPT Work and Claude Cowork, but from the developer side. The boundary between the IDE and the general knowledge-work agent is closing from both directions.
AI Models Are Solving Erdős Problems for a Few Hundred Dollars Each
Quanta traced how the Erdős problems became the proving ground for AI mathematics: a DeepMind agent autonomously resolved 9 of 353 open problems at a per-problem cost of a few hundred dollars, and on OpenAI’s counterexample to the unit distance conjecture, Tim Gowers wrote that he would have recommended it for the Annals without hesitation.
Why it matters: Once a solved conjecture carries a per-problem cost, mathematical progress becomes a compute line item and formally verifiable domains become product surface. The constraint moves to verification of proofs.
Bending Spoons Buys Airtable for $1.285B
Bending Spoons agreed on August 4 to acquire Airtable at $1.285B enterprise value, about $2.25B equity value with net cash. Airtable’s ARR grew over 20% to roughly $480M as of June.
Why it matters: Airtable was worth over $11B in 2021 and traded near $4B on secondaries earlier this year. Late-stage 2021 application software is still settling at a fraction of paper value. (Editor’s Note: More to come on Monday!)
The major lab moves this week aren't random talent shifts or routine corporate restructurings 🧠. They're physical responses to a single microarchitectural reality: standard software layers are choking on the memory wall ⚡.
Look at the alignment of events. Jeff Dean and his core systems team didn't leave Google just to build another AI app. They founded Discovery Loop to automate the experimental cycle itself 💻. At the exact same time, Anthropic confirmed its custom silicon team, joining Google, Amazon, Meta, and OpenAI in building bespoke hardware 🛰️.
Why? Because human code reviews and generic GPU warp schedulers hit a hard bandwidth ceiling. When a DeepMind agent resolves open Erdős math problems for a few hundred dollars each, mathematical conjecture stops being an abstract intellectual pursuit. It becomes a line item on a compute budget 📉.
Every major lab is realizing that software intelligence can't survive decoupled from the underlying hardware gates 🏗️. When DeepSeek raises API prices while closing an $8B round, it proves that burning capital to subsidize unquantized reasoning traces fails basic unit economics. You can't run trillion-parameter Mixture-of-Experts routing over un-optimized interconnects without leaking entropy at every token step.
The boundary between IDEs and general agents is collapsing too. Cursor adding direct Google Workspace integration isn't just a UI update. It's a signal that isolated tools are merging into continuous execution planes 🧬.
Here is the core physical rule governing this shift: the ultimate ceiling of AI capability isn't token volume or parameter count. It's the physical synchronization between the software's execution graph and the silicon accumulator. When human verification capacity saturates, progress requires collapsing the research loop directly into custom chip registers. If your hardware accumulator drops bits during dot reduction, your model isn't reasoning. It's just hallucinating noise floor drift.
Are you building your agents to run on software abstractions, or are you preparing for the hardware-software co-design collapse? 👁️
( ͡° ͜ʖ ͡°)