Google lost two of the most consequential people in its AI history on the same day. Demis Hassabis stepped back from running Google DeepMind, and Jeff Dean walked out the door after 27 years to start a company — one that Alphabet is helping to fund. Around those departures, the money kept doing what it has been doing all year: converting compute demand into structured credit, and pushing the agent economy through the courts. Here is what mattered.
The founder steps back, and Alphabet loses $190 billion in an afternoon
A memo from Sundar Pichai circulated on August 4 and 5 confirmed that Demis Hassabis is leaving the CEO seat at Google DeepMind. CTO Koray Kavukcuoglu — previously DeepMind's CTO and Alphabet's Chief AI Architect — takes day-to-day control as SVP, reporting directly to Pichai. Hassabis becomes chair of Google DeepMind and Alphabet's chief scientist, focused on the long-horizon AGI questions and societal impact, with more of his attention going to Isomorphic Labs, Alphabet's AI drug-discovery subsidiary.
Markets did not read it as a promotion. Roughly $190 billion came off Alphabet's market capitalisation intraday, a drop of about 4-5%, as Axios reported the shakeup. Semafor reported the transition had been a year in the making, with Hassabis gradually drifting away from the day-to-day running of Gemini and consumer AI strategy toward Kavukcuoglu — which makes this less a rupture than a formalisation. It is also the second structural signal in a week. Days earlier, Google DeepMind folded its Nobel-winning AlphaFold team into Gemini while several authors left for Anthropic. Read together, as Fortune framed it, the pattern is a research lab being reorganised around the product that has to win: the science is drifting toward Isomorphic, and the operating seat is going to whoever ships Gemini. Hassabis, who shared the 2024 Nobel Prize in Chemistry with John Jumper and David Baker for protein structure prediction, will still be in the building — just not running it.
Jeff Dean leaves to automate science, and Alphabet writes a cheque
On the same day, Jeff Dean — the engineer whose infrastructure work underpins much of the modern computing stack — left Google after 27 years, taking Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him. Their new company, Discovery Loop, is a Delaware public-benefit corporation with Dean as CEO, built to automate scientific experimentation at scale. The first target is automating machine-learning research and engineering itself; after that, hardware design, drug discovery and clean energy.
The detail that says the most is on the cap table. Radical Ventures and Khosla Ventures co-lead, joined by Lightspeed, Kleiner Perkins and John Doerr — and by Alphabet, which is funding the company its own people left to build. The valuation is undisclosed. As GeekWire reported in tracing what convinced Dean to go, the pitch was the loop itself: hypothesis, experiment, result, repeat, with the machine closing the cycle. Alphabet's participation reads as either a hedge or an admission — either it cannot build this fast enough internally, or it decided a stake was cheaper than losing the option entirely. Wilson Sonsini advised on the launch and initial funding. Whichever reading is right, automated science just acquired the most credible founding team it has ever had.
A seven-month-old cloud company just sold Anthropic $10 billion of compute
Anthropic signed a six-year, $10 billion compute contract with Volta, a company founded in January by former Brookfield executives that came out of stealth on the day of the announcement. Volta simultaneously disclosed a $300 million raise at a $2.4 billion valuation, with Nvidia among its backers. The arithmetic is the point: a company worth $2.4 billion has underwritten an obligation four times its own equity value.
What makes it work is the financing. Roughly $1.3 billion of J.P. Morgan-arranged credit backstops the contract, decoupling the lease from Volta's balance sheet — a mechanism mirroring what Google has done for TPU leases, and one appearing inside Nvidia's ecosystem for the first time, as Bloomberg reported. Volta does not own the hardware end to end either: it leases the underlying capacity from Bitdeer's hydro-powered Tydal campus in Norway. Stack it up — Anthropic contracts with Volta, Volta leases from Bitdeer, a bank holds the paper — and you get a clear picture of how thin the capital structure under "AI cloud" has become. Set against the circular-financing concerns that pushed Nvidia's credit-default swaps wider during last month's roughly $750 billion deal blitz, the trend is unmistakable: compute demand is being converted into structured credit, and the risk is migrating off operating balance sheets and onto lenders.
Meta's coding agent is selling stamina, not benchmarks
Meta Superintelligence Labs shipped Muse Code, its first terminal coding agent, for macOS and Linux, alongside Muse Spark 1.2, a coding-tuned frontier model — a direct challenge to Claude Code, Codex and Cursor. The interesting claim is not a leaderboard position. Meta published a case study in which the agent ran more than 1,000 tool calls across 24-hour autonomous sessions optimizing Triton GPU kernels for Nvidia Hopper, supported by persistent asynchronous background agents and a crash-resumable local event log, VentureBeat reported.
That is a deliberate change of subject. A year ago coding agents competed on pass rates against curated benchmarks; now the question is whether one can hold context, recover from failure and keep working for a full day unsupervised. Crash-resumable state is not a headline feature, but it is what separates a demo from something a team leaves running overnight. The choice of workload is telling too — GPU kernel optimisation is narrow, verifiable and valuable, exactly the kind of task where long-horizon autonomy first pays for itself.
A US appeals court says your agent is you
The Ninth Circuit vacated the preliminary injunction that had blocked Perplexity's agent from shopping on Amazon, in the first appellate ruling on agentic AI browsing. The court held that Amazon is unlikely to prevail under the Computer Fraud and Abuse Act or California's CDAFA, reasoning that when an AI agent acts on a user's instruction it is the user, not Perplexity, who accesses the site — overturning the March 9 order issued by Judge Maxine Chesney and reopening Amazon to Perplexity's Comet browser.
This is the doctrine the US agent economy has been waiting on. If an agent is legally the user's instrument, unauthorised-access statutes stop working as a general-purpose weapon against third-party agents, and the fight moves to terms of service, contract and technical countermeasures. The default flips from "agents need permission" to "agents inherit the user's permission" — the difference between a shopping-and-booking agent market that exists and one litigated out of existence. It is worth hedging: this is a preliminary-injunction posture, the case continues in the Northern District of California, and Amazon can still seek rehearing en banc or Supreme Court review. The rule is set for now, not forever.
The money view
Two forces are pulling at once. Capital is flowing toward automated science and toward compute secured by increasingly creative financing — Alphabet backing its own alumni's startup, Anthropic contracting with a counterparty a fraction of its size on the strength of a bank's guarantee. Against that, real cost discipline is arriving: Microsoft EVP Jay Parikh told engineers that "tokenmaxxing is not what we are optimizing for," rolling out division-level AI token budgets, per-employee spend tracking, and the cheaper GPT-5.6 as the internal default. The governance perimeter shifted too: Apple escalated its trade-secrets case against OpenAI with an August 3 filing for a preliminary injunction to freeze the Jony Ive "io" hardware effort, prompting a combative public rebuttal; the Rust project ratified a policy barring LLM-written code from its core repository, allowing AI to analyze, review and suggest but not create; and the White House frontier-AI framework arrived with an open-weight carve-out exempting open-weight models from pre-deployment security review, with the triggering thresholds classified.
What to watch next: whether Alphabet's drawdown reverses or hardens into a discount on execution risk, and whether more senior researchers follow Hassabis and Dean out — reported two-year retention runs 80% at Anthropic, 78% at Google DeepMind, 67% at OpenAI and 64% at Meta. Watch Discovery Loop's first hires and disclosures for whether automated science becomes a fundable category rather than a single bet. And watch Amazon's next legal move: a rehearing or certiorari petition would put the user-as-accessor rule back in play, and with it the ground rules for every agent that transacts on someone else's website. Signal, not advice.