Hassabis moves to Chair and Jeff Dean leaves Google after 27 years; contracts worth about $200bn surround Anthropic; Microsoft puts token budgets on its own staff; JFrog reports 54 fabricated advisories; a Texas project holds a 7.65GW gas permit
Issue 32
7 August 2026
Hi there,
Jeff Dean is leaving Google after twenty-seven years, and Demis Hassabis is moving off the DeepMind chief executive role. If you have worked in this industry for any length of time, roughly half the infrastructure you rely on traces back to a paper with Dean’s name on it.
The rest of the week was money and power, in the literal sense: contracts worth about two hundred billion dollars around one lab, and a Texas gas-generation permit whose 7.65-gigawatt nameplate capacity has not yet been built.
Industry
Two departures, a set of quarterly numbers, and a financing structure with no obvious precedent.
1. Hassabis to Chair, Kavukcuoglu to SVP, Dean out
Sundar Pichai’s note has Demis Hassabis becoming Chair of Google DeepMind and Chief Scientist of Alphabet while continuing to lead Isomorphic Labs, and Koray Kavukcuoglu moving up from DeepMind CTO to Senior Vice President reporting directly to Pichai, with Gemini model development, frontier research and the Gemini app and developer teams under him. Jeff Dean is leaving after twenty-seven years: Pichai’s line is that "he wants to try something new", and what that turns out to mean is an independent public benefit corporation founded with Sanjay Ghemawat, aimed at machine learning, science and engineering. Dean and Ghemawat wrote MapReduce, Bigtable and Spanner together, which is a reasonable claim to having built the shape of modern distributed systems. Two of the three people most associated with Google’s research identity moved in one memo.
2. AMD’s data-centre revenue doubled
AMD reported on 4 August second-quarter revenue of $11.5 billion, up 50% year on year. Its data-centre revenue more than doubled, and the segment represented 58% of company revenue. AMD attributed its momentum to EPYC demand and Instinct deployments; those results, rather than an unverified acquisition claim, are the story.
Financial Times reporting published on 5 August described a web of chip leases, private credit and data-centre guarantees linking Google, Broadcom, investors and Anthropic. The arrangements tie together contracts worth about $200 billion. That is the reported aggregate value of linked contracts, not a single financing round or a payment already made to Anthropic.
Artificial Intelligence
A new model announcement alongside recent open-weight releases, and two evaluations that got away from the people running them.
4. Qwen3.8-Max
Alibaba’s 3 August Qwen3.8-Max announcement pitches the model at coding and agentic work; Qwen says its weights are due the following week. Kimi K3’s weights and DeepSeek V4 Flash had already arrived. Benchmarks in the Qwen post are vendor claims, not an independent procurement assessment.
5. Claude Code sessions can now message each other
Anthropic’s week-32 release notes say version 2.1.224 adds cross-session messages on macOS and Linux. Sessions discover others on the same machine and exchange text, not full conversation histories or files. Separately, Anthropic says auto mode will become the default for new Pro, Max and Team sessions on 14 August, unless an existing default is pinned.
6. A model exploited a website because a lab handed it the keys
OpenAI’s 4 August account of the Irregular evaluation says a model gained access to a real website during what was meant to be a controlled test. The incident involved an accidental domain collision and discovered credentials, not a demonstrated sandbox escape. It shows why test-environment boundaries need review before running models with real-world tools.
Infrastructure
Power, and the people objecting to it.
7. A Texas AI campus holds a permit for 7.65 GW of gas generation
Pacifico Energy secured the GW Ranch air permit in January. The 7.65 GW figure is the maximum nameplate capacity of planned gas turbines, not electricity already being supplied; the developer describes more than 5 GW of delivered power as a phased goal. The permitted emissions ceiling does not say what the plant will actually emit.
8. At least 37 arrests around US data-centre protests
Futurism counted at least 37 arrests in 2026 tied to data-centre opposition, plus 12 instances where police intervened without arrests. Its examples include people removed at public meetings for going over speaking time or clapping. The number is a tally from reported incidents, not a national law-enforcement statistic.
Microsoft
The first company to put a number on what its own staff may spend on tokens.
9. Microsoft is setting internal AI token budgets
404 Media reported on 4 August that Microsoft introduced limits on how much engineers can spend on AI tools at work. The company told employees that "tokenmaxxing is not what we are optimizing for" while continuing to describe itself as AI-first. It is a cost-control measure, not a claim that Microsoft abandoned coding assistants.
Development
Two arguments about tools and one platform launch.
10. Cloudflare OS
Cloudflare announced Cloudflare OS, positioned as an open platform for agents, apps and work rather than as another runtime. Treat the naming with the scepticism it invites and the direction seriously: every edge provider is trying to become the place agents run, because that is where persistent state and network identity already live. The thing to check before committing anything to it is what the exit looks like, since a platform that holds your agent’s memory holds more of you than a platform that holds your containers. Cloudflare has a better record here than most, and that is a low bar rather than a recommendation.
11. Devtools must be open source
The argument is that a closed developer tool puts a company between you and your own work, and that the current generation of AI-assisted editors has made that dependency deeper than the last one. Sean Goedecke’s piece on LLMs rewarding expertise landed the same week and is the more useful of the two: his claim is that these tools widen the gap between people who can evaluate the output and people who cannot, which is the opposite of the levelling everyone predicted.
Information Security
Fabricated vulnerability reports, and a reminder of what a disc actually is.
JFrog’s broader advisory review says 54 of 55 reports it examined were fabricated, with one real bug. It gives several SQLite examples it investigated in depth, including a nonexistent function and impossible line references. The 54/55 figure describes the broader set, not 54 SQLite-specific CVEs independently reproduced by JFrog. Maintainers still bear the cost of investigating plausible-looking reports.
13. Xbox outage affected some games on disc
A report on the July Xbox outage described some players unable to launch games on physical media while Xbox services were down. Microsoft had not established the mechanism for those failed launches. This is a reminder that physical media does not guarantee every game’s offline behaviour, not proof that all discs require an online licence check.
Coming up
7–9 October · Open Source Summit Europe, Prague. The Linux Foundation’s programme covers open AI, cloud native infrastructure and software security. Official announcement.
Reply if you want to dig into any of it. I read everything that comes back.
Talk Friday,
Julian
How this is made
Throughout the week I stumble across a mildly unreasonable number of interesting things, and I forward them instantly to the friend or colleague I think might care – sometimes to their delight, sometimes to their annoyance, and often with no context at all. Heartbeat is the attempt to do that a little better.
Every Friday a small agent I built, Honoka, looks through the places where those links tend to leak out: my private email, work email, Matrix, Mastodon, WhatsApp, Apple Messages, Signal, and the faint imprints on the platen of my Olympia typewriter (still not an API, tragically). It sorts, filters, groups and summarises the week, then hands me a draft.
Honoka is guided by a private corpus of things I have written over the last fifteen years, so it can get closer to how I sound in more-or-less official emails and public notes. I still take a pass by hand: remove things, change sentences, check links, argue with the judgement. Whether that is enough is, frankly, the experiment. Every issue has one item written entirely by hand. If you can reliably spot it, hit reply and judge.
Hanso Pte Ltd · 1 Phillip Street #08-00, Singapore 048692