Some weeks pile up announcements; this one redraws the map. In a few days, four of Google's most renowned brains changed sides, OpenAI released its first in-house chip, China retook the lead in supercomputers without a single American chip, Anthropic accused Alibaba of the biggest plundering ever carried out against Claude, premium AI subscriptions turned out to be loss-making, and SpaceX's relapse on the stock market woke the word nobody wants to say: bubble. Here is the week of 23 to 26 June 2026 in full — more than fifty verified facts, and what they change for you.

In this issue
  1. The talent war ignites
  2. Models & agents
  3. SpaceX and the compute empire
  4. Chips, data centres & energy
  5. Money, jobs and the bubble
  6. Science is genuinely advancing
  7. Society, regulation & geopolitics
  8. Robotics & physical AI
  9. What to take away

1. The talent war ignites

John Jumper, 2024 Nobel Prize in Chemistry for AlphaFold.
John Jumper, Nobel laureate in Chemistry for AlphaFold, leaves Google DeepMind for Anthropic

Source: CNBC, June 2026.

This is the story of the week, and it reads like a haemorrhage for Google. John Jumper, co-recipient of the 2024 Nobel Prize in Chemistry for AlphaFold, is leaving Google DeepMind after almost nine years for Anthropic. A few days earlier, Noam Shazeer, co-lead of Gemini and co-author of "Attention Is All You Need", left for OpenAI — less than two years after his return through the acquisition of Character.AI (2.7 billion). And on 24 June, two more contributors to Gemini and AlphaFold, Jonas Adler and Alexander Pritzel, also joined Anthropic. Four major departures in one week. Recruiting a living Nobel laureate says a lot about Anthropic's ambition: to push Claude towards the life sciences, DeepMind's historic territory. All this while Anthropic has, since late May, led enterprise adoption (34.4% against 32.3% for OpenAI) and been valued at 965 billion after its Series H.

John Jumper tells the story of AlphaFold, from grand challenge to Nobel · official Nobel Prize channel

Also this week, on the lab side

2. Models & agents

Sam Altman, CEO of OpenAI, speaking at TED.
OpenAI: GPT-5.5-Cyber outperforms Claude Mythos 5 on the cyber benchmark

Source: Neowin, June 2026.

The model race is not slowing down. OpenAI published GPT-5.5-Cyber, specialised in security, which reaches 85.6% on CyberGym against 83.8% for Claude Mythos 5 — reserved for vetted defenders. Google DeepMind natively integrated computer control into Gemini 3.5 Flash (the agent sees the screen, clicks, fills in forms) at a far lower cost than the competition. And Japan's Sakana AI launched Fugu, a "conductor" that distributes each task across a pool of interchangeable models — a bulwark against vendor lock-in.

Mistral presents OCR 4, its self-hostable document extraction model · official Mistral AI channel

The rest of the releases

3. SpaceX and the compute empire

SpaceX no longer just launches rockets: it now sells AI compute · official SpaceX channel

SpaceX has crossed a threshold: it is commercialising its computing power. Its Colossus data centre signed a contract worth up to $6.3 billion with the startup Reflection AI ($150M/month until the end of 2029, on Nvidia GB300 GPUs in Memphis), joining compute agreements estimated at ~45 billion (Anthropic) and ~30 billion (Google). Straight afterwards, the group raised 25 billion in bonds — demand close to 90 billion — and declares more than 100 billion in cash. AI compute has become a commodity sold through colossal multi-year contracts, and SpaceX is building a new annuity from it. (Its 60-billion acquisition of Cursor, for its part, we have already decrypted — see our dedicated article.)

4. Chips, data centres & energy

Cabinets of a Chinese supercomputer — China retakes the top of the Top500 without GPUs.
LineShine retakes the global lead in the Top500: 2.2 exaflops, 100% Chinese CPU

Source: Top500, June 2026.

The June Top500 places the supercomputer LineShine first in the world: 2.198 exaflops/s, the first system to exceed 2 exaflops in double precision without any GPU, on entirely Chinese CPUs. A stinging rebuttal to the American export-control strategy. Meanwhile, OpenAI unveiled with Broadcom Jalapeño, its first in-house chip (an inference ASIC, deployment in late 2026) to reduce its dependence on Nvidia.

Infrastructure and energy

5. Money, jobs and the bubble

Oracle's headquarters — 21,000 jobs cut in a year, linked to AI.
Oracle cut 21,000 jobs in a year, explicitly linked to AI

Source: CNBC, June 2026.

The economic fact of the week is a hard one: premium AI subscriptions do not hold up. A SemiAnalysis study puts the real cost of a $200/month ChatGPT Pro at nearly $14,000 of compute at maximum use, and Claude Max at ~$8,000. The plans become loss-making as soon as use is moderate, because agentic workflows consume up to 1,000 times more tokens. The startup Lindy has in fact migrated 100% of its traffic from Anthropic to DeepSeek V4, for ~90% savings. Arbitrage between models becomes a major margin lever.

Jeff Bezos — his startup Prometheus raises 12 billion for physical-world engineering.
Prometheus, of which Jeff Bezos is co-CEO, raises 12 billion for the physical world's "general engineer"

Source: TechCrunch, June 2026.

Capital, acquisitions and jobs

6. Science is genuinely advancing

Beyond commerce, the week delivered tangible results. An immunologist used GPT-5 Pro to unlock a T-cell puzzle that had gone unanswered for three years: the model proposed a testable hypothesis outside the researcher's field of expertise. And Talos (Microsoft Research, open-source) delivered 241 new rare disease diagnoses across nearly 5,000 patients, by reanalysing genomic data already collected.

“The week was not decided by who has the biggest model. It was decided by who keeps their brains, who masters their compute, who keeps their financial promises, and who has the right to deliver. Raw power is no longer enough.” Otium · 26 June 2026

7. Society, regulation & geopolitics

The trust gap is widening: according to Pew Research, only 16% of Americans expect a positive effect from AI within twenty years, against 40% who anticipate a negative one. Any visible AI deployment must now come with transparency, not a mere promise of efficiency.

8. Robotics & physical AI

A collaborative humanoid robot — the market is shifting from demonstration to deployment.
Morgan Stanley doubles its forecast: 50,000 humanoids delivered in China in 2026

Source: CNBC, June 2026.

The humanoid market is shifting from demonstration to deployment: Morgan Stanley doubled its forecast of deliveries in China to 50,000 units for 2026 (market estimated at 15 billion by 2030). In the United States, Agility Robotics is aiming for the Nasdaq via a 2.5 billion SPAC — the first listed pure player in humanoid robotics.

What to take away

Read these facts together, and a single line emerges: raw power is no longer enough. Value is shifting towards what surrounds it — brains (the talent war), compute and energy (Colossus, LineShine, Jalapeño, Microsoft's 2 GW), economic sustainability (subscriptions that do not hold up, the bubble that worries), and the right to deliver (cyber under controlled access, export control, the AI Act). The "best model" does not win if it does not keep its engineers, does not control its costs, or does not have permission to be deployed.

For an executive, the course of action is confirmed. Do not lock yourself in with a single supplier: the price gap with open models (the Lindy/DeepSeek case) and regulatory volatility make lock-in costly. Keep control of the layer that chooses and routes models (the Sakana lesson). And treat the choice of an AI as a governance decision — real costs, jurisdiction, compliance — not just a benchmark. That is precisely the role of a studio like ours: helping you keep control of that layer, whatever the model of the moment.

To go further: the Chinese 295-billion AI infrastructure plan, the Google-A24 pact, the first case won through an AI lawyer, and the previous week's recap.