Google Reorganizes AI Coding 'Strike Team' in Bid to Close Gap with Anthropic
0xBroomberg
Google is upgrading its months-old AI coding strike team into a formal mid-training unit, just as two star researchers defected to OpenAI and Anthropic in the same week — Alphabet fell 5% in a single session, and the coding gap is becoming an organizational crisis.
What exactly changed in this reorganization?
Google split model training from two stages — pre-training and post-training — into three: pre-training, mid-training, and post-training. This means → a new middle layer now refines models with high-quality, specialized data, especially for coding ability.
The coding strike team was a temporary task force. It is now a formal mid-training team, and its scope has expanded from pure coding to business scenarios like presentation generation.
Post-training, in turn, narrows its focus to user interaction and experience. In plain terms = one team makes the model smarter, the other makes it easier to use — no more overlap.
Why did two key researchers leave?
Noam Shazeer — co-author of the original Transformer paper — abruptly left on June 18 and joined OpenAI. Sources say the trigger was a cut to his GPU compute quota and a forced team merger. Google had paid $2.7 billion less than two years ago to bring him back from his own startup.
Days later, DeepMind VP John Jumper announced he was joining Anthropic. Jumper shared a Nobel Prize with DeepMind CEO Hassabis for leading AlphaFold's protein-structure work; before leaving, he had already been moved into the coding strike team.
This reflects a deeper tension: Google's sheer scale makes internal resource fights fierce, and top researchers who cannot get the compute or autonomy they want will be poached.
How did the market react?
The two departures hit confidence hard. Alphabet dropped 5% on Monday — its steepest one-day fall in over a year — then slid another 1% over the next two sessions.
Zoom out, though, and Alphabet is still up nearly 30% since the start of 2026. This means → the market still backs Google's long-term AI position, but short-term worries about talent loss and execution are mounting.
Why is compute allocation Google's unique headache?
Google has the most vertically integrated AI stack in the industry: in-house TPU chips, frontier models, a cloud business, and ad-revenue monetization. But owning everything also means internal teams compete for the same resources.
The sharpest conflict: Google Cloud's major customers — Anthropic included — and Google's own AI teams fight over the same pool of compute. In plain terms = Google must serve paying cloud clients and feed its own models at the same time, and it cannot shortchange either side.
Shazeer's departure is this tension made personal — his compute quota was adjusted, and he voted with his feet.
Why does the coding race matter so much?
Coding is one of the strongest monetization verticals in AI right now. Anthropic's lead here has pushed its annualized revenue to $47 billion, more than triple the figure in February.
Google's prior bet was that a strong foundation model would naturally produce strong coding ability. Reality disagreed: developers criticized Gemini 3.5 Flash on answer quality and pricing, and the first release of its Antigravity coding tool drew mixed reviews.
The next flagship, Gemini 3.5 Pro, was announced for a June launch but has yet to ship. Testers reported the version available at the time was unlikely to surpass Anthropic's current Mythos model. This reflects that Google's coding deficit is not just a product-cadence issue — it points to a fundamental training-methodology gap, which is exactly what this reorganization aims to fix.
Content is for reference only, not financial advice.