AI News June 26 2026 — Google Loses 4 AI Researchers in One Week, $270B Wiped, Gemini 3.5 Delayed

1. The Four — What Each Researcher Actually Did at Google

The four researchers heading to Anthropic are not a random sample of Google DeepMind staff. Three of the four — Jumper, Adler, and Pritzel — worked together on the AlphaFold protein structure prediction research that earned the 2024 Nobel Prize in Chemistry. The team that created the most significant scientific AI achievement of the decade is now largely at Anthropic.

2. Why Four Researchers Are Leaving at Once — The Real Reasons

Reason 1: Pre-IPO equity. Both Anthropic and OpenAI are on the cusp of going public. Signing on before an IPO — even at a senior level with existing Google RSUs — offers a rare asymmetric payoff: substantial equity that could appreciate dramatically in a public listing. The exits highlight the pressure Google faces from two startups that are on the cusp of going public, offering even well-heeled employees at Big Tech firms the chance at a rare payday by signing on before an IPO.

Reason 2: Compute resource politics. Shortly before Shazeer announced his plans to join OpenAI, computing power dedicated to one of his projects was reassigned to a London-based team at Google DeepMind. The move was made in an attempt to boost collaboration across teams and streamline Google's work on pre-training. The internal competition for GPU access — in a company with 194,000 employees — creates friction that smaller, more focused labs do not have.

Reason 3: The AlphaFold team is going where John Jumper is going. Adler, Pritzel, Jumper and Shazeer worked on AlphaFold research alongside each other. Key members of Jumper's team on the protein-folding research have exited Google DeepMind in recent months. Research teams move together. When the team leader with the Nobel Prize announces he is going to Anthropic, the researchers who built that work with him follow. This is a team migration, not individual career decisions.

Reason 4: The 11x asymmetry. DeepMind engineers are nearly 11 times more likely to leave for Anthropic than the reverse, according to a 2025 industry analysis from the venture capital firm SignalFire. This has been true for years. What changed this week is that it became visible enough to move Alphabet's share price.

3. Google's Response — Strike Team Expansion and Gemini 3.5 Delay

Google is expanding the scope of its months-old AI coding strike team to "midtraining" — an acknowledgement that losing Adler (AI coding) and Pritzel (pretraining) creates specific capability gaps that need to be addressed with dedicated teams. The expanded strike team targets exactly the two domains the departed researchers were working in.

Gemini 3.5 Pro has been delayed to July for "final adjustments." The timing is uncomfortable: the delay arrives the same week the Gemini team lost four senior contributors. Analysts at Jefferies maintain a Buy rating on Alphabet with a $445 price target and describe the talent departures as "noise." DeepMind CEO Demis Hassabis at Cannes: "There's a lot of movement between all the leading labs, and we attract our fair share of top talent. We have the largest and broadest research team in the industry." Google's total workforce stands at approximately 194,700 — a 5% year-over-year increase. The bench is deep. But the names leaving are not anonymous.

4. Jalapeño Update — Production Workloads Running, Report Coming

Following Tuesday's Jalapeño announcement, OpenAI has confirmed that engineering samples are running GPT-5.3-Codex-Spark at production target frequency and power. The chip is functioning as intended in lab conditions. The detailed technical performance report — with hard benchmark numbers, memory bandwidth figures, and throughput comparisons — is expected "in the coming months." Initial deployment remains targeted for end of 2026 at small prototype scale, with gigawatt-scale data center deployment planned for 2027+ with Microsoft and other partners.

The 50% cost savings figure cited by Broadcom CEO Hock Tan remains the only public performance data point. AVGO shares are up approximately 10% year-to-date and have multiplied nearly 7x since end of 2022 — Broadcom's bet on custom AI silicon is being validated by the market even before Jalapeño ships at scale. Full Jalapeño analysis: OpenAI Jalapeño chip deep-dive →

5. Sail Raises $80M — The Software Inference Play

Sail, a startup whose software optimizes how AI models run on existing chips, emerged from stealth with $80 million in seed and Series A funding led by Kleiner Perkins at a $450 million valuation. The timing relative to Jalapeño's announcement is striking: on the same week OpenAI announced it is building custom hardware to reduce inference costs by 50%, a startup raised $80M on the premise that software alone can achieve similar efficiency gains without new hardware.

Sail's value proposition: most enterprises cannot build their own custom ASICs — but they can run Sail's software layer on their existing Nvidia infrastructure to improve utilization and reduce effective inference cost. The custom silicon approach (Jalapeño, Google TPUs, Amazon Trainium) is available only to hyperscalers. The software optimization approach is available to every enterprise running AI on Nvidia GPUs. Both markets are enormous, but they serve different customers at different scales.

The Week in Context — What Changed for Google

What Google actually lost this week is not five researchers out of 194,700 employees. It is the entire senior leadership of the AlphaFold team — the scientists who created the research that earned the company its first Nobel Prize — relocating to Anthropic. It is the co-author of the Transformer architecture paper heading to OpenAI. And it is the Gemini 2.5 contributor who cites AI safety as his reason for leaving going to the lab that built its brand on safety. Each departure is individually explainable. Together, they tell a story about where the most ambitious AI scientists believe the most important work is being done.