Google's AI Brain Drain Deepens as Qualcomm Makes Its $14B Move Against Nvidia — June 25, 2026

⚡ Top Story

Google's AI Brain Drain Hits Critical Mass: Four Foundational Researchers Exit in One Week

In the most concentrated talent departure from a single AI lab in recent memory, Google DeepMind and Google Research lost four foundational AI researchers to direct competitors in seven days. Transformer co-author Noam Shazeer announced his move to OpenAI on June 18. John Jumper — Nobel laureate and AlphaFold co-creator — confirmed his departure to Anthropic on June 20. Bloomberg and TechCrunch reported on June 24 that Jonas Adler and Alexander Pritzel — both key contributors to Google's Gemini model and collaborators with Jumper on protein-folding research — are also headed to Anthropic. The week's departures strip Google of talent central to three of its most significant AI achievements: the attention mechanism, AlphaFold's protein structure prediction, and Gemini's foundational training.

Why it matters: This isn't routine turnover. Shazeer co-authored "Attention Is All You Need" — the paper that gave us the transformer architecture every major AI lab now runs on. Jumper led the work that won the 2024 Nobel Prize in Chemistry. Adler and Pritzel's protein-folding background directly extends Anthropic's biological AI capabilities at a moment when biotech is one of the highest-value enterprise AI verticals. Alphabet shares dropped as much as 1.2% intraday on the Adler/Pritzel report.

Sources: Bloomberg: Google Poised to Lose Two More High-Profile AI Staffers to Anthropic · TechCrunch: AI researchers continue to leave Google for its rivals · CNBC: John Jumper to leave Google DeepMind for Anthropic · Search Engine Journal: Google Loses Two Top AI Researchers

🔬 Research & Papers

No landmark papers with confirmed June 24–25 publication dates cleared source validation for this edition. Readers tracking frontier research should check arXiv cs.AI — June 2026 listings directly.

Worth tracking from earlier this week: "Quantized Reasoning Models Think They Need to Think Longer, but They Do Not" (arXiv:2606.00206) finds that post-training quantization causes reasoning models to generate longer but less accurate chains-of-thought — in up to 52% of failures, models reach the correct intermediate answer but fail to output it as final. A training-free logit penalty on overthinking markers reduces chain-of-thought length by 12–23% while preserving accuracy. Directly relevant to production deployments of quantized frontier models.

Source: arXiv:2606.00206

🏢 Industry & Startups

Qualcomm Acquires Modular for ~$3.9 Billion

At its Investor Day on June 24, Qualcomm announced a definitive agreement to acquire Modular Inc. — the AI infrastructure startup founded by Chris Lattner (creator of Swift, LLVM, and MLIR) — in an all-stock transaction valued at approximately $3.92 billion. Modular's core product is the MAX platform: a compiler toolchain and inference engine designed to run AI workloads across heterogeneous hardware — CPUs, GPUs, and custom accelerators — without CUDA-style rewrites. The deal is expected to close in H2 2026. Qualcomm also raised its FY2029 non-handset revenue target to $40 billion at the same event.

Sources: Qualcomm IR: Qualcomm to Acquire Modular · Modular blog: Qualcomm to Acquire Modular · QZ: Qualcomm acquires Modular for $3.9 billion · HotHardware: Qualcomm Acquires Modular to Challenge Nvidia

Meta Plans AI-Powered Prediction Market App

Documents obtained by NPR (June 24) reveal that Meta is developing a standalone prediction market app competing with Kalshi and Polymarket. Internally codenamed "Antwerp" and "FBForecast", it will use Llama to automatically generate questions from trending topics. Meta is positioning the product as an AI-differentiated entry into a market some analysts project could reach $1 trillion.

Source: KCBX/NPR: Meta plans to release AI-powered prediction market app

🛠️ Tools & Releases

GPT-5.6 Misses Its June Window; Prediction Markets Shift to July

OpenAI's anticipated GPT-5.6 did not launch within the June 22–28 window that prediction markets had assigned ~83% probability two weeks ago. As of June 25, Polymarket odds for that window have collapsed to approximately 18%, with new contract volume pricing a July 2026 launch. OpenAI has issued no official announcement. The model — described internally by Chief Scientist Jakub Pachocki as a "meaningful improvement" over GPT-5.5, with a 1.5M-token context window and a redesigned training pipeline — has no release notes or product page as of today.

Gemini 3.5 Pro: 5 Days to Google's Self-Imposed Deadline

Gemini 3.5 Pro — featuring a 2M-token context window — remains in limited Vertex AI enterprise preview as of June 25. Google CEO Sundar Pichai committed explicitly at Google I/O (May 19) to a June 2026 general availability launch. Five days remain on that self-imposed deadline.

Sources: Polymarket: GPT-5.6 Released On? · TechTimes: GPT-5.6 Launch Window Starts Monday

🌏 Global AI & Geopolitics

Anthropic Export Ban — Day 13: Congress Demands Commerce Explanation by Tomorrow

The US Commerce Department's export control directive on Fable 5 and Mythos 5 remains in force. A bipartisan group of four members of Congress sent a formal letter to Commerce Secretary Howard Lutnick on June 18 demanding a written explanation of the technical and legal basis for the controls by June 26 — tomorrow. No public explanation from Commerce has been released. If Commerce responds publicly, it would be the first time the government formally articulates what triggered the ban — consequential for every AI lab operating under export-control uncertainty.

Sources: Anthropic: Statement on US Government Directive · CSIS: Department of Commerce Restricted Access to Anthropic's Latest Models

⚡ Energy, Infrastructure & Chips

Qualcomm's Dual-Track $14B Bet Against Nvidia

In a single day (June 24), Qualcomm committed to a strategy that could total more than $14 billion in two moves: the confirmed $3.92B Modular acquisition (software/compiler layer) and ongoing advanced negotiations to acquire Tenstorrent for up to $10B (AI chip hardware, RISC-V architecture). Tenstorrent — led by former Apple chip architect Jim Keller — designs AI accelerators on the open RISC-V instruction set, independent of Nvidia's proprietary software ecosystem. If both deals close, Qualcomm would have assembled: an open compiler stack (Modular MAX) + RISC-V AI accelerators (Tenstorrent) + edge AI silicon (Qualcomm's existing lineup) — potentially the most complete Nvidia-alternative AI stack with commercial credibility.

⚠️ The Tenstorrent acquisition is in talks and unconfirmed. Reuters originally reported Tenstorrent discussions June 15–16; the June 24 TechTimes analysis frames both moves as a unified Qualcomm Investor Day strategy.

Sources: TechTimes: Qualcomm Bets $14 Billion Cracking Nvidia's AI Monopoly (June 24) · Yahoo Finance: Qualcomm in Talks to Acquire Tenstorrent · The Register: Qualcomm circling Tenstorrent in $10B RISC-V play

🔒 Safety, Alignment & Ethics

Why the Anthropic Ban Happened: NSA Red-Team Reporting Gets Wider Coverage

Multiple tech outlets reported this week on the classified rationale behind the Anthropic export ban. Senator Mark Warner publicly disclosed that NSA Director General Joshua Rudd told him Anthropic's Mythos model "broke into almost all of our classified systems, not in weeks, but in hours" during an authorized red-team exercise. This is the first time the underlying technical justification for the ban has been publicly attributed to a named senior government official. Important caveats: there is no official NSA statement confirming the claim; the Economist editor who originally published the quote clarified it "should not be read literally" and that Mythos worked alongside other tools under specific conditions. Whether Commerce Dept's congressional response (due June 26) confirms or reframes this account is the week's most consequential pending public disclosure.

⚠️ Senator Warner's public statement is the only source; the NSA has not confirmed. Exercise conditions and tool configurations are not public.

Sources: TechSpot: Anthropic's Mythos AI reportedly cracked NSA classified systems in hours · Tom's Hardware: Anthropic's Mythos reportedly breached NSA classified systems · The CyberSec Guru: Mythos 'Hacked the NSA'? What's Actually Confirmed

📊 Numbers & Signals

4 — foundational Google AI researchers who announced departures in 7 days: Shazeer (→ OpenAI), Jumper (→ Anthropic), Adler (→ Anthropic), Pritzel (→ Anthropic)
~18% — current Polymarket odds for GPT-5.6 launching June 22–28 (down from ~83% two weeks ago)
5 days — until Google's self-imposed June 30 Gemini 3.5 Pro GA deadline
2 days — until GPT-4.5 retires from ChatGPT (June 27; API unaffected)
$3.92B — Qualcomm's all-stock acquisition of Modular (June 24)
Up to $10B — reported value of Qualcomm's Tenstorrent talks (unconfirmed)
$14B+ — combined value of Qualcomm's two anti-Nvidia moves in one day
$40B — Qualcomm's FY2029 non-handset revenue target (announced June 24)
Day 13 — Anthropic Fable 5/Mythos 5 export ban (no resolution)
June 26 (tomorrow) — Commerce Dept deadline to respond to congressional letter on Anthropic ban

🧠 Worth Thinking About

The Google talent wave — Shazeer, Jumper, Adler, Pritzel, all leaving within a week — is easy to read as a morale story. The more structurally significant read is what it says about research gravity in the post-GPT-4 era. These researchers are not leaving AI; they are concentrating at Anthropic and OpenAI — the two labs organized specifically around the next capability threshold. Google DeepMind, despite AlphaFold, Gemini, and a Nobel, may be experiencing what happens when an incumbent's research magnetism weakens not because it fails, but because the frontier moves faster in structures built around fewer constraints. Google's core products and user base are still growing. But the people whose methods those products depend on are now at its competitors. The asymmetry between institutional commercial success and research gravitational pull is the real story underneath the departures.

🏛️ Government & Regulation

Congressional Deadline: Commerce Dept Must Respond on Anthropic Ban Tomorrow

Four bipartisan members of Congress sent Commerce Secretary Howard Lutnick a formal demand on June 18 for a written explanation of the Fable 5/Mythos 5 export controls. The statutory response deadline is June 26. If Commerce responds publicly, it would be the first time the government formally articulates the legal and technical basis for the ban — directly consequential for every AI lab, regulator, and developer trying to understand where the export-control line sits.

Active Compliance Calendar (Updated June 25, 2026)

June 26 (tomorrow): Commerce Dept response deadline to congressional letter on Anthropic ban
June 27 (2 days): GPT-4.5 retires from ChatGPT (API unaffected)
June 30 (5 days): Google's self-imposed Gemini 3.5 Pro GA deadline
August 1 (~37 days): Treasury, NSA, and CISA must finalize classified benchmarking for "covered frontier models" under Trump's June 2 AI Executive Order
August 2 (~38 days): EU AI Act Article 50 transparency obligations take effect
August 26 (~62 days): o3 retires from ChatGPT (API unaffected)
January 1, 2027: Colorado SB 26-189 (revised AI transparency framework) takes effect

🔭 Frontier Lab Dispatch

No new verified posts from OpenAI, Anthropic, Google DeepMind, Meta AI, or leading Chinese labs were confirmed with a June 25 publication date at time of writing.

Google: No new technical blog posts confirmed for June 25. The Adler/Pritzel departure (Bloomberg/TechCrunch, June 24) is the dominant Google story today; no official response from Google.

OpenAI: GPT-5.6 did not launch as of June 25. Last confirmed official post: GPT-5.5-Cyber (June 22).

Anthropic: No new official posts. Fable 5 and Mythos 5 remain under the US export ban. Fable 5 API access requires usage credits at $10/M input tokens / $50/M output tokens.

Meta: Planning AI-powered prediction market app "Antwerp" / "FBForecast" using Llama (reported June 24 — covered above in Industry & Startups).

🔗 Quick Links

Tier 1 — Frontier AI Labs / Official

Tier 3 — Tech & AI News Media

Tier 4 — Research

Tier 5 — Policy & Governance

CSIS: Department of Commerce Restricted Access to Anthropic's Models — What Comes Next?