1. Anthropic Signs $9.1B, 20-Year Data-Center Lease With Bitcoin Miner Riot Platforms
Riot Platforms disclosed a 20-year lease covering 191 MW at its Rockdale, Texas campus, running through June 2048 and worth roughly $9.1 billion, rising to about $16.1 billion if two five-year extensions are exercised. Riot named the counterparty only as a "leading frontier AI lab"; Bloomberg identified it as Anthropic, and neither company confirmed. First 96 MW is expected to land in December 2027, with the full 191 MW by June 2028, backed by a $573 million interim facility from Morgan Stanley. The deal is the clearest sign yet that bitcoin miners are becoming the landlords of the AI buildout.
2. Google DeepMind Loses Both Its CEO and Its Chief Scientist in One Day
Demis Hassabis handed over day-to-day control of Google DeepMind to become chairman of the unit and chief scientist of Alphabet, with CTO Koray Kavukcuoglu taking operational command. The same day, Jeff Dean left after 27 years to co-found Discovery Loop, a public benefit corporation aimed at AI for science and engineering, taking Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him. Google will invest in the new venture. Alphabet shares fell about 4% on the news.
3. Gemini Crosses 1 Billion Monthly Active Users
Sundar Pichai announced the Gemini app passed 1 billion monthly active users, matching the milestone ChatGPT reached in June. Google cited 63% engagement on voice features, more than 150 million images generated per day, and over 100 million active iOS users. The number puts Google at genuine parity with OpenAI on consumer reach for the first time.
4. Meta Ships Muse Glimmer, a 30B Multimodal Model Under Apache 2.0
Meta released Muse Glimmer, a 30-billion-parameter dense multimodal model licensed Apache 2.0 and tuned for local agentic tool use, coding, and LLM-as-judge work. It carries a 131K context window and supports over 100 languages. Quantized to 4-bit it fits under 20 GB and runs on a single consumer GPU with a reported 3.1x speedup — squarely aimed at the run-it-yourself tier rather than the frontier.
5. OpenAI's GPT-5.6-Cyber Finds Two Unknown Chrome V8 Zero-Days
OpenAI's specialized security model surfaced two previously unknown vulnerabilities in Chrome's V8 engine that could be chained to corrupt memory and escape the V8 heap sandbox. Google patched them under CVE-2026-15903. GPT-5.6-Cyber is the first model to reach the "High" cyber capability threshold under OpenAI's Preparedness Framework, and access is gated behind the newly expanded Daybreak Red tier.
6. Researchers Extract Secrets From "Encrypted" Chain-of-Thought Blocks
A disclosure this week showed that encrypted reasoning blocks — the hidden chain-of-thought that Anthropic, OpenAI and Google models return to clients — can be decoded by an attacker. Researchers recovered 367 personally identifiable artifacts and 182 credentials from 315,320 decoded reasoning blocks. It is a pointed reminder that "the model hides its reasoning" is a product decision, not a security boundary.
7. OpenAI's IPO Moves Into Range: Confidential S-1, September Target
OpenAI filed its S-1 confidentially with the SEC in May and is reportedly targeting a listing as soon as September, at a valuation range of roughly $852 billion to $1 trillion, with Goldman Sachs and Morgan Stanley leading. Reported figures put OpenAI near a $25 billion annualized run rate with 50 million consumer subscribers and 9 million business users — while still losing money on every dollar of revenue. The full public S-1 would land roughly 15 days before the roadshow, putting late August in play.
8. Capital Keeps Flowing to the Physical Layer of AI
Safe Superintelligence announced a reported $5 billion Nvidia-backed financing alongside a long-term partnership. Atoms, the physical-AI startup founded by Travis Kalanick, raised $1.7 billion led by Andreessen Horowitz, and Fireworks AI closed a $1.505 billion Series D. The pattern across August is consistent: the largest checks are going to companies that own a hard, physical or regulated layer — compute, energy, robotics — rather than an application sitting on top of someone else's model.
9. US AI Rules Consolidate Around Federal Preemption
The White House's National Policy Framework for AI, released in March, recommends against creating any new federal AI regulator, routing oversight instead through existing agencies and industry-led standards. It builds on Executive Order 14365, which directs agencies to preempt state AI laws via litigation and funding leverage, carving out only child safety, compute and data-center infrastructure, and state procurement. States are adjusting: Colorado repealed its AI act in May and replaced it with the narrower SB 26-189, effective January 1, 2027.
// KEY TAKEAWAYS
Two themes dominate this week. First, the money has moved from models to megawatts — Anthropic's $9.1 billion Riot lease and the funding rounds flowing to compute, energy and robotics show that the scarce asset is now physical, not algorithmic. Second, the people and the rules are both in motion: Google DeepMind lost its CEO and chief scientist on the same day, OpenAI is weeks from a public S-1, and Washington is consolidating AI oversight federally rather than building a new regulator. Meanwhile the security picture darkened on both sides of the ledger — a model found real Chrome zero-days, and researchers proved that hidden model reasoning leaks credentials.