〔Overnight Brief〕Astra Crosses 'Critical' Cyber, Anthropic Cuts Cache Reads 75%, California Locks Data-Center Power — Frontier Race Hits an Infrastructure Wall

0. Overnight Arc

The AI industry is splitting in a pattern consistent with Madrona's IA40 list, with three simultaneous contests running in parallel: a frontier capability race (OpenAI Astra crossing 'Critical' cyber, Google prepping Gemini 3.8 Flash for as early as Wednesday, Anthropic answering with Fable 5.1), an enterprise pricing-and-policy reset (Anthropic's 75% cache-read cut, data-retention revision, zero-retention security suite), and an infrastructure squeeze (California's seven data-center bills, DRAM tightness through 2027, Samsung's HBM5/zHBM roadmap) ^{[1][2][3][4][5][6][7]}. Capability and capital keep compounding while physical and regulatory constraints are catching up.

1. Frontier Model Race

• **[NEW] OpenAI Astra — 'Critical' cyber threshold:** OpenAI confirmed Astra has reached the Critical threshold under its Preparedness Framework for cybersecurity capability and will be available 'soon,' with stronger safeguards in place after the Hugging Face incident ^[1][8][9].
• **[NEW] Google counter-program:** Google is preparing a new model with stronger coding capability aimed at closing the gap with OpenAI and Anthropic, and plans to release Gemini 3.8 Flash as early as Wednesday ^[10][3].
• **[NEW] Anthropic Fable 5.1 / Mythos 5.1:** Released Tuesday with a 75% cut on cache-read pricing, improved coding and research capability; invite-only Mythos 5.1 for pre-selected users; Anthropic framed it as a direct response to enterprise pricing pressure ^[4][11].
• **[NEW] Madrona IA40 split:** Madrona's annual IA40 list shows the AI industry 'splitting in two,' consistent with the divergence between the hyperscaler frontier race and the long tail of applied AI ^[2].

2. Enterprise Pivot: Data, Safety, Pricing

• **[NEW] Anthropic enterprise reset:** Revised data-retention policy after customer pushback, now allowing enterprise customers to retain data in their own cloud while still under Anthropic safety review; also launched a frontier enterprise security suite with zero data retention and abuse detection ^[12][4][13].
• **[NEW] Cache economics as the new battleground:** Anthropic product lead Dianne Penn said cache reads can account for half or more of token consumption in complex long tasks, explaining why the 75% cut is the central commercial move of the Fable 5.1 release ^[4].
• **[NEW] OpenAI in healthcare:** New Epic EHR integration connecting supported Epic environments to ChatGPT, plus a plugin linking to nine additional industry data sources, positioning ChatGPT inside clinical workflows ^[14][15].

3. Infrastructure: Memory, Power, Capital

• **[NEW] Samsung HBM roadmap:** Targeting HBM5 at 2x HBM4E performance, +20% perf/W, -20% thermal resistance; the post-HBM5 'zHBM' targets 8x HBM4E performance, +3x perf/W, and -75% to -90% thermal resistance, framed for AI chip high-performance, low-power needs ^[6].
• **[ONGOING] DRAM supply:** GigaDevice vice chairman and GM He Wei said DRAM prices will continue upward in H2, and that 2027 incremental capacity is limited so supply tightness will not meaningfully ease and the upcycle is long ^[5].
• **[ESCALATED] California data-center regulation:** In the final week of session, California state legislators pushed through seven bills requiring energy and water consumption reporting, environmental review before construction, and a special electricity tariff to prevent cost-shifting to other ratepayers ^[7].
• **[NEW] Capital flows:** SB Energy, the SoftBank-controlled AI infrastructure platform with Nvidia backing, filed for a Nasdaq IPO targeting a $5-7B raise (size per reports, single-source) ^[16]. SpaceX reorganized its data-center construction leadership, replacing multiple heads and pulling in executives from rockets and satellite internet (per people familiar, single-source) ^[17]. Sunshine Co.'s unit signed a 633M yuan five-year server-leasing contract with a counterparty disclosed only as 'Party A' ^[18].
• **[NEW] China state-backed compute network:** Southern Power Grid, China Telecom, and State Power Investment formed a '1+M+N' national compute-network joint body; the power-industry node is positioned as the first industry-level hub integrating power, carbon, compute, data, intelligence, and network ^[19].

4. Policy Tensions and What Decides Next

• **[NEW] Open-weight vs testing frameworks:** Meta CEO Mark Zuckerberg said countries should not restrict open-weight AI models, while Google DeepMind CEO Demis Hassabis told a G20 tech meeting that countries should establish AI testing frameworks — two governance philosophies aired within hours of each other ^[20][21]. Treated as live tension, not resolved.
• **[NEW] CrowdStrike–Google Cloud:** CrowdStrike announced an expansion of its Falcon platform within the Google Cloud enterprise AI ecosystem ^[22].
• **[NEW] Goldman semiconductor debates:** Ahead of its September 8-11 Communacopia + tech conference in San Francisco, Goldman laid out five core debates across AI compute demand, equipment cycle, memory supply-demand, analog recovery, and EDA AI commercialization, framing AI capex resilience as the through-line ^[23].
• **[NEW] China AI quick hits:** Tencent Hunyuan released a Hy4 preview lightweight version; Alibaba Cloud cut Qwen3-VL-Rerank pricing; Doubao phone cleared network-access approval for September sale; iFlytek open-sourced two on-device general models; BYD's next-gen 4D mmWave radar chip entered mass production; Manus resumed independent operation; Volcano Engine released AgentSentry; China led the world's first legged-robot international standard ^[24].
• **Falsifiable test:** whether Anthropic's 75% cache-read cut forces OpenAI and Google to match within a week ^[4]; whether California signs the seven data-center bills into law ^[7]; whether SB Energy's IPO prices in the announced $5-7B range ^[16]; whether Google releases Gemini 3.8 Flash on Wednesday as planned ^[3].