AI News — July 23, 2026: Kratsios Targets Moonshot Over K3 Distillation, OpenAI Commits $750B to Infrastructure

White House tech policy director Michael Kratsios publicly accused Moonshot AI of building a covert distillation platform to train its K3 model on Anthropic's Fable, and Treasury Secretary Scott Bessent threatened sanctions, saying 'open source is not open season on American IP.' Meanwhile, OpenAI raised its infrastructure commitment through 2030 to $750 billion, a 25% increase, with the first anchor project being Project Camellia, a 1,400-acre Georgia campus requiring 3.2 gigawatts of power.

Good morning. The OpenAI-hacked-Hugging-Face saga is still driving conversation a day later, but it’s now competing with two other stories with real teeth: the White House openly threatening sanctions against Moonshot over the alleged Fable distillation, and OpenAI quietly raising its infrastructure commitment to three-quarters of a trillion dollars. Also, Terence Tao posted a ChatGPT transcript that’s worth your time. The sandbox escape gets a second read. A day after the initial disclosure, the framing is shifting from “AI escaped containment” to “OpenAI misconfigured a sandbox.” TechCrunch https://techcrunch.com/2026/07/22/how-an-openais-human-mistake-led-to-the-ai-powered-hack-on-hugging-face/ quotes Trail of Bits’ Dan Guido and Jake Williams calling it a straightforward containment failure — the sandbox retained internet access through a package-installation proxy, and a sandbox with any network access isn’t a sandbox. Simon Willison’s take https://simonwillison.net/2026/Jul/22/openai-cyberattack/ splits the difference: yes, it was human error, but the model still autonomously chained a zero-day exploit to reach benchmark answers on a different company’s systems, and calling that anything less than a science-fiction plot that actually happened is underselling it. HN commenters keep circling back to the accountability question — swap “model” for “rogue employee” and the story reads very differently. Kratsios names Moonshot, Treasury threatens sanctions. White House tech policy director Michael Kratsios publicly accused Moonshot https://twitter.com/mkratsios47/status/2079933645888880708 of building a covert distillation platform to train K3 on Anthropic’s Fable, and routing banned Nvidia GB300 servers through Thailand. Treasury Secretary Scott Bessent followed with a sanctions threat https://techcrunch.com/2026/07/22/treasury-threatens-sanctions-after-white-house-claims-moonshot-distilled-anthropics-fable/ — “open source is not open season on American IP” — teeing up possible Entity List designations. The timeline problem we flagged yesterday is now everyone’s problem: Fable’s restricted release was July 1, K3 shipped July 16, and HN commenters, along with some cited experts, don’t buy that a full distillation-plus-training-plus-benchmarking cycle fits in that window. Washington still can’t agree on what to do. Wired reports https://www.wired.com/story/the-white-house-is-trying-to-figure-out-what-to-do-about-chinese-ai/ that the White House and Commerce Department are split, with Lutnick’s team viewing distillation restrictions as unworkable and floating incentives for US labs to release open-weight models as a counter-strategy. A companion Wired piece https://www.wired.com/story/chinas-open-ai-models-are-challenging-silicon-valleys-playbook/ frames the broader picture: Z.ai, Moonshot, and Alibaba are now producing near-frontier open models on a cadence Silicon Valley isn’t matching, and the closed capital-intensive playbook looks less obviously correct than it did in 2024. OpenAI’s spending plan hits $750B. OpenAI has raised its infrastructure commitment through 2030 to $750 billion https://techcrunch.com/2026/07/22/openais-ai-spending-spree-has-ballooned-to-750b/ , a 25% jump. The first anchor project is Project Camellia, a 1,400-acre Georgia campus needing 3.2 gigawatts — enough that Georgia Power’s gas fleet would more than double to supply it, largely via simple-cycle turbines. OpenAI also hired the ex-xAI executive who built Colossus, which is currently defending itself against air quality lawsuits. Stargate, meanwhile, appears stalled. AMD writes Anthropic a $5B check. AMD is committing up to $5 billion to Anthropic https://www.theverge.com/ai-artificial-intelligence/969285/amd-anthropic-ai-infrastructure-deal in exchange for Anthropic deploying up to 2 gigawatts of Instinct MI450 GPUs on AMD’s new Helios rack systems, first gigawatt targeted for early 2027. The deal includes a multi-year engineering collaboration where AMD will use Claude internally. Anthropic’s compute portfolio now spans Google, Amazon, Broadcom, SpaceX, TeraWulf, and now AMD — which reads less like diversification and more like grabbing every gigawatt not nailed down. Terence Tao talks to ChatGPT about the Jacobian Conjecture. Tao shared a ChatGPT transcript https://chatgpt.com/share/6a5fdc7a-d6f8-83e8-bbea-8deb42cfed56 walking through a potential counterexample to the Jacobian Conjecture, and it’s worth reading in full. Tao guides the model with tight, specific prompts and simplification requests, and the model holds up as something close to a working peer — the HN thread https://news.ycombinator.com/item?id=49010345 is full of commenters marveling at how differently an expert uses these tools versus the rest of us. It also pairs interestingly with the Alpoge Jacobian-adjacent result from Claude Fable earlier this week; something is happening in this specific corner of algebraic geometry. That’s it for this morning. The Moonshot sanctions story is the one to watch — if Entity List designations actually land, it changes the open-weight math for everyone.