Saturday brings a mix of infrastructure chess, open-model bragging rights, and courtroom tensions. Meta is reportedly in talks to sell compute capacity to Anthropic, Moonshot’s Kimi K3 is now the largest open-weight model anyone has announced, and Apple is widening its legal net around OpenAI hires.
Infrastructure deals: Meta may lease $10 billion in AI compute to Anthropic
The reported deal would turn a rival into a customer
The biggest strategic story of the day is a reported deal that would make Meta a landlord in the AI infrastructure market. According to the New York Times and multiple follow-on reports, Meta is in talks to lease computing power to Anthropic in a transaction that could be worth up to $10 billion over two years. The arrangement would let Anthropic pay monthly increments for access to Meta’s data center capacity, a structure that mirrors the immense compute crunch every frontier lab is facing.
The deal is still early and terms are fluid, but the implications are clear. Meta has spent heavily on GPUs and custom silicon, and selling excess capacity to a rival would mark a meaningful shift toward an AI-cloud business model. For Anthropic, the agreement would add another source of compute beyond its existing Amazon and Google partnerships.
- Anthropic is said to have proposed the arrangement in June, and Meta is now weighing it.
- If finalized, the deal would be about one-third the reported scale of Anthropic’s earlier SpaceX compute agreement.
- The news lifted Meta’s stock on Friday after a wobbly open, a sign investors are receptive to infrastructure monetization.
Sources: The Chosun Daily, Reuters via Yahoo Finance
Open models: Moonshot’s Kimi K3 claims the 2.8-trillion-parameter crown
A 2.8T-parameter open model with native vision and a 1M context window
Moonshot AI announced Kimi K3 on July 16, calling it the world’s first open 3T-class model. The headline number is 2.8 trillion total parameters with a sparse MoE design that activates only 16 of 896 experts per token, about 1.8% of the pool. It also ships with native vision and a 1-million-token context window.
Moonshot’s own benchmarks place K3 behind Claude Fable 5 and GPT-5.6 Sol but ahead of most other tested models, including Claude Opus 4.8 max and GPT-5.5 high. More interesting than the leaderboard is the project’s scope: Moonshot says an early version of K3 handled the majority of the team’s GPU-kernel optimization work, built a Triton-like compiler called MiniTriton from scratch, and even designed a chip in a 48-hour autonomous run.
- Full weights are promised by July 27, 2026.
- The model is available now on Kimi.com, Kimi Work, Kimi Code, and the Kimi API.
- Case studies include reproducing computational astrophysics research and building interactive visual narratives from web data.
Sources: Kimi blog, Simon Willison’s Weblog, BBC
Legal pressure: Apple expands its OpenAI trade-secret fight
Preservation letters go to about 40 former Apple employees now at OpenAI
Apple filed a trade-secret lawsuit against OpenAI last week, and on Friday the dispute escalated. The company sent legal preservation letters to approximately 40 former Apple employees now working at OpenAI, asking them to save documents and communications that could be relevant to the case.
The original suit named two hires: Tang Yew Tan, OpenAI’s Chief Hardware Officer and former Apple VP of Product Design, and Chang Liu, a former iPhone engineer. Apple now appears to be signaling that it believes more employees may have shared confidential information. OpenAI has called the lawsuit meritless, but the preservation letters put a broad swath of its hardware team on notice.
- Apple has said more than 400 former employees now work at OpenAI, so the targeted group is roughly 10% of that total.
- The letters could also serve as a warning and a basis for future discovery requests.
- The core allegation is that Apple secrets were used in OpenAI’s consumer hardware efforts.
Quick hits
NVIDIA posts two July 17 releases. The company published a Hugging Face guide on scaling image and video fine-tuning with NeMo Automodel and Diffusers, and MarkTechPost reports NVIDIA released Nemotron 3 Embed, an open embedding collection whose 8B checkpoint ranks first on RTEB at 78.46.
OpenAI details GPT-Red. On July 15-16, OpenAI described an internal automated red-teamer trained via self-play reinforcement learning. The model is used to find prompt-injection vulnerabilities and has already hardened GPT-5.6, cutting direct prompt-injection failures by roughly 6x versus GPT-5.5.
Anthropic’s Fable 5 guardrails draw fire. Microsoft CEO Satya Nadella reportedly criticized Anthropic’s Fable model in an internal Copilot meeting, saying its refusals feel “editorially controlled” and don’t make sense for a creation tool. The criticism highlights tension between safety-first labs and enterprise customers who want fewer restrictions.
Nationwide data-center protests are scheduled for today. The group Humans First has organized demonstrations in more than 50 U.S. cities on July 18 to oppose AI data center expansion, pointing to local water, power, and land-use concerns.
Google’s Gemini 3.5 Pro still has not shipped. Reports on July 17 indicate Google has delayed its flagship model beyond its previously rumored July 17 target, with some outlets now pointing to late July or August. Because the delay and the lack of launch remain unconfirmed by Google itself, we are keeping this in quick hits rather than the lead section.
Sources: Hugging Face / NVIDIA, MarkTechPost, The Hacker News, Quartz, Detroit Free Press, Mashable
Recent context
- July 14: Anthropic launched Claude for Teachers, giving verified U.S. K-12 educators free premium Claude access through June 2027. (Anthropic)
- July 9: OpenAI shipped GPT-5.6 (Sol, Terra, Luna) to general availability across ChatGPT, Codex, and the API. (OpenAI Newsroom)
Compiled from primary sources, major outlets, and release trackers on July 18, 2026.