1. Moonshot Releases Kimi K3 — the Largest Open-Weight Model Ever Built
Beijing-based Moonshot AI launched Kimi K3 on July 16: a 2.8-trillion-parameter mixture-of-experts model activating 16 of 896 experts per token, with native multimodal understanding and a one-million-token context window. Self-reported benchmarks put K3 ahead of Claude Opus 4.8 (max) and GPT-5.5 (high) on most tests while trailing Claude Fable 5 and GPT-5.6 Sol; it does take the top spot on the Frontend Code Arena benchmark. It is roughly 75% larger than DeepSeek's V4 Pro, and full open weights are due July 27.
2. 29 Countries Sign On to China-Led World AI Cooperation Organization
At the World AI Conference in Shanghai on July 16, twenty-nine countries signed the founding agreement for WAICO — an independent intergovernmental body headquartered in Shanghai and pitched as an alternative to Western-led AI governance forums. Xi Jinping delivered his first-ever WAIC keynote and pledged 5,000 AI training places for Global South countries. WAIC 2026 itself opened with roughly 300 product debuts and 112 forums.
3. Meta in Talks to Lease Anthropic up to $10 Billion of Compute
Meta Platforms is negotiating to rent out AI computing capacity to Anthropic in a deal worth as much as $10 billion over two years, first reported July 17. Anthropic proposed the arrangement in June and would pay Meta in monthly increments. The talks are a stark signal of how scarce frontier compute has become — rival labs are now buying capacity from each other — and would open a new hosting business line for Meta.
4. Anthropic Signs a $19 Billion, 20-Year Data Center Lease in Kentucky
TeraWulf secured Anthropic as the anchor tenant for its Justified Data campus in Hawesville, Kentucky — a converted Century Aluminum smelter site. The 20-year lease covers up to 401 megawatts and is expected to generate about $19 billion in contracted revenue. TeraWulf shares jumped 17% on the news, completing the company's pivot from Bitcoin mining to AI infrastructure. It also sold its 50.1% stake in the Abernathy joint venture to a Fluidstack-led group.
5. Claude Fable 5 Settles Into Max and Team Premium From July 20
After three extensions in five weeks, Anthropic has decided where its Mythos-class frontier model lands. The free promotional window for paid subscribers closes July 19 at 11:59 PM PT; from July 20 Fable 5 becomes a permanent part of Max and Team Premium plans, where subscribers can spend up to 50% of their weekly limit on it. Other tiers move to usage credits at standard API rates, with Standard Enterprise capped at $2,000/day on auto-reload.
6. UK AI Security Institute: Open Models Are Now Only 4–7 Months Behind on Cyber
In its first public analysis of the open/closed capability gap in frontier cyber skills, the UK AI Security Institute found that GLM-5.2 and DeepSeek V4-Pro match closed frontier models released 4 to 7 months earlier — down from a 6-to-10-month gap through most of 2025. On the long-horizon cyber ranges the lag sits at the wide end of that band; on narrower short-horizon tasks GLM-5.2 performs comparably to Opus 4.6.
7. US Navy Approves an 'AI-First Fleet' Strategy
Acting Secretary of the Navy Hung Cao signed the Department of the Navy's Strategy to Weaponize Data and Artificial Intelligence, announced July 16 and effective immediately. The roadmap calls for a data-ready, AI-enabled Navy and Marine Corps, with the explicit framing that slow adoption is a bigger risk than imperfect systems. Cao described the goal as being able to "out-learn and out-fight any adversary."
8. Sakana AI Trains Networks Without Backpropagation
Sakana AI published Error Diffusion, a biologically motivated learning rule that trains Dale-compliant dual-stream networks — separate excitatory and inhibitory populations with non-negative weights — entirely without backpropagation. The method reaches 96.7% on MNIST and 61.7% on CIFAR-10, ahead of other backprop-free alternatives. Far from frontier scale, but a rare result showing gradient-free training can be made to work under hard biological constraints.
// KEY TAKEAWAYS
This week the story is compute and geopolitics more than raw capability. Anthropic is buying power on two fronts — a $19B, 401MW Kentucky campus and talks to rent up to $10B from a direct rival — which tells you the scarcity is real. Meanwhile the open-weight frontier keeps closing: Kimi K3 is the largest open model ever released, and the UK AISI now measures the open/closed cyber gap at 4–7 months, down from 6–10. Governance is fragmenting too, with 29 countries signing a Shanghai-headquartered rival to Western AI bodies and the US Navy formalising an AI-first doctrine.