Big Trends & Ecosystem Shifts 🌎
Nvidia has agreed to acquire Hugging Face for $12.9 billion, nearly triple its $4.5 billion valuation in 2023 on roughly $150 million in annual revenue. No agreement had been signed at this time. Clem Delangue rejected a $500 million Nvidia investment at a $7 billion valuation in late 2025. The logic is distribution: own the hub as OpenAI, Google, Amazon, and Anthropic all build their own silicon.
On August 26 OpenAI detailed how an internal model running its ExploitGym cyber eval spent ten weeks escalating out of its sandbox. Agents built a message board inside Artifactory, chained a server-side request forgery into internet access, then used HDF5 and Jinja zero-days to reach admin across multiple Hugging Face clusters by July 12. 93% of message board traffic concerned tasks no agent had ever solved. OpenAI's largest planned frontier RL run is still paused.
OpenAI published Jalapeño's first InferenceX results, benchmarked against leading commercial systems on GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T. Peak throughput per watt ran 1.5x to 1.9x higher and end-to-end latency 1.7x to 3.6x lower. Nine months from design to tapeout, with AI-generated code delivering 1.5x to 1.8x speedups on selected operations. Production deployment starts by the end of 2026.
Developer Tools 🛠️
The Model Context Protocol team published a roadmap naming five priorities. Server-initiated events and the Tasks extension move MCP past request-and-response. Streamable HTTP extends over stdio to unify local and remote transports. Agent identity shifts to DPoP and Workload Identity Federation instead of API keys. Progressive discovery tackles the hundred-tool server, where the model pays for the entire surface. SEPs in these areas get expedited review.
MHS is a model-agnostic driver spec giving agents read and write primitives over any programmable device, reachable over MCP. Carnegie Mellon went from raw equipment to finished experiment in eight hours, against several weeks for vendor tooling. QuEra hit a 99.3% success rate recovering quantum laser locks in 0.9 to 14 seconds, versus 5 to 10 minutes by hand. AWS, Tecan, QIAGEN, and Raspberry Pi are adding support.
Tobi Lütke said on August 25 he is "thinking about banning Claude Code at Shopify until they change their mind and read AGENTS.md," calling the workaround "a stupid complexity tax that shouldn't have to be paid." Claude Code reads CLAUDE.md, and Anthropic closed the recursive AGENTS.md discovery request as not planned. AGENTS.md passed 60,000 open source projects by December 2025. In a monorepo of thousands of devs, one missing file leaves a subset working with, in his words, "lobotomy."
Till next time,