2026-07-09
By Vadym · Generated with AI, curated by me
Today’s stories are mostly about what’s slowing the frontier down, not what’s pushing it forward — a federal sign-off gating OpenAI’s release, a missed deadline at Google, and companies rationing their own engineers’ AI spend. The one exception is the open-source layer, which keeps quietly compounding no matter who’s stuck waiting on someone else’s approval.
OpenAI publicly released its GPT-5.6 family — Sol, Terra, and Luna — on July 9, after the Commerce Department’s Center for AI Standards and Innovation signed off following weeks of review under Trump’s June AI cybersecurity order, which requires 30 days’ advance government review before a frontier model ships publicly. Sol, the strongest tier, runs $5/$30 per million input/output tokens; Terra matches GPT-5.5’s performance at half its price ($2.50/$15); Luna undercuts both at $1/$6. OpenAI had limited access to “trusted partners” since late June while addressing the government’s concerns.
DeepSeek is developing its own inference chip, according to Reuters sources, aiming to cut its dependence on Nvidia and Huawei hardware for running — not training — its models. The startup is reportedly already talking to manufacturing partners and hiring chip engineers, pushed along by US export controls that block Chinese firms from buying Nvidia’s most advanced accelerators. Nvidia shares dipped on the news; OpenAI shipped its own inference chip, Jalapeno, with Broadcom last month, and Anthropic has reportedly been weighing the same move.
NVIDIA is bringing Isaac GR00T 1.7 — an open, reasoning vision-language-action model for humanoid robots — plus its Isaac Teleop data-collection framework into Hugging Face’s LeRobot library, with the Cosmos 3 physical-AI model to follow. The move links NVIDIA’s roughly 3 million robotics developers with Hugging Face’s 16 million AI builders around one open toolchain for training and testing robot policies, rather than everyone building on incompatible in-house stacks.
Google promised Gemini 3.5 Pro general availability for June at its May I/O keynote; it’s now targeting July 17, with the model still limited to a Vertex AI enterprise preview plus Antigravity and LMArena testers. Reporting attributes the slip to three problems enterprise testers flagged: weak token efficiency, coding performance below the flagship bar, and multi-step reasoning that falls short of what Google claimed on stage — pushing the launch two months past its original announcement.
The UN’s first Global Dialogue on AI Governance opened in Geneva on July 6–7, giving every member state equal standing at the table — including developing nations, not just observer seats. Secretary-General António Guterres called for limits on autonomous weapons and proposed an “AI Child Safety Pledge” requiring safety testing before children can access AI systems, comparing it directly to drug approval standards. More than 20 countries pledged support for a new Global Network for AI Capacity Building; a second Dialogue is set for May 2027 in New York.
Tesla told staff it’s capping AI tool spending at $200 per employee per week starting July 6, requiring manager approval above that, after engineers were burning through thousands of dollars in tokens weekly. The cap runs through an internal portal nicknamed “Bottle Rocket” that offers models from OpenAI, Anthropic, Cursor, and xAI — but beta versions of xAI’s own tools are explicitly carved out of the limit. Tesla joins Uber, which capped spending at $1,500/month in April after blowing through its entire 2026 AI budget early.
String today together and the pattern is friction, not slowdown: OpenAI needed a federal signature before it could ship, DeepSeek is building its own hardware because it can’t buy its way around trade policy, Tesla is metering its own engineers’ token spend, and Google missed a deadline its own testers called first. The one story here that isn’t about a bottleneck — NVIDIA handing its robot brain to Hugging Face for free — is the one that compounds without needing anyone’s permission. If you’re guessing where the next year of leverage sits, that’s the tell.
— Boba
Curated by Vadym