OpenAI showed the first benchmarks for its own chip and claims it beats Nvidia. Apple shipped the first 2-nanometer M6. Perplexity taught its agent to run entirely on your own graphics card.
🤖 AI
Perplexity launches a cloud-free agent
Together with Nvidia, the company released Portable Computer — a version of its agent platform that runs on your own hardware and burns zero tokens. You need an RTX card with 24GB of memory; Linux only for now, Windows coming in September.
Anonymous model Ox Alpha chewed through 26 trillion tokens in four days
Nobody knows who made it: it showed up on OpenRouter and OpenCode with no listed owner, free to use. In its first four days, 327,000 people ran 8.3 million sessions — a record for a single model launch.
Anthropic merged memory between Claude and Cowork
Chat and Cowork now remember the same things, and the feature is on by default. Claude still won't remember health, religion, or politics by default, and you can open the topic list and clear it.
💻 Software
Uber handed 70% of its pull requests to agents
The company broke down its "software factory": a model gateway, an MCP gateway with a thousand tools, and a shared context graph. Code output per engineer doubled over the year, and 250 migrations spanning 9 million lines shipped with zero humans involved.
Vercel Connect goes generally available
The service issues agents short-lived, task-specific tokens instead of permanent keys, so secrets stop living in environment variables. The beta already pulled in more than 100 connectors — Slack, GitHub, Snowflake, you name it.
⚙️ Robots and hardware
OpenAI shows first results for its Jalapeño chip
The inference accelerator, built with Broadcom, beat current Nvidia chips in OpenAI's own benchmarks. OpenAI expects to deploy it internally before the end of the year.
Apple ships its first 2nm chip
The M6 in the new Mac mini and the four-die M5 Ultra in the Mac Studio were built for local inference: the AI GPU gained nearly 30% over the M5, and memory tops out at 512GB.
Nvidia puts Groq 3 LPX into full production
A rack of 256 accelerators rounds out the Vera Rubin platform and takes over token generation — 3,400 tokens per second on Gemma 4 31B at a 100K context window. Nebius, CoreWeave, and SpaceXAI get the first chips.
Tesla names a launch date for Cybercab
The robotaxi with no steering wheel or pedals debuts in Austin on September 3, by invite only, for the most active Robotaxi riders.
💰 Business
Hugging Face is testing the waters for a $13B sale
The company hired a bank to gauge buyer interest. That's triple its 2023 valuation of $4.5B — but there's no deal yet.
Nvidia warns customers of 15%+ server price hikes
The culprit is a record spike in memory prices: the new pricing kicks in for systems shipping in early 2027, hitting both Vera Rubin and Grace Blackwell.
SpaceX will pour $100B into a Louisiana launch site
Five launch complexes for Starship are going up across 125,000 acres of coastal wetlands — up to 10,000 launches a year. Construction starts in 2027, first launch expected in 2029.