Two things shipped yesterday that point the same direction: OpenAI showed benchmark numbers for silicon it designed itself, and Perplexity shipped an agent that runs entirely on a machine you own with no token meter attached. For three years the assumption has been that serious inference happens in someone else's data center on someone else's GPUs, billed per token. Both of yesterday's announcements are attempts to break a different part of that sentence.
Inference Moves Off the Cloud
Perplexity and Nvidia launched Portable Computer, a fully local agent with no token costs
It runs on an Nvidia DGX Spark or any Linux box with an RTX GPU and at least 24GB of VRAM — roughly a 3090 or newer — using Qwen 3.8 27B or Perplexity's post-trained PPLX 27B, with Nvidia's Nemotron 3.5 Lightning coming. The numbers are the interesting part: 85.4% for PPLX 27B on Perplexity's own 53-task Local Knowledge Work Bench, 66.7% on BrowseComp at 51% less wall time, and 59.6% on Terminal Bench 2.1 locally versus 73.0% when it escalates to the cloud at $0.415 a task. Available now to Pro, Max, and Enterprise subscribers; Windows in September. I covered Nvidia's talks to invest in Perplexity at $30B+ on Sunday — this is the product side of that relationship, and note who benefits when local agents run "always on": Nvidia's own director of developer technology describes the demand for tokens as insatiable.
OpenAI's Jalapeño chip beat Nvidia Blackwell on an inference benchmark
Tested on SemiAnalysis' InferenceX benchmark, the Broadcom-co-designed accelerator came out ahead of a Blackwell system on both tokens per user and throughput per kilowatt. The design targets prefill and communication as the bottlenecks, keeping KV cache placement local rather than shuttling it around. Hardware head Richard Ho calls it "a very significant performance advance." Timeline matters more than the benchmark: announced October 2025, small-volume deployment at the end of this year, meaningful volume in 2027. One benchmark from the chip's own maker is not a market shift.
Business & Industry
OpenAI lost its head of data centers
Chris Malone left last week after joining in March 2025, following nearly five years at Meta and over a decade at Google. He ran execution on the Stargate buildout with Oracle, Nvidia, SoftBank, and Microsoft, and the company says it recently reorganized infrastructure so he reported to VP Sachin Katti rather than president Greg Brockman. That's at least 13 executive departures this year — COO Brad Lightcap and CRO Denise Dresser both in August, product chief Fidji Simo and ethics head Chloé Bakalar in July.
Funding
Generalist hit a $3B valuation on a $600M Series B
The 2024 robotics startup — founded by ex-DeepMind researchers Pete Florence and Andy Zeng with former Boston Dynamics engineer Andrew Barry — extended a $400M June round at $2B with roughly $200M more led by 8VC. Its Gen 1.5 model learns tasks from 3-to-12-second video demonstrations. Bezos Expeditions, Nvidia, and Fei-Fei Li are in. For scale, Skild AI sits at $14B and Physical Intelligence at $11B.
Stability AI raised $76M from three major record labels and EA
Universal, Sony, and Warner Music all put money in alongside Electronic Arts, AMD Ventures, and Pacific Alliance Ventures, bringing Stability's total to $232M. That's the unusual line: the labels suing generative AI companies elsewhere are funding this one, which follows partnership deals with each of them signed in late 2025. Stability largely won Getty Images' UK copyright case; the US suit is still open.
Runable raised $21M at a $65M valuation
A 15-person Bengaluru team with roughly 1.7 million registered users and a $2M annualized run rate reached within three weeks of turning on payments. Susquehanna Venture Capital and Nexus Venture Partners co-led. The honest detail TechCrunch got them to admit: gross margins are currently negative because they subsidize inference, and the plan is that token prices fall faster than usage grows.
Skipped as already covered: the Hugging Face sale talks, Alibaba's Wan3.0 launch, and XPeng's robotics round all ran Monday with nothing new since.
Sources
- VentureBeat — Perplexity partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs
- TechCrunch — OpenAI's Jalapeño chip is built for fast inference at scale, benchmarks show
- TechCrunch — OpenAI loses a top data center exec as stream of high-profile departures continues
- TechCrunch — Robotics startup Generalist reaches $3B valuation, sources say
- TechCrunch — Stability AI, maker of image generator Stable Diffusion, raises $76 million in fresh funding
- TechCrunch — Runable hits $21M to bet AI agents can go from building businesses to growing them
