Physical Superintelligence raises $58M after its AI plotted a star route
A physics lab staffed by virtual physicists came out of stealth with $58 million, a new world model put cameras back in the director's hands, and a startup that tunes open models for a company's own traffic closed a $40 million Series A.
Physical Superintelligence (PSI) launched on September 1 with $58 million in seed funding led by Breakthrough Energy Ventures, and its first public result is an interstellar one: the trajectory for the Fermi Explorer Mission, a nonprofit probe that aims to launch toward Alpha Centauri by the end of 2029. The Cambridge, Massachusetts company was founded by Matt Pines, Alex Klokus and Alexander Wissner-Gross to build what it calls virtual physicists — a platform named Emmy, after Emmy Noether, that decomposes a research problem into trees of verifiable hypotheses and tests them in parallel against a large library of simulations. On the Fermi mission PSI says it validated the physics and found a substantially more efficient path within the mission's mass and budget limits. Mission co-founder Philip Johnston told MIT Technology Review his team spent a year failing to find a route for a small solar-powered spacecraft costing $15 million, and that PSI's system ran largely on its own for three days on roughly a billion tokens, with an astrophysicist on staff steering it, requesting a cost analysis, and checking the output for errors.
The interesting part is not the star shot — a probe that takes up to 80,000 years to arrive is a symbol, not a transportation plan. It is the commercial thesis underneath it: PSI's first paying work is designing and retrofitting AI data centers and AI factories, terrestrial and orbital, across power, cooling, network and compute. An AI that can solve a multiphysics design problem under hard mass constraints is the same tool that decides how a gigawatt of GPUs gets cooled. Launching with a Breakthrough Energy-led seed and an interstellar headline is a very effective way to recruit physicists for what will mostly be infrastructure work.
World Labs, Fei-Fei Li's spatial intelligence company, unveiled Atlas, a world model pretrained from scratch to natively handle text, images, video and 3D. Atlas is a multimodal autoregressive diffusion transformer whose inputs are grounded in 3D space to form a shared spatial context, and its headline capability is precise camera geometry as a native input type — you specify a camera position and angle rather than describing a shot in words. It generates new views from one or a handful of reference images, reconstructs spaces as point clouds or 3D Gaussian splats, and the company claims it beats specialist models on both camera-conditioned generation and sparse-view 3D reconstruction. The most useful frame on this: more input images means less imagination — with one photo Atlas invents what it can't see, and by three it's reconstructing faithfully.
The pitch is control, not novelty. World Labs explicitly casts Atlas as putting you "in the director's chair" rather than pulling a slot machine, and it leans on the same trick that makes modern video models cheap at inference and servable at scale: it's an autoregressive transformer, so it can borrow LLM serving machinery. For robotics it doubles as a simulator, generating the RGB and depth frames a robot's body-mounted cameras would see from a reconstruction built out of a couple dozen phone-video frames. Atlas is in early access with select partners only, so the benchmarks are the company's own for now.
Wafer raised a $40 million Series A, co-led by Marathon and Chemistry, to automate the work of tuning open-source models for a specific company's traffic. A source put the valuation at over $200 million. The company says most inference optimization today is manual, service-heavy, and done once before deployment, and argues it should be continuous: Wafer learns from a workload's traffic patterns and performance constraints, then searches for the best deployment across the model, engine, kernels and hardware. Round participants include Wing, AMD Ventures and Outset Capital, with existing investors Fifty Years and Y Combinator returning; the angel list includes Jeff Dean, Guillermo Rauch and Matthew Prince. It published benchmark claims reading like hardware marketing — Kimi K3 at roughly 952 tokens per second per node on AMD, GLM5.2 at 2,626 tokens per second per node on MI355X at what it says is less than half the cost of Blackwell.
This is the unglamorous layer where the token economy actually gets decided. As agentic workloads pile up, the difference between a hand-tuned deployment and a default one is the difference between a viable margin and a burned one, and "performance per dollar" is now the metric every inference vendor sells against. The tell on how hot this lane is: Wafer reportedly turned down acquisition offers to take the round.
What to watch: whether PSI's Emmy shows up in a data center design contract before it shows up in a physics journal, and how fast Atlas moves from early access to a public API.
If an AI picks the trajectory, who signs off on the mission — the physicist who steered it, or the lab that built it? Tell us in the comments.
