FieldAI set to raise $700M at a $10B robot-brain valuation
Robot brains are starting to be priced like frontier-model labs. Friday's news window pairs an eight-billion-dollar valuation jump in physical AI with an open training stack for trillion-parameter models and a new voice product from AI music's most litigious player.
Robotics startup FieldAI is set to raise $700 million at a $10 billion valuation, quintupling its mark in just over a year. A person familiar with the deal said the company has signed a term sheet, so the financing has not formally closed, and FieldAI did not comment; the round's lead investor is not yet known. The 2023 startup, founded by former NASA Jet Propulsion Laboratory autonomy researcher Ali Agha, builds what it calls a "universal general-purpose brain" — one model family meant to run across humanoids, robot dogs, drones and industrial rovers operating in messy, unpredictable places like construction sites and mines. It was previously valued at $2 billion after raising more than $400 million last year from investors including Jeff Bezos' family office, Laurene Powell Jobs' Emerson Collective and Khosla Ventures, and it has added at least $35 million in revenue and customer contracts since June to pass $135 million across more than 30 customers. The number that matters is the relative one: rival Physical Intelligence sits at roughly $11 billion and Skild AI above $14 billion, so $10 billion puts a general-purpose-brain bet squarely inside the pack rather than above it — investors are now underwriting robot foundation models with the same conviction (and the same impatience for real revenue) they once reserved for chatbots.
Ai2 open-sources Olmo-core 3, a training stack that takes mixture-of-experts models to a trillion parameters. The Allen Institute for AI's framework grows the expert pool from eight to 128 while still activating only four experts per token, and Ai2 reports 52,000 tokens per second per GPU on eight NVIDIA B300 GPUs for a 47-billion-parameter model — about 2.7 times its own earlier implementation, which managed 19,400. The team also trained a 1.2 trillion-parameter configuration across 512 GPUs as a capacity demonstration; the code ships under Apache-2.0. The pitch is that trillion-parameter training stops being the exclusive property of labs with hyperscaler budgets, though Ai2's own caveat is worth keeping: the headline benchmarks used random routing, so this is a systems measurement, not proof that a trained model beats anything. We covered Cloudflare's open decision models built on Qwen weights earlier this week — Cloudflare open-sources the Clef models, aiming straight at Jev — the open-stack trickle keeps turning into a flood.
Suno launches Speech (beta), turning scripts into voiced tracks with music baked in. The public beta, live on web and mobile, generates spoken voiceovers from a script or prompt with optional background music in a single pass — Suno calls it the first audio model to produce voice and music together as one cohesive track — and the company is candid that accents can wander and dramatic pauses may be very dramatic. Alongside the launch, chief executive Mikey Shulman said Suno's last disclosed numbers were two million subscribers and $300 million in annual revenue and that the company is "far beyond that by now," in a Bloomberg conversation. It is a product story with a legal shadow: the music generator is still defending suits from major labels while it expands into the voice territory ElevenLabs and others already occupy.
What to watch: whether FieldAI's term sheet closes on the reported number, and who ends up leading the round.
Is a $10 billion valuation for one robot brain a signal about robotics or about how much money still needs a home? Tell us in the comments.
