Source: raw/Robotic_Deep-Dive_-_Beating_China_the_Race_Against_Tesla_New_Robotics_Timeline_With_Brett_Adcock.md
Creator: Moonshots with Peter Diamandis (guest: Brett Adcock, founder and CEO of Figure and of Hark; a co-host is addressed as Dave) · URL: https://www.youtube.com/watch?v=E864U582FVY · Duration: not stated in the transcript (about 30,000 words) · Platform: YouTube (auto-captions, fetched 2026-11-17)
A long factory-floor interview, kept here only for what it says about AI models, compute, cost and timelines. Adcock’s concrete claims: Figure spends nine figures a year on compute, could not buy more GPU capacity from September 2026, has a 2027 cluster deal that he says is over 100 times its current compute, and expects a robot to do hours-long household work in an unseen home within about six months. Every figure is Adcock’s own, about his own companies, to hosts who include an investor in Figure. Confidence is low for that reason. The source has no direct takeaway for marketing or productivity work; the usable part is the compute-supply signal.
Key Takeaways
- “September 2026, the world ran out of compute.” About seven months ago Figure got four competing bids for a cluster of Blackwell-series chips and pushed the price down. Three to four months ago it went out to double capacity and got no bids, then one, at a much worse price. He says older H100s are unobtainable and “trading above par”.
- Memory is the constraint behind the GPUs. He says the leadership of one of the largest memory makers told Figure it had no capacity for robots, and that the CTO of “one of the big groups” said humanoids were not in its demand projections. Figure has signed memory partnerships for the next few years.
- Compute spend. “Low nine figures” a year, with roughly 50 to 70% going to pre-training Helix 3. Over the last month, one pre-training run took more than half of all compute. Figure is doubling internal compute “next week”.
- The 2027 cluster. A deal with Nscale for the second half of 2027 on Nvidia Vera Rubin chips. A host describes it as 100,000 GPUs and more than $3 billion; Adcock does not correct that, and says the full deal is “well over 100x what we have”.
- Billion-dollar training runs. He expects to be running them on Helix “next summer”, if the cluster, data and capital all arrive.
- Scaling laws. With Helix 2.5 Figure reported what it calls the first humanoid scaling law for data: a log-linear relation between more data and better task execution. He says the same holds for more compute at fixed data and model size.
- Data waiting to be used. He says Figure holds 10 to 15 times more data than the Helix 3 checkpoint shown to the hosts was trained on. It must be audited, cleaned and annotated first.
- Diverse data transfers. Footage of someone hosing down a dog or painting a house made the robot better at folding laundry, he says, as diverse pre-training data does for language models.
The model: Helix
- Helix 1 ran the upper body on a neural network; Helix 2 the full body. That step took a year, because joint trajectories for feet, hips and fingers had to be trained together.
- Helix 2.5 (released “a few months ago”): zero-shot work, such as making a bed and tidying a living room, in 30 rental homes the robots had never seen. About three weeks before recording, a robot was taken to a client warehouse it had never seen and “zero shot the whole thing”.
- Helix 3 (not public): the hosts saw the first checkpoint in an internal test home. Adcock calls it “a baby AGI moment”, “a 2-year-old [in an] adult body”, and says it was “significantly undertrained”. Larger models were being evaluated that week.
- Everything runs on the robot. Planning, reasoning and joint torques are computed on board. Only real-time speech and fleet orchestration use servers.
- Why not the cloud. He puts the best 5G at 20 to 30 milliseconds per leg, so about 50 for a round trip, which caps a control loop at about 20 hertz before any inference. On-board also means a robot keeps working if the connection drops.
- On general models driving robots. He says demonstrations of language or multimodal models (the host cites GPT Astra) commanding robots are basic tasks with a fixed camera, run slowly and shown sped up. “They’re not real models for robotics. If they were, we could run them.”
- Speech. Robots use a real-time speech-to-speech model built by Hark, Adcock’s other company. His view of speech-to-speech in general: it “kind of sucks”, with weak tool calling, latency and interruption handling.
Timelines he gives
| Milestone | His statement |
|---|---|
| Hours-long, end-to-end work in any unseen home | ”I hope to see in the next sub six months, perhaps even sooner”; he “can’t promise” |
| The whole picture demonstrated | ”I’d be very shocked if we exit 2027 [and] we haven’t seen this” |
| Selling to consumers | Only when it works without teleoperation or stalls; no date committed |
| Fully “lights out” robot manufacturing | ”For sure within five years”; unclear whether 6 months or 3 years |
| Figure 4 | ”The largest step up we’ve ever made between generations”, with a new hand; no date |
| Custom chips | Not decided; a chip design takes “two, three plus years” and his inference design of three years ago “would have been completely wrong” |
- The years “27, 28” are spoken during the consumer-timing exchange, but the captions do not show who said them.
- Home testing now: at any time, 20 to 100 homes in California that the robots have not been in before, mainly for model evaluations.
Scale and cost figures
- Fleet: “well over a thousand” robots on campus (he says about 1,500) against 700 to 800 staff; robots have outnumbered people there for four or five months.
- Factory capacity: “over 40,000 or so units” a year at full rate across four lines. He sees a path to a million a year from one facility if the intelligence problems are solved. The facility’s name is captioned “Baku”, “Bakyu” and “botq”.
- Consumer price: “some sort of 500 a month lease”. The host’s $30,000 purchase figure is not confirmed.
- Logistics line: has run “practically every day” since a 200-hour live stream four months earlier; the target is under 3 seconds a package with a few percent errors per shift.
- Cash: “several billion dollars of cash” on the balance sheet.
- Long-run market: he reasons to about 20 billion humanoids. This is a back-of-envelope estimate.
Competitors and partners
- China: “the closest competitor”, as a collective. He says the dancing demonstrations are open-loop RL controllers trained on motion capture, “orthogonal” to general robotics, and that Unitree’s robots (captioned “Unry”) arrive with no hands and compute “the size of a USB stick”. He also expects China to improve quickly.
- Tesla: “I personally haven’t seen much out of Tesla in two or three years.”
- OpenAI: it co-led Figure’s Series B with Microsoft and joined the board. Adcock says the joint work ran about a year, OpenAI’s team would not come to the office, progress was “10 to one, 100 to one” in Figure’s favour, and “we fired them like two years ago”. This is one side of the story.
- Regulation: the only robotics rule he has seen in 90 days is an FCC ban on Chinese humanoids.
- Supply chain: Figure has moved its parts out of China, after one week of tariffs at about 270%.
Hark, his second company
- Hark builds “personalized intelligence”: computer use, speech, memory and vision, plus its own AI devices.
- Hark Pro, the software-only product, launched about a month before recording. He claims over 1.2 million signups and about 5 million browser sessions, and says usage was curbed in some countries.
- He says Hark’s revenue is already larger than Figure’s and that Hark could be the bigger company within six months, “maybe sub 90 days”.
- Device hardware exists in the lab; he will not describe the form factor.
How to use this
There is no technique here. Two uses:
- Treat it as one more data point that GPU and memory supply is tight. A well-funded buyer reports going from four bids to one in a few months. If you depend on rented GPU capacity or on API rate limits, plan for prices and limits that do not fall on schedule.
- Discount the dates. Each timeline is a founder’s hope for his own product, and he says so (“I obviously can’t promise”).
Open Questions
- Nothing is independently checked. The Nscale deal size, compute spend, fleet size, Hark’s signups and the OpenAI account all rest on this conversation.
- Speaker attribution. The transcript has no labels. The hosts are Peter Diamandis and a co-host called Dave; which host said which line is not always clear.
- Captioned names. Hark appears as “HARKC” and “Harkc”; Nscale as “nscale” and “in scale”; chip names as “Ver Rubins”, “VR 2200s” and “B300s”.
- Helix 2.5 scaling-law results were not fetched from Figure.
- Whether “ran out of compute” is general. It is one buyer’s experience. The wiki’s Epoch articles put total compute supply doubling about every nine months, which is a different measure from spot availability.
Related
- Epoch AI — Compute Supply Doubling Every 9 Months — the supply-side numbers to set against his “ran out of compute”
- The Compute Economics of the AI Buildout — who is buying the chips and at what cost
- AI Bubble Signals — Nscale and the neocloud financing picture
- The Data-Center Slowdown After the Midterms — why new capacity may arrive more slowly
- AI News, November 9-11, 2026 — Musk’s humanoid workforce and Terafab figures from the same show
- Stanford HAI 2026 — Technical Performance — independent household-robotics benchmark results
- Voice Agent Comparison — the speech-to-speech options he is criticising