Robot embodiment
Software agents can loop all day on a laptop. The next decade-long frontier, in the panel-class read, is bodies — humanoids in factories and streets, and a US–China race over who owns the stack.
With the pure-software AI wave maturing, several Moonshots and All-In-adjacent panels converge on the same rotation: capital and talent should move to embodied AI — humanoid and general-purpose robots — on a “10-year theme, just like biotech.” One investor framing on the wiki: “the window of opportunity for pure software AI is probably another couple of years and then the self-improvement loop is going to take over.” College CS majors, the same sources claim, are “shifting back to material science, MechE.”
Andreessen Horowitz issued what the wiki quotes as a “wake-up call”: America needs to work with allies to build a “defensible AI robotics stack” — “This is the time, but … not much time.” China’s playbook, in that telling, mirrors solar, EVs, and 5G: an “AI+” five-year plan integrating AI through the physical economy, including 150+ Chinese humanoid-robotics companies, and robots building solar panels to generate energy to build more robots — physical recursion described as the “innermost loop.”
Three jobs at once
Robots are framed three ways simultaneously on the wiki. They are the innermost loop — “the robots that build the data centers and the fabs that build the chips that host the models.” They are a new real-world training-data source as labs move teams toward embodiment; OpenAI redeployed its SORA video team under Aditya Ramesh toward an in-house robotics effort, read partly as a grounding play. And they are capital assets whose economics differ from fast-depreciating GPUs — robots “will not have the same depreciation schedule.”
Ray Kurzweil names embodiment as one of two remaining AGI blockers: “robotics is behind large language models … we don’t have robotics that can do that [clean up after dinner] at any price [and it] needs to be made less expensive. I think that will come about 2029.” Deployment friction is the counterweight. The US “still struggling to get Waymos” in many cities is a leading indicator; humanoid general manipulation may hit a self-driving-style “last 10% is a decade” wall.
Concrete datapoints the wiki carries, all flagged as capability snapshots not settled forecasts: Tesla targeting roughly 10 million square feet of Optimus production capacity with Optimus 3 expected; Figure running continuous package sorting for about eight days; Hyundai’s Atlas learning soccer. Wissner-Gross helped form “ProRL” to put humanoid and non-humanoid robots on US streets.
America needs a wake-up call on robotics … the US must work with allies to build a defensible AI robotics stack.
Andreessen Horowitz framing, Moonshots, 2026-06-06
August 2026 brought a single-source researcher post claiming Gemini 3.7 Flash scored 92% on a physical tool-use benchmark where Gemini 3.6 Flash scored 32% three weeks earlier. The wiki treats those figures as the researcher’s claimed scores on X — not a Google announcement, not a model-card number, bench name unstated. Do not promote them as official robotics progress.
Where the model thinks
The panel argument was about who owns the stack. A September SemiAnalysis note, paywalled on the cost sheet, asked a different question: where does the model actually run?
Ivan Chiam’s read, as the wiki has it: frontier robot models are still small — 5 to 14 billion parameters, names like π0.7 and DreamZero. Hierarchical vision-language-action systems can put the slow planning layer on a datacenter GPU and keep the high-frequency action loop on the robot. Jetson Thor, Nvidia’s on-device board, cannot run DreamZero in real time. A B300, in that note, serves seven robots under a 1.6-second motion budget. Past about five to seven robots, shared silicon and DRAM look cheaper than a GPU on every chassis. The bill of materials is behind the paywall. This is a where-the-flops-live argument. It does not settle the China-stack race, and it does not move Kurzweil’s 2029 dinner-cleanup date.