Summary

Figure’s bet is that humanoid economics improve when one general-purpose hardware platform serves most human work. Adcock rejects customer-specific hardware because high NRE, maintenance burdens, and changed observation spaces impede transfer learning; standardization is the path to “twenty thousand dollar levels,” while Figure’s near-term scale target is 100,000 units.
Deployment should arrive unevenly: indoor workforce work first, then homes, outdoor jobs, and space, though not strictly sequentially. Adcock places this broad rollout within the next 10 years but distinguishes demonstrations from scaled integration; Khosla expects farm capability within five years and considers 1% penetration “the job done,” with broad adoption taking longer.
Khosla sees fusion’s bottleneck less in physics than in siting, transmission, and politics. He is “a little less bullish” on fission because permitting may outlast developing fusion from scratch, yet sees an “easy path” to 5,000 fusion plants by 2050; 50 MW Realta units could occupy substations, while Commonwealth is building a 500 MW reactor.
AI and robots can deliver abundance while widening income disparity enough to make unmodified capitalism unworkable. Khosla forecasts job displacement, productivity, and GDP growth together; rather than prescribing one UBI model, he says countries must decide how to share abundance, pairing techno-optimism with “care and caring.”
The opportunity can be enormous even if most funded companies die. Khosla estimates perhaps 80% of AI investments will lose money, but exponential winners leave aggregate gains positive, with three to five winners per field and perhaps half a dozen robot companies; Adcock calls humanoids “winner take most” and says 95% of such deep-tech efforts go bankrupt.
Figure treats security and civilian positioning as constraints on the product, not public-relations add-ons. Its 2022 mandate rules out military work because the “Terminator vibe” could damage the larger commercial market; Helix kitchen demonstrations ran offline with weights on an onboard GPU, while firmware enforces actions the robot “ultimately can never do.”
Healthcare exposes the gap between model capability and regulated physical autonomy, even as AI could multiply scientific capacity 10–100x within five years. Khosla expects healthcare expertise to become free “well before 2030,” but unsupervised robot surgery to remain more than 10 years away—possibly 15 or 20 depending on the FDA—while AI scientists make research 10x cheaper and 100x more abundant.

Deep dive

1. One chassis is the route to $20,000 humanoids

Adcock says Figure—under three years old, with a few hundred engineers—chose the minimum number of companies to ship robots to. Its “magic number” is 100,000 robots, which it has across its first two groups; the focus reflects limited bandwidth even as Figure works in healthcare, construction, other industries, and homes.
Customization creates “high NRE scenarios,” reliability and maintenance burdens, divergent product paths, and weaker transfer learning when embodiment or observation space changes. Adcock instead wants one platform that increasingly moves, touches, and feels like a person while performing most human tasks end to end.
The economic objective is “twenty thousand dollar levels,” with foreseeable development coming through product revisions rather than adjustments to the main chassis. Khosla invokes Henry Ford—“any color as long as it’s black”—while allowing that specialized robots may emerge for superhuman jobs such as lifting a car without a jack.

2. Capability arrives before society absorbs it

Adcock’s rollout is “not quite sequential”: workforce robots appear in vast numbers first, followed by homes, outdoor work, and space on uneven, overlapping timelines. Figure’s immediate task is shipping 100,000 indoor units, lowering costs, building the data pipeline, and achieving full autonomy with no human intervention; homes and outdoors are part of the next 10 years.
Agriculture illustrates the timing gap. Adcock expects indoor environments first because weather makes outdoor operation harder; Khosla expects farm capability within five years but distinguishes that milestone from adoption: at 1% penetration, he considers the technical job done, while societal diffusion may take much longer.
Adcock’s dream is a humanoid flying an Archer aircraft amid a pilot shortage, and he believes robots will eventually do “everything a human can. Physically.” But demonstrating a robot driving or farming and integrating it into society at scale belong on different timelines.
Space appeals because it is “such a pernicious place for humans”: Adcock imagines mechanical embodied agents and autonomous specialists on Mars and the Moon, though Earth comes first. Khosla argues exploration—not resource extraction—will motivate expansion, and recalls telling “Sam” that a self-replicating AI probe would leave Earth in 15 years, only to hear: “You’re way too conservative.”

3. Fusion can localize generation, but permitting remains the risk

The named fusion efforts include Commonwealth, Helion, and Taifun; Khosla first refers to another effort as “Rialta” and later calls the 50 MW builder Realta Fusion, amid what he calls emerging “sovereign fusion technology.” He stresses that major technological transitions remain exposed to regulatory forces and political pressure, citing the Screen Actors Guild’s AI agreement as an example of rules that could drive customers out of business if followed.
Despite investing in TerraPower with Bill Gates, Khosla has become “a little less bullish” on fission because community objections, environmentalist objections, and NIMBYism may make siting slower than developing fusion from scratch. By contrast, he sees an “easy path” to 5,000 fusion plants by 2050, “maybe more.”
Reactor size changes grid economics: Commonwealth is building at 500 MW, while Realta targets 50 MW—small enough, Khosla argues, to place where a substation sits today and eliminate some transmission needs. More lines will still be needed, supplemented by superhot geothermal and, if cheap energy permits, transmission tunnels underground.

4. Abundance widens inequality and concentrates venture returns

Challenging Kathy Woods’s historical claim that technology creates more jobs than it destroys, Khosla says displacement will happen. His earlier forecast remains abundance, productivity, and GDP growth alongside increasing income disparity—the outcomes economists love coupled with a distributional problem that policy must address.
Khosla does not offer UBI as a universal prescription. He says policy must determine “how do we plan to share the abundance more broadly,” with different countries choosing different paths because “the current capitalist system unmodified” will not work; dystopic elements, he says, will also be societal choices. His formulation is techno-optimism with safety-minded “care” and “caring for those left behind.”
His venture lesson is staged fundraising: pressure teams to test assumptions incrementally, while his own analysis found that companies raising less money were more likely to produce more absolute dollars. Diamandis offered a different timing view—raising a lot early and scaling later once there is a path—while Khosla agreed that capital is needed at the scaling stage. Khosla’s conclusion is that complacency is innovation’s enemy.
On AI commoditization, Khosla expects perhaps 80% of investments to lose money, yet more money to be made than lost because of exponential winners. He foresees three to five winners per area and perhaps half a dozen humanoid companies—probably fewer than today’s car companies in 20 years; Adcock’s harsher version is “winner take most,” with billions required for manufacturing and data and 95% of such deep-tech efforts failing. Khosla therefore “only care[s] about 2030,” not today’s headlines.

5. Regulation slows physical autonomy as AI accelerates science

The robot-care question was framed by critical healthcare-worker shortages and boomers entering the healthcare system; Diamandis recast it as whether robot surgeons would become the baseline. Khosla expects healthcare expertise to become free within a couple of years, “well before 2030,” but interventional medicine such as catheter procedures will move more slowly. Unsupervised robot surgery is more than 10 years away; whether it takes 15 or 20 depends chiefly on the FDA.
AI safety remains “a real problem to worry about,” not something Khosla dismisses despite his optimism and respect for warnings from Yoshua Bengio and Geoffrey Hinton. He expects major explainability advances within three to five years—and within two to three years, claims that models are merely black boxes or parrots may begin to fade.
Figure’s 2022 mandate excludes military applications because humanoids already carry a “Terminator vibe,” and Adcock sees the civilian market as much larger. Cybersecurity is architectural: Helix kitchen robots ran offline with neural-network weights on onboard GPUs, while Figure works to deny kernel root access and encode actions the robot can never perform in nonvolatile firmware.
Khosla’s most aggressive forecast is the number of scientists multiplying 10–100x within five years through AI PhD chemists, biologists, and physicists. Agents would generate and test hypotheses without human intervention, making research 10x cheaper and 100x more abundant. Diamandis imagines AI models designing experiments and robots running them lights-out; Adcock agrees more generally that agents on both the physical and digital sides will increasingly do what humans do.

All-In Live from Austin: Colin and Samir, Chris Williamson, and Bryan Johnson Eiso Kant (CTO poolside) - Superhuman Coding Is Coming!