20VC: Unconventional's Naveen Rao on robotics and $10TRN inference

Summary

• Rao expects spare US power capacity to become scarce within a few years, forcing AI growth toward the slower pace of infrastructure construction. He cites roughly $50 billion per gigawatt: building 10 gigawatts annually would already mean half a trillion dollars, while claims of 40 gigawatts are implausible today. He regards current financing concerns as “a blip,” with his confidence resting on continued demand for AI services.
• The hardware opportunity is to redesign the model and machine together around energy efficiency. Rao argues that recent gains largely came from smaller data types, while physical efficiency improvements stalled. Unconventional’s premise is that “we don’t treat the model as sacrosanct”: changing algorithms, architecture and physical implementation together can unlock gains unavailable through software optimization alone.
• Rao forecasts a $10 trillion inference market within 10 years, while expecting pressure on inference operators’ margins. Falling token prices squeeze the layer mapping models onto existing hardware; improving energy-to-intelligence conversion offers more room for differentiation. In three years, he expects open models to handle 80% of tokens, but spending to split roughly 50/50 between open and frontier models.
• AI’s immediate bottleneck is shifting from producing artifacts to judging whether they work. Rao describes reducing a research question from roughly two weeks to an hour, then warns: “I think we’ve moved the problem from producing to verification.” Specifications and experienced human judgment remain essential, which is why he expects development output to rise faster than spending shifts from people to tools.
• Rao’s attack on SaaS narrows considerably when Stebbings brings up businesses delivering outcomes through software and people. He endorses that model, including AI-assisted legal work, and says he would send a provider 80% of his legal workloads if he could gain confidence in its results. The unresolved question is whether existing firms retain distribution or vendors eventually sell those outcomes directly.
• Rao thinks robotics will be 100 times larger than agentic coding. His timetable is working prototypes for some tasks in two to three years and an emerging market in five to seven; Stebbings remains wary of application timing, particularly humanoids. Rao currently favors the technology stack and expects model developers to become integrated robot businesses: “You’re not gonna have a separation.”
• Rao now prioritizes technical progress over market validation when demand for the solved problem appears assured. For trillion-dollar markets, he calls price “almost irrelevant,” while Stebbings presses on fund concentration, LP expectations and limited diligence time. Rao’s underwriting principle is to back “someone to drive the bus,” because the original technical narrative will be wrong and the team must navigate the changes.

Deep dive

1. Spare electricity is masking a harder construction problem

• Rao cites estimates of roughly $50 billion to build one gigawatt, versus $20–30 billion for a semiconductor fab. Recent data-center expansion has drawn on existing headroom: he puts average US grid output around 900 gigawatts, against potential peaks of 1.2–1.3 terawatts. Accessing that headroom is faster than constructing new supply.
• Even conservative AI growth, he argues, makes excess capacity scarce within the next few years. Expansion then converges toward the construction rate. Claims of building 40 gigawatts annually get a categorical rejection today; 10 gigawatts would be “amazing,” and at his cited cost would require half a trillion dollars.
• Rao thinks China surpassed US capacity in 2023 after building six times faster for 20 years. Central planning helps giant infrastructure rollouts, he says, while the US retains an advantage in opening new technological fields. He also rejects dismissing Chinese AI work as merely distillation: genuine innovation is happening there.
• On Stebbings’s debt-cycle concerns, Rao calls this “a blip,” resting on continued AI demand. Anecdotally, bespoke behind-the-meter electricity already commands premiums. His illustrative 5x compute-revenue-to-electricity-cost ratio allows 30–40% higher power prices. A Bloom Energy project involving a couple of gigawatts in Arizona might help, but its approach and economics remain unproven.

2. Energy-first computing requires changing the model itself

• Rao separates computing into physical devices, computing architecture and models. Physical efficiency gains have largely stalled; recent improvements mostly came from moving from 32-bit to 16-bit, 8-bit and 4-bit data. Moving and computing fewer bits helps, but leaves another opportunity: enforcing locality and reducing data movement at the model level.
• “We don’t treat the model as sacrosanct,” Rao says of Unconventional’s approach. Computing historically optimized speed: ENIAC calculated artillery trajectories faster than humans, and successive machines had to beat predecessors. With transistor improvements slowing, he wants every design decision evaluated for an energy advantage.
• His co-design example is sparsity. Software might deliver an illustrative 5% efficiency gain on a standard GPU, but off-chip random-access memory limits how far sparse connectivity can help. Brains co-evolved their physical and intelligence layers; hardware expressly built to exploit the model’s structure could produce much larger gains.
• Rao says specialization does not mean the market for horizontal providers like Nvidia will be small: “The market is going to be huge.” He recalls graphics developers resenting Jensen Huang’s AI emphasis at GTC in 2014, when Rao started Nervana. Huang recognized the opportunity and executed well. Rao’s own criterion is now singular: run neural networks “very, very efficiently, very, very scalably.”

3. Verification becomes the constraint after production accelerates

• Rao’s concrete productivity example comes from coupled oscillators in Unconventional’s first chip. Understanding noise propagation across four or six hops once meant waiting for researchers and simulations, perhaps two weeks altogether. Asking a model, inspecting its code and iterating now gives him an answer in an hour.
• That speed creates more artifacts requiring inspection. “I think we’ve moved the problem from producing to verification.” Stebbings suggests owning verification means owning automation; Rao is unconvinced. “You have to own the spec as well,” including whether the problem is useful and whether the finished result actually implements the intended specification.
• His caution about agents centers on judgment. Neurosurgery requires seven years of training after medical school because choosing the right action takes experience. Models generate possible solutions, but selecting the good one remains poorly understood across tasks; successive decisions can produce results that look robust while being completely wrong.
• Stebbings predicts people will eventually trust agents more than themselves. Rao agrees that may happen, but says it is not today’s reality. He expects expertise and training provenance to retain value: an intelligent machine’s usefulness will depend on how it was taught, much as a surgeon’s or chef’s does.

4. Data becomes defensible through curricula and specialization

• Rao expects data businesses to specialize around important problems. Bulk observations initially looked commoditized; curated coding examples and reasoning data are different. He attributes much of current model capability to data selection, loss functions and the curriculum connecting pre-training, reinforcement learning and supervised fine-tuning. “That curriculum” is the “secret sauce.”
• He rejects the assumption that reinforcement learning is more efficient. Measured by improvement per unit of compute or dollars, he thinks it is less efficient, although it keeps capabilities advancing and should remain important for several years. Gathering enough pre-training data to solve every task is impractical, and current models do not use that data efficiently enough.
• The tennis-serve analogy explains RL’s appeal: demonstrate a good serve, let the learner build a target, then judge attempts against it. Speed and whether the ball lands in bounds provide a verifiable reward. Connecting actions to measurable outcomes supplies a learning structure for tasks that cannot simply be covered by collecting more examples.

5. The transformer is an interim architecture, while labs race to ship

• “I do think we will be in a post-transformer world,” Rao says, a view he only fully understood a couple of years ago. He sees current reasoning as an inefficient combination of contextual retrieval, sampled traces and selecting among outputs. Brains do not reason by producing a huge trace and feeding it into another model; he expects better algorithmic approaches.
• Frontier labs face a different incentive: being one month late can be disastrous. To Rao, xAI’s rapid arrival at the frontier shows how much progress can come from resources, experienced teams and executing an established paradigm. That differs from taking the risk of inventing a new one.
• He places deeper architectural bets, including work associated with Yann LeCun and Fei-Fei Li, closer to startups. The analogy is drug development: startups take major risks, incumbents improve incrementally and acquire promising results. For founders, this raises the importance of vision; execution alone was more sufficient in the pre-AI software market.

6. Expanding inference demand does not protect every layer’s margins

• Rao’s demand thesis begins with competition: organisms, companies and countries benefit from greater intelligence because their environments are adversarial. Energy powers that intelligence, and hardware converts between them. His definition of hardware includes chips, systems, networks and cloud infrastructure; doing that conversion better is where he expects value to accrue.
• Some intelligence is already commoditizing. Six-month-old models, including Chinese open models, remain useful, and Rao expects the frontier to be available in open models within six months. That can create a huge market with thinner margins. He does not expect model companies to own the stack’s economics; he says Nvidia currently takes the lion’s share.
• Inference providers earn their place by making existing models reliable, efficient and available with uptime guarantees. But falling willingness to pay per token forces them to extract more from hardware, with less remaining margin in that intermediary layer. “I actually think the hardware is the place to be over the next 10 years.”
• In three years, Rao expects open models to carry 80% of token traffic and frontier models 20%, with spending roughly 50/50. His analogy is the mobile-phone market: he describes the iPhone as approximately 20% of the total market but 80% of the margin. A small set of difficult tasks should support premium frontier-model pricing.

7. Outcome businesses survive Rao’s critique of classic SaaS

• “If you’re a pure software investor, you’re kinda fucked,” Rao says, targeting the classic SaaS model. Stebbings challenges the definition: AI vendors still sell enterprise software. Rao concedes the distinction and endorses businesses combining software and services to deliver outcomes. Harvey illustrates using AI plus human verification to perform legal work faster with fewer people at the same or near the same quality.
• Stebbings’s pushback is direct: “Aren’t you just investing in a law firm?” Rao agrees it is effectively a better version of one. Existing firms provide distribution, but that channel may not endure. If he could gain confidence that a direct provider’s solution works, Rao says he would send it 80% of his legal workloads, including NDAs and contract reviews.
• Owning proprietary intelligence requires scale, Rao cautions. A large firm such as Cooley can justify customization; a small Nebraska practice wants an out-of-the-box service. He rejects the idea that every company will improve its own models using historical data. Competitive pressure encourages big firms to invest, while smaller firms may remain service customers.

8. Cheaper development raises expectations as well as output

• Stebbings cites Benioff discussing $300 million of coding-tool spending against a $6 billion developer salary budget: 5%. Does that become 30%, 50% or 1%? Rao thinks the percentages may remain “sort of stable,” because experienced people are still needed to recognize good output and verify it.
• His example holds annual development spending at $10 million while delivery accelerates. Customers once accepted yearly releases; now they want bugs fixed within a week, eventually an hour. AI plus humans enables that pace. The omitted variable in simple replacement calculations, he argues, is demand changing as supply becomes cheaper.
• He extends the argument to construction: robots doing 80% of tasks might shorten a skyscraper project from two years to eight months. The result would not simply be the same construction completed faster; he says twice as many skyscrapers would be built, with humans and robots sharing the work.
• Despite that productivity, Rao would avoid investing in agentic coding outside model providers. He expects value to migrate toward model capabilities, with less accruing to IDEs or harnesses. His evidence is internal switching: employees move fluidly between tools, suggesting limited stickiness.

9. Robotics opens a much larger physical market

• Stebbings avoids robotics applications because he sees five-plus years of timing risk, especially for humanoids; non-humanoid machines have a clearer path. Rao initially disagrees strongly, then finds their timelines fairly close: working prototypes for some tasks in two to three years, with a market emerging in five to seven. Exponential development, he argues, makes distant-looking capabilities arrive unexpectedly quickly.
• Rao forecasts AI models becoming a trillion-dollar market within the next few years and inference reaching $10 trillion within 10 years. Reducing token-delivery costs by one order of magnitude could effectively expand the market 10x “in some ways,” with potential applications increasing more than 10x. He invokes Jevons paradox to explain the expanding usage.
• “I actually think robotics is gonna be 100 times larger than agentic coding.” Existing industrial robots mostly execute rote human-defined tasks; Rao expects dynamic physical problem-solving. Billions of machines could notice something broken, clean it up or replace it. Those services’ costs will fall substantially, he stresses, rather than drop to zero.
• His investments favor the technology stack over applications today. Rao does not think the value will be in the model alone: companies building models may ultimately sell integrated robots performing particular tasks well. “You’re not gonna have a separation.” Asked where he would put all his money outside Unconventional, he chooses Physical Intelligence for tackling the conversion of a specified task into motor output.

10. Technical milestones replace market validation in Rao’s underwriting

• Rao’s changed mind over the last 12 months concerns progress metrics. Market validation fits products without deep technical roots; curing cancer does not require proving buyers want a cure. For robotics, computing and biology, he wants investors to assess progress toward solving the technical problem when demand for the solution is already evident.
• Asked whether he cares about price given the market size, Rao says, “Not really.” He wants a portion of his wealth exposed to robotics and regards entry price as “almost irrelevant” against trillion-dollar-plus markets. Stebbings counters that LPs still care deeply about price and ownership; Rao expects attitudes to shift as established firms show similar trends.
• Stebbings makes concentration concrete: a $20 million space investment from a $400 million fund risks 5% on technology he does not understand. Rao says early technical narratives will be wrong; investors back “someone to drive the bus, not the bus itself.” The team must make the pivots. Both acknowledge that two or three meetings make this difficult.
• Steve Jurvetson stands out among Rao’s roughly 1,000 investor meetings. During Nervana’s early days, they discussed numeric representations, exchanged papers and debated annotations like technical colleagues. Jurvetson had independently thought about the opportunity. Rao’s broader point: large successful bets require non-consensus thinking because consensus is already old.

11. The endgame is a better machine for scientific discovery

• With unlimited resources, Rao would build manufacturing and move closer to raw materials. Supply constraints and general-purpose components limit optimization; he would rather make exactly what his computer requires. That ambition extends down to processing and slicing silicon ingots, because he sees the computer’s value as delivering intelligence.
• His final enthusiasm is scientific reach: AI lets humans approach questions that previously exceeded the available pool of capable researchers. He mentions deeper physics and theoretical constructs such as wormholes as questions to investigate, including whether they are real or exploitable. He rejects human-extinction framing and describes expanding human understanding as the opportunity: “It makes me feel like a kid.”