20VC: Why Robotics Will be 100x Bigger than Coding Agents | Inference Will Be a $10TRN Market: What No One Sees | Why Software Investors are F***** Today with Naveen Rao, Unconventional

xAI

Across the shows: 1 speaker in other episodes, 1 bullish and 0 bearish calls, past 30 days

• Naveen Rao · Unconventional AINeutral

  xAI reached the frontier quickly through resources and assembling known pieces, not deep innovation.

  Elon Musk launched xAI with a clear paradigm and hired people, and with ample resources it got to the frontier fast. He reads this as evidence that there was not much deep innovation, but an assembly of existing pieces run by a team that knows how to execute. He contrasts this with startups pursuing deeper paradigm changes.

  “He got to the frontier pretty fast because he had the resources. So what that points out to me was that there wasn't a ton of, like, very deep innovation.”at 34:28

NVIDIA

Across the shows: 35 speakers in other episodes, 30 bullish and 12 bearish calls, past 30 days

• Naveen Rao · Unconventional AIBullish

  Nvidia captures the lion's share of AI stack margin today, and model companies will not own it.

  Rao argues that the margin in the AI stack will not accrue to model companies. He notes that this is already untrue today, because Nvidia owns the margin of the stack and takes the lion's share. He says commoditization of models and inference pushes value toward the hardware layer.

  “I actually don't think the model companies will be the ones who own the margin of the whole stack. It's not even true today. Like, NVIDIA owns the margin of the stack today.”at 39:56

Harvey AI

• Naveen Rao · Unconventional AIBullish

  Harvey was the first to disrupt law-firm work by pairing AI with human verification, delivering faster at similar quality.

  Under the old model, lawyers reviewed and wrote contracts by hand. Harvey showed that AI can produce the work with human checking, so fewer humans deliver it faster at the same or near-same quality. Rao says this model will disrupt that market, and he sees the outcome-delivery pattern spreading to other engineering-style tasks.

  “I would say Harvey was the first example of this for law firms, where the old way was I have ... a set of lawyers who go and look at contracts ... and I will disrupt that market.”at 43:12

Physical Intelligence

• Naveen Rao · Unconventional AIBullish

  Physical Intelligence is solving the core problem of converting task definitions into motor output for robots.

  Asked which company he would back with all his money if it were not his own, Rao picks Physical Intelligence. He says it is solving something fundamental about how a task is defined and then executed. He calls converting a task into motor output the key problem to solve for the physical world right now.

  “Probably like a physical intelligence. They are solving, in my view, something fundamental about how you define a task and then execute it.”at 55:53

Enterprise AI adoption & AI software

• Naveen Rao · Unconventional AI

  AI has shifted the bottleneck from producing outputs to verifying them.

  AI can generate code, models and other artifacts quickly, but each output must be checked to be correct before it can be trusted. At his company, verification is becoming a major bottleneck because so many artifacts are produced. He says the constraint has moved from producing work to verifying it, and he has also argued there is an 'AI slop' problem where much output is hard to verify.

  “doing that actually starts to become a big bottleneck because you're producing so many artifacts. I think we've moved the problem from producing to verification.”at 26:13

• Naveen Rao · Unconventional AI

  AI agents should not yet be trusted over humans for decisions; long decision chains go off the rails.

  Models are stochastic, so they produce many candidate answers and pick one by a heuristic that is not well understood. Each additional decision in a chain adds the risk of error, and outputs can look robust while being wrong. Human judgment is built from years of training, so agents lack the experience base needed for expert decisions. He expects this to change only after more breakthroughs.

  “And so I don't think we should trust these things. I think the longer and longer decision upon decision upon decision chain you build, the more off the rails it can go.”at 29:11

• Naveen Rao · Unconventional AI

  AI training data curation is valuable and will not commoditize as datasets get higher quality.

  Early data gathering was a commodity, but higher-quality datasets for coding and chain-of-thought carry a lot of value. Most model capability gains come from curating data and mixing training methods, so the data layer will matter more over time. Data companies are specializing, and specialization in important problems creates defensible markets.

  “But as you start to construct higher quality datasets... you actually start to have a lot of value in the data... So the data layer, I think, is gonna be pretty important going forward. I don't see that commoditizing.”at 32:14

• Naveen Rao · Unconventional AI

  Most companies will buy out-of-box AI solutions rather than build their own models; only large players will customize.

  Building your own intelligence needs scale. A large firm such as Cooley can invest in it, but a small firm cannot. Smaller firms will use a service and augment their staff with it, which creates a long tail for out-of-box solutions. Large players will customize because they can justify the cost, and smaller firms will follow competitive pressure.

  “So I don't see this as every company will use AI in a way of, leveraging their past data to make their own models better. They're not gonna do that.”at 45:57

• Naveen Rao · Unconventional AI

  Standalone agentic coding is overrated because value moves to model providers, not to tools like IDEs.

  Rao argues that agentic coding tools lack stickiness. People move between tools easily, and value comes from the model's capabilities rather than the harness or IDE. He would invest in agentic coding only as a market that sits on top of model providers. He calls quick automation of standard human tasks the worst place to invest right now.

  “I actually do think that, like, [agentic] coding stuff is quite overrated in terms of a market because I think it moves to the model provider.”at 59:35

AI: overall outlook

• Naveen Rao · Unconventional AI

  We will move into a post-transformer world; transformers will not be the end-all architecture.

  He expects future AI to move beyond transformers. Architectures will continue to be re-optimised around specific workloads and co-designed with hardware, so hardware becomes the place where the next decade's gains come from. He presents this as a long-run view on where AI is headed.

  “I do think we will be in a post transformer world. I don't think the transformers would be all end all.”at 0:13

• Naveen Rao · Unconventional AI

  Reinforcement learning with verifiable rewards is the path forward, because pretraining alone cannot gather enough data.

  Humans learn skills through an RL loop in which they judge themselves against an internal model of good performance. Data cannot cover every task in the world, and models are not efficient enough to learn from it alone. Setting up verifiable rewards for the important tasks lets a model learn any general task by checking whether its actions led to good outcomes.

  “we cannot gather enough data to solve every task in the world. We simply can't. You can't pre train your way out of this by just gathering data.”at 35:32

• Naveen Rao · Unconventional AI

  The transformer is not the end point; reasoning will move to a more efficient post-transformer architecture.

  Transformers do reasoning by sampling multiple traces, comparing them, and feeding long traces back in. Rao says this is extremely inefficient and does not resemble how the brain reasons. He argues there are far more efficient and higher-performing algorithmic routes to reasoning, so the field will move past transformers.

  “I do think we will be in a post transformer world. I don't think the transformer is the be all end all.”at 36:55

AI chips & semiconductors

• Naveen Rao · Unconventional AI

  Physical chip efficiency gains have largely stopped; future gains come from model-hardware co-design and specialized chips.

  Moore's law and Dennard scaling, which were physical shrinks, have largely stopped, and the last few GPU generations show little efficiency gain from the physical layer. Most recent gains came from lower-precision data types (32 to 4 bit). The bigger remaining opportunity is co-designing the model and the architecture to reduce data movement. Specialized hardware for workloads, like the brain's co-evolved sparsity, is the future of compute, and GPUs already move this way.

  “I think one thing I wish I'd known is the fact that we don't really see physical efficiency gains anymore.”at 11:25

• Naveen Rao · Unconventional AI

  Chips won't commoditize because intelligence has infinite demand and sells as energy-to-intelligence conversion.

  Intelligence has no demand ceiling because the environment is adversarial, and every organism and nation competes for better intelligence. Intelligence needs energy, and converting energy to intelligence is hardware across chips, systems, cloud and networks. Providers that convert energy to intelligence more efficiently capture more value, so he does not expect commoditization soon.

  “So I actually don't see chips being commoditized. ... intelligence, to me, it will have infinite demand.”at 21:20

• Naveen Rao · Unconventional AI

  Hardware is the place to be over the next decade, because the hardware layer keeps value that the infrastructure layer loses.

  Rao expects model and inference-service margins to compress. He argues the hardware side will not collapse because there is still a lot of efficiency to gain in converting energy into intelligence. He says there is more room to squeeze performance out of hardware than out of the infrastructure layer, so hardware keeps its advantage.

  “I actually think the hardware is the place to be over the next ten years.”at 41:10

Venture capital & startups

• Naveen Rao · Unconventional AI

  Startups take the big technical risks, while big AI labs do incremental work and buy startups once they show promise.

  Big AI labs such as OpenAI, Anthropic and Google cannot afford big risks yet. Startups pursuing deeper paradigms, such as those led by LeCun and Fei-Fei Li, take those bets. Big companies then build the incremental improvements and buy the startups when they prove promising. Rao compares this to drug development.

  “Like, all the big risks are taken by startups, and all the incremental stuff is done by big companies. The big companies buy the startups when they show promise.”at 35:01

• Naveen Rao · Unconventional AI

  With trillion-dollar TAMs, entry price is almost irrelevant, and traditional VCs wrongly call these deals overpriced.

  Rao says that when the addressable market is trillions of dollars, the price paid for a stake matters little. Value-focused VCs see these companies as overpriced. Stebbings adds that LPs lag in valuing this, and Rao says they will shift once big firms show the same trend.

  “guys, when we're talking about trillion plus dollar TAMs, the price is almost irrelevant.”at 55:12

• Naveen Rao · Unconventional AI

  For deep-tech startups, market validation is the wrong metric; technical progress is what investors should track.

  Market validation suits products serving a clear purpose in SaaS. For deep technical problems, such as curing cancer, the market is obviously there, so investors only need to validate progress on the technical problem. Rao says investors are split between those investing heavily regardless of valuation and those who see that as undisciplined, and he thinks the second camp is wrong.

  “you don't need to validate the market. You just need to validate the progress on the technical problem.”at 57:31

Software & SaaS

• Naveen Rao · Unconventional AI

  Pure software investors are in trouble; hardware is the place to be over the next decade.

  He argues that intelligence is produced by converting energy in hardware, so value accrues to whoever controls the energy-to-intelligence stack. He believes the transformer architecture will not be the endpoint, and that the hardware layer is where the value sits over the next ten years. Pure software plays are therefore exposed.

  “I would say if you're a pure software investor, you're kind of fucked, honestly.”at 0:22

• Naveen Rao · Unconventional AI

  Pure SaaS will stop being where investors earn margin, and software will not be a huge market in twenty years.

  The SaaS boom from about 2010 came from cheap cloud infrastructure, which fragmented markets as small teams could launch and test products quickly. Rao says applications are shifting toward the intelligence layer. Companies that deliver an outcome through software plus services will replace classic SaaS, so pure software investors will struggle.

  “SaaS, I don't think is gonna be where you gain margin anymore. ... I just don't see software being a huge market in twenty years.”at 42:03

AI infrastructure & data centers

• Naveen Rao · Unconventional AI

  The inference market will reach $10 trillion within ten years as intelligence has infinite demand.

  Intelligence has effectively unlimited demand because the adversarial environment rewards better decision-making, from bacteria to nations. Converting energy into intelligence is the hardware stack (chips, systems, cloud, networks), and the better a provider does that, the more value it captures. That makes inference a very large market, which he projects at $10 trillion within ten years.

  “The inference market is gonna be a $10,000,000,000,000 market within ten years.”at 0:22

• Naveen Rao · Unconventional AI

  AI growth will exhaust spare US grid capacity within a few years; building new power is slow and hard.

  Conservative AI growth rates will consume the grid's excess capacity (about 900 GW average output versus roughly 1,300 GW peak) within a few years. Once that slack is used, data center buildout is limited by the speed of new construction. Building one gigawatt costs around $50B, more than a semiconductor fab, and the US has underbuilt energy infrastructure while China built about six times faster over two decades. He says 40 GW in a year is impossible and that 10 GW would be amazing.

  “If we look at growth rates, even the much more conservative growth rates of AI, we're gonna be out of energy pretty soon in the next few years.”at 7:58

AI & jobs

• Naveen Rao · Unconventional AI

  AI turned a two-week modelling task into an hour, letting researchers explore far more ideas.

  A question he would have sent to his research team, which took about two weeks including the code and checks, now takes an hour with a model. He can iterate on the generated code and cover much more ground per unit time. He sees this as a real productivity gain in R&D, although verification remains a challenge.

  “I get an answer in an hour. So something that would have taken me two weeks to think through has now taken me an hour.”at 24:14

• Naveen Rao · Unconventional AI

  Share of spend on developers stays roughly stable; AI raises output and demand rather than shrinking budgets.

  The hard part is verifying what counts as good output, and that requires top developers. Those people will use AI to produce more, so total spend on humans plus AI stays roughly the same while output grows. Cheaper supply also raises demand: customers want bugs fixed in a week, not a year. Rao uses the skyscraper example, where robots cut build time from two years to eight months, so more buildings get built.

  “I actually think the percentages will remain maybe sort of stable because we end up with this verification problem.”at 47:23

Frontier AI labs & model race

• Naveen Rao · Unconventional AI

  Reinforcement learning is more compute-inefficient than pretraining, but will stay central for several years.

  Measured by how much the model improves per unit of compute or per dollar, RL is less efficient than pretraining. It still keeps moving capability forward, and he expects it to stay very important for the next several years, with specialization building in the middle layers and perhaps pretraining.

  “Well, I don't know that it's more efficient. Actually, I think RL is more inefficient. If you look at it in terms of how much better the model gets per unit compute or per unit dollar, it's actually more inefficient.”at 33:39

• Naveen Rao · Unconventional AI

  Big labs cannot afford long-term paradigm bets; startups take the big risks and big labs buy them.

  Labs such as OpenAI, Anthropic, xAI and Google are racing, and being a month late on a model is a disaster, so they cannot take big bets on new paradigms. Deep innovation is being pursued by startups, such as the efforts of Yann LeCun and Fei-Fei Li. Big companies then take the incremental work and buy startups when they show promise, a pattern similar to drug development.

  “Like, all the big risks are taken by startups, and all the incremental stuff is done by big companies. The big companies buy the startups when they show promise.”at 35:01

AI pricing & unit economics

• Naveen Rao · Unconventional AI

  Inference providers' margins will shrink as the price per token falls, since they mainly map models onto hardware.

  Inference providers such as Fireworks take an open model and make it reliable and cost-efficient on infrastructure. Their value is mapping the model onto physical hardware. Rao expects the price customers pay per token to fall. Providers will then have to squeeze more margin out of the hardware, and their margins will fall. He sees value in that layer in the future but a much smaller share of the profit.

  “Now I think there will be value there in the future, but the price of inference, like what people are willing to pay per token, will drop.”at 40:39

• Naveen Rao · Unconventional AI

  The inference market becomes a $10 trillion market within ten years as token costs fall.

  Rao says AI models alone will be a trillion-dollar market within a few years. Labor and value generation are shifting into technology, which creates very large addressable markets. He argues that cutting the cost of delivering a token by one order of magnitude effectively multiplies the market more than tenfold, which he calls Jevons paradox. Cheaper intelligence unlocks many new uses.

  “And the inference market is going to be a $10,000,000,000,000 market within ten years.”at 52:14

Robotics & autonomous vehicles

• Harry Stebbings · 20VC

  Robotics application layer is too early; Stebbings avoids the timing risk and prefers robotics datasets and infrastructure.

  Stebbings says he does not like robotics investments because the field is far from where it needs to be and timing matters. He sees the infrastructure and datasets behind robotics as interesting. He sees application-layer companies such as robot cleaners and robot builders as too early. He calls humanoid robots obviously far off.

  “I think we are so far from where we need to be, and I do think that timing is important.”at 50:27

• Naveen Rao · Unconventional AI

  Robotics will be about 100x larger than agentic coding because physical-world problem solving is barely tapped.

  Industrial robots today are rote; they execute trained tasks without solving problems dynamically. Future robots will diagnose and fix problems in the physical world, such as replacing a broken part, and the cost of that work will fall greatly though not to zero. Rao expects working prototypes for some tasks in 2-3 years and a market emerging in 5-7 years. He argues the physical-world market will therefore dwarf the virtual software market. He says he invests in companies building robot models and expects them to sell end-to-end devices.

  “So I actually think robotics is gonna be a 100 times larger than agentic [coding].”at 53:01

US–China AI & chip race

• Naveen Rao · Unconventional AI

  The US is far ahead on tip-of-the-spear AI innovation; AI was born there, while China does good work but not frontier.

  China has centralised planning and built infrastructure much faster, and it is doing real innovation, not just distillation. However, the most advanced innovations that open new areas have come from the US, and the US should leverage that advantage. He contrasts infrastructure scale with frontier innovation, where the US leads.

  “when it comes to innovation, tip of the sphere sort of things, like, The US is far and away better. AI was born here, not there.”at 7:17

AI bubble & tech valuations

• Naveen Rao · Unconventional AI

  The AI compute slowdown is a blip, not a bubble bursting, because demand has barely started.

  He says there is plenty of demand for AI and that unresolved AI problems, once solved over the next couple of years, will open up new demand types nobody foresees. He compares it to demand for curing cancer, which has zero adoption risk. He sees the current concern as a short pause, not a collapse of the buildout, and frames the energy shortfall as a pricing signal, not a bubble.

  “So my sense is that this is a blip. It's not some kind of bubble being burst. It's sort of like if you said, hey, is there demand for people curing cancer? Of course there is.”at 9:19

Power demand & the grid

• Naveen Rao · Unconventional AI

  Data centers are paying well above grid rates for power via bespoke behind-the-meter deals, creating two-tier pricing.

  He sees anecdotal evidence, without hard numbers, that data centers pay far more per kilowatt-hour through bespoke behind-the-meter deals than grid users pay. Data center compute margins are high enough to absorb 30–40% higher electricity costs, so the value of power to data centers exceeds its value for general use. He cites Bloom Energy's proposed multi-gigawatt Arizona deal as a possible technology path but says it is unproven and its economics differ from the grid.

  “I do see evidence through anecdotes at least. I don't have specific numbers to quote around this because every one of these behind the meter deals is bespoke, and people are paying way more per kilowatt hour than they would off the standard grid.”at 9:53

Open-source AI models

• Naveen Rao · Unconventional AI

  Open models will carry about 80% of token traffic in three years, though frontier models keep roughly half of spend.

  Intelligence is partly commoditizing. Six-month-old Chinese open models are already useful, and the frontier will reach open models within about six months. Rao expects spend to follow the mobile phone pattern: a small number of high-value problems pay premium prices for frontier models, while routine tasks run on open models. He expects roughly a 50/50 split on spend but only about 20% of tokens going to frontier models.

  “I think it'll be 80%. ... But in terms of tokens, it'll be only 20% to the frontier models.”at 1:00:24