Unconventional AI's Naveen Rao: US Spare Power Capacity Runs Out Within Years, and the Next Efficiency Gains Will Come From Co-Designing Model and Machine, Not From Silicon
Podcast interview with the Unconventional AI founder, October 24, 2026
Naveen Rao, founder of Unconventional AI, argues that the AI buildout is about to run into a hard physical limit. The US has been feeding data centers from slack in the existing grid, and he expects that slack to be gone "in the next few years." Rao sold MosaicML to Databricks for $1.3 billion. Unconventional AI has since raised hundreds of millions of dollars, most recently at a valuation the host put at close to $5 billion. In a wide-ranging podcast conversation, Rao also made a technical case for why his company's energy-first approach to chip design has room to run after the end of physical scaling.
The Easy Power Has Been Taken
Rao's central point on energy is that investors are treating gigawatts as an abstraction. He cited industry figures of roughly $50 billion to build one gigawatt of capacity, against $20 billion to $30 billion for a leading-edge semiconductor fab. "People are throwing around the word gigawatt like it's easy," he said. "Guys, this is not easy."
According to Rao, the industry has so far found power by running existing infrastructure at higher utilization. The US grid averages around 900 gigawatts of output but can peak at 1.2 to 1.3 terawatts, and that gap is what data centers have been absorbing. He expects that cushion to be exhausted soon, at which point supply growth converges to the pace of new construction. He was skeptical of claims that the US could add 40 gigawatts in a year: "We have never done that." Ten gigawatts would be a strong outcome, and he put that at roughly half a trillion dollars of spend.
He noted China passed the US in total capacity in 2023 after building about six times faster for two decades. He acknowledged the advantages of central planning for infrastructure rollouts, but argued the US retains the edge in frontier innovation.
Rao sees early evidence of tightness in pricing. Behind-the-meter deals are bespoke, so he has no hard numbers, but he said buyers are paying well above standard grid rates, creating a two-tiered market in which electricity is worth more to data centers than to general users. He cited his own rough, admittedly invented math that compute can be sold at around 5x the cost of the electricity behind it, which would let an operator pay 30% to 40% more for power and still earn attractive margins. He was cautious on Bloom Energy's multi-gigawatt Arizona project, saying that "if that works, amazing," but that it "hasn't really been proven out yet" and its economics differ sharply from grid power.
On whether this signals an AI bubble, Rao said the pullback in debt-fueled enthusiasm is "a blip. It's not some kind of bubble being burst." He admitted he is "pretty AI-pilled" and that the real test is demand for AI services, which he does not think has been scratched.
Efficiency Gains Have Moved Up the Stack
The most useful technical framing came when Rao described three layers of computing efficiency. The first is physical, covering Moore's Law and Dennard scaling. He said those gains have "largely stopped," and the last few GPU generations show little improvement from the silicon itself. The second layer is computing architecture, where most recent gains came from shrinking data types from 32-bit to 16-bit, 8-bit and 4-bit. The third, which he called largely unexploited, is enforcing locality and reducing data movement at the model level.
Rao said his main regret from two years ago is not appreciating how much "alpha" remains in the top layer. The shift is that a computer can now be treated as a device built solely to deliver intelligence, so the model, architecture and physical layers can be co-designed. "We don't treat the model as sacrosanct, and we're energy first in our mindset, meaning that everything we do has to give us an energy advantage," he said. He noted that this is unlike every prior computing generation, which was designed for speed, a lineage he traced back to ENIAC.
He argued that software-only improvements are capped by the hardware beneath them. Sparsity is his example: a GPU with standard off-chip memory can capture some benefit, but the gains are limited because the hardware was not designed for it. Hardware built to exploit sparsity, as the brain evolved to do, could yield "massive gains." He expects this to drive continued specialization of chips for particular workloads, naming Etched, Fractile, Olix and Cerebras among others pursuing it.
Why Rao Does Not See Chips Commoditizing
Asked whether specialization threatens Nvidia's addressable market, Rao credited the company with executing well and recognizing the importance of AI early. He noted Jensen Huang was talking about artificial intelligence at GTC in 2014, to the annoyance of graphics specialists. He did not see custom silicon from hyperscalers as evidence of commoditization. His reasoning is that intelligence has effectively infinite demand because the world is adversarial, that energy must be converted into intelligence, and that hardware, meaning the chips, systems, cloud and networks, is the conversion mechanism. "The better I do that, the more value I accrue."
He was more candid about where commoditization is happening. Energy already is a commodity. A portion of intelligence is following, because open models, including Chinese ones, that are six months old are "still pretty darn good," and the frontier reaches open models within that window. He drew the analogy to tomatoes: commoditized but still a big market, with margins that fall. Nvidia, he said, takes the lion's share of margin across the stack today, and he does not expect model companies to own it.
For inference providers such as Fireworks and Baseten, Rao expects price per token to fall and margins to compress, since they map open models onto hardware and the model itself is a commodity. His conclusion: "I actually think the hardware is the place to be over the next 10 years."
Market Size, Open Versus Closed, and What Rao Thinks Is Overrated
Rao said the inference market will reach $10 trillion within 10 years, and that AI models themselves will be a trillion-dollar market within a few years. He invoked Jevons paradox: a 10x efficiency gain in token delivery roughly 10xes the market because the base of applications expands. He went further on robotics, saying it will be "100 times larger" than coding because the physical world is untouched by intelligence. Industrial robots today, he said, are rote rather than intelligent. He expects working prototypes in two to three years and a real market in five to seven. He invests in the technology stack rather than applications, expects model builders to end up selling end-to-end devices, and said he is more interested in companies solving "the conversion of a task to motor output," such as Physical Intelligence.
On token mix, he forecast that 80% of token traffic will flow through open models within three years, but that spend will split roughly 50/50. He used the smartphone analogy: Apple has about 20% of units but 80% of margin, and frontier models will command premium prices for a small number of important problems while routine tasks go to open models.
He was negative on two categories. Customer support automation, a "standard thing done by humans" that startups try to automate, is "the worst place to invest right now." He also called agentic coding tools "quite overrated" as a standalone market, arguing that value accrues to the model providers and that harnesses and IDEs are not sticky. "People move between these things pretty fluidly." He acknowledged Cursor's exit but said the model is what matters.
Verification, Not Creation, Is the New Bottleneck
Rao said AI has made it far cheaper to explore ideas. A noise-propagation question that once took his research team two weeks now takes an hour. But he sees an "AI slop problem" in which the volume of output outruns the ability to check it. "We've moved the problem from producing to verification." He cautioned that owning verification does not by itself mean owning production, because the specification and end-to-end process matter too.
He does not think agent decisions deserve the trust given to human experts yet. Models are stochastic, the mechanism for selecting a good output among many is not well understood across tasks, and long chains of decisions can look robust while being "completely wrong." He expects a few more breakthroughs before that changes, and eventually expects people to trust their agents more than themselves.
Labs, Reinforcement Learning and the Post-Transformer View
Rao argued that data and training curriculum, not architecture, have driven most recent capability gains, and that the sequence from pre-training to reinforcement learning and fine-tuning is "the secret sauce." He said data providers are specializing rather than commoditizing as they move toward higher-quality datasets. He was candid that reinforcement learning is less efficient per unit of compute or dollar than pre-training. Its value is that it works with verifiable rewards on tasks where data cannot be gathered at scale.
He suggested the large labs are not taking deep architectural risks because a one-month lag on a model is a "disaster." He cited xAI, which reached the frontier quickly by hiring and executing, as evidence that the work is more integration than invention. He expects deeper innovation to come from startups, such as those led by Yann LeCun and Fei-Fei Li, and large companies to buy them, as in drug development. He also said the industry will move past the transformer, which he called an inefficient way to achieve reasoning because it relies on sampling and comparing multiple traces.
Software, Services and Jobs
Rao delivered the bluntest line on investing: "If you're a pure software investor, you're kind of f***ed." He argued that the SaaS boom since roughly 2010, built on frictionless delivery, is ending as applications become about the intelligence layer, and he does not see software as a huge market in 20 years. What replaces it, in his view, is outcome-based businesses combining AI with humans, with Harvey as the first example in law. He conceded the host's point that this is largely a better version of a services firm. He is skeptical that most companies will build their own models from proprietary data, calling that "a bit techno-optimist" because it takes scale only large firms have.
On whether AI shrinks developer spend, Rao expects budgets to stay roughly stable, with output rising instead. Customers will demand bug fixes in a week and then an hour, and humans are still needed to verify. He compared this to robots cutting skyscraper construction from two years to eight months, in which case the industry builds twice as many skyscrapers rather than the same number faster.
Rethinking Venture Risk
Rao said the biggest change in his thinking in the past year concerns investing metrics. Market validation is a SaaS-era habit, he argued, and for deep-technology problems where demand is obvious, such as computing, robotics and biology, what matters is progress on the technical problem. "The market is very clearly there." He acknowledged that growth investors who call this undisciplined are one camp, with Unconventional AI squarely in the other. He added that when the market is trillion-dollar scale, "the price is almost irrelevant."
Asked what he would do with unlimited resources, he said he would build his own manufacturing, moving closer to raw materials and designing fabs around his needs rather than adapting a commodity process. He compared this to what Elon Musk is doing. He also said Europe's investor base is too conservative to back ambitious ideas, which he sees as producing a "commoditization of ideas" there.