Naveen Rao's Unconventional AI Strategy Explained

In 1945 the United States government built a machine to solve one problem: calculate artillery trajectories faster than a room full of humans with slide rules. It was called ENIAC, for Electrical Numerical Integrator and Calculator, and it set a pattern computing would follow for the next eighty years (14:03). Every machine since has tried to beat the one before it at speed. CPUs beat human arithmetic. GPUs beat CPUs at rendering graphics. Naveen Rao, who sold his AI startup MosaicML to the data platform company Databricks for $1.3 billion and now runs the AI infrastructure company Unconventional AI, thinks that eighty-year race just changed its rules. The new constraint is not speed. It is energy.

The Grid Is the Ceiling

Rao's argument starts somewhere unglamorous: the power grid. Building one gigawatt of new energy capacity, roughly enough to run a very large data center, costs about $50 billion (05:09). That is more than the cost of building a semiconductor fab, he points out, and fabs are famously expensive. For the last several years, the AI industry has not really built new energy. It has squeezed more hours out of existing plants, pushing the US grid closer to its peak output of roughly 1,300 gigawatts instead of its typical 900 (05:53). That slack is close to gone. Rao expects the country to burn through spare capacity within a few years, at which point AI's growth will be capped by how fast anyone can physically build power plants, not by chip supply or model quality (07:55).

China, he notes, overtook the United States in total energy capacity in 2023, after two decades of building new capacity roughly six times faster (06:15). Some of that was dirty coal power, and Rao does not pretend otherwise. But the lesson he draws is not about fuel source. It is about commitment. The US innovates better at what he calls the tip of the spear, pointing out that AI itself was invented here, not in China. What the US has not done well is build boring infrastructure at scale, and that gap is about to matter more than any algorithm.

Scarcity in energy changes the economics of compute in a way most people miss. Companies running AI workloads already pay well above the standard grid rate, because the margin on a gigawatt of compute is roughly five times the cost of the power that feeds it (10:57). That gives data center operators room to pay 30 to 40 percent more per kilowatt hour and still profit, which is why behind the meter power deals, where a data center buys electricity directly from a generator instead of the public grid, have become common even though nobody quotes public prices for them (10:18).

Why Rao Says Software Investors Are in Trouble

The energy story feeds directly into Rao's view of computing architecture, and why he thinks the investing playbook of the last fifteen years is running out of road. For decades, computers got more efficient mainly because transistors got physically smaller, a trend called Moore's Law. Rao says that trend has largely stalled (11:38). The efficiency gains of the last few chip generations have come instead from a different layer: shrinking the size of the numbers a chip computes with, from 32-bit down to 8-bit and smaller. The next big gain, he argues, will come from designing the model and the physical chip together, rather than treating the model as fixed software running on off-the-shelf hardware (13:24).

"We don't treat the model as sacrosanct, and we're energy first in our, in our mindset." (Naveen Rao, [13:24])

That shift in mindset is also why Rao believes demand for intelligence has no ceiling. His reasoning is almost biological: life has always competed for scarce resources, from two bacteria competing for a patch of sunlight to nations competing for energy and influence today. Intelligence, in his view, is the tool that wins those contests, which means demand for it never saturates.

"Intelligence, to me, it will have infinite demand. There is no limit of demand of intelligence because the world that we live in, the universe is, uh, adversarial." (Naveen Rao, [21:35])

If intelligence is the product and energy is the raw material, hardware, the machinery that converts one into the other, becomes the place where value piles up. That is the logic behind Rao's blunt verdict on software investing.

"I would say if you're a pure software investor, you're kind of fucked, honestly." (Naveen Rao, [00:19])

His point is not that software stops mattering. It is that the shortcut of the last cycle, spin up a product on a cloud platform, test a market cheaply, scale a SaaS (software as a service) business to $100 or $200 million in revenue, depended on a frictionless, commoditized delivery layer. Rao thinks that layer is being replaced by one where the intelligence itself, and the hardware producing it, carries the margin. He points to the legal tech company Harvey as an early example: instead of selling cloud software to law firms, it replaces part of the actual work of reviewing contracts, with humans verifying the output (43:07).

That points to a new bottleneck. As AI produces more code, more legal drafts, more research, the limiting factor stops being production and becomes verification, the work of checking whether an AI-generated answer is actually correct (26:40). Rao says his own company runs into this constantly. He thinks it is also why today's dominant AI architecture, the transformer, which generates answers by retrieving and comparing patterns from training data rather than reasoning from first principles, will eventually be replaced (37:47). "I don't think the transformer is the be-all, end-all," he says (36:50).

All of this adds up to two large bets. The first is on inference, the process of running a trained AI model to generate an answer, which Rao expects to become a $10 trillion market within ten years, up from a category that, a decade ago, nobody would have guessed could be worth a trillion dollars at all (52:23). The second, more surprising bet, is on robotics.

"I actually think robotics is gonna be 100 times larger than agentic coding." (Naveen Rao, [53:11])

Agentic coding tools, AI systems that write and execute software with minimal human input, have been one of the hottest venture categories of the past two years. Rao thinks that market is small next to what happens when intelligence moves into the physical world: robots that diagnose a broken part and fix it, or a construction site where machines cut the time to build a skyscraper from two years to eight months, because far more projects can run at once, not because each one moves faster alone (49:34). He expects working prototypes within two to three years and a real market within five to seven (51:19).

None of this makes Rao a doomer. Asked about fears that AI poses an existential risk to humanity, he was dismissive, calling that worldview the stupidest he has heard and arguing that AI's real role is to let a small number of working scientists tackle problems, in physics, in medicine, that have simply been too hard for the available pool of human minds to approach (65:30). Whether or not the energy and hardware bets land exactly as he describes, that is the scale he is playing for: not a better app, but a bigger set of questions the species can finally afford to ask.