Naveen Rao | All In Summary

The Energy Wall Is Coming in ~3 Years

Naveen lays out a stark quantitative case that AI's growth trajectory will collide with global energy capacity far sooner than most assume. Using Google's own disclosed usage as a data point, he shows the math doesn't work: "Per month, they cross 3.2 quadrillion tokens... if I just take 10 joules per token of energy... this is 12 gigawatts. The US puts about 40 gigawatts of energy into data centers today... we're under 100 gigawatts of data center energy in the world today. 12 gigawatts is going into one company just for AI services" 00:04:54. His conclusion: "we're going to run out of energy pretty fast in like three years or so is my estimate" 00:05:23.

Data Center Economics Have Shifted From Space to Power

The bottleneck driving infrastructure decisions has fundamentally moved. "Over the last several years, it used to be about floor space... Then it was about networking equipment. Then it became about GPUs. Today it's about energy. First you think about energy. I get the energy contract and then I have to figure out how to fill it" 00:05:52. He quantifies this further: "About 50% of the cost of serving a token... every time you try something on ChatGPT, 50% of that cost is energy" 00:06:21.

Biology Proves Radical Efficiency Is Physically Possible

Naveen uses animal brains as an existence proof that today's compute paradigm is needlessly wasteful. "The human brain... runs on about 20 watts of energy... a monkey's brain... runs on one watt... a squirrel... their brain runs on eight milliwatts of energy. You could run over a hundred squirrel brains on your phone" 00:06:50. The core inefficiency is data movement, not computation itself: "the human cortex... only moves about 16 billion bits per second... A GPU or a high-end computing system moves nearly 30 trillion bits in and out of memory per second" 00:08:18.

Von Neumann Architecture Is an Unquestioned 80-Year-Old Assumption

The talk's central technical thesis is that modern computing inherited a structural flaw from its earliest design that nobody has seriously challenged. "The operation of that computer in 1940, 1945, is actually very similar to how they operate today. There's not a huge paradigm shift... That operation creates a machine that just requires a lot of movement" 00:09:13. He notes this original design goal — speed, not efficiency — still governs chip design today: "That computer in 1945... was built to do it faster than the alternative... This computer is twice the speed of that computer. That's how we sell computers. But it doesn't contemplate energy efficiency" 00:09:13.

Moore's Law Efficiency Gains Have Ended, Forcing a Rethink

"Moore's Law... making transistors smaller has largely ended. So we're not seeing efficiency gains just from making transistors smaller. So we need to rethink the problem a bit" 00:10:12. This is presented as the structural reason a new computing paradigm (rather than incremental fab improvements) is now necessary.

Dynamical/4D Computing as a New Paradigm

Rather than iterating on GPUs, Unconventional AI built a physically different machine using oscillator-based dynamical systems (inspired by synchronizing metronomes, bird flocking, ant colonies) that fuses memory and compute. "What we built is what's called a dynamical computer, which actually has compute and memory in one thing. We don't have a memory interface. Each individual computing element is a memory" 00:16:53. He frames the physical dimensionality: "we use the time dimension and the dynamics, and we actually use the physical three dimensions of die stacking... So we have three dimensions from the physical, and we have one dimension in time... we call this 4D computing" 00:17:19.

Sparsity as a "Holy Grail" Free Lunch

Naveen describes discovering that removing connections between compute elements (sparsity) doesn't just save cost but actually improves performance and trainability — a rare combination. "It turns out you can actually not only throw away some of the connections, but you can actually get better behavior out of the whole system. It actually becomes more trainable... So this is kind of a holy grail" 00:14:33.

Jevons Paradox Will Make This the Largest Market in History

Rather than a 1000x efficiency gain shrinking the AI market, Naveen argues it will explode it. "There's a concept called Jevons paradox where when you drop the underlying cost of an asset, you actually consume more than the drop of that asset. So if you make something half the price, you'll consume more than 2X... I think this will create the largest market that humanity has ever seen" 00:19:06.

Contrarian Perspectives

Selling Nirvana Systems to Intel Was a Mistake

Naveen openly states he moved too early on his first AI chip company, a rare admission of a suboptimal exit from someone now running a similarly ambitious venture. "I actually founded the first AI chip company called Nirvana Systems... I think I sold the company way too early to Intel" 00:02:05.

AI Doomerism Is Wrong — This Is Human Evolution, Not a Threat

Against the grain of much AI-safety discourse, Naveen positions himself explicitly as an "anti-doomer" who sees AI as an evolutionary leap rather than existential risk. "I'm the opposite of a doomer. I think AI is one of those transformational technologies that humanity has ever created and will enable us to get to that next level of evolution, which I'm here for. And this is sort of the anti-doomer conference" 00:00:43.

You Don't Understand Something Until You Can Build It

A philosophical stance with practical implications: theoretical modeling of intelligence (neuroscience, ML theory) is insufficient — real understanding requires physical instantiation. "I don't feel like we truly understand something until we can create it. We've gotten a lot better at creating intelligence systems. However, they do it in a kind of inefficient way" 00:07:49. This justifies building actual dynamical silicon rather than just simulating it.

Big Data Centers Are a Transitional Phase, Not the End State

Contrary to the industry consensus of ever-larger gigawatt-scale data center buildouts (hyperscaler capex race), Naveen predicts fragmentation and decentralization once efficiency improves by orders of magnitude. "I think what will be interesting is that we'll see the shift going from big, big data centers with gigawatts to many small data centers all over the place. I think this is a good thing. It actually makes things that are more environmentally friendly, more local, more adaptive" 00:18:12.

Abstractions in Computing Are Lossy and Should Be Collapsed, Not Stacked

The industry has built decades of software/hardware abstraction layers (digital logic, floating point, linear algebra) as unquestioned progress. Naveen argues each layer is actually a source of inefficiency that should be stripped away rather than built upon. "Each one of these abstractions actually is lossy. It means it's inefficient. It doesn't contemplate all the complexity underneath it... We find an abstraction of the physics of the semiconductor and connect that to the neural network" 00:10:42.

Companies Identified

Unconventional AI — Naveen Rao's current company building a fundamentally new computing architecture ("dynamical computer" / 4D computing) aimed at 1,000x power efficiency gains over GPUs. Mentioned as the core subject of the talk, with a working physical chip prototype built in five months. "This is actually the first physical dynamical computer ever built. We did this in five months... we taped out the design, meaning we sent it to the fab on June 1st. The chip is back in our lab, and we actually have results from it" 00:15:30.

Nirvana Systems — The first AI chip company, founded by Naveen Rao in 2014, before AI was a mainstream category; later acquired by Intel. Mentioned as his origin story in AI hardware. "I actually founded the first AI chip company called Nirvana Systems. So, this was in 2014... there was no AI, or at least not in the common vernacular" 00:02:05.

Intel — Acquired Nirvana Systems; Naveen ran the AI group there through 2020. Mentioned as context for his career trajectory.

Databricks — Acquired Naveen's GPU-platformizing infrastructure company (MosaicML, implied) in 2023; that business unit reportedly now represents a significant share of Databricks revenue. "We decided to actually join forces with Databricks. That was in 2023. And actually, that's a quarter of the total revenue of Databricks today" 00:02:59.

Google — Cited as the data point proving AI's energy problem is real and urgent, given its disclosed token throughput. "Per month, they cross 3.2 quadrillion tokens... I never even think in quadrillions, but that's the world we're in today" 00:04:54.

NVIDIA (referenced via Jensen) — Implicitly discussed as the incumbent hardware giant whose GPU-based, von Neumann architecture approach Unconventional AI is positioning itself against. "You heard from Jensen up here, like the largest company in the world, a hardware company because of AI" 00:02:05.

People Identified

Naveen Rao — Founder/CEO of Unconventional AI; previously founded Nirvana Systems (first AI chip company, sold to Intel), ran Intel's AI group, then built GPU infrastructure platform later acquired by Databricks (now ~25% of its revenue). Identified as a repeat deep-tech founder with a track record of building category-defining infrastructure companies at each phase of the AI hardware/software stack. "Naveen is kind of definitionally outlier founder. When I came there, we had about a $20 million business and it was, you know, $700 or $800 million when I left" 00:00:09.

Jensen Huang (referenced, not present) — Held up as the benchmark example of hardware value creation via AI, framing NVIDIA as proof that hardware bets on AI can create the most valuable company in the world. "You heard from Jensen up here, like the largest company in the world, a hardware company because of AI" 00:02:05.

Operating Insights

Build Cross-Disciplinary Teams That Don't Naturally Speak the Same Language, Then Force Translation Layers

Naveen describes the central organizational challenge of Unconventional AI: combining theoretical scientists with chip engineers who have no shared vocabulary, and deliberately building a translation layer (a Python-based library) to bridge them. "We got people from that world, and then we got people who actually build chips, and they don't talk to each other... That's actually one of the most challenging things about this company is the span of talents that we have is so big that getting them to kind of all coordinate and build one thing is actually pretty hard" 00:21:50. His solution was explicit infrastructure: "we actually have built a set of libraries in Python... it's a language of sorts that allows you to kind of express time varying elements that have stochastic behavior" 00:22:12.

Structure R&D as a Literal Top-to-Bottom Stack With Handoffs

Rather than a flat research org, Unconventional AI is organized in an explicit pipeline from abstract theory to physical product, forcing each stage to validate the one before it. "We start with theorists... They come up with... concepts that we think would effectively give us more power efficiency... We then translate that into models that do real things, trained on real data and evaluated against real criteria... Then eventually, we have to actually build something physical" 00:03:57.

Compress Timelines by Revising Public Goals Downward as Evidence Comes In

Naveen publicly moved up his own five-year milestone based on faster-than-expected progress, using AI itself as an accelerant for solving deep scientific/hardware problems — a signal for how founders should recalibrate roadmaps aggressively rather than anchoring to original plans. "The goal has been within, it was initially within five years to get to 1,000x power efficiency. I've actually revised this to three and a half years because things have gone faster than we anticipated. We've actually solved very deep scientific problems quicker because of AI" 00:03:29.

Minimize Migration Pain to Drive Adoption of a Fundamentally New Substrate

Rather than requiring the market to rewrite everything, Naveen designed the product so existing model architectures port over at the model layer, not the low-level operations layer, sharply lowering the switching cost despite the underlying hardware being entirely different. "We actually don't port at the operations layer, you port the model layer. So yes, the existing models will work, but there's a fair bit of compute required to make that transition happen" 00:20:55.

Overlooked Insights

The "Intelligence per Watt" Framing Implies Today's Frontier Models Are ~10 Billion Times Off the Efficiency Frontier

This number is stated almost in passing but is staggering and reframes the entire AI infrastructure conversation: current compute is not merely inefficient, it is ten orders of magnitude away from the thermodynamic limit that biology already approaches within 1-2 orders of magnitude. "Today we're on the far left of this graph, and we're about 10 billion times away. That's one with 10 zeros after it from that thermodynamic limit" 00:17:47. This single data point suggests the AI hardware investment landscape (GPUs, ASICs, even current AI chip startups) is optimizing within a paradigm that is fundamentally, not marginally, wasteful — which is a much stronger claim than "GPUs could be more efficient," and implies enormous asymmetric upside for anyone who actually solves the substrate problem rather than incrementally improving silicon.

Sparsity Breaking the N-Squared Scaling Wall Is Understated as a Cost Story But Is Actually an Architecture Story

Naveen mentions sparsity almost as an efficiency footnote, but the deeper implication — barely dwelled on — is that it solves the fundamental scaling bottleneck (n² connection growth) that constrains how large any densely connected computing or neural system can practically get, while simultaneously improving trainability. "This doesn't scale very well. We call this n-squared scaling... it turns out you can actually not only throw away some of the connections, but you can actually get better behavior out of the whole system. It actually becomes more trainable" 00:14:33. This is a rare case where a cost-reduction technique (removing connections) also removes a structural ceiling on scale — a combination that, if it generalizes beyond this chip, has implications well beyond Unconventional AI's own hardware, potentially affecting how any large-scale neural or physical system should be designed.