Yesterday, I spent the afternoon at NVIDIA’s headquarters in Santa Clara. It’s hard not to walk through those glass corridors without feeling that you’ve stepped inside the engine room of a new industrial age. Banks of machines hum like power plants. Engineers move with quiet precision, orchestrating what feels less like software development and more like manufacturing reality itself.
That impression captures Jensen Huang’s vision perfectly. The founder and CEO of NVIDIA has become the defining industrialist of our time—not because he makes chips, but because he builds the infrastructure of intelligence. In a recent conversation with Sequoia Capital’s Konstantine Buhler at Citadel Securities’ Future of Global Markets 2025 conference, Huang dismantled one of the laziest tropes of our age: that we are living through an AI bubble.
We aren’t, he argued, because AI is already producing hundreds of billions of dollars in tangible returns. It is the invisible operating system behind modern capitalism—the intelligence layer driving search, recommendation engines, advertising, and logistics. If the dot-com boom was speculation on digital potential, the AI boom is monetized cognition in action. The returns are real, measurable, and accelerating.
Huang’s framing of this transformation is simple but profound. The world’s data centers are no longer just server farms; they are AI factories—industrial complexes that transform electricity into intelligence. What matters now is not how many chips you own, but how much thinking your infrastructure can perform per watt. “Throughput per unit of energy governs your revenue,” he likes to say. In his view, the defining metric of the coming decade won’t be teraflops—it will be tokens per megawatt.
That insight reverses the logic of the last fifty years of computing. Power, not compute, becomes the ultimate bottleneck. Each incremental gain in energy efficiency translates directly into business growth. The company that can produce three times the intelligence per joule of energy doesn’t just save power—it generates three times the economic output. Cheap cognition demands new strategy. NVIDIA’s systems now scale from desk-size accelerators to gigawatt-scale installations: full-stack, rack-scale intelligence factories capable of producing cognition at industrial intensity.
This is why talk of a bubble misses the point. When Apple’s privacy changes in 2022 disrupted Meta’s advertising models, the company lost hundreds of billions in market value almost overnight. It recovered that loss not with marketing spin or layoffs, but by rebuilding its recommendation systems using AI trained on NVIDIA GPUs. That pivot restored more than a trillion dollars in market cap.
I started my career in the late nineties during the first dotcom boom. By 2000, the entire internet industry was worth $30 billion and mostly unprofitable. Today, AI already powers hundreds of billions in recurring revenue for hyperscalers like Google, Amazon, and Meta. This isn’t Pets.com—it’s Ford Motor Company, circa 1913, wiring up the first assembly lines at its Highland Park, Michigan plant.
And if Huang is right, we are only at the beginning. He predicts two new trillion-dollar industries emerging from the AI revolution: digital labor and physical AI. Inside NVIDIA, every software engineer already works alongside an AI collaborator. Productivity has surged. Digital workers that don’t just assist humans but act alongside them as peers. In the near future, enterprises will employ both biological and digital staff: some licensed from large model providers like OpenAI or Harvey, others fine-tuned internally using proprietary data and organizational culture. The IT department, he suggests, will evolve into the HR function for digital employees—onboarding, training, and evaluating agentic systems just as it does people.
The same logic extends beyond the screen into the physical world. Autonomous vehicles, warehouse cobots, delivery drones, surgical systems—these are all forms of embodied intelligence. If an AI can generate a realistic video of a person picking up a bottle, Huang asks, why can’t it control a robot to do the same? NVIDIA’s Omniverse platform, a photorealistic simulation environment, allows robots to train safely in virtual space before entering the real one. Every physical AI, he explains, requires three computers: one for training, one for simulation, and one for operation. In other words, AI doesn’t just learn—it practices, rehearses, and then performs. Together, digital labor and physical AI touch nearly $100 trillion of global economic activity.
Behind this transformation lies a deeper philosophical shift—from retrieval to generation. The computers of the past fetched stored information; the computers of the future will generate it on demand. When you ask an AI a question today, it doesn’t look up an answer—it creates one, conditioned on context and intention. “Everything we just did in this conversation was generated,” Huang said. That’s not marketing hyperbole; it’s a description of a new computational reality. The world is moving from stored intelligence to living intelligence—systems that think, reason, and respond in real time.
But even revolutions face limits. For AI, that limit is power. Every industrial era eventually confronts its energy constraint: coal for steam, copper for electricity, now electricity itself for computation. No matter how advanced your architecture, you can’t exceed the capacity of the grid. That is why energy efficiency has become the new arms race. NVIDIA’s vertically integrated design—co-developing chips, networks, and software—breaks through the slowing of Moore’s Law by delivering tenfold performance improvements at constant power. Sustainability, in this sense, is no longer an ethical imperative; it is a competitive one. The winners of the AI era will be those who can convert electrons into intelligence with the least waste.
The logic of the AI factory also extends to geopolitics. “No country can afford to outsource its data and import its intelligence back,” Huang warned. Every nation will soon need its own sovereign AI—trained on domestic data, aligned with local values, and secured by national infrastructure. As I have written about previously, across Europe, Asia, and the Middle East, governments are already building national AI factories. What oil refineries were to the twentieth century, AI factories will be to the twenty-first—the infrastructure of cognitive sovereignty.
And underpinning all of it is a new economic truth: as the marginal cost of intelligence approaches zero, the constraint on productivity shifts from labor to energy. AI doesn’t replace human work; it multiplies it. We are entering an age of infinite labor, where digital and physical agents expand the productive frontier far beyond biological limits. In such an era, the scarce commodity will not be intelligence but judgment. When machines can generate infinite content and action, the differentiator becomes what we choose to create, constrain, or believe.
That was my lingering thought as I left NVIDIA’s campus and watched the setting sun glint off the mirrored façade of the building—a modern cathedral to computation. The future no longer belongs to those who simply own machines, but to those who can orchestrate cognition at scale.
If you want to understand the real metric of power in this new world, forget headcount, profit, or even compute. Ask instead: how much intelligence can you generate per watt?
