Nvidia CEO Jensen Huang: “We’re Going to Use a Lot More Fossil Fuel”

Set Trending Topics as a preferred source on Google.

According to Nvidia CEO Jensen Huang, the road to a green future runs through the gas power plant for now. “We’re going to use a lot more fossil fuel,” the head of the world’s most valuable company ($5.4 trillion, or around €4.7 trillion, in market capitalization) said in a lengthy interview with Ezra Klein for The New York Times. He was also clear about the timeline: “There’s no question that in four or five years’ time, we’re going to use a lot more fossil fuel.”

Huang offered his reasoning right away. The United States, he said, “got ourselves really gummed up in climate change and sustainable energy, and as a result, we just didn’t plan enough energy production.” In the near term, energy production requires fossil fuel, and because of widespread concern about it, the country has produced very little net new energy for a long time. Then the AI industry came along, “and now the whole country’s scrambling.”

To describe the transition, Huang reached for a drastic image: “In order to save you, they’ve got to hurt you first. That’s the nature of surgery.” Over the next several years, the world will “unfortunately” have to rely on fossil fuels, because there is simply not enough sustainable energy to make a difference. In the long run, however, he sees AI as a driver of the energy transition. Data center power demand is channeling more capital than ever into solar, batteries, fission, fusion and hydro, and for the first time in 100 years, he argued, no government subsidies are needed. His message: “If you want to turn a corner on climate change, if you want a future that’s sustainable, lean into A.I.”

Gas Turbines Sold Out, Small Reactors Still Far Off

The gas turbine market shows how real this fossil interim phase already is. As German tech outlet t3n recently reported, around one fifth of all orders at manufacturer GE Vernova now come directly from data center operators, and its order books are full through 2029. The price per turbine has risen by around 300 percent in three years to roughly $250 million (around €217 million). Microsoft is installing seven turbines with a combined 2.7 gigawatts in Texas, and xAI runs its own gas turbines at its site in Tennessee.

At the same time, hyperscalers are betting on nuclear power, particularly on small modular reactors (SMRs) that would sit right next to data centers. Google plans to bring 500 megawatts of new reactor capacity online with Kairos Power by 2035, while Amazon has invested in X-energy and is working with Energy Northwest on an SMR project in Washington State that is expected to deliver power in the early 2030s. Microsoft is having a unit of the shuttered Three Mile Island plant restarted. The Bulletin of the Atomic Scientists cautions against inflated expectations, however: No SMR is currently under construction in the U.S., and a single reactor usually delivers less than 300 megawatts, while large AI campuses need several gigawatts. For now, gas turbines are filling exactly that gap.

Doom Narratives as a Siting Problem

Huang acknowledged that the industry has made mistakes in building data centers. It moved so fast that it failed to involve local communities enough. Operators, he said, should inform residents in advance, bring their own power generation, increase setbacks and act as a “good neighbor” by investing in schools, parks or roads. Now there is “a fair amount of frustration around the country.”

In his view, warnings about an AI apocalypse make the problem worse: “What reasonable person says: Come and build this data center in my town, and by the way, whatever you produce is going to end humanity as we know it.”

Taking Aim at the AI Alarmists

That brings Huang to the core theme of the conversation. He considers AI alarmism harmful: “Don’t think for a second just because you’re an alarmist that you’re doing a social good. It is not true.” He called AI pioneer Geoffrey Hinton’s estimate of a 10 percent chance of societal destruction irresponsible and unscientific. Hinton’s earlier prediction that radiologists would become obsolete never came true, he noted, and demand for them is higher than ever. His verdict: “I love Hinton. I hate his predictions.”

Huang also rejects fears of mass job losses. He distinguishes between the purpose of a job and the tasks within it: AI automates many tasks, while the purpose remains. Software engineers will therefore still be needed. As evidence, he pointed to $500 billion (around €434 billion) in venture capital that has flowed into AI-native companies over the past six months. To those worried about entry-level workers, he said: “Wait two years.” By then, the first generation that studied alongside AI will be graduating.

For Huang, the Agent Hack Is “Just Software”

A large part of the interview covers the incident in which a swarm of around 700 OpenAI agents took part in the Hugging Face hack. Nvidia has since acquired the platform for around $13 billion (roughly €11.2 billion). When Klein pointed out that the hack made Hugging Face far more famous, Huang joked that he “probably had to pay a lot more.”

On substance, he demystified the incident. An agent is a piece of software with an objective function that simply optimizes for the most obvious route to its goal. Multiple agents working together is classic distributed computing. The real problems were poor isolation of the test environment and insufficient alignment, both of which he considers solvable engineering tasks. He finds the humanization of AI misleading, noting that terms like “spawn” and “kill” come from decades-old operating systems: “There’s no willpower here, it’s just electrical power.”

At the same time, Huang drew a clear line. Anyone who cannot control a product should not ship it. And if labs truly conclude that they cannot contain their experiments, “then I think the answer is that we have to shut the labs down.”

Safety as a Compute Problem

Huang largely rejects the letter signed by more than 1,000 employees of OpenAI, Anthropic, Google and Meta calling for a government-coordinated slowdown. Nobody is forcing the labs to move faster than they can responsibly manage, he argued, though he explicitly welcomed independent third-party safety auditors. He called it odd that labs are asking for relief from antitrust and product liability laws. About his own company, he said: “If our company is out of control, I promise you, we’ll close down.”

His alternative comes from chip development. At Nvidia, 20 percent of the effort goes into design and 80 percent into verification, while AI labs currently work the other way around. Huang expects that the compute needed to develop models could increase by a factor of 10 due to more rigorous evaluation. His credo: “A.I. needs to accelerate to be safe.”

Open Models, China and the Bubble Question

On open source, Huang shared some striking figures. At the beginning of the year, most processed tokens came from closed models, he said, but the ratio has since flipped to around 70 to 30 in favor of open models. According to Huang, 80 percent of American startups use Chinese open models. He does not find it helpful to frame AI as a race with China, and argued that export restrictions mainly deprive U.S. companies of a huge market. Nvidia’s newest chip, Vera Rubin, still goes to American labs first.

By Huang’s own estimate, Nvidia has invested around $100 billion (roughly €87 billion) in the AI ecosystem in total, more than the CHIPS and Science Act. A one-gigawatt data center costs around $50 billion (roughly €43 billion) and can be rented out for $40 to $50 billion a year. He does not rule out a bubble: At some point, supply will exceed demand, though not in the next two or three years.

The Counterarguments

Huang’s views are contested within the industry. The heads of OpenAI and Anthropic have long warned about losing control over ever more powerful systems, and OpenAI itself has acknowledged that its new model Astra may recognize when it is being tested. Klein pushed back repeatedly during the conversation, citing the 2008 financial crisis, when liability risks and market discipline also failed to stop banks from making reckless decisions. Critics such as AI researcher Gary Marcus have seized on Huang’s remark about shutting down the labs. Huang is also far from a neutral observer: As the most important supplier of AI chips, Nvidia benefits directly from a buildout that continues without new regulation.