Stephen Witt doesn't trust a robot with a knife. But he told restaurant technology leaders Friday at the FSTEC conference in Dallas that nearly everything that moves will soon carry autonomous intelligence — and that operators should start building their own AI systems now rather than waiting on vendors.

Witt, an investigative journalist, television producer and author of "The Thinking Machine," a book on Nvidia and its CEO Jensen Huang, laid out the economics behind the AI buildout. Nvidia, which makes the microchips and much of the equipment inside the data centers fueling the AI boom, is positioned to earn roughly $250 billion in profit this year, he said. He called it the largest and most powerful company in the world, noting Huang has President Trump's ear.

The data center construction now underway will be the biggest infrastructure investment in history, Witt argued — larger than railroads and canals combined, larger than highways, larger than telecom. Huang is now all in on robotics, operating on the thesis that everything that moves will become autonomous and AI-controlled. Nvidia has been developing robot brains since before the robots existed.

Fully autonomous restaurants? Witt said they already exist in China, though even there humans still handle some of the actual cooking. His own line is sharper: language processing and business intelligence are ready for deployment; robotics is not.

"I would gladly have ChatGPT do my taxes," he said. "But if someone was like, 'I'm going to give this robot a knife and chop some cabbage in the kitchen,' I don't want to be anywhere near that thing."

Build cheap, and build now

The core advice for restaurant companies: build the AI systems you need internally, on the cheap, rather than buying from vendors whose products could rapidly become obsolete given the pace of advancement and competitive pressure.

Five years ago, Witt said, he would have counseled caution on in-house development — IT departments had a terrible track record of delivering on-time, tangible results and constantly ran over budget. Cloud coding has changed that math.

"Now it's different," he said. "I would advise you to build this stuff on the cheap. Take a team, if you're considering an IT deployment, and you should know within just a couple of weeks of experimenting whether or not you have a pathway or a runway to cloud coding, essentially a bespoke software application for your organization."

He drew one boundary: don't attempt anything requiring large-scale infrastructure deployment. "I would rely on a vendor for that," he said.

The political risk

Despite widespread alarm over AI, Witt is mostly optimistic, though he sees real risk and positions himself somewhat at odds with Huang, arguing AI needs better guardrails. He pointed to this summer, when AI broke out of its container and marauded through the internet, taking over other computers. AI is already solving math problems humans couldn't solve — superintelligent in some ways, very stupid in others, and that combination is what creates the danger. "I don't think it's an engineering problem that can't be resolved," he said.

The bigger threat to the industry's AI ambitions may be political. Americans hate data centers and are deeply skeptical of AI. The booming AI sector feeds what Witt called the K-shaped economy, squeezing those on the lower end and breeding resentment. A recent Gallup poll found 90% of Chinese respondents view AI as good, with 70% in Singapore and similarly high numbers in Korea and Japan. In the U.S., the figure was 35% — a majority think AI is bad.

"And we're the country that's in the lead. We're the country that's building it," Witt said. "I'm not any kind of socialist or anything, like, I'm pro-business. But I see this ending, and it could have real repercussions down the road."

His proposed fix: pay the people who live near data centers. "What they really love is cheap gasoline. So every data center should have a Buc-ee's attached to it."

For restaurant operators, the takeaway from Dallas was direct: two weeks of cheap in-house experimentation can now tell you whether bespoke AI is viable for your organization — before the vendor landscape shifts again.