The man who sells chips to the AI industry thinks the AI industry should talk less. Andrew Feldman, CEO and co-founder of Cerebras Systems, told the audience at Disrupt 2026 that "some of the leaders in the AI world maybe shouldn't be allowed to talk in public." It is a funny line from someone whose business depends on the thing being discussed. It is also a real argument, and worth taking apart.
The doom problem
Feldman's complaint is about tone. Public conversation about AI has drifted toward containment failures, models misbehaving, and insiders trading estimates of the percentage chance that AI wipes out humanity by the end of the decade. Apparently this is now routine dinner conversation.
The mood has gone mainstream enough that Anthropic CEO Dario Amodei got parodied on "Saturday Night Live." The fictional version summed up the discourse with the line, "We do not condone what we are doing." When your risk messaging becomes a sketch premise, the messaging has arguably escaped your control.
Feldman's view is that the negativity is crowding out the other half of the story. He thinks AI leaders should spend more time describing what the technology could deliver, even while they hold concerns about its risks. His example was a world in which your grandkids "never meet anybody who dies of cancer."
The interested party
Feldman has a stake in this, and the source is upfront about it. Cerebras counts OpenAI, Block, Figma, the pharmaceutical company GSK, and Lovable among its customers. A chipmaker whose clients are building and deploying AI is not a neutral referee of the doom debate, and nobody should pretend otherwise.
But he did not argue that risk is imaginary. His position is closer to a shrug of realism. In his words, powerful technologies come with the question of what happens when they land in the hands of bad people, and you cannot have one without the other.
His prescription has three parts:
- Say it plainly. The technology is powerful.
- Be thoughtful. Treat that power with care.
- Regulate reasonably. The operative word being "reasonably," which is doing a lot of work and which he did not define.
The community problem
The more concrete part of his talk concerned data centers. Feldman said some of the public negativity traces back to how the industry handled the buildouts that AI demand set off. In his telling, the industry spent the past few years being poor at working with local communities, particularly in the United States.
He invoked the power industry, including oil and gas, as a point of comparison, and said the data center industry did not engage communities the way it should have. He called that a problem, which is an unusually direct word for an executive whose sector needs a great many permits.
He also pointed abroad. The US, he said, is dealing with NIMBY-ism ("not in my backyard!"), while other countries are encouraging companies to build. He named Brazil, Canada, Finland, Iceland, Norway, Mexico, and Sweden.
The subtext is plain enough. If communities at home keep saying no, the compute goes somewhere that says yes.
The efficiency worry that isn't
Cerebras competes with Nvidia, so the obvious question is what happens if everyone gets better at doing more with less. If models and chips both grow more efficient, does demand for chips shrink?
Feldman was not concerned. His answer was that hardware vendors can make compute more efficient, while model makers do their own work to make models more efficient in pursuit of the same goals. He framed it as a division of labor, with progress requiring all of it to come together. The report does not include a deeper argument for why total demand holds up, so what we have is his confidence rather than his math.
What to take from it
Three claims are worth separating, because they differ in how checkable they are.
- The tone claim. Doom talk is crowding out the upside. This is a judgment call, and reasonable people will land in different places.
- The community claim. The data center industry underinvested in local engagement. This is the most actionable point, and it is an admission from inside the industry.
- The demand claim. Efficiency gains won't dent chip demand. This is a forecast from a vendor, so apply the usual discount.
The awkward part of the first claim is that the people most associated with the risk talk are often the ones building the models. Feldman's suggestion that some of them stay off the microphone is a joke, but it points at a real tension. The industry wants credit for taking safety seriously and also wants the public to stay calm about it.
His most practical advice may be the least glamorous. Before anyone promises a cancer-free generation, they might try attending the zoning meeting.