The chip maker building the hardware for the AI future thinks the people running the biggest AI labs need to shut up about doom.
The Summary
- Cerebras CEO Andrew Feldman told TechCrunch Disrupt that AI doomsday warnings from industry leaders are drowning out the technology's benefits
- Feldman predicts personal AI assistants will become household appliances, running partly local and partly in the cloud
- His bet: The people selling picks and shovels have a clearer view than the gold miners warning about the mountain collapsing
The Signal
Andrew Feldman has a front-row seat to the AI buildout. His company sells the specialized chips that train frontier models, the hardware layer underneath the chatbots and coding assistants everyone's using. So when he says the industry's apocalypse talk is off base, it's worth parsing why.
At TechCrunch Disrupt, Feldman argued that AI leaders' fixation on existential risk is obscuring the technology's practical upside. Translation: Sam Altman and Dario Amodei keep talking about AGI timelines and risks while Feldman is watching actual deployment patterns. His read on where this goes is more prosaic and probably more accurate.
"We're nearing a future where everyone has a personal AI assistant at home."
Feldman's vision isn't cloud-only assistants locked in data centers. He sees a hybrid model: part local, part cloud. Think less "rent your brain from OpenAI forever" and more "your device handles the routine stuff, calls out for the heavy lifting." This maps to what we're already seeing with on-device models in phones and the push for edge inference.
The hardware matters here. Cerebras builds wafer-scale chips designed to train models faster and more efficiently than GPU clusters. If personal AI assistants are coming, someone needs to make the silicon that powers them without burning through a small nation's energy grid. That's the infrastructure layer of Web4, the part that makes "everyone has an agent" economically feasible instead of a thought experiment.
Key deployment assumptions:
- Personal assistants become as common as smart speakers or thermostats
- Hybrid architecture splits workload between local devices and cloud compute
- Cost and energy efficiency improve enough to make this accessible, not luxury-tier
Feldman's pushback on doom scenarios matters because it reveals a fault line in how AI gets built versus how it gets marketed. The CEOs racing to AGI have an incentive to talk up both the promise and the peril. It keeps regulators engaged, justifies safety teams, and positions their companies as responsible stewards of powerful tech. Meanwhile, the infrastructure players see demand patterns that look less like runaway superintelligence and more like everyone wanting a better version of Siri.
The Implication
The doom narrative serves the frontier labs. It makes them gatekeepers, the only ones responsible enough to build the dangerous thing safely. Feldman's counter-narrative is simpler: AI is becoming a utility, and utilities don't end the world. They get regulated, commoditized, and eventually boring.
Watch where chip makers place their bets. If Cerebras and competitors start optimizing for edge deployment and hybrid models, that tells you more about the next five years than any white paper about alignment. The people building the hardware see local agents coming. The question is whether the model makers will let them run without a subscription.