Cerebras CEO Andrew Feldman sees home AI assistants coming soon

Andrew Feldman, CEO and co-founder of Cerebras Systems, gave a short answer when asked at TechCrunch Disrupt whether personal AI assistants will end up as just another household appliance: "Actually, I don't think that's very far away." He said it on Tuesday 13 October in San Francisco, on the first day of an event that runs until the 15th.

The prediction is not new in the industry, but it comes from someone whose business is precisely the hardware that runs models. Cerebras makes wafer-scale chips for training and inference, so its view of where AI will run has direct implications for its own market. The interview is covered by TechCrunch.

What Feldman said

Feldman described a hybrid architecture. "Just like your cell phone, they'll do some work on device and some work going back to the cloud," he explained. In his view, these assistants will build up personalised knowledge about each user, operate within a security perimeter and have access to credit cards to buy things on the user's behalf.

As a sign that consumers are already moving in that direction, he cited Meta's Muse and Instinct, a startup valued at 10 billion dollars. He himself uses "a collection of agents" to manage his calendar and prepare for calls, and said these systems "will make a meaningful impact in making our lives better."

He was explicit about Cerebras' role: it does not aim to build the consumer product. "I like building the platforms that other people's ideas take flight on," he said.

Why it matters

The split between device and cloud is the most concrete part of the pitch. Small models running locally solve latency, cost and privacy for frequent tasks, while long reasoning or searches still need data centres. If that hybrid model wins, demand for cloud inference does not disappear: it changes shape, with shorter, more frequent requests that are more sensitive to response speed. That is exactly the kind of workload Cerebras has positioned its hardware for.

The other part, payments, is the most delicate. An assistant that shops for you needs more than a good model: scoped permissions, a log of every action and a clear way to revoke access. Feldman spoke of security perimeters but did not detail what they would look like.

What we already see in the Claude ecosystem

Much of what Feldman describes already exists as separate pieces for anyone working with Claude. MCP servers let a model query calendars, email or external APIs with permissions the user defines. In Claude Code, hooks such as PreToolUse can intercept an action before it runs and ask for confirmation or block it, which is a practical version of that security perimeter. And subagents make it possible to split tasks, such as preparing for a meeting, across specialised agents, something very close to the collection of agents Feldman says he uses.

The difference from his vision is the audience. Today these setups are built by developers and technical teams; the leap he anticipates is for them to work at home without anyone having to touch a configuration file.

Who it is useful for

Product teams designing consumer assistants: the hybrid architecture forces a decision about what is processed locally and what leaves the device, and that decision affects privacy and cost.

Agent developers: if delegated payments are going to be common, authorisation controls should be designed from the start, not bolted on later.

Anyone following the AI hardware market: Feldman's pitch confirms that chipmakers see consumer inference as their next big source of volume.

Our reading is that the technical side is more mature than the trust side. Running a hybrid assistant is feasible today; handing it a credit card without supervision is another conversation, and the limits should be set by the user, not the manufacturer.