Cerebras Backer Pivots to 'Physical AI' at Mayfield

Adit Singh, an early investor in Cerebras Systems, has graduated to the next rung of the VC ladder: he is now an infrastructure partner at Mayfield, where he will focus on semiconductor, cybersecurity, and physical AI investments. This is the kind of career trajectory that requires no actual exits, no demonstrable returns, and absolutely no accountability—just a willingness to rebrand your thesis every eighteen months and call it vision. Singh's new portfolio mandate is instructive primarily for what it reveals about how capital allocates itself when it has exhausted traditional categories and must invent new ones.

Let us be clear about what Mayfield's infrastructure partner will actually be doing: he will be writing checks into semiconductor companies and hardware ventures, a category that has consumed tens of billions of dollars from venture capital over the past five years with wildly uneven results. The semiconductor and AI chip space is crowded with well-capitalized players—Nvidia, AMD, custom silicon makers, and dozens of venture-backed startups burning capital at rates that would make a 2000-era B2B SaaS company blush. Into this arena, Singh brings his Cerebras track record, which is to say: he backed a company that promised to revolutionize AI computing through custom architecture and has spent years trying to find customers willing to buy its chips at premium prices. The fact that he is now positioned as an infrastructure expert rather than questioned about that investment tells you everything about VC amnesia.

Singh's pivot to Mayfield arrives precisely at the moment when 'physical AI' has ascended from Reddit thread to legitimate venture category. This is not a coincidence. In VC, terminology is destiny. When a thesis stops working under its old name, the solution is not to admit defeat—it is to rebrand. 'AI infrastructure' was hot for a while, then 'edge AI,' then 'sovereign AI,' and now 'physical AI.' Each phrase means roughly the same thing: we will give you money to build hardware and hope someone buys it before your runway ends. 'Physical AI' is simply the latest linguistic container for the same bet that Cerebras, Graphcore, and a hundred other chip startups have been making for years: that custom silicon beats general-purpose processors for AI workloads. The category name changes; the capital flows in the same direction.

The press release framing is pure VC boilerplate masquerading as insight. Singh will focus on 'infrastructure' investments, which means he will evaluate companies based on whether they are building chips, networks, or security tools that claim to be foundational to AI. This is not strategy; this is bucket-filling. Mayfield is placing a bet that the next decade of venture returns will come from the industrial backbone of AI—the picks-and-shovels play—rather than from applications that might actually generate revenue. It is the same bet that has been made repeatedly: that whoever controls the hardware layer wins. History suggests otherwise. The companies that have actually built durable AI businesses—OpenAI, Anthropic, Perplexity—are ultimately constrained by the chips they can rent or buy, not by the availability of venture capital for new chip designs.

What could go wrong with Singh's new mandate? Almost everything. Semiconductor design cycles are long; customer adoption is uncertain; the competitive landscape is brutal and capital-intensive; and the venture model assumes returns that may never materialize. Cerebras, despite years of funding and hype, remains a boutique player in a market dominated by Nvidia's margin-generating installed base. Other venture-backed chip companies have burned through hundreds of millions with limited commercial traction. The pattern is consistent enough that adding a new buzzword category does not change the underlying physics: custom silicon is expensive to design, expensive to manufacture, expensive to iterate, and only valuable if customers actually adopt it. Singh's track record suggests no special insight into whether any given semiconductor bet will overcome these headwinds.

What the Singh-to-Mayfield move really reveals is the current state of VC capital allocation in 2026: when the easy deals are exhausted and the narrative is tired, the solution is to hire a veteran who has seen the category before, give him a new title, and let him deploy fresh capital into the same infrastructure race. Mayfield is not making a breakthrough bet; it is deploying capital into a well-established category under a freshly minted name. Singh is not bringing secret knowledge about semiconductor success; he is bringing credibility, networks, and a willingness to believe that this time the hardware thesis will work. Call it what it is: recycling.

In VC, past performance does not predict future results, but it does predict future capital allocation. Singh backed Cerebras; Cerebras is still trying to prove its business model; and Singh is now positioned to write checks into the next dozen companies attempting the same bet. The house always wins.