Unable to load audio. Please try again later.
Capital and talent have come apart across the AI infrastructure stack. Where they diverge tells you which layers are mispriced.
Every new gigawatt of datacentre capacity depends on transformer and grid hardware. That hardware depends on medium-voltage solid-state power conversion. Worldwide, that is the specialism of roughly 630 engineers. Zeki found them across more than 5,000 deep-tech companies. Fifteen employers hold 64 per cent of them. Two — Tesla and GE Vernova — hold more than half of that.
You can raise a billion dollars in a quarter. You cannot raise a power electronics engineer. That is the constraint underneath the constraint. Packaging, memory, interconnect — every bottleneck in AI infrastructure resolves, eventually, into a headcount. Those headcounts are small, and the capital map and the talent map have stopped agreeing on where the people should be.
One well, one sink
Thirteen per cent of Intel’s hardware engineering talent has left in three years, and the rate is not slowing.
They did not scatter. Intel is now the largest single insider source for the inference chip challengers — Tenstorrent, Cerebras, SambaNova, Groq — and, separately, the largest single feeder into edge AI silicon, where roughly nine per cent of the workforce has passed through it.
The second figure is the strange one, because those edge companies otherwise share almost nothing. Computer-vision firms recruit from chip incumbents. On-device LLM firms recruit from AI accelerator startups. Always-on sensing firms recruit from somewhere else entirely. Four sourcing strategies, one common origin. A single company’s restructuring is staffing an entire challenger class.
Then it drains again. Across Tenstorrent, Cerebras and SambaNova, departures to Nvidia run at seven to eleven per cent of all workforce movement. Arrivals from Nvidia run between half a per cent and three.
Three to one, one direction, five years running, against 1,800 open engineering roles still posted. Groq is the extreme rather than the exception: since the December 2025 partnership, 69.5 per cent of its outflow has gone to Nvidia.
Which produces a diligence test that takes one query to run. Can this challenger hire senior people out of Nvidia? Almost none can.
So why is anyone walking the other way?
Because on one layer, they are — and it is the layer with no money in it.
Power and grid scale-ups raise $1.12m per head. AI silicon raises $5.99m, more than five times as much. Yet the power layer carries nearly twice the share of former founders and CTOs, 12.3 per cent against 6.5, and matches AI silicon on doctorates.
The most entrepreneurial talent in AI infrastructure is choosing the layer the market prices lowest.
Edge AI silicon is the exact inverse. It raises less than half what other AI chip companies raise per head — and it also hires later-career engineers less often, recruits from top labs less often, and grew headcount by a median nine per cent last year against 22. Median experience at hire is 85 months against 121. Both gaps point the same way.
One of those is a mispricing. The other is a trap. On a cap table they look identical.
Hiring commits before revenue arrives
The reason talent leads is mechanical. A company decides what it is going to be, then hires for it, then finds out whether it was right. The hiring is visible from the outside about two years before the answer is.
Two patterns make that legible.
Edge AI companies carry 20 per cent of staff in go-to-market roles against 12 per cent across AI hardware generally — and almost none show meaningful revenue. Heavy commercial hiring in a fragmented market measures coverage cost, not traction. The honest signal is narrower: field applications and solutions engineers, who are only needed once something is being deployed.
And the gap between the two is datable. DG Matrix hired its first commercial staff in 2023, ran pilots from 2024, delivered in June 2026 against a $20m order. Three years, near enough. That interval is stable enough to date a company’s revenue from its org chart.
Hailo, in four moves
April 2024. Hailo is valued at $1.2bn. It has raised over $340m, fields one of the largest teams in edge AI, and has shipped more than 500,000 chips into real products, including Husqvarna mowers and the Raspberry Pi AI HAT.
August 2025. Its vice president of sales leaves for BrainChip.
February 2026. A senior field application engineer leaves for SiMa.ai.
June 2026. The company announces a 50 per cent workforce cut. The valuation is under $500m.
The financials caught up in 2026. The people had already left, and they left for direct competitors in the same small cohort — not for a different industry, not for more money elsewhere. Attrition that is senior and directional is not turnover. It is a verdict.
Capital tells you what a market believes. Hiring tells you what the people closest to the work believe. When those two come apart, the people are usually right first.
What this article leaves out
Every figure above is cohort-level. Zeki’s diagnostics run at company level: which firms are hiring against their next problem rather than their last round, which ones can pull senior engineers out of Nvidia, and where each sits on the clock between first commercial hire and first revenue — including companies that have neither yet.