Every major tech company is now building its own AI chip.
Google is developing two new inference chips with Marvell — a custom memory processing unit and a next-generation TPU. Intel joined Elon Musk's Terafab project alongside SpaceX and Tesla. Tesla's AI5 chip just completed tape-out, with AI6 and a new Dojo supercomputer generation already in development.
Three companies. Three simultaneous bets on custom silicon. All of them making the same calculation: NVIDIA dependency is a strategic risk, and the organizations that control their own inference hardware will run AI workloads faster and cheaper than the ones that don't.
That calculation is probably right. It also creates a talent problem that almost no one is solving for.
Why Everyone Is Leaving NVIDIA
The NVIDIA dominance story is real and it's expensive. H100s and H200s are in short supply, command premium pricing, and put every company's AI infrastructure roadmap at the mercy of one supplier's allocation decisions.
Custom silicon breaks that dependency. A Google TPU optimized for Google's inference workloads doesn't compete on NVIDIA's timeline. A Tesla AI chip tuned for Tesla's training runs costs a fraction of what commodity GPU compute would. And a Terafab-scale project — AI chips built by the company that also runs the rockets — represents a vertical integration play that would have seemed absurd three years ago.
This isn't just big tech chasing efficiency. It's big tech deciding that hardware is a core competency, not a purchase order. That decision is spreading.
The Talent Layer Nobody Talks About
Here's the part that doesn't make the press release: custom silicon requires people. Specifically, it requires chip design engineers, hardware verification specialists, and inference optimization engineers who understand how to map AI workloads to custom architectures.
These people are extraordinarily scarce.
Software AI engineers are scarce. Agentic AI architects are scarce. But hardware AI engineers — people who can design or optimize inference chips, write firmware for custom neural processors, and validate silicon before it ever runs a model — are a completely different tier of rare. Traditional chip design is a 5-10 year career path. AI-optimized chip design barely has a career path yet.
Universities aren't producing enough of them. Boot camps don't touch this stack. And the companies that already have these people — NVIDIA, AMD, Apple, Qualcomm, Broadcom — are not in the habit of letting them leave.
What This Means Outside of Big Tech
Most companies reading about the Terafab project or the Google-Marvell partnership are thinking: that's a hyperscaler problem. My company buys compute; we don't make it.
That's accurate today. It's a shorter-term position than it looks.
As custom inference chips become the standard for enterprise AI workloads — as they will, because the cost curve demands it — organizations are going to need engineers who can evaluate, integrate, and optimize AI systems at the hardware layer. Not build chips from scratch, but understand what the hardware is doing well enough to architect around it. That's a meaningfully different skill set than what most technical hiring processes are screening for.
The companies building hardware expertise now — even at the integration and evaluation level, not the design level — are going to have a structural advantage as the hardware landscape diversifies. The ones that don't are going to be dependent on vendors to explain what's happening in the machines running their most critical workloads.
That's not a comfortable place to be when those workloads are making decisions that affect your operations, your customers, and your margins.
The Hiring Signal Hidden in the Chip News
When you watch Google partner with Marvell, Intel join Terafab, and Tesla tape out a new inference chip in the same week, the obvious story is about chips. The actual story is about a category of technical talent that is about to become the most contested in the market.
Hardware AI engineers are already employed. They're not on job boards. They're not responding to generic outreach. Getting to them requires sourcing capability that most internal teams don't have — and moving fast, because the companies with deeper pockets and longer runways are already in the market.
The organizations that start building those sourcing relationships now — before the bidding wars start — are going to be in a materially better position than the ones that wait for the market to tell them there's a problem.
The market will tell you. It just won't wait for you to be ready.
VC5 Consulting specializes in hard-to-source technical talent — the engineers who don't appear in the first three searches. If you're building AI infrastructure and need people who understand the hardware layer, let's start there.