The $1 Billion That Wasn't Supposed to Exist
Let's be honest: nobody expected this.
Not in 2023, when Gavin Uberti and Robert Wachen were eating ramen in a San Francisco co-living space, running on fumes and a $125,000 runway. Not in 2024, when every VC they pitched said, "Great idea, but GPUs are good enough." And certainly not in early 2025, when they were one payroll cycle away from shutting down.
Yet here we are. Etched, a company that didn't even have a website until last month, has booked $1 billion in contracts for AI inference systems—before shipping a single rack. They've raised $800 million across four unannounced rounds, and now sit at a $5 billion valuation. And they didn't do it by building a better GPU. They built something entirely different.
This isn't a story about disruption. It's about invisibility.
Etched didn't come out of stealth. They were never in it.
They were just… building.
And now the market is catching up.
The Quiet $800 Million
The $500 million round closed in December 2025? Unannounced. No press release. No Crunchbase update. Just wires moving.
That's not a PR strategy. That's a survival tactic.
In 2023, when Uberti and Wachen started pitching, the AI chip space was already crowded. Cerebras had just gone public. Groq was raising $300 million. NVIDIA's H100s were flying off the shelves. Why would anyone bet on a third-party inference chipmaker?
The answer, apparently, was a 30-page memo. And a lot of silence.
The investors who did say yes—Jane Street, Hudson River Trading, Two Sigma, Ribbit Capital, and a handful of angels including Andrej Karpathy and Geoffrey Hinton—weren't betting on a product. They were betting on a process.
Etched's team didn't just have hardware chops. They had obsession. Brian Loiler, ex-NVIDIA platform lead, joined after 22 years at the company. Mark Ross, ex-CTO of Cypress, had shipped five systems that each generated over $1 billion in revenue. Wayne Cao, who led iPhone and Pixel production, had seen what happens when you don't control your supply chain.
They didn't want to be another chip company. They wanted to be the company that redefined what a chip company could be.
And they did it by refusing to talk.
The Two Breakthroughs Nobody Saw Coming
Etched's product isn't a chip.
It's a cluster.
Frontier inference clusters. That's what they call them. And the name isn't marketing. It's a technical manifesto.
They're not optimizing for FLOPs. They're optimizing for throughput under real load. That's the difference between a spec sheet and a server rack.
Two things make their system unique:
Low Voltage Inference (LVI)
Most AI chips throttle under load. The more math you ask them to do, the hotter they get, the slower they run. It's physics. It's unavoidable.
Etched cracked it by running their math blocks at half the voltage of conventional chips. That's not a tweak. That's a rewrite of the transistor-level design. They rebuilt the power delivery, the clocking, the thermal management—all from the ground up.
The result? 80%+ peak FLOPs utilization on trillion-parameter MoE models. No throttling. No thermal throttling. No magic. Just engineering.
Cluster Scale Memory (CSM)
Memory is the silent killer of AI inference. HBM is slow. SRAM is expensive. You pick one, you sacrifice the other.
Etched built a hybrid memory pool that spans multiple chips using a proprietary interconnect. It's not optical. It's not 3D stacked. It's something new—low-latency, high-bandwidth, and designed to keep the decode pipeline fed without bottlenecks.
The outcome? Latency cuts of 40–60% on long-context workloads. That's not incremental. That's transformative.
And here's the kicker: they didn't build this in a lab. They built it with customers. Dozens of engineers lived overseas for months, working side-by-side with TSMC, hyperscalers, and AI labs to co-design every layer—from the transistor to the token.
The $1 Billion in Contracts
You can't book $1 billion in contracts for a product that doesn't exist.
Unless you're Etched.
They didn't pitch a roadmap. They didn't show a prototype. They showed results.
Early customer tests, run in real data centers, showed their systems outperforming H100 clusters on cost-per-token by 3x. Power efficiency? 2.5x better. Latency? Half.
And they didn't need to convince anyone. The customers came to them.
One major cloud provider, already running 20,000 H100s, asked for a pilot. Then two. Then a full-scale deployment. They didn't wait for a press release. They didn't wait for a conference keynote. They just signed the contract.
The $1 billion? That's not revenue. That's commitment. That's a guarantee that when these racks ship this summer, they'll be filled with real workloads from real AI companies.
And that's what scares NVIDIA the most.
Not a better chip.
A better partnership.
This shift away from single-supplier dependence mirrors what OpenAI and Broadcom demonstrated with their Jalapeño custom chip—see OpenAI's Jalapeño Chip and the End of Nvidia's Monopoly Era—but Etched is taking it further by building the entire stack in-house rather than relying on a partnership model.
The Real Threat Isn't the Chip
NVIDIA doesn't lose because someone builds a faster GPU.
They lose when someone builds a system that doesn't need a GPU.
Etched isn't competing on specs. They're competing on ecosystem.
They're vertically integrated: design, packaging, cooling, software, manufacturing—all under one roof. Engineers sit next to thermal experts. ASIC designers talk to supply chain managers. No silos. No handoffs. No delays.
They've opened a factory in Taiwan. Built a test house in San Jose. Hired 400 engineers from NVIDIA, Google, Broadcom, and TSMC.
They didn't raise $800 million to build a product.
They raised it to build a machine.
And that machine? It's already running.
The broader pattern of AI chip startups navigating NVIDIA's dominance is complex—some, like Groq, were acquired for their IP and talent (After Losing Its CEO and Core IP to NVIDIA, Groq Bets on a Cloud Pivot), while others like Etched are building entirely independent architectures from scratch.
The Quiet Revolution
This isn't a startup story.
It's a quiet revolution.
No IPO. No keynote. No investor call.
Just a team of 400 engineers, working in silence, building something the world didn't know it needed.
And now, the world is catching up.
Etched didn't come out of stealth.
They were never in it.
They were just building.
And now? The inference race has a new winner.
Not because they had the best tech.
But because they had the patience to build it right.