From a Friend's Towel-for-a-Blanket to $10.3 Billion
Robert Wachen landed in the Bay Area with no apartment lined up, no office booked, and parents back home who weren't sure this Harvard-dropout-startup thing was going to work out. He slept on the floor of a friend's house that was about to be sold, using a folded towel as a blanket.
Three years later, Etched — the AI chip startup Wachen cofounded with Gavin Uberti and Chris Zhu — closed a $300 million Series C at a $10.3 billion valuation. The round was led by Sequoia, with Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital also participating. Sequoia calls it the highest valuation ever for a company at a Series C they've led.
Wachen told TechCrunch he keeps a real pillow now. Multiple pillows, actually. He laughed about it. But the point isn't the pillow — it's that the trajectory from towel-on-floor to $10 billion happened fast, and not because conditions were favorable. For most of Etched's early life, the opposite was true.
What Etched Actually Does — the Misconception First
Before getting into the money, it's worth clearing something up, because Etched keeps having to do this.
The company does not make chips that only run specific large language models. That's the perception — a lingering one — rooted in Etched's origins as a startup building silicon purpose-built for transformer-based AI. Back in 2022, when the founders dropped out to do this, transformer-optimized chips sounded niche to the point of foolishness.
The product today is different. Etched's systems can run any AI model: Mixture of Experts models like DeepSeek and Qwen, which split tasks across many smaller sub-models rather than one big one, and even non-transformer architectures like Mamba, which uses a fundamentally different design called a state-space model. "The systems can run any AI model," Wachen told TechCrunch plainly.
They sell full systems — not bare chips. They call them frontier inference clusters: co-designed chips, custom racks, cold plates, interconnects, power delivery, and software, all tuned to run frontier model inference better than what's available today. Customers aren't buying a component; they're buying a complete compute environment.
The Two Technical Bets at the Core of Etched
Inference — the computing that happens after you hit enter on a prompt — splits into two stages, and Etched designed a new component for each.
Low Voltage Inference for the Prefill Stage
The prefill stage is the compute-heavy part. Your prompt comes in, context gets parsed, math gets done. It's intensive. Most chips throttle here: as utilization climbs, heat builds, clock speed drops, and sustained throughput ends up less than half of peak FLOPs on paper.
Etched built a chip that runs its math blocks at under half the voltage of conventional AI chips. Less voltage means less heat, which means the chip can pack more transistors and run at much higher utilization without thermal scaling down. The company says their systems sustain above 80% peak FLOPs on trillion-parameter sparse MoE models — without throttling — which is a number general-purpose GPUs can't touch in real-world inference conditions.
This required redesigning the entire stack: splittable math arrays, novel circuit techniques, tiling and scheduling algorithms, power delivery networks, voltage regulator architecture, and the packaging itself. It wasn't a firmware fix. It was a ground-up architecture decision.
Cluster Scale Memory for the Decode Stage
The decode stage is where the model generates the actual answer — one token at a time. Less compute-intensive than prefill, but wildly memory-hungry. The problem: existing AI chips using HBM memory can't achieve SRAM-level speeds because the memory subsystem and interconnect architecture creates bottlenecks. SRAM-only chips solve the speed problem but sacrifice capacity and FLOPs density.
Etched built a third option. They created a proprietary ultra-low-latency interconnect — not optical, not 3D-stacked DRAM — that lets many chips share a common high-bandwidth memory pool across the full compute cluster. Their HBM/SRAM hybrid design, which they call Cluster Scale Memory (CSM), solves both the capacity and latency problems simultaneously. The result is a decode pipeline that stays fed without the bottlenecks that kill interactive application performance.
To understand how much memory architecture matters in practice, The Memory Tipping Point explores why AI's soaring bandwidth demands are reshaping the entire semiconductor supply chain.
Taken together, LVI and CSM are why customers with access to the systems are signing contracts. Andrej Karpathy from Anthropic, Noam Brown from OpenAI, and Geoffrey Hinton — Hinton, the godfather of deep learning — all tried the hardware privately before their names appeared on Etched's investor list.
Why AI Developer Tools Startups Face the Same Problem Etched Did
The reception Etched got in 2023 wasn't unique to chips. Across the AI developer tools landscape — startups in India, the US, and Europe building specialized infrastructure — founders keep bumping into the same wall: investors and buyers convinced that general-purpose solutions (read: GPUs, cloud APIs) are good enough, and that purpose-built tools are too narrow to matter.
Etched's AI hardware journey is a sharp illustration of that broader dynamic. They wrote a 30-page memo arguing that AI would eventually need specialized compute — not just general-purpose GPUs — and every major investor they pitched passed. Every single one. The company ran month-to-month, close to running out of cash entirely.
The same logic plays out in AI developer software tools: narrow, specialized products get dismissed as edge cases until the market matures and the edge case becomes the norm. What changed for Etched wasn't the idea — that was always right. What changed was the market's ability to see it.
Now the AI hardware ecosystem has multiple well-funded players: Groq (which eventually pivoted after losing core IP to Nvidia), Cerebras (which went public), and Etched, which took the fully vertical path. Each made different bets about how to compete with Nvidia's GPU dominance. Etched's bet — own the entire system stack — is the most capital-intensive, and apparently the one that's resonating right now.
The broader investor thesis has flipped entirely. In 2026, the question isn't whether AI needs specialized compute. It's who's going to provide it at scale.
$800 Million Raised, $10.3 Billion Valuation, $1 Billion in Orders
The numbers stack up quickly. Etched has raised $800 million across four rounds, almost all of it unannounced as it went in. The most recent prior round was a $500 million tranche that closed in December 2025 at a $5 billion post-money valuation, led by Stripes, with VentureTech Alliance, Jane Street, Hudson River Trading, Two Sigma, and Ribbit Capital participating.
The $300 million Series C announced in July 2026 — at $10.3 billion — doubled the valuation in roughly seven months. The lead investor shifting to Sequoia signals something: this is no longer a hardware bet being carried by quant funds and high-net-worth angels. It's mainstream VC conviction.
Angel investors read like a who's-who of the AI research community: Karpathy, Hinton, Fei-Fei Li, Arthur Mensch from Mistral, Scott Wu from Cognition, Noam Brown from OpenAI, Aidan Gomez from Cohere, Tri Dao (the researcher behind FlashAttention). Stanley Druckenmiller and Peter Thiel are on the cap table. Dylan Field from Figma. Amjad Masad from Replit.
These aren't passive checks. Several of them tried the hardware in person before writing a wire. Wachen told TechCrunch the company landed famous investors by showing them private demos at the office — not with decks, with hardware.
As of June 2026, Etched had already booked $1 billion in customer orders. Its A0 silicon came back from TSMC — manufactured under TSMC's Emerging Businesses Group, on the N4P process node — earlier this year. The first full rack-scale systems are being tested with early customers, with first rack shipments expected in summer 2026.
For a look at how another AI startup secured big-ticket funding on specialist positioning, Sarvam's $234M raise shows how India-focused AI developer tools startups are now attracting serious institutional capital.
The Team That Built It
The founders are still young — Uberti (CEO), Wachen (COO/President), and Zhu (co-founder) were all Thiel Fellows, a program that pays students $100,000 to drop out and start a company. They were joined by CTO Mark Ross, formerly CTO of Cypress Semiconductor, who shipped five systems generating over $1 billion in revenue each — all on A0 silicon, which is the first-pass chip that usually has bugs.
Brian Loiler, VP of Platform, spent 22 years at NVIDIA and built the HGX and DGX systems from scratch. Wayne Cao, VP of Production, has led supply chain ramps for the original iPhone, MacBook Air, Pixel, and Chromebook. Saptadeep Pal, VP of ASIC and Architecture, was on the NVIDIA H100/A100/V100 architecture team and co-founded Auradine.
This is unusual for a startup at this stage. You'd normally expect a founding team and a collection of early hires still working things out. Etched has genuine hardware industry experience at the VP level — people who've seen what it looks like when chips actually ship at scale.
Wachen was candid about how far they still have to go: "I think we still have to be humbled by what it will take to actually get to scale." That's not false modesty. Chip manufacturing at gigawatt scale is a different category of problem than building a prototype that works in a lab. Early-stage engineering success and mass-production success are two completely separate mountains.
Operations Today — and What's Next
Etched's San Jose office runs a 2-megawatt data center. In parallel, the company opened a new 80,000 square-foot, 10-megawatt facility in Milpitas, about 15 minutes away, built to support continuous deployment of first-gen hardware and rapid prototyping of future generations. They've also opened a factory in Taiwan and built a test house and NPI prototyping lab in San Jose.
The team is 400+ engineers drawing from NVIDIA, Google TPU teams, Broadcom, SK Hynix, and TSMC. Etched describes itself as vertically integrated — math block designers sit next to inference engineers, thermal experts next to global supply managers. The mission: reach gigawatt-scale inference compute. Their website states it plainly: "run the world's inference."
Google is reportedly pursuing a similar concept with its Frozen v2 chip for Gemini — the idea of etching a specific model's architecture into silicon. When Etched was founded, that idea was considered far too narrow. Now Google is apparently exploring it for one of the world's most commercially important AI models. The validation couldn't be much clearer.
Etched is pre-mass-production. Rack systems are shipping and being validated, but not yet rolling off the line in volume. With $10.3 billion in valuation, a billion dollars in committed orders, and working silicon already back from TSMC, though, the gap between "promising startup" and "real infrastructure company" is closing.
No GPUs Required — and a Market That's Finally Listening
Etched's core claim is deliberately provocative: systems that speed up inference on any AI model without requiring GPUs. That's the pitch. Based on what early customer results and the investor response suggest, it's holding up.
Nvidia isn't sitting still. Its cumulative data center revenue is projected to exceed $500 billion by end of 2026. It dominates training compute and a large slice of inference. But inference is specifically where challengers have the best opening — it's the bottleneck frontier AI companies actually pay to fix, the cost center that eats margins when you're serving users at scale.
Etched isn't trying to beat Nvidia at training. They're not in that fight. They're competing on the inference problem that actually keeps AI product teams awake: how do you serve a billion tokens a day without burning through margin, while waiting for the next GPU generation to land?
That framing — specialized compute for a specific, painful, financially significant problem — is why Etched has found both VC conviction and committed customer orders before the first mass-produced rack has shipped. It's also why the AI developer tools and AI hardware investment market globally treats Etched as a genuine signal, not a side bet.
The three Harvard dropouts who ran chip-design software off a garage server, rebooted via a spouse's walk across the house, and slept on floors in half-furnished rooms — they might actually pull it off.
Etched's $1 Billion Inference Bet covers how this same team forced Nvidia to reconsider its inference roadmap long before this Series C closed.