.png)
Today, Etched has launched, announcing $800M in cumulative funding. The progress since Primary led the seed round less than 40 months ago is remarkable; we have invested in every round and refreshed our underwriting each time.
Remarkably, the thesis has not changed: first, inference would be the central system through which AI turns model capability into revenue, and therefore be one the largest markets in the AI revolution; second, serving frontier models at scale would require frontier inference clusters with extraordinary speed, tokens-per-watt, and ROI.
But the company has evolved constantly. From a founding team in their dorm rooms at Harvard with no money, team, partners, customers, or technology, to a 400+ person talent juggernaut, the best partners you can get in the semiconductor industry, over $1B in orders, and customers testing rack scale systems. This is skeptic-defying magic at scale.
Our conviction has grown as the thesis has become reality. Today, we celebrate Etched’s A0 tapeout for frontier inference clusters that deliver uncanny performance, and we are excited for the world to see what is coming.
Brian has led these investments and sits on the Board at Etched; he wrote a reflection on the past 40 months of watching the team, especially the founders, execute. Check it out on X here.
What follows is our view on their accomplishments and vision, through the lens of our investment memos for the company.
Seed, March 2023
AI Phase: Chat
Estimated monthly tokens: 2 trillion
NVIDIA’s data center run rate: ~$15B
When we led the seed round in Etched, we had never made an investment in semis, although we had been thinking deeply about Nvidia, AI, and the looming GPU bottleneck. We wrote this post the same month we met the Etched team in which we pondered, “What’s next?... Fit-for-purpose hardware built with LLMs in mind.” This led to Brian being quoted in the WSJ, which was serendipitously published the day we met Gavin: “There’s new openings for attack and opportunity for those players because the types of chips that are going to most efficiently run these algorithms are different from a lot of what’s already out there.” The problem was becoming clear to us, but we had not met the team to usher us into this new era.
While unconventional to the max for the semiconductor industry, Gavin seemed special. We commissioned experts from ASIC startups and scaled incumbents to read Gavin’s white paper and spend time with him on the phone. Every time, they started off as non-believers, borderline indignant, and left curious, impressed, and rooting for Gavin. (We have now seen this same pattern play out again and again for 40 months.) Mark Ross, former CTO of Cypress Semiconductor who would later become the CTO of Etched, helped us understand the strategy and Gavin’s hunger to learn. He retold his experience of reading the first draft of the white paper, and his feedback to Gavin, which was, “People buy systems, not chips. You need to build the whole thing.” Within days, Gavin had worked out a new overview for the Etched strategy.
We came to be believers in his raw horse power. In our memo, we quoted Gavin’s manager at OctoML: “I’ve had exposure to a lot of incredible engineers, including being Mark Zuckerberg’s professor at Harvard, a close friendship with the Head of Research at OpenAI, and a decade at Google as a Principal engineer. Gavin is the best software engineer I’ve ever met.”
Seed investing comes down to the scale of the market and the talent of the founders, and we decided to make a bet on Etched. The first thing to derisk: can Etched build a world class team?
Series A, June 2024
AI Phase: Reasoning
Estimated monthly tokens: 40 trillion
NVIDIA’s data center run rate: $90B
By the time Etched raised its Series A, as we said in our memo, “We believe AI compute is the great market opportunity of our generation.” And Etched was on a tear.
- The critical technical team members were in place. Mark Ross had pivoted from advisor to CTO. Saptadeep Pal joined as Chief Architect. Ajat Hukkoo joined as VP of hardware.
- Early simulations showed better than expected performance
- The team had secured capacity at TSMC on their 4nm node
- Etched co-founder and President, Rob Wachen, had built a commercial machine and closed tens of millions in purchase orders
- Rob’s deal-making superpower had emerged around raising capital as well, with over $100M brought into the business from strategic investors
It was clear the team and commercial motion, enabled by a superior technology in an exploding market, were coming together. There was work ahead to get the system up and running.
Demand was accelerating as usage kept growing and reasoning models increased the token per task count exponentially. The compute crunch was becoming more real each day. Multi-silicon strategies for hyper scalers and the labs were starting to get serious. And more startups focused on inference were emerging.
The Etched thesis, at a high level, was clearly right. But could they deliver? Semis is notoriously unforgiving. Supply chain challenges, software complexity, sharp-elbowed competition, failed tapeouts, and more will routinely set companies back years, not weeks or quarters.
We knew that risks were everywhere, but our conviction in the team had only grown. Notably, Rob was emerging as a generational leader, capable of seeing multiple steps ahead to technical, commercial, and cultural challenges, and he had the wherewithal to call it out, develop plans, galvanize talent, and plow through any setback. He was an incredible complement to Gavin’s deep technical nature, giving Etched the necessary technical and commercial depth to solve their next set of challenges.
The inference market was clearly at a tipping point. The potential for what Etched could be was growing in front of us. But the challenges were abundant, the competitors were growing, and the skeptics were not appeased. The next thing to de-risk became product superiority and rapid commercial growth.
Series B, December 2025
AI Phase: Agents
Estimated monthly tokens: 3 quadrillion
NVIDIA’s data center run rate: $205B
By now, the centrality of the inference market had gone mainstream. Jensen, who had for years scoffed at the idea of anything beyond GPUs and their incredible flexibility with CUDA, was talking about inference constantly, saying things like “compute is revenue,” and buying Groq IP and talent to improve their inference offering. Hardware providers were being stacked up on the Pareto curve, measuring throughput versus interactivity, with the goal to be to push the curve up and to the right across all workloads. To meet the demands of inference, many looked to disaggregation, watching companies choose to win at either pre-fill or decode, then needing to partner, or sell, to complete the stack.
Etched remained contrarian to the market in their technical bet: disaggregation was not necessary if the system was built for AI inference from the start.
Etched made two critical decisions that would architect their system to be massively performant across both prefill and decode: Low Voltage Inference (LVI) enables an order of magnitude throughput in the same energy footprint as existing hardware while their Cluster Scale Memory (CSM) is a novel memory subsystem that allows for extremely fast latencies for users. You can build one system to solve for the multitude of demands from inference if you build that system – chip to rack – from the bottom up.
This team is ready for the road ahead. Etched’s leadership expanded materially beyond architecture as challenges moved away from chip design into systems and software. David Munday joined to lead software, after having built Google’s TPU software stack from v1 through v5. Brian Loiler joined as VP of Platform after two decades at Nvidia scaling HGX and DGX systems. Wayne Cao joined as VP of Production after a career at Google and Apple as a critical component needed to meet demand.
Now, the bet is execution: can Etched scale their systems into the real world? The team is ramping their office in Taiwan, customers are testing deployments, and racks will be installed at scale this fall. Soon, the most powerful and important AI workloads will be running on Etched hardware. The bet is scaling, across every part of the system, all at once.
Game time!
Fast forward to today: Etched is building frontier inference clusters. They co-design chips, racks, software, and manufacturing methods to deliver rack-scale inference systems with unprecedented cost-efficiency and performance. Their chips are working after a successful A0 (first-try) silicon tapeout and they’ll be shipping their first-generation rack-scale product to customers this summer while kicking off production to fulfill $1B+ in customer contracts.
Etched and its market have in some ways grown up together, and we’ve been lucky enough to have a front row seat to this technological renaissance through the eyes of Etched. Etched is a company that has always surprised us on the upside, and we are confident that this trajectory will continue for many years to come.
Today, as Etched announces its fundraise and product, we celebrate all of these accomplishments, while looking ahead to the next challenges and milestones to come, more aware and in awe of them than ever.
Congratulations to the whole Etched team!