aipod / invest-like-the-best / 2026-06-30-etched-building-ai-hardware-t…
ILTB · Jun 30, 2026

Etched - Building AI Hardware to Make Inference Faster and Cheaper

🎙 Patrick O'Shaughnessy · Rob Wachen · Gavin Uberti
34 claims11 predictions34 facts6 notable quotes raw JSON ↗

Patrick O'Shaughnessy interviews Etched co-founders Gavin Uberti and Rob Wachen, who dropped out of Harvard in 2023 to build inference-specific AI hardware, taped out a working chip on their first attempt, and have raised $800 million against more than $1 billion of customer demand for their first rack-scale product. They explain their two core technical bets — low-voltage inference (running at under half the voltage of any other AI chip to escape thermal throttling) and cluster-scale memory (a fully custom interconnect that cuts chip-to-chip latency by more than 5x so a whole cluster's memory behaves as one pool). The conversation covers their company-building system: pairing industry legends like Nvidia rack veteran Brian Loyler with driven young engineers, extreme vertical integration ('the best vendor is no vendor'), and 'prefetching' the schedule (700+ FPGAs, chipless racks shipped early) that took them from silicon back to inference running in a rack in 40 days — plus near-death moments from a 50-picosecond clock-alignment crisis to scraping together a $103 million Series A after every major Valley investor passed. They close with predictions: inference becomes the biggest market in the world and eventually the majority of global GDP, all future AI chips go low-voltage, more agents than humans doing knowledge work by 2027, and inevitable trillion-dollar token-factory data centers.

SEGMENTS

2:21
Facing Down Industry Skepticism

Patrick recounts that the consensus was two 21-year-olds could not build a better AI chip; Rob argues the semiconductor industry is built on general-purpose 'buffer' that a focused design can exploit, and Gavin tells how skeptic Mark Ross demanded a white paper and simulation, then became Etched's full-time CTO.

7:43
Technical Bets: Low-Voltage Inference and Cluster-Scale Memory

The product is a full inference rack, not just a chip. They walk through prefill vs decode, why MFU and thermal throttling motivate running at under half the voltage of other AI chips, and a fully custom interconnect that cuts chip-to-chip latency by more than 5x so a cluster's SRAM and HBM behave as a single memory pool.

14:27
Why Inference Is the Bottleneck

Rob and Gavin argue faster decode compresses the wall-clock time of long-horizon tasks, that billions of concurrent users are impossible on today's infrastructure, and that token production needs iPhone-like economies of scale.

18:23
Founder Origin Stories

Rob describes surviving stage-4 bone cancer and the GPT-4V moment that convinced him inference infrastructure was the thing to build; Gavin describes kernels work at XNOR at 17 and winning world robotics championships with Sanford as a two-person team.

28:41
Talent: Legends Paired with Chips-on-Shoulders

Etched's bimodal talent philosophy: project-based recruiting to land the best person in the world for each hard problem (like Brian Loyler of Nvidia's HGX/DGX), paired with driven young engineers; the company's contrarian status self-selects truth-seeking hires.

33:59
Velocity, Risk, and Prefetching

Stories of extreme urgency: shipping a dozen engineers to Bangalore for six months to rescue a lagging vendor, 24-hour development handoffs, and 'prefetching' the schedule (700+ FPGAs, chipless racks shipped to customers, thermal mock chips) that produced a 40-day silicon-to-racked-inference bring-up.

40:55
Supply Chain, TSMC, and Power

Gavin argues TSMC's real moat is service and partnership; Rob frames power availability and time-to-power as the binding constraint and explains interactivity curves, with Etched claiming an order of magnitude more concurrency at a given speed; Gavin notes chip-to-chip latency has physical room to fall from 4,000ns toward 2-3ns.

49:33
Software: Kernels-First and Game Selection

Etched bet on kernels-first programming over graph compilers, targeting the under-100 models that matter with physics-level primitives; tools are designed for AI-model users first, high-frequency trading talent joined, Codex brought up GPT-OSS overnight, and Rob argues in-house hyperscaler chips fail because they aren't existential.

56:31
Near-Death Moments: the Clock Crisis and the $103M Series A

The hardest technical episode — aligning two on-chip clocks to within 50 picoseconds after the chip came back producing wrong results, solved in two dark weeks — and the hardest financial one: every major Valley investor passing before a snowball of soft commits reached $103 million; Patrick reflects on contrarian investing and damning the base rate.

1:16:37
The Future: Model-Hardware Co-Evolution and Token Factories

Gavin argues machines don't think like people (math is cheap, memory dear) so models should be designed for enormous compute; Rob describes dynamism in MOE models, forecasts productivity measured as agents per megawatt with more agents than human knowledge workers by 2027, and both predict trillion-dollar token-factory data centers; Rob closes with the kindest-thing story from his cancer treatment.

NOTABLE QUOTES

“You kind of have to be sick in the head to join our company.”
Rob Wachen · 33:08 — Captures the self-selecting talent filter of a pre-product semiconductor startup run by 24-year-olds taking on the biggest companies in the world.
“if you have the most performant product and you can't produce it, then you're just a podcast”
Rob Wachen · 1:15:31 — Distills the production-is-the-product thesis — performance without supply-chain-scale availability is worthless.
“We're still in the very early days where the general purpose system, relatively small one, is kind of hand crafting these tokens, like they made screws back in the Renaissance.”
Gavin Uberti · 16:59 — The pre-industrial analogy for today's token economics — the argument for industrial economies of scale in AI inference.
“The best ability is availability.”
Rob Wachen · 1:13:36 — The lesson of Etched's past year: demand is so intense that having chips available today beats having marginally better chips later.
“He leads back and he's like, the puzzle begins.”
Rob Wachen · 1:11:51 — The silicon-validation legend's reaction to an all-red first wafer at 2-3 AM — the find-a-way mentality the founders say defines the company.
“we measure productivity as a society as GDP per capita. But really, it's going to look much more like agents per megawatt”
Rob Wachen · 1:21:27 — Reframes national productivity for an AI-agent economy — the episode's most quotable macro prediction.

PREDICTIONS ON THE RECORD

Dated and falsifiable — full record at /api/episode/<id>.

Rob Wachen · within roughly a decade · 20:53
Inference will become the biggest market in the world.
“it feels like we are on a decade march for inference to become the biggest market in the world”
Rob Wachen · unspecified (future chip generations) · 9:38
All AI chips in the future will be low-voltage chips, cramming far more flops into the same silicon without thermal throttling.
“we think all AI chips in the future are going to be low voltage chips”
Rob Wachen · a few years · 14:49
Within a few years, giant models will serve billions of users, with inference clusters growing from 8 chips to tens of thousands of networked chips.
“A few years from now, there's going to be giant models serving billions of users.”
Gavin Uberti · unspecified · 50:02
Coding models will eventually do all kernel generation and become superhuman at it.
“when the models keep getting smarter, it will eventually do all of it. It will become superhuman”
Rob Wachen · unspecified · 53:41
Whoever produces the most tokens will be the most valuable company in the world.
“Whoever produces the most tokens is going to be the most valuable company in the world.”
Rob Wachen · by 2027 · 1:21:27
In 2027 there will be more agents doing knowledge work than humans — making 2026 the second-to-last year a majority of the workforce is human.
“I think in 2027, you're going to see there's going to be more agents doing knowledge work than humans.”
Rob Wachen · 10+ years · 1:21:27
Inference will become the majority of global GDP, with national productivity measured as agents per megawatt.
“I firmly believe we are on a global march of inference becoming majority of global GDP.”
Gavin Uberti · unspecified (a matter of time) · 1:23:41
There will be a trillion-dollar individual data center — a giant mega token factory serving one or a handful of models for a massive number of users.
“Absolutely. It is a matter of time.”
Rob Wachen · soon · 46:36
Feats like Cursor's coding agents building an entire browser in a week will soon take under an hour, done by massive parallel agent fleets.
“that will soon happen in under an hour”
Gavin Uberti · unspecified · 1:16:59
Models will reach the point of attending to a billion tokens of context — every book ever written in short-term memory.
“I would love to be able to talk to a machine that was able to go attend to every book ever written in short-term memory. I think you're going to get to a point where you can.”
Gavin Uberti · 20-30 years · 48:33
Chip voltages will go much, much lower than today's low-voltage designs — inevitable on a 20-30 year horizon.
“You could go much, much lower. It's very challenging, but when I think about 20, 30 years in the future, then I think it's inevitable.”

WHO CLAIMED WHAT

Every position taken in the episode, attributed to its holder, with conviction (0–1) and the verbatim receipt.

Rob Wachen — 20 claims

assertionconviction 0.854:21
The entire semiconductor and data-center industry is built on general-purpose 'buffer', so designing for a specific use case lets you change constraints and compound many small wins into a radically better inference system.
“the entire semiconductors and data center industry is built on buffer”
assertionconviction 0.859:38
Adding more flops to today's GPUs yields no more performance because the chips thermal-throttle; the thermal problem must be solved first, which is why Etched runs at much lower voltage.
“If I just add more flops to a GPU today or another AI chip, I'm not actually going to get more performance because it's just going to thermal throttle.”
assertionconviction 0.814:49
Faster decode compresses the wall-clock time of long-horizon AI tasks — a year of inference-time compute into a month — which accelerates scientific innovation and technology proliferation.
“if you have way faster decode speed, you can compress that into a month”
assertionconviction 0.8514:49
It is not possible today for a billion people to use AI models concurrently; serving global scale requires multiple orders of magnitude more infrastructure across the whole stack.
“today it's just not possible for a billion people to use these models concurrently”
assertionconviction 0.8520:53
Every GPU, TPU, and AI chip serving today's models was fundamentally designed before ChatGPT and merely retrofit, so an entirely new wave of architectures is coming.
“Every GPU, every TPU, every AI chip that was serving these models were just fundamentally built before this and are retrofit to serve these modern models.”
assertionconviction 0.820:53
AI inverts software economics: the incremental-user cost of goods for every software company will no longer be near zero but a function of inference.
“The cost structure of every software company in the cogs is not going to be like zero anymore for an incremental user.”
opinionconviction 0.8527:27
Extreme vertical integration — 'the best vendor is no vendor' — yields both more performance and faster time to market.
“it's also the best vendor is no vendor”
assertionconviction 0.727:27
Etched is the only startup currently building both its own rack and its own chips, and it did both at the same time.
“I think we're the only startup right now that's building its own rack as well as its own chips.”
opinionconviction 0.8529:09
For the hardest technical problems, the difference between the number-one person in the world and the number ten or hundred determines whether the problem is solvable at all.
“the number one guy in the world versus the number 10 guy versus the number 100 guy, huge difference in whether it's actually possible to solve the problem”
opinionconviction 0.8531:34
The winning talent model pairs industry legends who know what good looks like with extremely driven young engineers who take aggressive risks — neither works without the other.
“you need the extremely driven people that just keep asking why and don't know where the bodies are buried to like take tons of aggressive risks. And then you need the people who've seen scale and still have the startup scrappy mentality”
opinionconviction 0.8538:34
Willingness to spend money should be extremely high when the ROI is clear, because with over a billion dollars of daily revenue in the category, every day not shipping forfeits enormous opportunity.
“So your willingness to spend money should be extremely high if you can get a very clear ROI out of it.”
assertionconviction 0.842:51
Power availability and time-to-power are the binding constraints on AI buildout, so the industry must extract far more tokens per megawatt.
“One of the things that matters a ton is power availability and time to power.”
assertionconviction 0.844:06
At a given level of interactivity, Etched's hardware delivers roughly an order of magnitude more concurrency, which translates directly into tokens per watt and tokens per dollar.
“our hardware is going to generally be able to get you an order of magnitude, more concurrency at a given level of interactivity”
opinionconviction 0.8551:31
The right game selection was to skip building an arbitrary graph compiler and instead bet that under 100 mathematically similar models would matter, accelerating those with physics-level primitives.
“we envisioned a world where there was going to be under 100 models that actually mattered”
opinionconviction 0.8555:43
In-house hyperscaler and lab chips (TPU, MTIA, Maia, Jalapeño) are not existential to their companies, which is why the best chip in the world is built by a company that only builds that chip.
“It is completely unsurprising that the best chip in the world is built by a company that only builds that chip. It's Nvidia, right?”
assertionconviction 0.81:18:31
Hardware that accelerates dynamism — per-token, per-user adaptive compute and memory in MOE-era models — is extremely important, and pre-dynamism hardware carries large overheads doing it.
“The hardware that really accelerates these types of very dynamic computations is extremely important”
assertionconviction 0.851:24:56
Because producing intelligence is dramatically cheaper than its value, we are in a many-year, probably many-decade supply shortage of tokens, and anything that produces tokens will be extremely valuable.
“we are in a many year, probably many decade supply shortage of these tokens”
assertionconviction 0.851:13:36
The next-generation design priority is simplicity and producibility at gigawatt scale, because availability decides deployment — the best ability is availability.
“We need to be able to be building a product that is producible at gigawatts per month from the limit.”
assertionconviction 0.91:13:36
Running a majority of the world's tokens requires exactly three things: the most flops in a given power budget, the lowest chip-to-chip latency (biggest scale-up domain), and maximum production volume.
“if you want to run a majority of the tokens in the world, you need to do three things. You need to build a chip with the most flops and a given power budget. You need to build a chip that has the lowest latency between other chips, so the biggest scale-up domain possible, and you need to produce as much of it as possible”
assertionconviction 0.81:15:31
AI chip supply is not zero-sum: because Etched built gen-1 on a different supply chain than Nvidia's Rubin, customer deployments are additive — two gigawatts rather than a choice between one or the other.
“So it actually is not a zero-something. It's a positive something.”

Gavin Uberti — 12 claims

opinionconviction 0.8512:29
The right decode metric is the memory bandwidth of the full scale-up cluster, not of a single chip.
“People often ask how much memory bandwidth is on your chip. You should be asking how much memory bandwidth is on your full scale up cluster.”
opinionconviction 0.816:59
Token production should reach iPhone-like economies of scale, where more money does not buy better tokens and everyone can access the best models.
“I wanted to live in a world where you have the same economies of scale for token making that you do for making, say, iPhones or cars”
assertionconviction 0.8524:26
In kernels engineering the math is relatively easy; data movement within and across chips is what actually determines high-speed decode.
“the math is relatively easy, but to get high-speed decode, the thing that matters is data movement”
opinionconviction 0.8526:53
You win by shipping, and a relentlessly focused team can build the best product in the world with far fewer people than the incumbents' tens of thousands.
“You can do the best product in the world with far fewer people.”
opinionconviction 0.632:47
Etched's contrarian status self-selects deeply diligent, truth-seeking hires, and announcing more of the product and specs risks losing that filter.
“it's one of the things I worry about as we announce more and more of the product and the specs. We may lose some of this if we're not very careful”
opinionconviction 0.8537:01
Fast, delegated decision-making — being right most of the time immediately — beats waiting for the perfect answer every time; speed wins.
“I would have much, much rather be right most of the time and give an answer immediately than wait every time for the perfect response”
opinionconviction 0.941:42
TSMC's real value is its service and collaborative partnership, not just its best-in-world technology — that is why it is number one and why it will win.
“the real value is all in the service. TSMC customer service is way, way better than I have seen at any other company in any other industry”
assertionconviction 0.850:02
Etched designs its profiling and debugging tools for AI-model users first and human users second, because models will do more and more of the kernel work.
“We think about it from how will the model use these tools? More than we think about how will humans use these tools?”
assertionconviction 0.851:16:59
Because chips invert the brain's cost structure — memory expensive, math cheap — next-generation models should be designed to use enormous compute: many copies, huge or numerous experts, giant context.
“for people storing data and loading memory is very cheap for neurons, and doing math is relatively expensive. And it is the exact opposite for chips”
assertionconviction 0.81:16:59
Math gets cheaper at a faster rate than memory gets cheaper, and there is a fundamental limit on any kind of DRAM device.
“math gets cheaper at a rate that is faster than memory gets cheaper”
opinionconviction 0.854:44
The natural business boundaries follow economies of scale: chip fabrication at the bottom and the model layer at the top have huge scale economies, so Etched fills the entire gap between them.
“We think the natural boundaries are on the chip side, on the bottom, and the model layer at the top. We'll fill the whole gap between.”
opinionconviction 0.81:25:39
Winning systems must get cheaper per token as they scale — serving 10x more tokens should not cost 10x more, thanks to economies of scale like cluster-scale memory.
“It must be some solution where if I want to go serve 10 times more tokens, then I get some economies of scale benefit”

Patrick O'Shaughnessy — 2 claims

opinionconviction 0.751:08:20
Outlier venture bets require ignoring base rates — if you invested on base rates you should do something else entirely, like an index fund.
“you kind of have to damn the base rate”
opinionconviction 0.81:11:21
A team's accumulated evidence of solving seemingly impossible problems is near-binary — companies either have it or they don't — and is one of the most interesting assets a company can have.
“the accumulated evidence of your guys' and your team's ability to solve seemingly impossible problems, is one of the most interesting things a company can have”

QUERY THIS EPISODE

agent access
# full structured record
curl 'https://aipod-two.vercel.app/api/episode/invest-like-the-best-2026-06-30-etched-building-ai-hardware-to-make-inference-faster-and-che'

# or over MCP
aipod_episode { "id": "invest-like-the-best-2026-06-30-etched-building-ai-hardware-to-make-inference-faster-and-che" }