ExpertsFMEtched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480]
Invest Like the Best with Patrick O'Shaughnessy · June 30, 2026
5 Selected Lessons:
case · 2 framework · 2 reasoning_pattern · 1
Prefetching: parallelize everything that doesn't require the bottleneck resource
Source
Invest Like the Best with Patrick O'Shaughnessy — Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480] (June 30, 2026) · at 33:14 ▸
When this applies
when a critical deliverable is on the critical path and many downstream workstreams nominally depend on it but could be de-risked in advance
The takeaway
Identify everything that can be done without the bottleneck resource and do all of it in parallel, using simulation/mocking to de-risk downstream dependencies early.
Why it works
By building mock/simulated versions of the critical dependency (FPGA emulation of the full chip, thermal mock chips, pre-built software stack) you can advance every downstream workstream to 'boring' readiness before the real dependency arrives, collapsing the integration time at the end.
One level up
Spending money aggressively to buy parallelism, when the ROI (time saved against daily market opportunity cost) is clear, is a rational bet, not recklessness.
Where it stops holding
This stops working when the ROI of speed isn't high enough to justify the spend, or when the mocked/simulated version diverges too much from reality to meaningfully de-risk the real dependency.
Detail
procedure:
- Build FPGA emulation of the full chip and run models on it before silicon arrives
- Build thermal mock chips matching expected hot spots to validate cold plates early
- Build and ship racks to customer data centers without chips, with all networking/CPU/storage pre-installed
- Build the entire software stack and run it on the FPGA cluster ahead of time
- Set up the full production line before the chip lands
benchmarks: {"etched_time":"40 days from silicon to inference in a rack","competitor_time":"10 months, publicly announced to investors"}
rationale: Every day without shipping leaves over a billion dollars of daily revenue opportunity in the category on the table, justifying high willingness to spend to compress the schedule.
In their words
We were able to do it in forty days. And it's because by the time the chip came back, everything was boring. The software was already written. The rack was already there. The production line was already set up.
Vertical integration boundary rule: only where economies of scale exist or it's existential
Source
Invest Like the Best with Patrick O'Shaughnessy — Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480] (June 30, 2026) · at 49:26 ▸
When this applies
when a company must decide how far to vertically integrate across a value chain with many layers
The takeaway
Vertically integrate only the layers that are both necessary to your core goal and exhibit real economies of scale; leave everything else to specialized partners.
Why it works
Integrating a layer is worth it only if it's both absolutely required to achieve your core scale goal and has real economies of scale; layers without scale effects (e.g., a small metal part) are 'noise' better left to others.
One level up
Focus all integration decisions around a single north-star metric (here: tokens online) and test every potential integration against whether it serves that metric better than the alternative use of resources.
Where it stops holding
This stops working when a non-integrated layer becomes the binding constraint on your core goal, at which point you should integrate it even if it previously seemed like 'noise'.
Detail
procedure:
- Define the single metric that determines company value (tokens produced/online)
- For each stack layer, ask whether integrating it meaningfully increases that metric
- Integrate layers with real economies of scale and that are required to reach target scale (chip design, rack design, model-level primitives)
- Skip layers with no scale effect or where customers already solve it well (e.g., data centers)
rationale: Economies of scale exist in designing models and chip fabrication but not in, e.g., building a small metal rack part — so the natural boundary is 'chip on the bottom, model layer on top, fill the gap.'
In their words
I think this comes down to economies of scale again. At certain parts of the stack, there are huge economies of scale and others there aren't.
Evaluate hardware by concurrency at a given interactivity, not raw speed
Source
Invest Like the Best with Patrick O'Shaughnessy — Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480] (June 30, 2026) · at 40:21 ▸
When this applies
when evaluating any system where you must guarantee a minimum service-level quality (speed/latency) while serving as many users/jobs as possible
The takeaway
Evaluate a system's capacity by fixing the required service-level quality and measuring maximum concurrency at that level, rather than by peak performance alone.
Why it works
Raw speed alone is meaningless if it can't be delivered at scale; the real constraint is how many concurrent users/agents a fixed power/resource budget can serve while still meeting the required speed threshold.
One level up
The correct capacity metric for a resource-constrained service is throughput-at-guaranteed-quality, not the headline peak number.
Where it stops holding
This stops working when demand is low enough that concurrency never becomes binding, or when quality requirements vary so much across use cases that a single interactivity curve isn't meaningful.
Detail
approach: Fix the required interactivity (speed) for your product, then ask how many concurrent users/agents you can serve at that speed within a given power budget.
signals:
- published interactivity curves (e.g. Blackwell vs AMD)
- tokens per watt
- tokens per dollar
rationale: As the AI infra market matures past the 'just needs to be fast enough' phase, the differentiator becomes order-of-magnitude more concurrency at a given interactivity level, which is what Etched claims its hardware delivers.
In their words
our hardware is going to generally be able to get you an order of magnitude more concurrency at a given level of interactivity.
Personal crisis as motivational fuel for mission-driven founding
Source
Invest Like the Best with Patrick O'Shaughnessy — Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480] (June 30, 2026) · at 16:53 ▸
When this applies
when deciding what large, difficult mission to commit years of your life to
The takeaway
A founder's personal stake in the outcome of a technology shift can be a durable source of motivation through years of difficulty that purely rational market analysis cannot provide.
Why it works
A near-death experience reset Rob's sense of what mattered and created a concrete, visceral proof point (AI correctly diagnosing his old photo) that made an abstract technology trend personally urgent.
One level up
The strongest long-horizon commitments often come from merging a rational market thesis with a deeply personal reason to care.
Where it stops holding
This stops working when the personal motivation isn't paired with the technical/market judgment needed to execute, or when it hasn't been tested against the realities of a multi-year grind.
Detail
situation: Rob survived stage IV bone cancer as a teenager with under 30% survival odds, then years later tested GPT-4V on an old photo of his undiagnosed tumor and it correctly flagged it as a likely tumor.
decision: He concluded AI was about to change everything but the infrastructure to serve it at scale didn't exist, and chose to work on inference hardware rather than models.
outcome: This became a founding motivation for Etched, compounding with his co-founder's separate realization that inference costs would dominate every software company's economics.
lesson: A vivid personal encounter with a technology's life-changing potential, combined with a clear-eyed view of an infrastructure gap, can be the seed of a durable founding mission.
In their words
And I just kind of sat there still. It was like, that took me six months. And, you know, yesterday, this feature wasn't there. Today, it's here.
Raising capital via survival-mode networking and a soft-commit snowball
Source
Invest Like the Best with Patrick O'Shaughnessy — Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480] (June 30, 2026) · at 56:03 ▸ · at 59:39 ▸
When this applies
when a venture needs an order-of-magnitude more capital than peers have ever raised at that stage, and consensus investors have already said no
The takeaway
When conventional fundraising channels reject an unprecedented ask, switch to exhaustive network activation and accumulate smaller commitments into a critical mass rather than seeking one big check.
Why it works
Going into 'survival mode' — exhaustively contacting every possible connection, being radically transparent about the ask and risk, and accumulating many smaller checks ($1M, $2M, $5M, $10M) — builds momentum (a 'snowball') that eventually crosses the needed threshold even when no single investor will write the whole check.
One level up
An unprecedented capital need for an unprecedented bet requires an unprecedented, non-standard process to meet it — the same boldness as the product thesis itself.
Where it stops holding
This stops working when the underlying team/thesis lacks enough credibility to attract even small early commitments, so the snowball never starts.
Detail
situation: Etched needed ~$100M to build chips, boards, interconnects, and cold plates after realizing a chip alone wouldn't suffice, while their biggest semiconductor Series A comparable was only $40-50M and had only $15M on hand.
decision: They wrote an exhaustive 30-page technical memo, got rejected by every major Valley investor, then entered 'survival mode,' calling every possible connection and being fully transparent about the risk and ask.
outcome: Checks accumulated from $1M to $10M until a board meeting revealed ~$103M in soft commits; they took it, and have since raised nearly a half dozen more rounds, many from the same investors doubling/tripling down.
lesson: An aggressive, unprecedented capital need can be met not by finding one believer but by activating an entire network and letting small commitments snowball into critical mass.
In their words
We had a board meeting. I show you the spreadsheet and we look at it and it's like a 103,000,000. It's like, these are all like soft commits. And we all look at each other and we say, we're gonna take it.