"The reality is the entire semiconductors and data center industry is built on buffer... everything is built to be general purpose for everything not just in the data center but IoT on the edge. And when you have a specific use case you're really trying to design for you can change the constraints a lot." - Rob Wachen [00:02:26]
"If we want way more flops because we want to run at way higher throughputs, we fundamentally need to solve the thermal problem before we even think about adding flops to the chip." - Rob Wachen [00:09:08]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"You kind of have to be sick in the head to join our company... you're going to convince your family to move to San Jose... for the semiconductor company run by two 24 year olds now, going against the biggest companies in the world... with a design that they're saying is not going to be like 10% better but it's going to be 10x better." - Rob Wachen [00:32:15]
"We know inference is going to be the biggest market in the world. Whoever produces the most tokens is going to be the most valuable company in the world." - Rob Wachen [00:54:12]
"I think this is the second to last year where a majority of the workforce is going to be human. I think in 2027 you're going to see there's going to be more agents doing knowledge work than humans." - Rob Wachen [01:25:05]
"We are on a global march of inference becoming majority of global GDP... right now we measure productivity as a society as GDP per capita, but really it's going to look much more like agents per megawatt or maybe agents per gigawatt." - Rob Wachen [01:24:40]
Speakers & Credentials
Patrick O'Shaughnessy (Host): Investor and host of Invest Like The Best. Early backer of Etched.
Rob Wachen (Guest): Co-Founder and CEO of Etched. Harvard dropout who originally ran an incubator (ProR / Pear VC) backing AI native tools like Cursor, realizing that global compute COGS for inference would scale unsustainably without custom hardware.
Gavin Uberti (Guest): Co-Founder and CTO of Etched. Harvard dropout and former kernel developer for Xnor (acquired by Apple for $200M) and OctoML (acquired by Nvidia). World FTC robotics champion in high school.
1. Executive Summary
The prevailing semiconductor architecture relies on generalized "buffer" designs accommodating edge, IoT, and traditional datacenter workloads, rendering legacy GPUs intrinsically unsuited for optimal, large-scale AI inference.
Etched is bypassing traditional scaling limits by decoupling prefill (compute-bound) and decode (memory-bound) phases, prioritizing low-voltage architectures that bypass thermal throttling to maximize Model Flops Utilization (MFU).
To solve the memory bandwidth bottleneck, Etched bypasses the standard 4,000-nanosecond GPU-to-GPU latency limitation by developing a proprietary cluster-scale interconnect stack that cuts latency by >5x, pooling memory across massive arrays of chips.
The startup achieves unprecedented operational velocity through extreme schedule parallelization ("pre-fetching"), building custom networking racks, thermal cooling plates, and software stacks using 700+ FPGAs before the first physical silicon arrives.
The founders operate under the macro-thesis that AI inference will eventually eclipse traditional GDP metrics, requiring a transition to "agents per megawatt" measurements as multi-agent frameworks replace human workforces by 2027.
In contrast to hyperscaler internal projects (Google TPU, Meta MTIA, Microsoft Maia) where failure is not an existential threat to core businesses, Etched leverages extreme risk-taking and hyper-focused alignment to outpace trillion-dollar incumbents.
2. Chronological Table of Contents
[00:00:00] Introduction & Defying Industry Skepticism
[00:06:19] Redefining AI Chip Architecture: Prefill vs. Decode Disaggregation
[00:08:22] The MFU Challenge & Low-Voltage Physics
[00:14:01] Accelerating Wall Clock Time for AI Agents
[00:18:53] Founder Origins: Stage 4 Cancer, GPT-4, and Robotics
[00:26:11] Vertical Integration (No Vendor is the Best Vendor)
[00:28:30] Bimodal Recruiting: "Legends" vs "Chips on Shoulders"
[00:34:07] Engineering Under Fire: The Bangalore Sprint & Pre-Fetching
[00:41:19] Navigating the Megawatt Shortage and TSMC Ecosystem
[00:54:12] Hyperscaler Apathy & The Value of Existential Threat
[00:57:18] The 50-Picosecond Phase Alignment Crisis
[01:03:11] Raising the $100M Series A (Ramen-to-Chip Threshold)
[01:14:00] Wafer Sort Crisis: The "Puzzle Begins" Mindset
[01:19:57] Next-Gen Architectures: MOEs, Dynamism, and the Future Data Center
3. Detailed Thematic Summary
Theme 1: Subverting General-Purpose Architecture & Physics of Inference
The Buffer Tax: The incumbent semiconductor industry operates on broad buffers. EDA (Electronic Design Automation) tools by default sign off on timing configurations assuming chips will run in freezing (0°C) temperatures [00:03:16]. By realizing AI data centers never drop below 80°C, Etched can discard useless constraints to recapture massive efficiency margins.
PD Disaggregation: Etched partitions the inference workload into two stages: Prefill (reading text/loading the KV Cache) and Decode (generating tokens). Because prefill requires massive flop density and decode requires high memory bandwidth, processing them in isolation prevents hardware misalignment [00:06:47].
Solving the MFU (Model Flops Utilization) Floor: GPUs typically only harness 20-50% of their advertised peak flops during real workloads due to extreme thermal throttling [00:08:39]. Adding more raw compute to a standard GPU architecture yields diminishing returns because the chip will simply self-regulate and underclock to avoid melting.
Dennard Scaling & Voltage Economics: Physics dictates that power draw is quadratically proportional to voltage (doubling voltage 4x's power consumption) [00:09:28]. While traditional experts claimed running inference at sub-GPU voltages was impossible, Etched studied Bitcoin miners (which run at <25% of GPU voltage) [00:09:56] and developed a low-voltage power delivery platform operating at less than half the voltage of incumbent AI chips [00:12:28].
Breaking the Interconnect Latency Trap: On standard NVIDIA Blackwell architecture, data hopping from one chip to another incurs ~4,000 nanoseconds of point-to-point latency, meaning scaling an 8x cluster yields far less than 8x token performance [00:11:21]. Etched bypassed this by custom-building everything above the second layer of Ethernet, reducing latency by a factor of 5x, pooling SRAM and HBM into a unified "Cluster Scale Memory" [00:11:53].
Production as the Product: Etched is unique as an AI hardware startup in building not just the silicon, but the server racks, custom networking arrays, logic boards, and liquid cold plates concurrently [00:26:26]. This verticalization guarantees that hardware integration overhead doesn't delay time-to-market.
The "Pre-Fetching" Strategy (Schedule Parallelization): To eliminate post-silicon lag, Etched spent massive capital running over 700 FPGAs networked as a full-reticle emulation of their final chip [00:39:05]. They built server racks and provisioned networking architecture without the CPUs, constructed a mock thermal die mirroring the expected hotspot footprint to perfectly calibrate the cold plate dynamics, and wrote the full inference stack before the silicon was printed.
40 Days vs 10 Months: By parallelizing the schedule, Etched went from receiving back physical silicon to running live inference in a production rack in 40 days [00:39:41]. Conversely, a leading competitor previously reported a 10-month lag for the exact same milestone [00:39:31].
Kernels over Compilers: Instead of building a flexible graph compiler for broad software adoption (PyTorch/CUDA/ONNX compatibility), Etched forced developers to write bare-metal kernels. High Frequency Trading (HFT) engineers flocked to Etched because HFT natively relies on direct hardware kernel manipulation to minimize latency [00:53:05].
The 24/7 Global Sprint Paradigm: Facing a crippling one-year schedule delay from a third-party vendor, Etched deployed a dozen of its top engineers to live in Bangalore, India for six months, establishing a rolling 24-hour development cycle with handoffs at 8 AM and 8 PM every single day to force the tape-out to completion [00:34:46].
Theme 3: Talent Topography & Existential Threats
"Legends" & "Chips on Shoulders": Etched utilizes project-based recruiting. To build their racks, they mapped NVIDIA's historic team structures and extracted Brian Lerer, the creator of Nvidia's HGX and DGX systems (responsible for tens of billions in quarterly revenue) [00:30:03]. They paired him with a raw, 22-year-old talent (Sanford) who designed a functioning thermal cold-plate rig from scratch in a single week [00:30:47].
The 50-Picosecond Impossible Problem: During FPGA validation, they discovered an analog logic defect causing back-pressuring faults across clock domains. Solving it required phase-aligning two distinct clock signals to a tolerance of 50 picoseconds (50 trillionths of a second), twice a billion times a second [00:58:06]. Despite senior engineers threatening to quit over the "impossible" task, the team leveraged forced drifting mechanisms to lock the phases over a highly stressful two-week sprint [00:59:19].
Hyperscaler Apathy: Etched poached a top architect from a major trillion-dollar hyperscaler. The architect's reason for leaving: internal projects like Google's TPU or Meta's MTIA are ultimately hobbies; the parent company will not die if the hardware fails [00:56:15]. Etched views existential risk as a vital mechanism for achieving elite output—without an overarching monopoly to fall back on, their flop density (FP8 x FP8) strictly outperforms internal incumbent projects [00:56:55].
Theme 4: The Economics of Tokens and Macro Intelligence Scaling
Interactivity vs. Concurrency: The initial AI hardware boom rewarded raw speed. The next era is bound by interactivity curves: how many concurrent users can a chip serve while maintaining an acceptable baseline speed inside a strict power limit (e.g., 100 megawatts) [00:45:07]. Etched's hardware offers a full order of magnitude increase in concurrency.
Agents per Megawatt as the New GDP: The founders argue that human productivity constraints mean tasks taking a year (e.g., Cursor building a browser in a week) will eventually take mere hours as high-speed inference allows parallelized agents to replace massive human workloads [00:47:56]. They project that by 2027, agents will outnumber human knowledge workers globally, fundamentally changing global economics [01:25:05].
The Dynamism Directive for MOEs: Future models will feature radically dynamic context utilization. Mixture of Expert (MOE) frameworks will allow for parameters and compute logic to be dynamically provisioned per token on the fly, sharing memory across massively distributed chip clusters, ultimately resulting in $100 billion mega-clusters dedicated entirely to inference factories [01:26:52].
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
Model Flop Utilization (MFU)
20% - 50%
Average proportion of advertised peak flops actually achieved by traditional GPUs during real workloads due to thermal throttling.
1. Prefill / Decode (PD) Disaggregation [00:06:47]
Application & Synthesis: Traditional AI chips handle the entire lifecycle of an inference request in a silo, processing both the ingestion of the prompt and the generation of the response. Etched separates these processes structurally because they require fundamentally different hardware optimizations. "Prefill" requires massive processing bandwidth (flops) to quickly encode context into the KV Cache, while "Decode" generates the sequence iteratively and is entirely bounded by memory bandwidth. By disaggregating these functions and shipping the loaded KV Cache over a high-speed custom network to a secondary chip optimized purely for generation, Etched avoids the tragic inefficiency of a hybrid, "master-of-none" architecture.
2. Dennard Scaling and Low-Voltage Economics [00:09:28]
Application & Synthesis: The semiconductor industry has long battled the physical law of Dennard Scaling, which shows power consumption is proportional to the square of operating voltage. As AI compute needs hyper-scale, incumbent GPUs merely pack more transistors onto dies, rapidly crashing into severe thermal limits that throttle down their clock speeds—achieving merely 20-50% MFU. Taking inspiration from the extreme operational frugality of Bitcoin miners, Etched applies low-voltage architectures to bypass thermal limitations entirely, shrinking power delivery overhead to a fraction of standard GPUs. This allows them to maximize flop utilization without catastrophic melting.
3. Schedule Pre-Fetching (Capital-Intensive Velocity) [00:38:26]
Application & Synthesis: Standard hardware startup wisdom is linear: design silicon, await tape-out, receive silicon, build testing boards, debug, write software, scale to rack. Etched uses "Pre-Fetching" to crush linearity by brute-forcing capital at the schedule. They constructed massive FPGA simulation farms to run full-reticle software emulation, assembled server racks devoid of active CPUs, and engineered custom cold plates optimized against simulated thermal footprints. By spending vastly more cash up-front to construct parallel dependencies, Etched compressed a typical 10-month post-silicon rack deployment cycle down to a blistering 40 days.
4. Bimodal Talent Architecture (Legends & Chips-on-Shoulders) [00:28:30]
Application & Synthesis: Building deep tech requires a fusion of domain mastery and naive aggression. Etched intentionally structured its engineering teams into a bimodal distribution. They use "project-based recruiting" to hunt down "Legends"—individuals who historically built the precise zero-to-one architecture for legacy giants (e.g., Nvidia's HGX team). They then pair these veterans—who provide the institutional map of where landmines reside—with "Chips-on-Shoulder" talent: hyper-competitive, deeply naive youth (often robotics champions) who don't know that specific tasks are "supposed" to take months. The result is a compounding operational velocity where the veteran prevents catastrophe, and the amateur redefines the timeline constraint.
6. Anecdotes
Stage 4 Cancer and the Image Diagnosis Epiphany [00:18:53]
Context: Rob recounts his high school battle with Stage 4 bone cancer, suffering a less than 30% survival rate. Years later, after GPT-4V (vision capability) released, Rob fed it the original photo of a bump on his back from years prior, and the AI immediately diagnosed a potential tumor requiring an MRI—a process that took his actual human doctors six agonizing months to conclude. Why it matters: This visceral reality underscored for Rob that intelligence had become artificially commoditized, and his realization immediately following—that he ran out of image credits—highlighted the urgent supply-chain constraint in inference compute, directly inspiring Etched.
The Bangalore 24/7 Red-Eye Sprint [00:34:07]
Context: Pre-tape-out, Etched realized an offshore vendor in Bangalore was severely behind, threatening a 12-month delay to the company's survival. Rather than fire the vendor (delaying a year) or wait (delaying a year), Gavin shipped a dozen elite core engineers directly to Bangalore for six months. They established a 12-hour dual-hemisphere shift with the remaining US team. Why it matters: It validates Etched's core philosophy of aggressive risk-taking and physical presence to crush latency. By standing over the shoulders of the offshore developers, they cut 12-hour email back-and-forths into instant, real-time architectural decisions.
The "Impossible" 50-Picosecond Clock Problem [00:57:18]
Context: During advanced FPGA emulation, Etched uncovered a fatal logic error where back-pressuring across clock domains caused catastrophic failures. The only physical solution was aligning two clock signals to a margin of 50 picoseconds. Senior engineers deemed this mathematically impossible and quit the company. Gavin’s remaining team decided to assume it was possible and applied a highly complex temporal drifting mechanism to intentionally misalign and walk the signals into perfect, permanent lockstep. Why it matters: This narrative serves as the ultimate litmus test for Etched's cultural filter—separating those constrained by legacy heuristics from those willing to aggressively rewrite physical logic.
The $100 Million Ramen-to-Chip Threshold [01:03:11]
Context: Etched found itself in a fatal financial trough in early 2024. Physical design agreements alone required $50M, and building the full required stack demanded an astronomical $100M Series A for two college dropouts with no tape-out. Silicon Valley universally rejected their 30-page memo. They mathematically reduced their budget to the raw floor, modeling a scenario where the founders took no salary, ate ramen, and took high-interest debt just to print the mask. Why it matters: It demonstrated a terrifying level of founder commitment, which organically created a snowball of smaller believers, eventually leading to a $103M soft-commit board meeting.
The TSMC Speaker Dinner Encounter [01:08:35]
Context: Before raising the hundred million, Gavin (22 years old) was the lone young speaker at a prestigious semiconductor summit. Seated next to an elite TSMC Vice President at dinner, the two began mapping out advanced tensor math and low-voltage logic on a scrap of paper. By the next morning, TSMC emailed Gavin stating they wanted to officially partner with Etched. Why it matters: TSMC is famously conservative, but this encounter proves how deep technical fluency bypasses traditional corporate bureaucracy, leading to TSMC granting Etched highly favorable emulator financing before the startup was financially solvent.
7. References & Recommendations
Companies & Institutions
TSMC (Taiwan Semiconductor Manufacturing Company): [01:08:35] Mentioned as the most critical supply chain partner for Etched. Gavin highlights their unparalleled customer service and willingness to run highly localized yield experiments on their own dime.
Nvidia: [00:30:03] The trillion-dollar incumbent. Referenced continually as the legacy giant burdened by generalized "buffer" designs, and the source from which Etched poached elite architectural talent (e.g., HGX creator Brian Lerer).
Xnor (Acquired by Apple): [00:22:47] Gavin's first employer at age 17 where he operated under a custom NDA to write highly optimized software kernels, forging his fundamental understanding of data movement.
OctoML (Acquired by Nvidia): [00:23:12] Another past employer of Gavin's where he specialized in kernel development.
Colossus: [00:15:54] Mentioned as the standard for massive AI training clusters (over 100,000 GPUs) to demonstrate the sheer power and scale needed in modern AI data centers.
Cursor: [00:21:29] An AI native coding environment. Mentioned by Rob as an incubation success from his ProR (Pear VC) days, and as an example of agents executing long-horizon tasks (building a browser in a week).
Anysphere: [00:21:29] Mentioned alongside Cursor as an early-stage AI company navigating the same macro compute explosion.
Cerebras: [00:43:42] Referenced by Patrick O'Shaughnessy regarding physical space constraints in data centers, validating Etched's focus on megawatt footprint density.
Technology & Silicon Concepts
KV Cache (Key-Value Cache): [00:07:01] The system memory framework inside LLMs that stores historical prompt data during prefill to be used efficiently during iterative decode generation.
Blackwell (B300 / Y) & Rubin: [00:11:21] / [00:43:36] Nvidia's next-generation GPU architectures. Etched contrasts its custom point-to-point interconnect latency against Blackwell, and notes they utilize a different supply chain node than Rubin to avoid direct bottlenecks.
HBM (High Bandwidth Memory) & SRAM (Static Random-Access Memory): [00:11:01] Core physical memory paradigms. Etched is pooling these resources across large arrays of chips to create a unified "Cluster Scale Memory" without latency penalties.
FPGAs (Field Programmable Gate Arrays): [00:39:05] Reprogrammable silicon used by Etched to emulate their final ASIC chip geometry to write and validate software prior to final TSMC physical tape-out.
Codex: [00:52:01] OpenAI's code model, highlighted by the Etched team who successfully used it to run an operating system from scratch based solely on hardware docs.
MOE (Mixture of Experts): [01:26:19] An LLM architecture where only a sub-fraction of a model's total parameters are activated per token, making dynamic routing on the hardware level essential for the next generation of inference.
Hyperscaler Custom Silicon (TPU, MTIA, Maia, Jalapeno): [00:56:15] The bespoke silicon attempts by Google, Meta, Microsoft, and OpenAI. Mentioned by Etched to point out that these internal ventures suffer from poor flop density because the parent companies lack an existential threat for failure.
People
Brian Lerer: [00:30:03] Creator of Nvidia's HGX and DGX ecosystems. Hunted down and hired by Etched out of near-retirement to bring elite, seasoned "Legend" experience to their rack architecture execution.
Sanford: [00:24:05] Gavin’s high school robotics partner who helped win global FTC championships. Recruited to Etched to build a functioning thermal logic cold-plate system from scratch in a single week.
Mark Ross: [00:04:14] Former CTO at Cypress Semiconductor. Initially skeptical of the founders, he challenged them to write a white paper and build functional simulations. He ultimately led early investments and joined as a full-time CTO.
Noam Brown: [01:23:25] AI researcher referenced for predicting that models will run increasingly long-horizon tasks, shifting evaluations from raw intelligence to massive multi-agent parallelized execution.
Theories, Concepts & Events
Unit Disc Conjecture: [00:46:02] A complex mathematical proposition solved by AI models, cited by Gavin as proof that accelerating wall-clock time in inference will radically compress the horizon of new scientific and mathematical breakthroughs.
Jul 16, 2026
Dr. Robert Wachter | A Giant Leap: How AI Is Transforming Healthcare... | 14 Jul 2026 | Talks at Google
"don't get me wrong US healthcare delivers miracles every day particularly when it comes to cutting edge and intensive care... but the health care system itself is a headache wrapped in red tape inside the nightmare that France Kofka himse…