{
 "episode_id": "invest-like-the-best-2026-06-30-etched-building-ai-hardware-to-make-inference-faster-and-che",
 "title": "Etched - Building AI Hardware to Make Inference Faster and Cheaper",
 "podcast": "invest-like-the-best",
 "published_at": "2026-06-30",
 "speakers": [
  "Patrick O'Shaughnessy",
  "Rob Wachen",
  "Gavin Uberti"
 ],
 "summary": "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.",
 "topics": [
  {
   "name": "Facing Down Industry Skepticism",
   "timestamp": "2:21",
   "summary": "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."
  },
  {
   "name": "Technical Bets: Low-Voltage Inference and Cluster-Scale Memory",
   "timestamp": "7:43",
   "summary": "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."
  },
  {
   "name": "Why Inference Is the Bottleneck",
   "timestamp": "14:27",
   "summary": "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."
  },
  {
   "name": "Founder Origin Stories",
   "timestamp": "18:23",
   "summary": "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."
  },
  {
   "name": "Talent: Legends Paired with Chips-on-Shoulders",
   "timestamp": "28:41",
   "summary": "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."
  },
  {
   "name": "Velocity, Risk, and Prefetching",
   "timestamp": "33:59",
   "summary": "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."
  },
  {
   "name": "Supply Chain, TSMC, and Power",
   "timestamp": "40:55",
   "summary": "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."
  },
  {
   "name": "Software: Kernels-First and Game Selection",
   "timestamp": "49:33",
   "summary": "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."
  },
  {
   "name": "Near-Death Moments: the Clock Crisis and the $103M Series A",
   "timestamp": "56:31",
   "summary": "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."
  },
  {
   "name": "The Future: Model-Hardware Co-Evolution and Token Factories",
   "timestamp": "1:16:37",
   "summary": "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."
  }
 ],
 "entities": [
  {
   "name": "Etched",
   "type": "company",
   "salience": "central",
   "role_in_episode": "The guests' AI inference hardware company — building a full rack (chip, boards, power delivery, interconnect, production) around low-voltage inference and cluster-scale memory."
  },
  {
   "name": "Gavin Uberti",
   "type": "person",
   "salience": "central",
   "role_in_episode": "Co-founder of Etched (Harvard dropout, ex-XNOR kernels engineer, world robotics champion); holder of the core technical claims on interconnect, kernels, and model-hardware co-evolution."
  },
  {
   "name": "Rob Wachen",
   "type": "person",
   "salience": "central",
   "role_in_episode": "Co-founder of Etched; stage-4 cancer survivor who ran the Prod incubator; holder of the market, production, and company-building claims and the boldest inference predictions."
  },
  {
   "name": "Patrick O'Shaughnessy",
   "type": "person",
   "salience": "central",
   "role_in_episode": "Host of Invest Like the Best and a multi-round Etched investor since 2023; discloses his bias and reflects on contrarian investing."
  },
  {
   "name": "Nvidia",
   "type": "company",
   "salience": "central",
   "role_in_episode": "The incumbent benchmark throughout — Blackwell latency, HGX/DGX heritage via Brian Loyler, Rubin supply chain, and the proof that the best chip is built by a company that only builds chips."
  },
  {
   "name": "TSMC",
   "type": "company",
   "salience": "discussed",
   "role_in_episode": "Etched's fab partner; Gavin argues its customer service, not just its technology, is why it is number one — it ran a yield experiment on its own dime and believed in Etched pre-Series A."
  },
  {
   "name": "Sohu",
   "type": "product",
   "salience": "discussed",
   "role_in_episode": "The name of Etched's first product — the complete inference rack whose design, bring-up, and production the episode centers on."
  },
  {
   "name": "Blackwell",
   "type": "product",
   "salience": "discussed",
   "role_in_episode": "Nvidia's GPU generation used as the comparison point for chip-to-chip latency (~4,000ns) and interactivity curves."
  },
  {
   "name": "Rubin",
   "type": "product",
   "salience": "mentioned",
   "role_in_episode": "Nvidia's next generation; Etched deliberately built gen-1 on a different supply chain (4nm vs Rubin's 3nm, different HBM)."
  },
  {
   "name": "Mark Ross",
   "type": "person",
   "salience": "discussed",
   "role_in_episode": "Former Cypress Semiconductor CTO who went from telling the founders it would not work to becoming Etched's full-time CTO after their simulation convinced him."
  },
  {
   "name": "Brian Loyler",
   "type": "person",
   "salience": "discussed",
   "role_in_episode": "Nvidia legend who built the HGX/DGX systems; recruited via project-based recruiting to lead Etched's rack effort."
  },
  {
   "name": "Sanford",
   "type": "person",
   "salience": "discussed",
   "role_in_episode": "Gavin's high-school world-robotics-champion partner who joined Etched and built a key cold-plate contraption in a week."
  },
  {
   "name": "OpenAI",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Maker of ChatGPT/GPT-4V (Rob's motivation moment) and of the announced in-house Jalapeño chip cited as non-existential."
  },
  {
   "name": "Google",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Cited for TPUs as the canonical non-existential in-house chip — its revenue comes from Search."
  },
  {
   "name": "Cursor",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Incubated at Rob's Prod incubator; its coding agents building a browser in a week is Rob's example of agentic work about to compress further."
  },
  {
   "name": "Prod",
   "type": "organization",
   "salience": "mentioned",
   "role_in_episode": "Rob Wachen's startup incubator whose alumni include Cursor, Mercor, and Etched itself."
  },
  {
   "name": "XNOR",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Gavin's first employer at 17 (kernels development); later bought by Apple for $200 million."
  },
  {
   "name": "Apple",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Acquired XNOR for $200 million; the iPhone is Gavin's model of economies of scale tokens should reach."
  },
  {
   "name": "Synopsys",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "EDA vendor that let pre-Series-A Etched pay for emulators over many years — an early supplier act of belief."
  },
  {
   "name": "Codex",
   "type": "product",
   "salience": "mentioned",
   "role_in_episode": "OpenAI's coding agent that got GPT-OSS running on Etched's stack from scratch overnight using only their docs."
  },
  {
   "name": "GPT-OSS",
   "type": "product",
   "salience": "mentioned",
   "role_in_episode": "The open model Codex brought up on Etched hardware completely by itself."
  },
  {
   "name": "GPT-4",
   "type": "product",
   "salience": "mentioned",
   "role_in_episode": "Its GPT-4V image feature instantly flagged Rob's pre-diagnosis tumor photo — the moment that convinced him inference infrastructure was the thing to build."
  },
  {
   "name": "Noam Brown",
   "type": "person",
   "salience": "mentioned",
   "role_in_episode": "His viral tweet about long-horizon task evaluation frames the case for faster inference; now an Etched angel."
  },
  {
   "name": "Jalapeño",
   "type": "product",
   "salience": "mentioned",
   "role_in_episode": "OpenAI's announced in-house AI chip (transcribed as 'Talapeño'/'Jalapeno'), cited as a non-existential side project."
  },
  {
   "name": "Harvard",
   "type": "organization",
   "salience": "mentioned",
   "role_in_episode": "The university both founders dropped out of to start Etched in 2023."
  },
  {
   "name": "Bangalore",
   "type": "place",
   "salience": "mentioned",
   "role_in_episode": "Where Etched shipped a dozen top engineers for six months to rescue a lagging vendor."
  },
  {
   "name": "Taiwan",
   "type": "place",
   "salience": "mentioned",
   "role_in_episode": "Home of Etched's factory (a few dozen people) and of TSMC, with 2-3 AM wafer-sort calls."
  },
  {
   "name": "Mercor",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Another company incubated through Rob's Prod incubator."
  },
  {
   "name": "Cypress Semiconductor",
   "type": "company",
   "salience": "mentioned",
   "role_in_episode": "Mark Ross's former company, which sold for nine billion dollars."
  },
  {
   "name": "inference",
   "type": "concept",
   "salience": "central",
   "role_in_episode": "The market Etched is built for — predicted to become the biggest market in the world and eventually the majority of global GDP."
  },
  {
   "name": "tokens",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The economic unit of AI; the episode's thesis is that whoever produces the most tokens wins, and token-making needs industrial economies of scale."
  },
  {
   "name": "low voltage inference",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Etched's first key technical bet — a new power-delivery mechanism running chips at under half the voltage of any other AI chip to escape thermal throttling."
  },
  {
   "name": "cluster scale memory",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Etched's second key bet — a fully custom interconnect that cuts chip-to-chip latency more than 5x so a whole scale-up cluster's memory acts as one pool."
  },
  {
   "name": "prefill-decode disaggregation",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The split of inference into prefill and decode clusters, transferring KV caches between them, which frames Etched's architecture."
  },
  {
   "name": "KV cache",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "The model's memory that prefill produces and decode consumes; transferred between clusters in PD disaggregation."
  },
  {
   "name": "vertical integration",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Etched's operating doctrine ('the best vendor is no vendor'), bounded below by chip fabrication and above by the model layer."
  },
  {
   "name": "prefetching",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Etched's schedule-parallelization doctrine: do everything possible before the chip lands (FPGAs, chipless racks, thermal mock chips, production line)."
  },
  {
   "name": "economies of scale",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Gavin's recurring frame — tokens should get cheaper with scale like iPhones, fabs, and steel plants; the boundary logic for what Etched builds itself."
  },
  {
   "name": "kernels",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Gavin's craft since XNOR; Etched's kernels-first software bet, with data movement (not math) as what matters, increasingly done by AI models."
  },
  {
   "name": "mixture of experts",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The model architecture (MOE) whose sparse-expert, distributed-brain shape motivates Etched's cluster-scale hardware and dynamism support."
  },
  {
   "name": "dynamism",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "The model trend of per-token adaptive compute and memory that Rob says pre-ChatGPT hardware handles with big overheads."
  },
  {
   "name": "GPUs",
   "type": "product",
   "salience": "discussed",
   "role_in_episode": "The incumbent hardware category: 20-50% MFU, thermal throttling, high voltages, slow chip-to-chip hops — the baseline Etched designs against."
  },
  {
   "name": "FPGAs",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "Validation platform — Etched put the full reticle chip on a 700+ FPGA cluster before silicon returned."
  },
  {
   "name": "AI models",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Discussed as co-evolving with hardware: under 100 models matter, MOE dynamism, billion-token context, and models designed around cheap math and expensive memory."
  },
  {
   "name": "AI chips",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The category under transformation — all designed before ChatGPT, predicted to all go low-voltage, with in-house hyperscaler efforts framed as non-existential."
  },
  {
   "name": "AI agents",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The demand driver — long-horizon agent teams whose wall-clock time faster inference compresses; predicted to outnumber human knowledge workers by 2027."
  },
  {
   "name": "AI adoption",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "Framed as early innings: only a few million paid users worldwide, one one-thousandth of the global population."
  },
  {
   "name": "knowledge work",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "The domain where Rob predicts agents outnumber humans by 2027."
  },
  {
   "name": "data centers",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The deployment unit — power-constrained today, predicted to reach trillion-dollar single-site token factories."
  },
  {
   "name": "power",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Rob's binding constraint: power availability and time-to-power, driving the race for tokens per megawatt."
  },
  {
   "name": "supply chain",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Discussed as partnership-driven (TSMC, memory vendors) and deliberately de-conflicted from Nvidia's (different node and HBM)."
  },
  {
   "name": "semiconductors",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "The industry whose general-purpose 'buffer,' capital intensity ($40-50M physical design, $40-50M record Series As), and validation tooling shifts frame Etched's opportunity."
  },
  {
   "name": "high-frequency trading",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "The only early believers in kernels-first programming — dozens of HFT engineers joined Etched."
  },
  {
   "name": "recruiting",
   "type": "concept",
   "salience": "discussed",
   "role_in_episode": "Etched's project-based recruiting system: map the hardest problems ever solved, find who did the zero-to-one, and persist through 20 conversations."
  },
  {
   "name": "Bitcoin miners",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "Proof point that sub-GPU voltages are physically possible — they run at under a quarter of GPU voltage."
  },
  {
   "name": "investing",
   "type": "concept",
   "salience": "mentioned",
   "role_in_episode": "Patrick's reflections on contrarian bets, damning the base rate, and perception gaps as opportunity."
  }
 ],
 "claims": [
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "semiconductors",
    "inference",
    "Etched"
   ],
   "timestamp": "4:21",
   "quote": "the entire semiconductors and data center industry is built on buffer"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "GPUs",
    "low voltage inference"
   ],
   "timestamp": "9:38",
   "quote": "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."
  },
  {
   "holder": "Gavin Uberti",
   "claim": "The right decode metric is the memory bandwidth of the full scale-up cluster, not of a single chip.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "cluster scale memory",
    "GPUs"
   ],
   "timestamp": "12:29",
   "quote": "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."
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "inference",
    "AI agents"
   ],
   "timestamp": "14:49",
   "quote": "if you have way faster decode speed, you can compress that into a month"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "inference",
    "AI adoption"
   ],
   "timestamp": "14:49",
   "quote": "today it's just not possible for a billion people to use these models concurrently"
  },
  {
   "holder": "Gavin Uberti",
   "claim": "Token production should reach iPhone-like economies of scale, where more money does not buy better tokens and everyone can access the best models.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "tokens",
    "economies of scale"
   ],
   "timestamp": "16:59",
   "quote": "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"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "AI chips",
    "GPUs"
   ],
   "timestamp": "20:53",
   "quote": "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."
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "inference",
    "AI adoption"
   ],
   "timestamp": "20:53",
   "quote": "The cost structure of every software company in the cogs is not going to be like zero anymore for an incremental user."
  },
  {
   "holder": "Gavin Uberti",
   "claim": "In kernels engineering the math is relatively easy; data movement within and across chips is what actually determines high-speed decode.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "kernels",
    "cluster scale memory"
   ],
   "timestamp": "24:26",
   "quote": "the math is relatively easy, but to get high-speed decode, the thing that matters is data movement"
  },
  {
   "holder": "Gavin Uberti",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "Etched"
   ],
   "timestamp": "26:53",
   "quote": "You can do the best product in the world with far fewer people."
  },
  {
   "holder": "Rob Wachen",
   "claim": "Extreme vertical integration — 'the best vendor is no vendor' — yields both more performance and faster time to market.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "vertical integration",
    "Etched"
   ],
   "timestamp": "27:27",
   "quote": "it's also the best vendor is no vendor"
  },
  {
   "holder": "Rob Wachen",
   "claim": "Etched is the only startup currently building both its own rack and its own chips, and it did both at the same time.",
   "stance": "assertion",
   "conviction": 0.7,
   "subjects": [
    "Etched"
   ],
   "timestamp": "27:27",
   "quote": "I think we're the only startup right now that's building its own rack as well as its own chips."
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "recruiting"
   ],
   "timestamp": "29:09",
   "quote": "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"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "recruiting",
    "Etched"
   ],
   "timestamp": "31:34",
   "quote": "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"
  },
  {
   "holder": "Gavin Uberti",
   "claim": "Etched's contrarian status self-selects deeply diligent, truth-seeking hires, and announcing more of the product and specs risks losing that filter.",
   "stance": "opinion",
   "conviction": 0.6,
   "subjects": [
    "Etched",
    "recruiting"
   ],
   "timestamp": "32:47",
   "quote": "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"
  },
  {
   "holder": "Gavin Uberti",
   "claim": "Fast, delegated decision-making — being right most of the time immediately — beats waiting for the perfect answer every time; speed wins.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "Etched"
   ],
   "timestamp": "37:01",
   "quote": "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"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "Etched",
    "inference"
   ],
   "timestamp": "38:34",
   "quote": "So your willingness to spend money should be extremely high if you can get a very clear ROI out of it."
  },
  {
   "holder": "Gavin Uberti",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.9,
   "subjects": [
    "TSMC"
   ],
   "timestamp": "41:42",
   "quote": "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"
  },
  {
   "holder": "Rob Wachen",
   "claim": "Power availability and time-to-power are the binding constraints on AI buildout, so the industry must extract far more tokens per megawatt.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "power",
    "data centers"
   ],
   "timestamp": "42:51",
   "quote": "One of the things that matters a ton is power availability and time to power."
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "Etched",
    "inference"
   ],
   "timestamp": "44:06",
   "quote": "our hardware is going to generally be able to get you an order of magnitude, more concurrency at a given level of interactivity"
  },
  {
   "holder": "Gavin Uberti",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "Etched",
    "kernels",
    "AI models"
   ],
   "timestamp": "50:02",
   "quote": "We think about it from how will the model use these tools? More than we think about how will humans use these tools?"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "Etched",
    "AI models"
   ],
   "timestamp": "51:31",
   "quote": "we envisioned a world where there was going to be under 100 models that actually mattered"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.85,
   "subjects": [
    "Nvidia",
    "Google",
    "OpenAI",
    "Etched",
    "AI chips"
   ],
   "timestamp": "55:43",
   "quote": "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?"
  },
  {
   "holder": "Gavin Uberti",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "AI models",
    "AI chips"
   ],
   "timestamp": "1:16:59",
   "quote": "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"
  },
  {
   "holder": "Gavin Uberti",
   "claim": "Math gets cheaper at a faster rate than memory gets cheaper, and there is a fundamental limit on any kind of DRAM device.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "compute",
    "AI chips"
   ],
   "timestamp": "1:16:59",
   "quote": "math gets cheaper at a rate that is faster than memory gets cheaper"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "AI chips",
    "dynamism",
    "mixture of experts"
   ],
   "timestamp": "1:18:31",
   "quote": "The hardware that really accelerates these types of very dynamic computations is extremely important"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "tokens",
    "inference"
   ],
   "timestamp": "1:24:56",
   "quote": "we are in a many year, probably many decade supply shortage of these tokens"
  },
  {
   "holder": "Patrick O'Shaughnessy",
   "claim": "Outlier venture bets require ignoring base rates — if you invested on base rates you should do something else entirely, like an index fund.",
   "stance": "opinion",
   "conviction": 0.75,
   "subjects": [
    "investing"
   ],
   "timestamp": "1:08:20",
   "quote": "you kind of have to damn the base rate"
  },
  {
   "holder": "Patrick O'Shaughnessy",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "Etched",
    "investing"
   ],
   "timestamp": "1:11:21",
   "quote": "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"
  },
  {
   "holder": "Rob Wachen",
   "claim": "The next-generation design priority is simplicity and producibility at gigawatt scale, because availability decides deployment — the best ability is availability.",
   "stance": "assertion",
   "conviction": 0.85,
   "subjects": [
    "Etched",
    "tokens"
   ],
   "timestamp": "1:13:36",
   "quote": "We need to be able to be building a product that is producible at gigawatts per month from the limit."
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.9,
   "subjects": [
    "inference",
    "AI chips",
    "Etched"
   ],
   "timestamp": "1:13:36",
   "quote": "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"
  },
  {
   "holder": "Rob Wachen",
   "claim": "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.",
   "stance": "assertion",
   "conviction": 0.8,
   "subjects": [
    "Etched",
    "Nvidia",
    "supply chain"
   ],
   "timestamp": "1:15:31",
   "quote": "So it actually is not a zero-something. It's a positive something."
  },
  {
   "holder": "Gavin Uberti",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "Etched",
    "vertical integration",
    "economies of scale"
   ],
   "timestamp": "54:44",
   "quote": "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."
  },
  {
   "holder": "Gavin Uberti",
   "claim": "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.",
   "stance": "opinion",
   "conviction": 0.8,
   "subjects": [
    "economies of scale",
    "cluster scale memory",
    "tokens"
   ],
   "timestamp": "1:25:39",
   "quote": "It must be some solution where if I want to go serve 10 times more tokens, then I get some economies of scale benefit"
  }
 ],
 "predictions": [
  {
   "holder": "Rob Wachen",
   "prediction": "Inference will become the biggest market in the world.",
   "timeframe": "within roughly a decade",
   "subjects": [
    "inference"
   ],
   "timestamp": "20:53",
   "quote": "it feels like we are on a decade march for inference to become the biggest market in the world"
  },
  {
   "holder": "Rob Wachen",
   "prediction": "All AI chips in the future will be low-voltage chips, cramming far more flops into the same silicon without thermal throttling.",
   "timeframe": "unspecified (future chip generations)",
   "subjects": [
    "AI chips",
    "low voltage inference"
   ],
   "timestamp": "9:38",
   "quote": "we think all AI chips in the future are going to be low voltage chips"
  },
  {
   "holder": "Rob Wachen",
   "prediction": "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.",
   "timeframe": "a few years",
   "subjects": [
    "AI models",
    "inference"
   ],
   "timestamp": "14:49",
   "quote": "A few years from now, there's going to be giant models serving billions of users."
  },
  {
   "holder": "Gavin Uberti",
   "prediction": "Coding models will eventually do all kernel generation and become superhuman at it.",
   "timeframe": "unspecified",
   "subjects": [
    "kernels",
    "AI models"
   ],
   "timestamp": "50:02",
   "quote": "when the models keep getting smarter, it will eventually do all of it. It will become superhuman"
  },
  {
   "holder": "Rob Wachen",
   "prediction": "Whoever produces the most tokens will be the most valuable company in the world.",
   "timeframe": "unspecified",
   "subjects": [
    "tokens",
    "inference"
   ],
   "timestamp": "53:41",
   "quote": "Whoever produces the most tokens is going to be the most valuable company in the world."
  },
  {
   "holder": "Rob Wachen",
   "prediction": "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.",
   "timeframe": "by 2027",
   "subjects": [
    "AI agents",
    "knowledge work"
   ],
   "timestamp": "1:21:27",
   "quote": "I think in 2027, you're going to see there's going to be more agents doing knowledge work than humans.",
   "due": "2027-12-31"
  },
  {
   "holder": "Rob Wachen",
   "prediction": "Inference will become the majority of global GDP, with national productivity measured as agents per megawatt.",
   "timeframe": "10+ years",
   "subjects": [
    "inference",
    "AI agents"
   ],
   "timestamp": "1:21:27",
   "quote": "I firmly believe we are on a global march of inference becoming majority of global GDP."
  },
  {
   "holder": "Gavin Uberti",
   "prediction": "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.",
   "timeframe": "unspecified (a matter of time)",
   "subjects": [
    "data centers",
    "tokens"
   ],
   "timestamp": "1:23:41",
   "quote": "Absolutely. It is a matter of time."
  },
  {
   "holder": "Rob Wachen",
   "prediction": "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.",
   "timeframe": "soon",
   "subjects": [
    "AI agents",
    "Cursor"
   ],
   "timestamp": "46:36",
   "quote": "that will soon happen in under an hour"
  },
  {
   "holder": "Gavin Uberti",
   "prediction": "Models will reach the point of attending to a billion tokens of context — every book ever written in short-term memory.",
   "timeframe": "unspecified",
   "subjects": [
    "AI models"
   ],
   "timestamp": "1:16:59",
   "quote": "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."
  },
  {
   "holder": "Gavin Uberti",
   "prediction": "Chip voltages will go much, much lower than today's low-voltage designs — inevitable on a 20-30 year horizon.",
   "timeframe": "20-30 years",
   "subjects": [
    "low voltage inference",
    "AI chips"
   ],
   "timestamp": "48:33",
   "quote": "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."
  }
 ],
 "facts": [
  {
   "statement": "Etched has more than $1 billion of customer demand for its first product and has raised $800 million to build it.",
   "subjects": [
    "Etched"
   ],
   "value": "$1B+ demand; $800M raised",
   "timestamp": "2:21",
   "quote": "They already have more than a billion dollars of customer demand for their first product and have raised $800 million to build it.",
   "source": "Patrick O'Shaughnessy"
  },
  {
   "statement": "Etched taped out a working chip on its first attempt — the first hardware company founded after ChatGPT to do so.",
   "subjects": [
    "Etched"
   ],
   "timestamp": "2:21",
   "quote": "taping out a working chip on their first attempt and becoming the first hardware company founded after ChatGPT to do so",
   "source": "Patrick O'Shaughnessy"
  },
  {
   "statement": "Etched was started in 2023 and its first product is a complete rack for inference — chip, boards, power delivery, interconnects, and manufacturing.",
   "subjects": [
    "Etched",
    "Sohu"
   ],
   "value": "founded 2023",
   "timestamp": "2:21",
   "quote": "They started the company in 2023 and the product is a complete rack for Inference.",
   "source": "Patrick O'Shaughnessy"
  },
  {
   "statement": "GPU model-flops-utilization (MFU) on real workloads is typically 20-50% of the advertised peak, and 100% is provably unreachable due to thermal self-regulation.",
   "subjects": [
    "GPUs"
   ],
   "value": "20-50% MFU",
   "timestamp": "9:38",
   "quote": "On GPUs, you often get somewhere between 20 and 50 percent, depending on the workload",
   "source": "Rob Wachen"
  },
  {
   "statement": "Bitcoin miners run at under a quarter of the voltage of GPUs, proving sub-GPU voltages are physically possible.",
   "subjects": [
    "Bitcoin miners",
    "GPUs",
    "low voltage inference"
   ],
   "value": "<25% of GPU voltage",
   "timestamp": "9:38",
   "quote": "Bitcoin miners run at under a quarter of the voltage of GPUs.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Point-to-point chip-to-chip hops on Nvidia Blackwell take about 4,000 nanoseconds.",
   "subjects": [
    "Blackwell",
    "Nvidia"
   ],
   "value": "~4,000 ns",
   "timestamp": "12:29",
   "quote": "on Blackwell chips, it can be about 4,000 nanoseconds to go point to point",
   "source": "Gavin Uberti"
  },
  {
   "statement": "Etched built a fully custom interconnect stack (everything above ethernet layer two) that cuts chip-to-chip latency by more than a factor of 5x.",
   "subjects": [
    "Etched",
    "cluster scale memory"
   ],
   "value": ">5x latency reduction",
   "timestamp": "12:29",
   "quote": "We can go ahead and cut this by more than a factor of 5x",
   "source": "Gavin Uberti"
  },
  {
   "statement": "Etched's first-generation low-voltage inference technology runs at under half the voltage of any other AI chip.",
   "subjects": [
    "Etched",
    "low voltage inference"
   ],
   "value": "under half the voltage",
   "timestamp": "13:57",
   "quote": "We're now bringing forward the first generation of this low-voltage inference technology, which is running at under half the voltage of any other AI chip.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Only a few million people worldwide use paid AI plans — about one one-thousandth of the global population.",
   "subjects": [
    "AI adoption"
   ],
   "value": "a few million paid users",
   "timestamp": "14:49",
   "quote": "there's only a few million users in the world using paid plans of AI models",
   "source": "Rob Wachen"
  },
  {
   "statement": "GPT-4V immediately identified Rob Wachen's pre-diagnosis back-bump photo as a possible tumor — a call that had taken his doctors six months.",
   "subjects": [
    "GPT-4",
    "Rob Wachen"
   ],
   "timestamp": "20:53",
   "quote": "It immediately says, this could be a tumor. You should get an MRI and immediately go to the doctor.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Rob Wachen was diagnosed with stage 4 bone cancer at the end of his sophomore year of high school and given under a 30 percent chance of survival.",
   "subjects": [
    "Rob Wachen"
   ],
   "value": "<30% survival odds",
   "timestamp": "20:53",
   "quote": "Stage 4 bone cancer was told I had under 30 percent chance of survival.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Rob Wachen ran the startup incubator Prod, whose earliest companies included Cursor (merged with Polysphere), Mercor, and Etched itself.",
   "subjects": [
    "Prod",
    "Cursor",
    "Mercor",
    "Etched"
   ],
   "timestamp": "20:53",
   "quote": "Some of the earliest ones being Cursor and Polysphere which merged and Mercor and Etched went through it",
   "source": "Rob Wachen"
  },
  {
   "statement": "Gavin Uberti's first job at 17 was kernels development at XNOR, which was later bought by Apple for $200 million.",
   "subjects": [
    "XNOR",
    "Apple",
    "Gavin Uberti"
   ],
   "value": "$200M acquisition",
   "timestamp": "24:26",
   "quote": "XNOR got bought by Apple for $200 million",
   "source": "Gavin Uberti"
  },
  {
   "statement": "Mark Ross, former CTO of Cypress Semiconductor (which sold for $9 billion), went from telling the founders it would not work to becoming Etched's full-time CTO.",
   "subjects": [
    "Mark Ross",
    "Etched",
    "Cypress Semiconductor"
   ],
   "value": "$9B (Cypress sale)",
   "timestamp": "6:02",
   "quote": "used to be CTO at Cypress Semi, that sold for nine billion dollars",
   "source": "Gavin Uberti"
  },
  {
   "statement": "Brian Loyler, who built all of Nvidia's HGX and DGX systems (about 80% of Nvidia's revenue, tens of billions of dollars a quarter), joined Etched to build its rack team.",
   "subjects": [
    "Brian Loyler",
    "Nvidia",
    "Etched"
   ],
   "value": "~80% of Nvidia revenue (HGX/DGX)",
   "timestamp": "27:27",
   "quote": "we brought over Brian Loyler, who built all of NVIDIA's HGX and DGX systems, which is like 80% of their revenue",
   "source": "Rob Wachen"
  },
  {
   "statement": "Etched has a factory in Taiwan with a few dozen people and a 2-megawatt data center on its office floor running 24-7 day/night development shifts.",
   "subjects": [
    "Etched",
    "Taiwan"
   ],
   "value": "2 MW office data center",
   "timestamp": "27:27",
   "quote": "We have a 2-megawatt data center on this floor, and we did 24-7 development cycles.",
   "source": "Rob Wachen"
  },
  {
   "statement": "When a vendor fell a year behind, Etched shipped a dozen of its top engineers to Bangalore for six months (Gavin lived there four and a half months) rather than accept the delay.",
   "subjects": [
    "Etched",
    "Bangalore"
   ],
   "timestamp": "34:51",
   "quote": "We went out and shipped a dozen of our top engineers across the world to Bangalore for six months.",
   "source": "Gavin Uberti"
  },
  {
   "statement": "Before silicon returned, Etched put the entire full reticle chip on a cluster of over 700 FPGAs and ran a dozen models with its full inference stack, and shipped chipless racks with networking, CPUs, and storage to customer data centers.",
   "subjects": [
    "Etched",
    "FPGAs",
    "prefetching"
   ],
   "value": "700+ FPGAs",
   "timestamp": "38:34",
   "quote": "we took over 700 FPGAs and put the entire full reticle chip on an FPGA cluster and ran a dozen different models with our full inference stack on them before the chips came back",
   "source": "Rob Wachen"
  },
  {
   "statement": "Etched went from getting silicon back to running inference in a rack in 40 days.",
   "subjects": [
    "Etched",
    "Sohu"
   ],
   "value": "40 days",
   "timestamp": "38:34",
   "quote": "We were able to do it in 40 days.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Another very famous AI chip company took 10 months to go from silicon back to running inference in a rack, as publicly announced to its investors.",
   "subjects": [
    "AI chips"
   ],
   "value": "10 months",
   "timestamp": "38:34",
   "quote": "There is another very famous AI chip company that took 10 months to go from getting their silicon back to having them running inference in a rack.",
   "source": "Rob Wachen"
  },
  {
   "statement": "TSMC ran a yield experiment on its own dime at Etched's recommendation and, when it worked, moved the rest of the line over.",
   "subjects": [
    "TSMC",
    "Etched"
   ],
   "timestamp": "41:42",
   "quote": "they will go run an experiment on their own dime",
   "source": "Gavin Uberti"
  },
  {
   "statement": "To fix a back-pressure failure across a clock domain crossing, Etched had to align two on-chip clock signals to within 50 picoseconds, on every chip, two billion times a second — solved in about two weeks.",
   "subjects": [
    "Etched",
    "Sohu"
   ],
   "value": "50 picoseconds",
   "timestamp": "57:50",
   "quote": "we had to go line up two clock signals on our chip to within 50 picoseconds",
   "source": "Gavin Uberti"
  },
  {
   "statement": "During chip bring-up, Etched ran about 30 different board experiments and three of them worked.",
   "subjects": [
    "Etched"
   ],
   "value": "3 of 30 experiments",
   "timestamp": "59:57",
   "quote": "30 different board experiments, and three of them worked",
   "source": "Rob Wachen"
  },
  {
   "statement": "Signing with a physical design vendor costs at least $40-50 million.",
   "subjects": [
    "semiconductors"
   ],
   "value": "$40-50M",
   "timestamp": "1:01:31",
   "quote": "We needed to sign an agreement with the physical design vendor, which will cost you at least 40, 50 million dollars.",
   "source": "Rob Wachen"
  },
  {
   "statement": "In early 2024 Etched projected spending $100 million in the following 12 months, against $15 million in the bank.",
   "subjects": [
    "Etched"
   ],
   "value": "$100M planned spend",
   "timestamp": "1:02:30",
   "quote": "we think we're going to spend $100 million in the next 12 months",
   "source": "Rob Wachen"
  },
  {
   "statement": "At the time of Etched's raise, the biggest semiconductor Series A had been around $40-50 million.",
   "subjects": [
    "semiconductors"
   ],
   "value": "$40-50M prior record",
   "timestamp": "1:02:30",
   "quote": "At the time, the biggest semiconductor fundraisers Series A was around $40, $50 million.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Etched's Series A came together as roughly $103 million in soft commits after every major Valley investor initially passed; the company has raised almost half a dozen rounds since.",
   "subjects": [
    "Etched"
   ],
   "value": "$103M Series A",
   "timestamp": "1:04:11",
   "quote": "We had a board meeting, I show you the spreadsheet and we look at it and it's like $103 million.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Synopsys let pre-Series-A Etched take its emulators on extremely favorable terms, paying over many years — effectively a big loan.",
   "subjects": [
    "Synopsys",
    "Etched"
   ],
   "timestamp": "1:05:59",
   "quote": "went ahead and let us get some of their emulators on extremely favorable terms where we'd pay over many years",
   "source": "Gavin Uberti"
  },
  {
   "statement": "In an internal experiment, Codex got GPT-OSS running on Etched's stack from scratch, overnight, using only their docs, completely by itself.",
   "subjects": [
    "Codex",
    "GPT-OSS",
    "Etched"
   ],
   "timestamp": "51:31",
   "quote": "we had Codex actually get GPT-OSS running from scratch just based off of our docs, completely by itself",
   "source": "Rob Wachen"
  },
  {
   "statement": "Dozens of people from high-frequency trading firms have joined Etched, drawn by its kernels-first, no-compiler philosophy.",
   "subjects": [
    "Etched",
    "high-frequency trading"
   ],
   "timestamp": "51:31",
   "quote": "we've had dozens of people from high-frequency trading join the team",
   "source": "Rob Wachen"
  },
  {
   "statement": "Chips built by labs and hyperscalers have lower FP8 flop density than Nvidia's B300.",
   "subjects": [
    "AI chips",
    "Nvidia"
   ],
   "timestamp": "56:10",
   "quote": "the flop density for, say, FB8 times FB8 is lower than the black hole B300",
   "source": "Gavin Uberti"
  },
  {
   "statement": "Etched's first-gen product uses a different supply chain than Nvidia's Rubin — 4nm versus Rubin's 3nm, and a different HBM.",
   "subjects": [
    "Etched",
    "Nvidia",
    "supply chain",
    "Rubin"
   ],
   "value": "4nm vs 3nm",
   "timestamp": "1:15:31",
   "quote": "We're on four nanometer Rubens on three nanometer.",
   "source": "Rob Wachen"
  },
  {
   "statement": "Patrick O'Shaughnessy has invested in Etched about five times since 2023; at the time it was his largest-ever first check.",
   "subjects": [
    "Patrick O'Shaughnessy",
    "Etched"
   ],
   "value": "~5 investments",
   "timestamp": "1:10:43",
   "quote": "I think I've invested five or so times in Etched.",
   "source": "Patrick O'Shaughnessy"
  },
  {
   "statement": "Cursor published that a group of its coding agents built an entire browser from scratch in a week.",
   "subjects": [
    "Cursor",
    "AI agents"
   ],
   "timestamp": "46:36",
   "quote": "Cursor published that they had a bunch of coding agents build an entire browser from scratch in a week",
   "source": "Rob Wachen"
  }
 ],
 "notable_quotes": [
  {
   "speaker": "Rob Wachen",
   "timestamp": "33:08",
   "quote": "You kind of have to be sick in the head to join our company.",
   "why": "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."
  },
  {
   "speaker": "Rob Wachen",
   "timestamp": "1:15:31",
   "quote": "if you have the most performant product and you can't produce it, then you're just a podcast",
   "why": "Distills the production-is-the-product thesis — performance without supply-chain-scale availability is worthless."
  },
  {
   "speaker": "Gavin Uberti",
   "timestamp": "16:59",
   "quote": "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.",
   "why": "The pre-industrial analogy for today's token economics — the argument for industrial economies of scale in AI inference."
  },
  {
   "speaker": "Rob Wachen",
   "timestamp": "1:13:36",
   "quote": "The best ability is availability.",
   "why": "The lesson of Etched's past year: demand is so intense that having chips available today beats having marginally better chips later."
  },
  {
   "speaker": "Rob Wachen",
   "timestamp": "1:11:51",
   "quote": "He leads back and he's like, the puzzle begins.",
   "why": "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."
  },
  {
   "speaker": "Rob Wachen",
   "timestamp": "1:21:27",
   "quote": "we measure productivity as a society as GDP per capita. But really, it's going to look much more like agents per megawatt",
   "why": "Reframes national productivity for an AI-agent economy — the episode's most quotable macro prediction."
  }
 ],
 "extracted_by": "agent",
 "prompt_version": "episode-extract@3",
 "_file": "invest-like-the-best-2026-06-30-etched-building-ai-hardware-to-make-inference-faster-and-che.json"
}