Gavin Uberti is Etched’s co-founder and CEO. He left Harvard initially to travel, then worked on microkernel and AI-code optimization before helping start a company focused on specialized inference hardware. This profile distinguishes his architectural thesis and Etched’s projections from independently verified product outcomes. — Pioneers of AI Interview.

Part 1: The Transformer Thesis
- On the dominance of transformers: Uberti’s founding bet was that transformer models would attract enough adoption to justify hardware built specifically for them; he presents this as a strategic thesis, not a certainty about all future AI. — Pioneers of AI Interview.
- On the end of general-purpose AI: Etched accepts that its transformer-only design cannot run many other AI architectures. Uberti frames that restriction as the trade-off for dedicating more chip area to the workload it targets. — Pioneers of AI Interview.
- On specialization: Uberti compares transformer specialization with Bitcoin-mining ASICs: as a workload becomes large and established enough, the cost of a custom chip may become worthwhile. He offered this as an analogy and prediction, not a universal rule. — Invest Like the Best 356.
- On the longevity of attention mechanisms: Uberti traced Etched’s bet to the rise of transformer models after earlier neural-network architectures; the design focuses on the common transformer computation rather than a single named language model. — Invest Like the Best 356.
- On historical hardware trends: His Bitcoin-mining example illustrates a possible path from flexible processors to task-specific hardware when a particular workload becomes economically large enough; he did not claim every computing field must follow it. — Invest Like the Best 356.
- On algorithmic overhang: Uberti argues that existing software, hardware support and research investment reinforce the transformer ecosystem, making a switch to an alternative architecture harder as model development becomes more expensive. — Pioneers of AI Interview.
Part 2: Challenging the Incumbents
- On NVIDIA's generalist approach: Uberti sees a general-purpose GPU’s flexibility as a design trade-off: hardware that supports many tasks dedicates less area to the specific matrix operations needed by his target workload. — Invest Like the Best 356.
- On wasted silicon: In a 2025 interview, Uberti estimated that only a small share of an H100 die was devoted to matrix-multiplication blocks; this was his argument for a narrower design, not proof of Etched’s delivered advantage. — Pioneers of AI Interview.
- On the CUDA advantage: Uberti calls CUDA a powerful Nvidia software moat, while arguing that a less-programmable, transformer-specific chip could need a simpler software stack. He also says Etched still had to make its own software production-ready. — Invest Like the Best 356.
- On market positioning: In the original 2023 interview, Uberti positioned Etched’s first specialized chip around inference, while noting that training would require different system and network design. — Invest Like the Best 356.
- On memory bandwidth: Uberti identifies memory movement as an inference constraint: batching requests can reuse model weights across users, but useful throughput still depends on adequate memory bandwidth. — Invest Like the Best 356.
- On incumbency inertia: In 2023 Uberti argued that Nvidia’s profitable general-purpose product line might slow its move toward a dedicated transformer ASIC; he acknowledged Nvidia could eventually compete. — Invest Like the Best 356.
Part 3: The Architecture of Sohu
- On design philosophy: Sohu was presented as a transformer-only inference system, trading general programmability for more compute devoted to that model family. Its comparative performance figures were then Etched forecasts. — Pioneers of AI Interview.
- On removing control flow: Uberti’s proposed specialization reduces some flexible memory-access and control features so more circuitry can serve the target computation; the claimed utilization gains were pre-silicon expectations in the 2023 interview. — Invest Like the Best 356.
- On matrix multiplication: Etched’s 2026 product update describes co-designed math arrays, circuit techniques, scheduling and power delivery for its inference clusters; these are company-reported design choices, not an independently verified throughput benchmark. — Etched Frontier Inference Clusters.
- On power efficiency: Etched says its low-voltage architecture aims to raise sustained compute density without thermal throttling; the published efficiency figures are company-reported early tests, not an external comparison. — Etched Frontier Inference Clusters.
- On the manufacturing process: Etched reported that its A0 silicon returned from TSMC’s N4P process in 2026 and that it was validating a rack-scale product with customers; the company’s innovation claim concerns the integrated design, not a new fabrication node. — Etched Frontier Inference Clusters.
- On memory walls: Etched describes a shared cluster-scale memory pool and an HBM/SRAM hybrid intended to reduce cross-chip access latency. Its update does not establish the original lesson’s specific KV-cache placement claim. — Etched Frontier Inference Clusters.
- On scale-out networking: The company says a low-latency, high-bandwidth interconnect links memory across its scale-up domain, making the rack architecture—not a standalone chip—the relevant unit of design. — Etched Frontier Inference Clusters.
Part 4: Software and Complexity
- On the compiler tax: Uberti argues that a less-programmable transformer chip could need a simpler compiler and software stack than CUDA, but he explicitly says Etched still needs reliable, production-grade software. — Invest Like the Best 356.
- On the history of TVM: Uberti’s earlier microkernel work involved hand-optimizing matrix and convolution operations; that experience helped him notice the cost of moving data through flexible chips. — Pioneers of AI Interview.
- On the developer experience: Etched had a software prototype running against a chip simulation in 2023; Uberti said the company still needed to iterate it into reliable production software before physical chips arrived. — Invest Like the Best 356.
- On ignoring legacy support: Limiting Sohu to transformers removes broad programmability, simplifying the target workload while sacrificing support for other models. It does not mean the company can dispense with software. — Pioneers of AI Interview.
Part 5: The Economics of Inference
- On the bottleneck of AI: Uberti distinguishes the recurring cost of running inference from training and argues that grouping many requests can amortize expensive model-weight reads; the eventual unit economics remain workload-dependent. — Invest Like the Best 356.
- On cost per token: Uberti’s goal is to reduce inference cost enough to broaden access to advanced models. His 2025 order-of-magnitude cost reduction was a forecast, not a measured outcome for deployed systems. — Pioneers of AI Interview.
- On compute constraints: He points to expensive agent workflows and advanced model access as examples of products constrained by inference cost, while treating cheaper compute as an opportunity rather than a guaranteed result. — Pioneers of AI Interview.
- On capital intensity: Uberti said Etched planned to spend much of its announced $120 million Series A on building a larger-than-normal initial production run, illustrating how capital-intensive deployment can be beyond chip R&D. — Pioneers of AI Interview.
- On democratizing access: For Uberti, lower-cost inference matters because it could make advanced AI accessible beyond well-funded companies and paying subscribers; this is an intended social benefit, not a claim already achieved. — Pioneers of AI Interview.
Part 6: Building a Hardware Startup
- On the difficulty of silicon: Uberti explains that a semiconductor design moves through architecture, RTL, verification, physical layout, masks and fabrication before production, making mistakes expensive to correct after tape-out. — Invest Like the Best 356.
- On pacing the team: Because specialized-chip competitors and customer demand were emerging, Uberti said Etched had to move quickly while still shipping working silicon; speed did not remove the need for verification. — Invest Like the Best 356.
- On hiring criteria: Uberti recounts recruiting experienced chip leaders after he and a cofounder built a functional simulation and technical white paper; the actual lesson is to combine young founders’ thesis with veteran execution, not to reject veterans. — Invest Like the Best 356.
- On tape-out pressure: His account of RTL verification and physical timing closure shows why a chip design must be tested extensively before it is committed to fabrication; the profile’s exact tape-out quotation is not authenticated. — Invest Like the Best 356.
- On IP blocks: Rambus says it supplied Etched with HBM memory-controller IP and integration services, allowing the team to focus its custom work elsewhere in the system. — Rambus–Etched Partnership.
- On building credibility: Uberti says the founders answered an experienced skeptic by producing a technical white paper and functional simulation, which helped attract an experienced architect and seed funding. — Invest Like the Best 356.
- On investor skepticism: The team’s initial technical pitch met skepticism because the founders had not built a chip before and would need substantial funding; Uberti says the simulation and early seed round helped overcome that hurdle. — Invest Like the Best 356.
- On focus: Uberti describes a hardware startup leader’s core responsibilities as setting a technical vision, placing capable people in the right roles and keeping the company funded. — Invest Like the Best 356.
Part 7: Leaving University and Taking Risks
- On timing: Etched’s transformer-specialization idea predates ChatGPT’s public launch; Uberti says that launch made customer demand and investor interest clearer, sharpening the timing of the company’s bet. — TechCrunch Found Interview.
- On the decision to drop out: Uberti says he left Harvard to travel, not to found Etched. Microkernel work during that period led him toward the hardware-efficiency problem from which Etched emerged. — Pioneers of AI Interview.
- On early validation: Uberti recounts raising an early $5.5 million after he and a cofounder built a functional simulation and white paper; the financing enabled recruitment but did not prove the commercial thesis. — Invest Like the Best 356.
Part 8: The Future of Real-Time AI
- On the latency barrier: Uberti separates time-to-first-token from the delay between later tokens and argues that reducing both matters for natural conversation and other real-time applications. — Invest Like the Best 356.
- On autonomous agents: Uberti uses costly agent workflows as a reason to pursue cheaper inference; he does not establish that every agent requires thousands of loops or that only Etched hardware can run them. — Pioneers of AI Interview.
- On new product categories: He expected lower latency and cost to enable interfaces beyond text chat, including real-time speech and more responsive products; these were prospective applications in the 2023 interview. — Invest Like the Best 356.
- On multimodal futures: Uberti names video generation, audio generation and multimodal inference as workloads where he expected transformer-focused hardware to help; product-specific speedups remained projections. — Pioneers of AI Interview.
- On systemic efficiency: Etched frames lower-voltage inference and rack-level thermal design as ways to improve sustained throughput per unit of power; its energy-efficiency comparisons remain company-reported. — Etched Frontier Inference Clusters.
- On the speed of thought: Uberti argued that faster token generation could shorten code-writing waits and make interactive AI feel more immediate. His 1,000-token-per-second example was an aspiration in 2023, not a delivered service benchmark. — Invest Like the Best 356.
- On real-time search: Uberti imagined search and document-query products that synthesize far more material when inference latency falls; he did not demonstrate the profile’s specific millisecond-scale Internet-search claim. — Invest Like the Best 356.