This podcast features Gavin Uberti, a 21-year-old Harvard dropout who founded Etched. He argues that the future of AI hardware isn't Nvidia's general-purpose GPUs but specialized chips (ASICs, chips designed for one task). He's bullish on these chips, saying they'll make AI much cheaper and faster, enabling real-time voice and robotics. Key mentions: Etched (building a chip 20x faster than Nvidia's H100), Nvidia (its GPUs risk being replaced by specialized chips), and AMD (its MI300X is a viable competitor to Nvidia).
21-year-old Harvard dropout Gavin Uberti founded Etched, a company that bets on specialized AI chips, arguing that the current general-purpose chips represented by Nvidia GPU will gradually be replaced by specialized chips hardcoded for specific model architectures (such as Transformer), which can s
21-year-old Harvard dropout Gavin Uberti is the founder of dedicated AI chip startup Etched. The core thesis of this interview is: the current general-purpose AI chips represented by Nvidia GPUs will be replaced by dedicated chips (ASICs) hardcoded for the Transformer architecture, delivering order-of-magnitude advantages in inference cost and latency. Gavin Uberti argues that the next phase of AI infrastructure investment will be "the largest construction since the Industrial Revolution," and that dedicated chips are the key to unlocking real-time AI applications (e.g., voice conversations, robotics) and even superintelligence.
Gavin Uberti believes that dedicated chips (ASICs) customized for Transformer models will thoroughly outperform general-purpose GPUs, following a historical path similar to Bitcoin mining ASICs replacing GPUs.
Uberti draws an analogy to the historical evolution of Bitcoin mining: from CPU to GPU, then to dedicated ASIC. Once ASICs appear, GPUs become commercially infeasible for that specific task. He argues that the AI chip sector is at a similar inflection point. The flexibility of GPUs is their advantage, but it also brings significant "waste." In an Nvidia GPU, only about 4% of the transistors are used for matrix multiplication (the core of AI computation); the rest supports its general-purpose nature (caching, I/O, instruction processing, etc.). A Transformer ASIC can devote over 90% of its transistors to matrix multiplication, achieving more than an order of magnitude improvement in raw compute power (FLOPS).
Uberti emphasizes that this improvement is not only in cost but also in latency. Dedicated chips can process an entire prompt "in milliseconds rather than seconds," enabling true real-time conversations, voice-to-voice interactions, and other applications. He states clearly that "when dedicated chips enter a domain, general-purpose chips will no longer be competitive, and the first mover is almost certain to win."
Uberti proposes a five-layer "AI infrastructure chain," arguing that each layer presents opportunities for disruption and innovation, rather than being entirely monopolized by existing giants.
The chain from top to bottom is:
1. Top-tier AI Models: e.g., OpenAI, Anthropic; their core moat is talent.
2. Hyperscale Cloud Providers: e.g., Google, Microsoft, Amazon; responsible for providing compute infrastructure.
3. AI Chip Designers: Currently Nvidia (general-purpose GPU), future will include Etched (dedicated ASIC) and AMD (general-purpose GPU).
4. Chip Manufacturing and Key Components: e.g., TSMC (manufacturing), Samsung and SK Hynix (HBM high-bandwidth memory).
5. End Applications: Uberti sees this as the biggest startup opportunity, where startups can build new AI-based interfaces (e.g., voice conversations, AI search, document analysis).
Uberti judges that the market will eventually become highly concentrated. Building a trillion-parameter model requires billions of dollars; the world cannot support 100 such companies, and there will likely be only about 3 left. These top-tier model companies will become "pure AI model producers" like TSMC, with most companies building applications by fine-tuning on top of them.
Uberti believes that the Transformer architecture has been "nailed down" due to its scale effects and ecosystem advantages, and will not be replaced in the future; meanwhile, the data hunger of AI models will be solved through video data and self-play.
Uberti points out that both hardware and software ecosystems have developed significant path dependence on Transformers. Nvidia has added hardware optimizations for Transformer in its next-generation GPUs and possesses highly optimized Transformer software libraries. This means that for any new architecture to replace Transformer, its performance advantage must be large enough to offset the huge disadvantages in hardware efficiency and software ecosystem—a "tremendous burden." He asserts that "Transformers are not just a fad; they have become the foundational architecture."
Regarding the data bottleneck, Uberti notes that humanity has not yet tapped the most valuable data source—video. Infants learn by observing the world, and future models should be trained on massive amounts of video data from the start. When that data is exhausted, the ultimate solution is self-play, where models generate their own data (e.g., AlphaGo) and self-evaluate to improve. This will be the "final frontier" to solve the data bottleneck.
| Position | Attitude (Bullish/ Risk Warning/ Neutral) | Key Data |
|---|---|---|
| Etched | Bullish (his own company) | Its chip will have "more than an order of magnitude" raw compute power compared to Nvidia H100, achieving "20x" latency improvement. |
| Nvidia | Risk Warning | Faces "innovator's dilemma"; high margins on H100 make it reluctant to introduce lower-margin dedicated chips; its CUDA software loses its moat in the dedicated chip era. |
| AMD | Neutral | Its MI300X will be a "viable competitor" to Nvidia H100. |
| OpenAI / Anthropic | Neutral | At the top of the value chain, core moat is talent; they will not monopolize all application layers. |
| Google / Microsoft / Amazon | Neutral | As infrastructure providers, they are customers of "AI model producers," not direct competitors. |
| TSMC | Neutral | Key link in chip manufacturing, but the entire chip design cycle (4-5 years) needs to be drastically shortened. |
| Samsung / SK Hynix | Neutral | The cost of HBM high-bandwidth memory can even exceed that of the chip itself; a critical and often underestimated link in the supply chain. |
| Bitmain / MicroBT | Bullish (analogy) | As successful examples of Bitcoin mining ASICs, their first-mover advantage made them giants "worth over $50 billion." |
1. "AI is undergoing the largest infrastructure build since the Industrial Revolution." (Gavin Uberti) — Support: Training and running future models will require building data centers at the 2-gigawatt scale, consuming energy equivalent to a large power plant, and unprecedented cooling and reliability challenges must be solved.
2. "Dedicated chips will follow the path of Bitcoin ASICs, completely replacing general-purpose GPUs." (Gavin Uberti) — Support: Once dedicated chips are mass-produced, their cost-effectiveness will make using GPUs "commercially infeasible," and the first mover will gain a huge advantage, forming a duopoly.
3. "The Transformer architecture has been 'nailed down'; replacing it requires overcoming enormous hardware and software ecosystem inertia." (Gavin Uberti) — Support: Nvidia is optimizing hardware for Transformer and has highly optimized software libraries; any new architecture must be "good enough to offset this disadvantage" in performance, which is nearly impossible.
4. "The ultimate solution to the data bottleneck is 'self-play,' not finding more online data." (Gavin Uberti) — Support: When text and video data are exhausted, models can generate their own outputs and self-evaluate (like AlphaGo) to create infinite training data.
5. "The application layer of AI is the biggest startup opportunity; new interfaces will go beyond chatbots." (Gavin Uberti) — Support: Real-time voice conversations, AI document processing, AI search, and other new interfaces will become mainstream, and startups building these applications (rather than model companies or cloud providers) will have the greatest room for innovation.
6. "Superintelligence is not a 'black-and-white,' 'overnight' switch, but a slow, gradual 'spectrum.'" (Gavin Uberti) — Falsification condition: If a model's capabilities suddenly jump by several orders of magnitude in a short period ("fast takeoff"), this judgment is falsified. Uberti believes such "fast takeoff" will not happen because each model upgrade requires building new data centers at scale.
7. "The AI chip design cycle must be shortened from the current 4-5 years to a few months." (Gavin Uberti) — Support: To meet the urgent demand for dedicated chips, the back-end chip design process (e.g., floorplanning) should be automated, and front-end design (RTL coding) should be assisted by AI models, drastically shortening iteration cycles.
8. "A good AI leader's core is 'setting the vision, finding the right people, and not running out of money.'" (Gavin Uberti) — Support: Uberti cites his own experience, explaining how as a non-technical leader he convinced industry veterans to join the team, and emphasizes that deep understanding of technical details is crucial for making the right decisions.