Brief Summary
This podcast episode explores Gavin Baker's investment thesis on AI infrastructure, highlighting his focus on the picks and shovels of the AI industry rather than software or chatbots. Baker, a seasoned investor with a 20-year track record, believes the biggest returns in AI lie in electricity, power, and silicon fabrication. He identifies key bottlenecks and constraints in AI infrastructure, viewing the current landscape as a generational buying opportunity rather than a bubble. The discussion covers his investment strategy, key holdings, and his perspective on why the AI boom differs from the dot-com bubble.
- Gavin Baker focuses on AI infrastructure investments.
- He believes the biggest returns are in electricity, power and silicon fabrication.
- He views the current AI landscape as a super cycle, not a bubble.
AI Infrastructure Thesis
Gavin Baker, a prominent AI investor, has been investing in major AI companies for 20 years. He was an early backer of NVIDIA and Cerebris, which recently IPO'd. Baker's thesis is that AI is in a super cycle, not a bubble, focusing on the infrastructure of AI, specifically watts, wafers, and tokens. He believes the most significant returns will come from electricity, power, and silicon fabrication, rather than SaaS or chatbots. Baker has invested $4.1 billion in this thesis, viewing the current AI landscape as a generational buying opportunity for AI infrastructure.
Gavin’s Track Record
Gavin Baker is the founder of Trades Management, an investment fund. He has invested in NVIDIA for 20 years, achieving significant returns. Early wins include companies like Cerebris, which recently IPO'd for a substantial amount. The discussion will cover Baker's portfolio and his guidance on the future of AI investment opportunities.
Bottlenecks in Chips
Trades Management has approximately $4 billion in assets under management. A significant position is in Astera Labs, which provides connectivity between GPUs in data centres. As AI clusters scale, the bottleneck shifts from GPUs to data transfer. Astera Labs addresses this by building the plumbing system needed for efficient data access and transfer. Another notable investment is Cerebrus Labs, which recently IPO'd for a substantial amount. Gavin believes NVIDIA will maintain its profit margins and demand, potentially reaching a $10 trillion market cap. Micron, a major memory maker, has also seen significant growth, reflecting the importance of memory in AI.
Unity and World Models
Unity Software, a 3D game engine, is another interesting investment. Unity is a world model builder with a deep understanding of physics, textures, and lighting. AI companies use Unity to simulate virtual environments for training humanoid robots. World models allow AI models to understand the world through simulated experiences, similar to a game character in a virtual environment. Gavin Baker uses a barbell approach, investing in both established players like Micron and NVIDIA, and forward-looking companies like Cerebrus and Unity.
Inference
Gavin's portfolio includes Positron, which creates inference chips, similar to Cerebrus. The AI infrastructure stack is shifting from pre-training to post-training, requiring more focus on inference. Inference is needed for AI models to understand new information and reason effectively. The revenue opportunity from inference is estimated to be 5 to 10 times greater than pre-training. NVIDIA is also creating GPUs aligned to inference. China has a unique opportunity to create inference-focused chips due to its abundance of energy and chip manufacturing capabilities. Gavin is betting on the U.S. infrastructure setup for inference.
Four AI Constraints
Gavin Baker has a put position in the QQQ ETF, indicating a bearish sentiment on the overall market. He is investing in key companies that address bottlenecks in AI. The four key constraints he is focusing on are: verticalised small language models (SLMs), sovereign infrastructure, performance per watt, and energy and space compute. Verticalised SLMs are optimised for specific enterprise data or local devices, addressing privacy concerns. Apple is expected to be a major player in locally run models. Sovereign infrastructure refers to the speed of physical deployment of hardware, which is a significant challenge.
Energy and Space Compute
Performance per watt is an important metric, as AI labs are increasingly focused on generating more tokens per watt. Cost is a significant factor, influencing decisions on which models to use. Gavin's investments in Cerebras, Astera Labs, and Positron reflect his focus on improving performance per watt. The fourth constraint is energy and space. Terrestrial grids limit energy, and there is resistance to new data centres. Solutions include portable energy sources and orbital compute. SpaceX is a key player in delivering payloads to orbit for space-based data centres.
This Isn’t Dot-Com
The dot-com bubble was fuelled by debt, with money being borrowed for unproven products. In contrast, the current AI super cycle is funded by real revenue, with companies like OpenAI and Anthropic generating substantial ARR. Major companies like Google, Microsoft, Amazon, and Meta are using their own cash reserves to invest in AI infrastructure. The constraints in AI are physical, limited by the supply of chips and manufacturing capabilities.
Supply Constraints
NVIDIA could sell $2 to $3 trillion of GPUs if TSMC could supply them. TSMC's limited capacity acts as a blocker to the rate of acceleration of a potential bubble. As long as TSMC remains constrained, the growth rate is sustainable. Demand for AI is exponentially increasing, outpacing the production supply of chips. A major memory manufacturer, SK Hynix, is receiving significant offers from Google and Microsoft to secure future supply.
Conclusion
The key limiting factors are watts and wafers. As long as these remain valuable and supply is limited, the AI industry will continue to thrive. Gavin's largest holding is a QQQ put position, reflecting a bearish market outlook. Other significant holdings include Astera Labs and Unity. Gavin's approach is a cautious, long-term, futuristic one, potentially offering exponential gains. The discussion concludes with a debate on whether the current AI landscape is a bubble, with the consensus leaning towards it being in the earlier stages, sustained by supply constraints.
