Gavin Baker, founder of Atreides Management, is making a concentrated bet that AI is not a bubble but an infrastructure supercycle driven by power, wafers and compute. His portfolio reflects that view. He owns companies tied to physical choke points in the AI stack, including Astera Labs, Cerebras, Micron, Nvidia, Unity and Positron, while also holding QQQ put options to protect against a broader market drawdown.
The thesis was outlined in a recent episode of Limitless Podcast. The discussion described Atreides Management as managing about $4.1 billion, and highlighted Baker’s long-running investment record in Nvidia over more than two decades, along with early backing of Cerebras. Instead of chasing the most visible application-layer names, he focuses on the “picks and shovels” side of AI: GPU connectivity, memory, inference chips, advanced manufacturing and power supply.
His focus is on supply constraints, not model hype
Baker’s case rests on hard bottlenecks rather than sentiment. He keeps coming back to two physical limits: power and wafers. As long as critical supply points such as TSMC, ASML, high-bandwidth memory and the grid cannot move into surplus quickly, he argues that AI capital spending is less likely to spiral into a replay of the dot-com era. The point is simple. If supply cannot ramp fast enough, overbuilding the entire market becomes much harder.
That leads him to companies solving specific infrastructure problems. Astera Labs is one example. In the podcast, it was framed as the “plumbing system” for GPU clusters. Once data centers scale to hundreds of thousands of chips, the problem is no longer only the GPU itself. It becomes about moving data at the right time, between the right nodes, and accessing memory efficiently. The show said Astera Labs at one point made up roughly 9% to 10% of the fund.
Cerebras and Positron represent another part of the thesis: inference infrastructure. Baker believes AI spending is shifting away from a world dominated by pre-training toward one that puts more weight on post-training and inference. According to the podcast, his view is that the cost or revenue opportunity from inference alone could be 5 to 10 times larger than pre-training compute spend. That helps explain why he keeps leaning into chips and systems designed for inference, not only training GPUs.
Unity, on-device models and the cost per watt question
Some names in Baker’s portfolio look less obvious at first glance. Unity is one of them. The podcast argued that the company’s 3D engine can also be seen as a world-modeling tool useful for simulated training environments and synthetic datasets for robotics. If AI development keeps moving toward embodied systems and richer interaction with physical reality, that capability may matter more than traditional software labels suggest.
Baker also talked about vertical small language models and on-device deployment. In that framework, users may not rely every day on a single general-purpose assistant. They may need personalized AI agents trained on their own data. The episode noted that he sees Apple as well positioned in a future where models run locally on devices. This again comes back to one metric: performance per watt. For AI labs, the practical question is increasingly how many tokens each watt of electricity can produce.
That cost pressure is already visible. The podcast said Microsoft and Uber had reduced use of Claude Code because annual budgets were getting consumed within months. For Baker, examples like that reinforce the same point: the companies that raise efficiency per watt and lower token costs are the ones more likely to win spending.
Long niche winners, short broad market risk
Baker does not treat a bullish AI view as a bullish call on the whole equity market. He separates the two. His use of QQQ put options shows that clearly. The episode described the QQQ put position as one of the largest in his portfolio, which signals a defined expression: go long the AI infrastructure names solving real bottlenecks, while hedging downside in the broader market.
That structure also explains why established names such as Nvidia and Micron can sit alongside more forward-looking positions such as Cerebras, Positron and Unity. One side of the portfolio leans into already proven hardware demand. The other reaches for bottlenecks that may become far more valuable over the next several years. That is the barbell.
Why he does not see a repeat of the dot-com bubble
When Baker responds to the AI bubble argument, he does not start with valuation rhetoric. He starts with funding structure and demand quality. The podcast said he views the 2000 internet bubble as heavily debt-fueled, with capital flowing into ideas and products that were not yet validated by real users. In contrast, today’s biggest buyers of AI chips and compute are cash-rich companies such as Google, Microsoft, Amazon and Meta, not fragile firms relying on aggressive leverage.
He also points to a hard limit that restrains excess. If TSMC’s advanced manufacturing capacity could expand overnight, companies like Nvidia and their customers might push capital spending much harder and drive the market closer to a true capex bubble. But that is not the current setup. Wafer capacity, memory supply and energy access are all constrained, and those constraints slow the pace at which any bubble could inflate.
The episode added more examples. ASML equipment was described as having a backlog of about five years. SK Hynix, a key supplier of memory used with Nvidia GPUs, was said to be receiving large proposals from Google and Microsoft aimed at locking in supply for the next three years. Different examples, same message: demand is rising fast, but supply is still tight.
In Baker’s framework, as long as power and wafers remain the two brick walls in front of the industry, AI infrastructure spending is not easily pushed into uncontrolled oversupply. As for whether the wider market can still fall, his QQQ puts already answer that question.