This interview says AI growth is limited by two physical things: electricity and chips. Manager Gavin Baker thinks power shortages ease by 2027-2028, with long-term solution being orbital compute (putting NVIDIA racks in space). He watches TSMC's capacity decisions as key to spotting a bubble. Mentions Anthropic (record revenue growth), NVIDIA (could sell $2-3 trillion GPUs), and Cerebras (doing wafer-scale computing, different approach).
Gavin Baker (Founding Partner and CIO of Atreides Management), in his sixth dialogue, focuses on the two physical constraints of "watts and wafers," arguing that they will dominate the next phase of AI development. Regarding electricity, he judges that the short-term shortage will ease in 2027-2028
Here is the English translation of your investment research notes, adhering to all specified rules.
Gavin Baker (Founding Partner & CIO of Atreides Management) focuses on the two physical constraints of "watts and wafers" in his sixth conversation, arguing they will dominate the next phase of AI development. On power, he believes the short-term shortage will ease by 2027-2028 as new power sources come online, with the long-term solution relying on orbital compute. On wafers, he emphasizes that this cycle is different from the internet bubble, and TSMC's capacity decisions are the most critical variable to watch. He notes Anthropic's record-breaking ARR growth and believes OpenAI and Anthropic should raise capital at higher valuations.
Gavin Baker believes capitalism will solve the power shortage, but the short-term bottleneck lies in permitting, not energy itself. He points out that heads of data center infrastructure at large PE firms (e.g., Blackstone, Apollo) have stated that the biggest constraint has shifted from energy and chips to "zoning and permitting." Many companies are waiting until after the midterm elections to act, to avoid becoming political targets. Although production capacity for key equipment like turbines (e.g., large blade casting) has been lost in the West for 80 years, he believes capitalism has the ability to solve these industrial engineering challenges over time.
He judges that the power shortage will begin to ease in 2027-2028 when new power sources come online. The long-term solution is orbital compute. Baker emphasizes this is not "the Pentagon in space," but rather deploying Blackwell racks (weighing 3,000 lbs, 8 feet tall) directly in space, connected via lasers to form a virtual data center. SpaceX engineers are highly confident about this, with the key being Starship's reusability. He refutes concerns about heat dissipation and maintenance, noting that SpaceX already operates the world's largest satellite fleet (98-99% of the global total), possesses the largest data center on Earth, and employs the best hardware engineers.
Regarding the impact of orbital compute on terrestrial data centers, Baker believes it is not bearish. He argues that humanity's demand for compute is infinite ("the US will suck up every last drop of energy"), reasoning tasks are suitable for space, while training tasks will remain on Earth for the long term. Therefore, terrestrial data centers will retain value within his lifetime. However, he warns that when a large amount of power capacity comes online, it will coincide precisely with the time orbital compute is proven viable, and related companies need to be wary of this overlapping effect.
Gavin Baker believes TSMC's capacity decisions are the single most important variable for determining whether AI is forming a bubble. Citing Carlota Perez's theory, he notes that every foundational new technology (railroads, canals, internet) in history has been accompanied by a bubble. However, this AI bubble has two major differences: 1) Construction capital comes primarily from operating cash flow, not debt; 2) GPU utilization is 100%, whereas in 2000, 99% of fiber optic cable was dark.
Baker judges that if TSMC fully meets Jensen Huang's demand, NVIDIA could sell $2-3 trillion worth of GPUs in 2026-2027, which would inevitably lead to oversupply. Therefore, TSMC's supply discipline is key to avoiding a bubble. However, historical patterns suggest that Intel and Samsung will eventually break discipline, forcing everyone else to follow. He believes whether TSMC can maintain its lead over Intel and Samsung (approximately 9-15 months), and the pace of its capacity expansion, defines the "Goldilocks zone" for observing a bubble.
Regarding Elon Musk's Terrafab, Baker believes it will succeed. The reasons are: 1) Collaboration with Intel, gaining access to 50 years of process knowledge; 2) Elon's reputation in hardware engineering will attract the "A-team" from equipment suppliers like ASML and KLA for priority support; 3) Elon will use methods like "Taiwan City" and "Japan City" to attract the world's best engineers. He believes Terrafab's long-term nature means it won't immediately threaten TSMC, but its execution speed (building a data center in 122 days) could disrupt the industry's pace.
Gavin Baker is open-minded about the core question of "whether frontier tokens will continue to capture the vast majority of economic value in the model layer," and considers this a necessary assumption for investors. He notes that currently, the vast majority of economic returns still come from frontier models (Anthropic, OpenAI, xAI), which is surprising. Using Google's Gemini 3.1 Pro as an example, he notes it was once "amazing" but is now "unbearable," indicating that the frontier's lead is not permanent.
Baker introduces the concept of the Pareto frontier (Intelligence vs. Cost) . Nine months ago, Google dominated this frontier; now, it is dominated by Anthropic, OpenAI, and xAI's Grok 4.3. He believes Google might be "subsidizing out of pride" to stay on the frontier. He warns that violating Richard Sutton's "Bitter Lesson" (that more compute always beats human algorithmic innovation) is the biggest risk in AI investing. However, he also presents a counterintuitive idea: when AI reaches ASI (Artificial Superintelligence), it might first improve efficiency through self-optimization, temporarily "violating" the Bitter Lesson.
Regarding continual learning, Baker considers this the third key question (the first two being the frontier token premium and the Bitter Lesson). If solved, it would lead to a "fast takeoff." Currently, humans are extremely efficient sample learners, while AI requires millions of trials and errors. He points out that once a model can dynamically update its weights, its capabilities will undergo a qualitative change.
Gavin Baker believes new chip companies should pursue "different and hard" architectures, rather than trying to build a "better GPU." He analogizes chip design to the "iron triangle" of tank design (firepower, protection, mobility), constrained by the laws of physics and TSMC's design rules. He argues that trying to build a "better GPU" (like Google TPU, Amazon Trainium, AMD) is like "pulling on Superman's cape," and NVIDIA, as a fast follower, can leverage its scale advantages and deep relationship with TSMC to crush competitors.
He proposes a rule: 1% market share is worth $100 billion, which is a decent venture capital return. But the key is doing "different and hard" things. He uses Cerebras as an example, whose wafer-scale computing is a "hard and fundamentally different" architectural choice, allowing it to do things others cannot. However, Cerebras also faces an I/O bottleneck (the "coastline" problem), which they are trying to solve with optical wafers.
Baker also points out that the disaggregation of inference into "prefill" and "decode" provides a richer canvas for chip design. Prefill is memory capacity-bound, while decode is memory bandwidth-bound. This disaggregation significantly extends the useful life of older GPUs (like Hopper, Ampere), as they can be used for prefill tasks. He believes this will be a major positive for private credit, as the financing term for GPUs could extend from 3-4 years to 10-15 years, thereby lowering financing costs.
Gavin Baker believes AI has "net destroyed value" in the application layer, and successful AI-native companies must focus on the "token path." He notes that, aside from a few companies like Cursor and Cognition, the AI application layer has not created value. Citing Jamin Ball's "token path" concept, he argues that software companies must be on this path (like Databricks), or they will struggle to survive unless they are in a highly vertical niche.
Regarding AI-native founders, Baker believes they face a fundamental challenge: is their idea obvious to the world before they build a scale advantage? He praises Cursor and Cognition for focusing on coding, citing Replit founder John Massad's view: "Coding is probably the shortest path to ASI and useful AI." He believes model companies might not enter overly niche markets, but founders must be wary that frontier model companies could quickly replicate their advantages through distillation techniques.
Baker also predicts that frontier model companies will face a new "prisoner's dilemma": whether to release models via API. If all frontier companies agree not to release, Chinese open-source models will fall behind; but once someone "defects" to generate revenue, others will be forced to follow. He believes NVIDIA might eventually launch its own frontier model to maintain control over the ecosystem.
| Ticker/Company | Analyst Stance | Key Data Points |
|---|---|---|
| Anthropic | Bullish | Added $11 billion ARR in a single month; burns ~80% less cash than OpenAI; if unconstrained by compute, URR (Unconstrained Revenue) could be $100-200 billion |
| OpenAI | Bullish | A "reference asset" alongside Anthropic; should be able to raise capital at higher valuations |
| NVIDIA | Bullish | Could sell $2-3 trillion in GPUs in 2026-27 if TSMC meets demand; valuation in early April was at its cheapest relative to the market in nearly 10-12 years |
| TSMC | Neutral (Key Variable) | Capacity decisions define the "Goldilocks zone" for avoiding a bubble; operates on a handshake, not a formal contract, with NVIDIA |
| Cerebras | Bullish (VC Perspective) | Did "different and hard" things (wafer-scale computing); succeeded only after three generations of chips; faces I/O bottleneck |
| Bullish | Has the most compute; TPU advantage is lost; YouTube data is valuable for robotics; GCP growth is strong | |
| Meta | Bullish | The only major internet company to truly transform into an AI-first company; Muse model is near the Pareto frontier |
| Amazon | Bullish | Trainium chip is an advantage; robotics in retail will drive P&L efficiency; Nova model is underrated |
| Microsoft | Neutral to Positive | Satya "flinched" in early 2025, losing allocation; now making correct but risky decisions (using compute internally vs. selling to OpenAI); Copilot underperforms due to compute constraints |
| SpaceX | Bullish | Operates the world's largest satellite fleet (98-99%); possesses the largest data center on Earth; has the strongest engineering team globally |
| Intel | Neutral | Provides process knowledge through Terrafab partnership; could break TSMC's supply discipline |
| Samsung | Neutral | Could break TSMC's supply discipline |
| AMD | Neutral | Performance of MI450 is unknown; has zero interaction with startups |
| Broadcom | Bullish | Is "everyone's favorite ASIC supplier"; deeply engaged with startups |
| Astera Labs | Bullish | Misclassified as a "copper loser"; its switch product definition makes it a winner |
| Cursor | Bullish | Focused on coding, achieved scale; one of the few successful AI application layer companies |
| Cognition | Bullish | Doing "truly different" things; achieved scale |
| xAI (Grok) | Bullish | Grok 4.3 is on the Pareto frontier; best and lowest-cost model among 500B parameter models |
1. Anthropic added more ARR in a single month than Palantir, Snowflake, and Databricks combined (Gavin Baker). Support: These three companies built their businesses over 10 years with thousands of employees, while Anthropic did it in one month – "something never seen before in the history of capitalism."
2. TSMC's capacity decisions are the single most important variable for judging the AI bubble (Gavin Baker). Support: If TSMC fully meets demand, NVIDIA could sell $2-3 trillion in GPUs, inevitably leading to oversupply; TSMC's supply discipline is the "Goldilocks zone" for avoiding a bubble.
3. Violating the "Bitter Lesson" is the biggest risk in AI investing, but ASI might temporarily violate it (Gavin Baker). Support: More compute always beats human algorithmic innovation. But when AI reaches ASI, it might first improve efficiency through self-optimization, temporarily "violating" the lesson.
4. New chip companies should pursue "different and hard" architectures, not try to build a "better GPU" (Gavin Baker). Support: NVIDIA, as a fast follower, can leverage scale advantages and TSMC relationships to crush competitors. Cerebras's wafer-scale computing is a success story, but requires the perseverance of three chip generations.
5. The "prefill/decode" disaggregation of inference extends GPU lifespan to 10-15 years and saves private credit (Gavin Baker). Support: Older GPUs can be used for prefill tasks, making them useful "until they melt." This extends financing terms from 3-4 years to 10-15 years, lowering financing costs from 7% to 5-6%, fundamentally changing the economics of compute financing.
6. AI has "net destroyed value" in the application layer; successful companies must be on the "token path" (Gavin Baker). Support: Even counting Cursor and Cognition, AI has destroyed trillions of dollars in application layer value. The highest value-creating companies today are those with "the most GPUs per human employee."
7. Frontier model companies will face a new "prisoner's dilemma": whether to release models via API (Gavin Baker). Support: If no frontier company releases, Chinese open-source will fall behind; but once someone "defects" for revenue, others are forced to follow. NVIDIA might eventually launch its own frontier model.
8. Elon Musk's Terrafab will succeed because it will attract the world's best engineers and the "A-team" from equipment suppliers (Gavin Baker). Support: Elon's reputation in hardware engineering, the partnership with Intel, and the ability to recruit talent using methods like "Taiwan City" make it unique.