How Will Watts and Wafers Shape AI Infrastructure?

TL;DR

AI infrastructure growth depends on securing enough electricity, chips, and compute capacity to meet rapidly rising demand. Gavin Baker argues that Anthropic’s extraordinary revenue expansion, reasoning models’ heavy inference requirements, and constrained GPU availability support continued investment, while capital intensity, geopolitical uncertainty, application-layer struggles, and shifting competition among major technology companies remain significant risks.

Transcript

What was happening in AI was I think the most extraordinary moment in the history of capitalism, the history of American business. Anthropic they added 11 billion of ARR. The three highest profile SAS companies founded in the last 10 12 years are Palunteer, Snowflake and Data Bricks. And these three companies spent 10 years building their businesse... Read More

Key Insights

• Anthropic’s reported growth is presented as historically unusual because it added $11 billion of annual recurring revenue in one month. Baker compares that increase with the combined businesses that Palantir, Snowflake, and Databricks spent roughly a decade building.
• Market drawdowns can create opportunity when business fundamentals remain strong and an investor deeply understands the affected companies. Baker viewed the March decline this way because AI adoption was accelerating while technology stocks were selling off and becoming relatively inexpensive.
• Reasoning models are more compute-hungry during inference than non-reasoning models. After DeepSeek, Baker observed GPU availability falling and rental prices rising, with prices in Asian AWS availability zones reportedly doubling before similar increases appeared in the United States.
• The key constraints on AI infrastructure are watts, wafers, and available compute. The episode frames electricity supply, semiconductor manufacturing, GPU lifespans, private credit, new chip companies, and even orbital data centers as parts of capitalism’s effort to expand capacity.
• Anthropic is described as more capital-efficient than OpenAI because its cost per token and cumulative cash burn appear materially lower at a roughly similar revenue scale. Baker estimates that Anthropic may have burned about 80 percent less, while acknowledging OpenAI secured more compute.
• Compute scarcity can directly restrict model quality and revenue. Baker says Claude was generating 70 percent fewer tokens for the same question, including on Opus, and argues that token quantity influences answer quality and thinking alongside the intelligence density of each token.
• Investor-friendly fundraising can preserve long-term access to capital. Baker points to Elon Musk’s practice of avoiding overly aggressive valuations, arguing that consistently making investors money can create the ability to raise substantial funding whenever uncertain conditions or expansion plans require it.
• The AI opportunity includes major risks beyond infrastructure capacity. The conversation covers application-layer struggles, open-source dynamics, cybersecurity, competition among Google, Meta, Amazon, and Microsoft, geopolitical risks surrounding AGI, and the potential for AI to improve biotechnology and extend human life.

Questions & Answers

Q: Why does Gavin Baker consider AI growth historically unusual?

Gavin Baker considers the growth historically unusual because Anthropic reportedly added $11 billion of annual recurring revenue in one month. He compares that increase with Palantir, Snowflake, and Databricks, three prominent software companies that employed thousands of people and spent about ten years building their businesses. The contrast illustrates the extraordinary speed and scale of current AI demand.

Q: How do watts and wafers constrain AI infrastructure?

Watts and wafers represent two central physical constraints on AI expansion. Watts refer to the electricity required to operate compute infrastructure, while wafers refer to the semiconductor manufacturing capacity needed for chips. The discussion also connects these constraints to GPU availability, new chip companies, manufacturing, longer GPU lifespans, private credit, and possible data centers in space.

Q: Why did DeepSeek increase the outlook for compute demand?

DeepSeek revealed how much more compute-intensive reasoning models can be during inference than non-reasoning models. Although AI-related stocks initially sold off, Baker observed GPU availability declining, GPU rental prices rising, DRAM prices increasing sharply, and prices in Asian AWS availability zones reportedly doubling. He interpreted those signals as evidence of stronger compute consumption rather than collapsing infrastructure demand.

Q: How does Anthropic compare with OpenAI on capital efficiency?

Anthropic is described as having a materially lower cost per token and using much less capital to reach a roughly similar revenue scale. Baker suggests Anthropic may have burned about 80 percent less money than OpenAI. OpenAI, however, has secured more compute and is pursuing multiple initiatives intended to improve its position, creating a different capital and infrastructure profile.

Q: How can limited compute affect an AI model’s performance?

Limited compute can force a company to reduce the number of tokens a model generates, which may weaken the depth or quality of its answers. Baker cites an analysis claiming Claude, including Opus, was producing 70 percent fewer tokens for the same question. He argues that token quantity affects answer quality and thinking, although intelligence density per token also matters.

Q: Why might AI companies avoid raising money at the highest possible valuation?

AI companies may choose a more moderate valuation to preserve strong investor returns and maintain dependable access to future capital. Baker argues that uncertain geopolitical conditions and the capital-intensive nature of AI make repeated fundraising strategically important. He cites Elon Musk’s investor relationships as an example of how fair pricing and sustained returns can create long-lasting financing flexibility.

Q: When can a technology market drawdown become a buying opportunity?

A drawdown can become an opportunity when an investor understands the affected companies, disagrees with the price movement, and sees business fundamentals strengthening rather than deteriorating. Baker says March fit this pattern because AI adoption was accelerating while the NASDAQ declined. He also believed technology had become unusually inexpensive relative to the rest of the market over the preceding decade.

Q: What risks could challenge continued AI infrastructure investment?

The discussion identifies capital intensity, constrained compute, electricity and chip supply, geopolitical uncertainty, cybersecurity, application-layer difficulties, open-source competition, and shifting relationships among major technology companies as important risks. It also raises geopolitical concerns connected to AGI and Taiwan. These issues can affect access to manufacturing, financing, infrastructure, customers, and the capacity required to serve rapidly expanding demand.

Summary & Key Takeaways

• Baker describes AI growth as an unprecedented business expansion, highlighting Anthropic’s reported addition of $11 billion in annual recurring revenue. He contrasts that single month with the decade Palantir, Snowflake, and Databricks spent building their businesses, arguing that the scale and speed of demand distinguish AI from earlier software cycles.

• The AI build-out is constrained by watts, wafers, GPU availability, and access to compute. Baker says DeepSeek initially triggered a market sell-off, but rising GPU rental prices, reduced availability, and more expensive DRAM suggested reasoning models were increasing inference demand rather than weakening the long-term case for infrastructure investment.

• The discussion connects AI infrastructure with valuations, capital efficiency, chip design, orbital compute, manufacturing, application-layer economics, open-source models, cybersecurity, and geopolitical risk. Baker distinguishes Anthropic from OpenAI on cost per token and capital consumption while emphasizing that uncertain conditions make continued access to investors and compute strategically valuable.