Gavin Baker is the founding partner and CIO of Atreides Management, joining host Patrick O'Shaughnessy for their sixth conversation on Invest Like the Best. Baker brings deep expertise in technology investing and historical market analysis to discuss the current AI revolution.

The conversation centers on 'watts and wafers' - the two physical constraints Baker believes will dictate AI's next phase. On power, he expects shortages to ease in 2027-28 as new energy sources come online, with orbital compute providing long-term solutions. On semiconductor capacity, he views Taiwan Semi's production decisions as potentially the most important variable preventing an AI bubble.

Baker draws extensively from Technological Revolutions and Financial Capital by Carlota Perez to frame how foundational technologies create bubbles, while referencing The Bitter Lesson by Richard Sutton regarding the primacy of compute over algorithmic innovation. The discussion covers everything from Anthropic's unprecedented growth to Elon's Terafab plans and the geopolitical implications of AI dominance.

March 2024: The Most Extraordinary Moment in Capitalism

Anthropic added $11 billion ARR in one month, equivalent to combining Palantir, Snowflake, and Databricks - companies that took 10+ years and thousands of employees to build their businesses.

"Nothing like that has ever happened in the history of capitalism. Forget my career. Just the flat out history of capitalism, the history of business. It's wild" - Gavin on Anthropic's growth.

The market sold off during this unprecedented AI acceleration, creating what Baker saw as a clear buying opportunity for those focused on fundamentals rather than macro fears.

DeepSeek's release actually increased compute demand dramatically, with GPU prices doubling in Asian availability zones and rental prices rising globally within days.

Watts: Power Constraints and Orbital Solutions

The watts shortage will likely begin alleviating in 2027-28 as new energy sources come online, though regulatory and zoning approvals have become bigger constraints than energy and chips.

Orbital compute consists of individual racks in space connected by lasers, not massive floating data centers - "A Blackwell rack weighs 3,000 pounds, is 8 feet high, 4 feet deep, 3 feet wide."

SpaceX operates 98-99% of all satellites in orbit and already runs 20-kilowatt Starlink satellites, making the jump to 100-120 kilowatt compute racks feasible.

Terrestrial data centers will remain valuable for training, while orbital compute will excel at inference workloads, with America consuming "as much compute as we can."

Wafers: Taiwan Semi's Bubble-Prevention Power

Taiwan Semi's capacity decisions may single-handedly prevent an AI bubble - "If Taiwan Semi did what Jensen wanted, NVIDIA could sell $2 trillion of GPUs in 26 or 27."

Drawing from Technological Revolutions and Financial Capital, Baker expects bubbles for foundational technologies but notes key differences: current buildout is funded by operating cash flows, not debt.

NVIDIA and Taiwan Semi operate without contracts, doing business on handshakes and what "seems fair" - a remarkable partnership built on trust and mutual benefit.

Elon's Terafab represents a joint SpaceX-Tesla venture to build America's largest semiconductor fab, recruiting top global talent with dedicated cultural enclaves.

The Bitter Lesson and Frontier Model Economics

The Bitter Lesson by Richard Sutton states that compute and data always outperform human algorithmic ingenuity, but this principle faces its biggest test with superintelligent AI.

Frontier tokens continue capturing the overwhelming majority of economic value at the model layer, surprising many who expected commoditization.

AI models have shifted to usage-based pricing, with consumer plans now "severely rate limited" - "To understand what Frontier AI is capable of today, you need to be on an enterprise plan."

The Pareto frontier for intelligence vs. cost is now dominated by Anthropic, OpenAI, and XAI's Grok, with Google losing its previous cost leadership position.

Chip Innovation and the Disaggregation Revolution

New chip companies must do something "different and hard" rather than trying to build better GPUs - "1% market share is going to be worth $100 billion."

The disaggregation of pre-fill (loading the cannon) and decode (firing) creates new opportunities for specialized chip architectures with different trade-offs.

GPU useful lives will extend to 10-15 years due to disaggregation, potentially saving private credit markets struggling with SaaS loan markdowns.

Cerebras exemplifies the "different and hard" approach with wafer-scale computing, taking three chip generations to achieve success through persistent engineering.

Big Tech's AI Positioning and Strategic Choices

Google maintains the largest compute installed base but lost TPU cost advantages, making this week's Google I/O a critical test of their frontier capabilities.

Zuckerberg deserves "immense credit" for making Meta truly AI-first internally, with Muse representing impressive progress from their AI research lab.

Microsoft's Satya Nadella went "from we're going to make Google Dance to being the product manager of Copilot in like three years" but is making courageous resource allocation decisions.

Amazon and NVIDIA show the most startup engagement by far, while AMD, Microsoft, and Meta have "essentially zero engagement" with the emerging ecosystem.