Link: https://youtu.be/mDG_Hx3BSUE

Duration: 151 min

Short Summary

This episode analyzes the massive capital expenditure and infrastructure expansion required for the AI industry, highlighting a $600 billion combined CapEx forecast for Amazon, Meta, Google, and Microsoft. Key speakers including Michael Burry and Elon Musk discuss the critical role of GPU depreciation cycles, memory constraints, and the manufacturing bottlenecks of EUV tools in shaping future compute capacity.

Key Quotes

1. "If you add up the big four—Amazon, Meta, Google, Microsoft—their combined forecasted CapEx this year that you published recently is $600 billion." (00:00:18)
2. "The cost to rent the compute that OpenAI and Anthropic will have this year to sustain their compute spend is $10 to $13 billion a gigawatt." (00:01:40)
3. "In a sense, Anthropic needs to get to well above five gigawatts by the end of this year." (00:05:55)
4. "Now, two years in, you're signing deals for two to three years at $2.40? Those margins are way higher. Now you can crowd out all of these other suppliers, whether Amazon had these, or CoreWeave, or Together AI, or Nebius, or whoever it is." (00:12:37)
5. "Ultimately by 2028 or 2029, the bottleneck falls to the lowest rung on the supply chain, which is ASML." (00:37:11)

Detailed Summary

AI Infrastructure and Capital Expenditure Analysis

Financial Landscape and Capital Expenditure

• The combined forecasted CapEx for Amazon, Meta, Google, and Microsoft is projected at $600 billion, with total American data center CapEx reaching roughly $1 trillion this year.
• Google's $180 billion CapEx includes spending on turbine deposits for 2028 and 2029, data center construction for 2027, and power purchasing agreements.
• OpenAI announced a significant capital raise of $110 billion, while Anthropic announced a $30 billion raise and added $4 billion to $6 billion in revenue over the last few months.
• Big Tech's capital expenditure in 2026 is projected to allocate 30% of the total $600 billion budget specifically towards memory requirements.

Compute Capacity and Hardware Evolution

• The cost to rent compute to sustain annual spend for OpenAI and Anthropic is estimated between $10 billion and $13 billion per gigawatt.
• Michael Burry argues that the depreciation cycle for a GPU is three years or less, impacting long-term deployment costs.
• An H100 GPU costs $1.40 per hour to deploy at volume over five years if the depreciation cycle is five years.
• By 2026, with Blackwell in high volume, the market value of an H100 is projected to fall to $1.00 per hour.
• Nvidia's Blackwell NVL72 implemented rack-scale scale-up where seventy-two GPUs connect to each other at terabytes per second, doubling previous memory capacity to twenty terabytes.

Manufacturing Constraints and Supply Chain

• Manufacturing one gigawatt of Nvidia's Rubin chip capacity requires approximately 55,000 wafers of 3 nm technology, alongside 6,000 wafers of 5 nm and 170,000 wafers of DRAM memory.
• ASML currently produces about 70 EUV tools, with production increasing to 80 next year and projected to reach slightly over 100 by the end of the decade.
• The cost of tooling required for one gigawatt is $1.2 billion, whereas the total economic CapEx for the data center is roughly $50 billion.
• Carl Zeiss, a critical optics supplier bottlenecking ASML, has a market capitalization of $2.5 billion and employs less than one thousand people specialized in lenses.

Market Dynamics and Strategic Decisions

• Nvidia is fracturing the neocloud industry by giving allocation to random neoclouds to prevent any single entity from controlling all compute.
• Google sold a million TPU v7 chips to Anthropic, while its own DeepMind lab lacked immediate access to this compute capacity.
• TSMC is prioritizing allocation for Amazon's Graviton CPU over its Trainium AI chip because the CPU business is viewed as more stable with long-term growth.
• Over 70% of AI data centers are currently located in America, while Australia, Malaysia, Indonesia, and India are seeing faster growth.

Power and Future Scalability

• Current critical IT capacity is 20-30 gigawatts, with a projection of 200 gigawatts by the end of the decade.
• Elon Musk targets 100 gigawatts of power in space annually by 2028 or 2029, while Sam Altman aims for 52 gigawatts by the end of the decade.
• Data centers are currently 3-4% of the US grid's power and are projected to reach 10% by 2028.
• A blockade of Taiwan that destroys fabs would reduce global incremental compute capacity from hundreds of gigawatts a year to approximately 10-20 gigawatts across Intel and Samsung.