$7.6 trillion is being spent on AI infrastructure between 2026 and 2031 — and Nvidia is forecast to capture roughly 75% of the entire compute layer. Here is where every dollar goes, and the accounting gamble hiding underneath it.
David and Sophia break down Goldman Sachs’ map of the AI build-out: $5.1 trillion for chips, $2.1 trillion for data centers, and $358 billion for the power to turn them on. They explain how Nvidia’s software moat produced roughly 75% gross margins on $80,500 chips, why server racks jumping from 15 kilowatts to 500+ kilowatts are forcing a total shift to liquid cooling, and why Meta is locking up 2,600 megawatts of nuclear power for 20 years.
Then the uncomfortable part: the $1.76 trillion depreciation swing that hinges on one question — how long before an $80,000 chip becomes a brick? Michael Burry’s short thesis says profits are inflated by over 20%. CoreWeave’s rental data says 2020-era A100s still earn 95% of their original price. And underneath it all sits a closed loop of circular financing where the money never leaves the ecosystem.
Chapters
00:00 The 40-homes-from-one-outlet problem
01:56 How $7.6 trillion breaks down
02:51 How Nvidia cornered 75% of compute
04:56 Training vs inference: AMD’s agentic AI angle
06:52 From 15kW to 500kW racks: the liquid cooling shift
09:30 Power is the gatekeeper: Meta’s nuclear deal
10:54 The $1.76 trillion depreciation gamble
15:00 The trillion-dollar closed loop
17:32 Copilot loses $20-80 per user; OpenAI’s $14B loss
19:29 The paradox that could obsolete it all
Key points
– $7.6 trillion projected global AI infrastructure spend, 2026-2031; $765 billion flowing in 2026 alone
– Nvidia forecast to capture ~75% of the $5.1 trillion compute layer, at ~75% gross margins
– AI racks now demand 500+ kilowatts, pushing the liquid cooling market toward $15.75 billion by 2030
– A 3-year vs 7-year chip lifespan assumption swings industry depreciation costs by $1.76 trillion
– GitHub Copilot reportedly lost $20-80 per user monthly; OpenAI projected to lose $14 billion in 2026
More from TechDaily: https://techdaily.ai
Thumbnail photo: Frontier supercomputer, Oak Ridge National Laboratory / OLCF, CC BY 2.0, via Wikimedia Commons.
#AIInfrastructure #Nvidia #TechNews
