Sep 25, 2026 · 19 min
Listen from 13:15
Listen at 13:15
AI’s chip glut may be hiding in plain sight
Monologue: The AI Data Center Overbuild
If hundreds of billions of dollars in AI hardware cannot run, the infrastructure boom may be building capacity before demand exists.
- 1Roughly $200–$300 billion in AI GPUs may be idle in warehouses or data centers without enough power.
- 2Ed Zitron argues hyperscalers such as Microsoft overstate how much AI capacity is actually operational.
- 3The resulting data-center overbuild could create a glut more severe than the dot-com-era fiber surplus.
Don't miss
Zitron’s estimate that $200–$300 billion in AI GPUs may be idle reframes the boom as a possible hardware glut.
The brief
Ed Zitron opens an investigation with The Guardian into where the AI boom’s chips actually are, finding hardware in warehouses and data centers that lack enough power to operate it.
The central tension is between the scale of AI infrastructure spending and the amount of computing capacity that is genuinely online, with Microsoft among the hyperscalers facing scrutiny.
Zitron estimates that roughly $200–$300 billion in AI GPUs may be sitting idle, turning a supposed capacity surge into a question about stranded hardware.
The warning extends beyond one company: a data-center overbuild could produce an infrastructure glut worse than the dot-com-era fiber surplus.
What was said on this episode
32 statements · 26 negative · 6 neutral
AI chips are primarily stored in warehouses or unpowered data centers.
“the answer is either in warehouses or unpowered data centers”
Listen at 4:46
Microsoft had 2.2 million GPUs in service, below common estimates
“Microsoft only had 2.2 million GPUs in service, far less than people believed”
Listen at 4:55
Microsoft’s chips represented approximately 1.993 gigawatts of capacity.
“the total power of the chips was estimated at about 1.993 gigawatts of capacity”
Listen at 5:04
The installed Microsoft GPUs cost approximately $50–60 billion.
“the overall cost of those GPUs installed somewhere in the region of $50 to $60 billion”
Listen at 5:09
Up to $106 billion of AI hardware is currently unused or warehoused
“I estimate that as much as $106 billion worth of GPUs and associated hardware are now sitting either unpowered in data centers, or in complete data centers, or in warehouses”
Listen at 5:45
Construction in Progress includes uninstalled GPUs and incomplete data centers
“companies bury their uninstalled GPUs and incomplete data centers”
Listen at 6:31
Named technology companies collectively report about $374 billion in construction in progress
“Google, Meta, Oracle, Amazon, SpaceX and Tesla, neoclouds like CoreWeave and IREN, and co-location companies like Core Scientific and Applied Digital all have about $374 billion of construction in progress”
Listen at 6:39
Google accounted for approximately $122 billion of construction in progress.
“A large chunk of that larger CIP number is Google's, which sits around $122 billion”
Listen at 7:08
Google accounts for approximately $122 billion of construction in progress
“Google's, which sits around $122 billion”
Listen at 7:11
Total AI construction in progress is estimated at $400–$500 billion
“the total construction and progress number is somewhere between $400 and $500 billion”
Listen at 8:00
Approximately half of sold AI hardware is currently warehoused
“approximately 50% of all AI hardware and chips that have been sold is being warehoused”
Listen at 8:07
Warehoused AI hardware will take over two years to deploy
“will take more than two years to fully ingest”
Listen at 8:14
NVIDIA generated $200–300 billion from hardware sold 24–36 months ahead.
“I think NVIDIA has made somewhere between $200 and $300 billion in revenue on stuff that's been sold probably 24 to 36 months in advance”
Listen at 8:27
NVIDIA generated $200–$300 billion from sales made 24–36 months early
“NVIDIA has made somewhere between $200 and $300 billion in revenue on stuff that's been sold probably 24 to 36 months in advance”
Listen at 8:27
Anthropic and OpenAI purchasing most capacity causes AI-compute scarcity
“any scarcity around AI compute is a result of most of the capacity being sold immediately to Anthropic and OpenAI”
Listen at 8:35
Available AI capacity is substantially below common perceptions
“the available AI capacity is much, much, much smaller than we think”
Listen at 9:09
Microsoft’s difficulty bringing AI capacity online implies others face similar difficulties
“if Microsoft, one of the most well-capitalized and experienced data center developers in the world, is having trouble bringing the majority of its AI capacity online, everyone is”
Listen at 9:23
If Microsoft struggles to bring AI capacity online, other companies likely do too.
“if they're failing, I think everyone is”
Listen at 9:49
Actual AI capacity coming online is likely several gigawatts below expectations.
“very little actual capacity is coming online, likely gigawatts less than people believe”
Listen at 9:59
Almost all newly available AI capacity flows to Anthropic and OpenAI
“almost all of it is flowing directly to Anthropic and OpenAI”
Listen at 10:03
Anthropic and OpenAI comprise 50–75% of backlogs but cannot pay for them
“anywhere from half to 75% of those backlogs are directly from Anthropic and OpenAI, neither of whom can actually pay for them”
Listen at 10:29
Anthropic and OpenAI cannot actually pay for their contracted infrastructure backlogs.
“neither of whom can actually pay for them”
Listen at 10:34
OpenAI’s compute spending represented 70% of Microsoft’s fiscal-2026 AI revenue
“OpenAI's compute spend, which made up 70% of Microsoft's AI revenues in fiscal year 2026”
Listen at 11:08
Actual AI capacity additions are at most tens or hundreds of megawatts quarterly
“the actual AI capacity is, at best, coming on at a rate of maybe tens or hundreds of megawatts a quarter”
Listen at 12:22
The AI infrastructure situation is much worse than the dot-com bubble
“this is magnitudes worse than the dot-com bubble”
Listen at 12:47
Planned AI infrastructure capacity is 5–50 times current demand
“there's anywhere from 5 to 50 times the amount of capacity in planning than there is demand for it”
Listen at 13:03
AI infrastructure will not be used at the scale being built
“There is no situation where this is going to be used at the scale it's built”
Listen at 13:15
The AI industry may not progress beyond Vera Rubin or Feynman GPUs
“I don't think we make it past Vera Rubin or Feynman”
Listen at 13:34
There will be no post-bubble GPU economy at the current construction scale
“there is no post-bubble economy for GPUs, at least at the scale they're being built”
Listen at 13:49
AI inference economics resemble airlines with high fixed costs.
“the cost of serving inference is similar to an airline with high fixed costs”
Listen at 13:54
AI operators must purchase hundreds or thousands of GPUs regardless of realized demand.
“You as an operator must buy capacity, hundreds or thousands of GPUs, over a certain period of time, whether or not you actually have the demand”
Listen at 13:59
Excess GPU capacity relative to forecast demand destabilizes service and forces customer rejection
“Anything above the demand you foresaw means an unstable and unreliable service and turning customers away at the door”
Listen at 14:12
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.