Everyone's talking about the AI data center buildout like it's just a money problem. It's not. There are 5 distinct hard constraints stacked on top of each other — and solving one just reveals the next.
Why there is no way the US can unlock 100GW of AI compute by 2030.
First, the scale of the gap.
The US has ~50 GW of data center capacity online today. Bain and McKinsey both put demand at ~100 GW by 2030.
That means adding ~10 GW per year for 5 years straight. The record year so far? About 2.5 GW actually delivered. We need 4x that. Every year.
The US has ~50 GW of data center capacity online today. Bain and McKinsey both put demand at ~100 GW by 2030.
That means adding ~10 GW per year for 5 years straight. The record year so far? About 2.5 GW actually delivered. We need 4x that. Every year.
View Tweet
Constraint #1: Grid interconnection.
Lets assume this gets solved with an interuptible tariff.
Data centers get a "qualified" yes which requires them to curtail 100-200 hours a year. Match with VPPs and you could see 10% rate decreases across the country.
energyempirepodcast.substack.com/p/the-ai-power…
Lets assume this gets solved with an interuptible tariff.
Data centers get a "qualified" yes which requires them to curtail 100-200 hours a year. Match with VPPs and you could see 10% rate decreases across the country.
energyempirepodcast.substack.com/p/the-ai-power…
Constraint #2: Transformers and Electrical Equipment
Every data center — grid-tied OR behind-the-meter — needs large power transformers to distribute electricity internally. No substitute.
Lets assume this also gets solved by importing 100GVA of transformers from India.
techcrunch.com/2026/04/15/thi…
Every data center — grid-tied OR behind-the-meter — needs large power transformers to distribute electricity internally. No substitute.
Lets assume this also gets solved by importing 100GVA of transformers from India.
techcrunch.com/2026/04/15/thi…
Constraint #3: GPU chips.
Blackwell (NVIDIA's current AI GPU) is sold out through mid-2026 with a 3.6 million unit backlog. TSMC's advanced fabs are running above 95% utilization. NVIDIA has locked up ~70% of next year's CoWoS packaging capacity. Enterprise buyers wait 36-52 weeks.
New fabs take 5 years to build.
Blackwell (NVIDIA's current AI GPU) is sold out through mid-2026 with a 3.6 million unit backlog. TSMC's advanced fabs are running above 95% utilization. NVIDIA has locked up ~70% of next year's CoWoS packaging capacity. Enterprise buyers wait 36-52 weeks.
New fabs take 5 years to build.
Constraint #4: HBM memory.
Every AI GPU requires High Bandwidth Memory. SK Hynix has "already sold out our entire 2026 HBM supply." OpenAI's Stargate project alone is targeting 900,000 DRAM wafers per month — roughly 40% of all global DRAM output.
New fabs don't open until 2027-28.
Every AI GPU requires High Bandwidth Memory. SK Hynix has "already sold out our entire 2026 HBM supply." OpenAI's Stargate project alone is targeting 900,000 DRAM wafers per month — roughly 40% of all global DRAM output.
New fabs don't open until 2027-28.
Constraint #5: Skilled labor.
Electricians, plumbers, and data center mechanical trades are in severe national shortage.
They're also being competed for by: semiconductor fab construction, grid hardening, EV charging rollout, and factory reshoring.
Electricians, plumbers, and data center mechanical trades are in severe national shortage.
They're also being competed for by: semiconductor fab construction, grid hardening, EV charging rollout, and factory reshoring.
So what's the realistic number?
Lets assume we solve the grid/transformer challenges. The other constraints limit us to:
2026: ~3 GW added
2027: ~5 GW
2028: ~7 GW
2029: ~9 GW
2030: ~10 GW
Cumulative: ~34 GW added
Nowhere close to the numbers being thrown around right now.
energyempirepodcast.substack.com/p/the-ai-power…
Lets assume we solve the grid/transformer challenges. The other constraints limit us to:
2026: ~3 GW added
2027: ~5 GW
2028: ~7 GW
2029: ~9 GW
2030: ~10 GW
Cumulative: ~34 GW added
Nowhere close to the numbers being thrown around right now.
energyempirepodcast.substack.com/p/the-ai-power…
The meta-point:
This isn't a funding problem.
$650B+ in hyperscaler capex is committed for 2026 alone.
It's a physical supply chain problem — transformers, chips, memory, skilled workers — where every layer has multi-year lead times that can't be money'd away.
The buildings are the easy part.
This isn't a funding problem.
$650B+ in hyperscaler capex is committed for 2026 alone.
It's a physical supply chain problem — transformers, chips, memory, skilled workers — where every layer has multi-year lead times that can't be money'd away.
The buildings are the easy part.
Generated by Thread Navigator
Press ⌘ + S to quick-export