The $650B
Capex Race
Who’s spending what, and is it the next fiber-optic bubble?
agent37.club
The Spending Explosion
Top hyperscalers: aggregate capex. Goldman Sachs projects $1.15 trillion from 2025–2027 — more than double the prior three years.
~75% of this spending is for AI infrastructure specifically.
Who’s Spending What (2026)
| Company | 2025 Capex | 2026 Guidance | Trajectory |
|---|
| Google | ~$93B | $175–185B | Nearly doubled. $240B cloud backlog. |
| Amazon | ~$131B | $200B | Stock fell 8% on the announcement. |
| Meta | ~$69B | $115–135B | $600B US data centers by 2028. |
| Microsoft | ~$80B | ~$145B | Q1 alone: $34.9B. Doubling DC footprint. |
| Oracle | ~$21B | ~$50B | Raising $45–50B in debt & equity. |
Plus: Stargate Project (OpenAI + SoftBank + Oracle): $500B committed, 7 GW planned capacity, 25,000+ jobs.
CoreWeave — The GPU Cloud Upstart
- IPO: March 2025, $40/share, raised $1.5B
- Stock hit $187 before correcting to ~$74
- Revenue: $229M (2023) → $1.92B (2024) → $3.52B (mid-2025)
- Q3 2025: $1.36B revenue (+133% YoY)
- Backlog: $55.6B (~10 years of visibility)
- 45,000 GPUs — largest private GPU provider in North America
- ~590 MW of connected power
The risk
Massive debt load relative to revenue. Key customer concentration (Microsoft). GPU depreciation mismatch: booking 5–6 year life for hardware that may be obsolete in 2–3 years. Michael Burry flagged this publicly.
The bull case
$55.6B backlog provides long visibility. GPU-as-a-service model captures demand from companies that can’t build their own data centers.
Data Center Economics 101
Build cost per MW
- National average: $11.7M per MW of critical load
- Range: $9.3M (San Antonio) to $15M (Reno)
- MEP systems: up to 50% of the budget
- Land: $244K/acre avg, up 23% YoY for large parcels
PUE (Power Usage Effectiveness)
- Industry average: 1.54–1.56 (46-56% overhead)
- Google: 1.09 (best-in-class)
- AWS: 1.15
- Microsoft: 1.16
TCO breakdown
| Component | % of Cost |
|---|
| Servers & IT equipment | ~40% |
| Power infrastructure | ~20% |
| Cooling systems | ~20% |
| Land, core & shell | ~20% |
82% of infrastructure cost is power distribution + cooling.
The Depreciation Mismatch
GPU hardware
- Physical lifespan under AI workloads: 1–3 years
- Google architect estimate: GPUs at 60–70% utilization survive 1–2 years
- Technology obsolescence: 2–3 years (NVIDIA ships new arch annually)
- Accounting depreciation: companies book 5–6 year useful life
- Amazon reduced server lifespan from 6 to 5 years
Data center building
- Operational life: 15–20 years
- Lease terms: typically 25 years
- Steel frame: can last 60 years
- The building outlives its contents by 5–10×
The question
If GPUs last 2 years but are depreciated over 6, are AI companies overstating earnings? This is the bear case in a single number.
The Liquid Cooling Transition
Why air cooling hit the wall
- Traditional rack: ~7 kW
- Air cooling limit: ~40–70 kW per rack
- NVIDIA GB200 NVL72: 132 kW per rack
- Next gen (expected <1 year): 240 kW per rack
- Rack density up 38% from 2022 to 2024
Cooling types
Direct-to-Chip
Cold plates on CPUs/GPUs. Removes 70–80% of heat. Up to 300 W/cm². Standard for AI by 2025–2026.
Immersion
Servers submerged in dielectric fluid. Up to 1,500 W/cm². Mainstream by 2027.
GB200 NVL72: 72 GPUs, 3,000 kg per rack, 13.5 TB HBM3e. Mandatory liquid cooling — deviation triggers 60% performance throttling.
The Physical Constraints
Power — the #1 bottleneck
- Grid access, not land, kills projects
- US power demand accelerating to 2.6% growth (was near-zero for decades)
- 2,600+ GW proposed generation waiting for grid connection
- Median interconnection wait: >4 years (was <2)
- Texas CenterPoint: 700% increase in large load requests in one year
Transformer shortage
- Large power transformer lead time: 128 weeks (~2.5 years)
- 30% supply deficit in 2025
- Each weighs 100,000+ lbs, costs up to $250K
- Example: two Silicon Valley DCs built but can’t operate — no transformers
Water
- Large data center: up to 5M gallons/day
- Water usage projected to increase 870%
Cloud GPU Pricing (2026)
| GPU | Major Cloud (On-Demand) | Boutique Providers | Trend |
|---|
| H100 | $3–4/hr | $1.49–2.99/hr | Falling to sub-$2 by mid-2026 |
| H200 | ~$10.60/hr | $3.80/hr | Softening as B200 ships |
| B200 | Up to $5.87/hr | Limited avail | Premium but 4–5× perf/$ vs H100 |
| A100 | Sub-$1/hr | Near commodity | Becoming nearly free |
Key event: June 2025 — AWS announced ~44% price cut on H100 instances. Each new GPU generation pushes the prior one 50%+ cheaper within 12 months. Spot instances: 40–60% cheaper than on-demand.
Bubble or
Bonanza?
Is this the fiber-optic bust all over again?
The 2000 Telecom Parallel
What happened
- Late 1990s: >$500B invested in fiber optic infrastructure
- Financed with debt, based on 1,000% annual traffic growth projections
- Actual traffic: ~100% growth/year (impressive, but not enough)
- By 2001: 95% of fiber was “dark” (unused)
- Telecom stocks lost >$2 trillion in market cap
But eventually…
- The overbuild drove bandwidth costs to near-zero
- Enabled: broadband, streaming, cloud computing, smartphones
- The infrastructure was early, not wrong
- Companies that built it went bankrupt; companies that used it became trillion-dollar platforms
The question: are today’s hyperscalers the builders who go bust, or the platforms who profit?
Bear Case vs Bull Case
Overbuild risk
- Bain: even aggressive adoption = $1.2T revenue vs $2T+ capex spending gap
- GPU depreciation mismatch (5-6 yr book, 2-3 yr real)
- Revenue hasn’t justified spending levels yet
- Amazon stock fell 8%, Google stock sank on spending announcements
- CoreWeave: massive debt relative to revenue
Not a bubble
- Power constraints prevent overbuild — can’t build if you can’t get power
- Contracted demand: Google $240B backlog, CoreWeave $55.6B
- Revenue is real — self-funding companies, not debt-fueled startups
- Supply-constrained, not demand-constrained
- NVIDIA CEO: $600B annual demand from DC customers
- AI compute demand growing ~4×/year
Why This Time Might Be Different
| Factor | 2000 Telecom | 2025 AI Infrastructure |
|---|
| Financing | Leveraged debt, startup equity | Self-funded from cash flow (mostly) |
| Revenue | Speculative, minimal | $Billions in cloud revenue already |
| Demand visibility | Projected, uncontracted | $240B+ contracted backlog (Google) |
| Build constraint | Could build freely (just lay fiber) | Power-constrained, transformer-short |
| Customer base | Undeveloped internet | Every Fortune 500 deploying AI |
| Key risk | Demand didn’t materialize fast enough | GPU depreciation + timing mismatch |
The nuance: infrastructure will likely prove valuable long-term (as fiber did), but near-term returns may disappoint. The risk is more about GPU asset stranding than total demand destruction.
The Investment Map
NVIDIA (NVDA)
92% discrete GPU share. Q3 FY26: $57B revenue (+62% YoY). Every dollar of capex flows through NVIDIA first.
Vertiv (VRT)
Pure-play DC infrastructure. Q4 orders +252% YoY. $15B backlog. Liquid cooling + power management.
Eaton (ETN) / Schneider (SBGSY)
Power management & distribution. Eaton: ~70% revenue from electrical. Schneider: more diversified.
Data Center REITs
Equinix (EQIX): Top 10 analysts rate “strong buy.” Doubling capacity by decade end. Digital Realty (DLR): FFO up 17% YoY.
Utilities
175% data center power demand growth by 2030. Constellation Energy (nuclear), Vistra (Texas), Dominion (Virginia).
Supply chain
Corning (fiber), Arista (networking), Quanta Services (construction), EMCOR Group.
The Liquid Cooling Opportunity
Key players
- Vertiv: Market leader, acquired CoolTera. Launched MegaMod HDX (2026)
- CoolIT Systems: Direct-to-chip cold plates, 300 W/cm²
- GRC: Immersion cooling pioneer, partnered with Vertiv
- Asetek, LiquidStack: Specialized cooling solutions
Every new AI GPU requires liquid cooling. There is no air-cooled path forward for AI training racks.
Stargate — The Largest AI Infrastructure Project
- $500B commitment over 4 years
- Joint venture: OpenAI (operations) + SoftBank (finance) + Oracle + MGX
- Announced January 21, 2025, by President Trump
- Flagship: Abilene, Texas + 5 additional sites
- Nearly 7 GW of planned capacity
- 25,000+ onsite jobs
- Already $400B+ committed for first 3 years
For perspective
7 GW = the output of 7 large nuclear reactors. More power than many small countries consume. A single facility could power a city of 500,000 people.
The skepticism
Where does 7 GW of power come from? Grid interconnection takes 4+ years. SoftBank’s “Vision Fund” track record is mixed. Is $500B actually realistic?
NVIDIA — The Toll Booth
- 92% discrete GPU market share (H1 2025)
- ~90% of AI chip market overall
- Q3 FY2026: record $57.0B revenue (+62% YoY)
- Data center: $51.2B (+66% YoY)
Revenue trajectory (DC segment)
Annual product cadence
- Hopper (H100, 2022)
- Blackwell (B200/GB200, 2024)
- Rubin (2026)
- Rubin Ultra (2027)
Every hyperscaler dollar flows through NVIDIA. The picks-and-shovels play in the AI gold rush. But competition is coming (AMD, custom silicon).
Numbers Worth Knowing
| Fact | Number |
|---|
| 2026 aggregate hyperscaler capex | ~$650B |
| Dark fiber in 2001 bust | 95% unused |
| NVIDIA GB200 NVL72 power per rack | 132 kW (vs ~7 kW traditional) |
| Power transformer lead time | 128 weeks (~2.5 years) |
| Texas grid interconnection increase | 700% in one year |
| GPU real lifespan vs book depreciation | 2–3 years vs 5–6 years |
| Google cloud backlog | $240B contracted |
| NVIDIA GPU market share | 92% |
| GB200 NVL72 rack weight | 3,000 kg (6,600 lbs) |
| Stargate project total commitment | $500B |
| Vertiv Q4 order growth | 252% YoY |
| Large data center water usage | Up to 5M gallons/day |
Key Takeaways
- $650B in 2026 — the largest infrastructure buildout since railroads
- Power is the real bottleneck — not money, not demand, not land
- Liquid cooling is mandatory — no air-cooled path forward for AI racks
- GPU depreciation is the hidden risk — 2-3 year real life, 5-6 year book life
- Not exactly 2000 telecom — self-funded, contracted demand, power-constrained
- But timing risk is real — near-term returns may disappoint even if long-term demand is right
- Every dollar flows through NVIDIA — 92% market share, annual product cadence
- Supply chain beneficiaries — Vertiv, Eaton, utilities, REITs, networking
$650B per year.
132 kW per rack.
128 weeks for a transformer.
The physical world is the real bottleneck.
agent37.club