AI Power Infrastructure
The Grid Bottleneck. AI Labs Gave Up on Grid Power.
Last updated: 14 March 2026 · YK Research
Contents
March 2026 Update. Current Prices vs Thesis
Thesis intact. Grid interconnection is still THE binding constraint. BYOG is playing out exactly as predicted. xAI went from dirt to 100,000 GPUs in 4 months. Meta's comparable facility is still stuck in the interconnection queue.
GE Vernova (GEV)
$805.02
Howmet (HWM)
$236.75
Bloom Energy (BE)
$154.51
Caterpillar (CAT)
$693.99
| Ticker | Rating | Current Price | Key Thesis |
|---|---|---|---|
| GEV | STRONG BUY | $805.02 | Backlog understated. 62 GW in slot reservations. 3+ years of visibility. |
| HWM | STRONG BUY | $236.75 | Market calls it aerospace. It's actually an AI infrastructure bottleneck. |
| BE | BUY | $154.51 | Permitting arbitrage is underpriced. 3-12 months vs 18-36 months. |
| CAT | HOLD | $693.99 | Diversified. Solar Turbines optionality is unpriced. |
Core Thesis: Grid Interconnection Is THE Constraint
The bottleneck isn't GPUs. It's power. Grid interconnection takes 5-7 years.
The ERCOT Reality
ERCOT Requests
78 GW
ERCOT Approved
1.1 GW
Request:Approval Ratio
70:1
Avg Queue Time
5-7 Years
Why This Isn't Fixable Soon
- Real-time balancing: The grid matches supply and demand every second. Large loads require multi-year engineering studies.
- Prisoner's dilemma: Speculative requests clog the queues. 70% of AEP Ohio requests don't even have land control.
- Transmission bottleneck: New generation needs new transmission lines. NIMBY + permitting = years of delay.
- FERC Order 2023: Meant to speed things up. Implementation takes 3-5 years. The current backlog isn't going anywhere.
🔍
“FERC reforms could cut queue times in half. Doesn't matter. Demand outpaces supply by 10x. The grid is sold out. BYOG is structural.”
The $5.2B Speed Arbitrage
Unit Economics for 200 MW Datacenter
Revenue per MW
$11M/year
Annual Revenue
$2.2B
Monthly Revenue
$183M
Gross Margin
~95%
36-Month Scenario Analysis
| Scenario | Power Online | 36-Mo Cumulative Profit | Arbitrage Value |
|---|---|---|---|
| BYOG (Onsite Gas) | Month 6 | $5.2B | — |
| Grid Power | Month 36 | $0.17B | — |
| Arbitrage Delta | — | — | $5.0B |
Wrong Frame vs Right Frame
Wrong Frame
Comparing $/MWh:
- Grid: ~$60/MWh
- Onsite gas: $65-91/MWh
- “Grid is cheaper.”
Right Frame
$65-91/MWh NOW vs $0/MWh for 3+ years:
- 6-month acceleration = $5.5M/MW
- CapEx premium = $500K-1.5M/MW
- Payback: 1-3 months
🔍
“xAI went from zero to 100,000 GPUs in 4 months. Meta's comparable build is still waiting for grid interconnection. Speed wins.”
AI Power Demand Explosion
Demand vs Supply Gap
The Demand Trajectory
| Year | US AI Datacenter (GW) | YoY Growth |
|---|---|---|
| 2023 | 3 GW | — |
| 2024 | 8 GW | +167% |
| 2025 | 15 GW | +88% |
| 2026 | 28 GW | +87% |
| 2030 | 110 GW | +41% CAGR '26-'30 |
67% CAGR. 3 GW to 110 GW by 2030. The grid cannot add this capacity. BYOG is the only path.
BYOG Technology Stack
Technology Decision Framework
| Technology | CapEx ($/kW) | Lead Time | Efficiency | Best Use Case |
|---|---|---|---|---|
| Aeroderivative GT | $1,700-2,000 | 18-36 mo | 36% | Bridge power, N+1 backup |
| Industrial GT | $1,500-1,800 | 12-36 mo | 34% | Cost-effective baseload |
| RICE Engines | $1,700-2,000 | 15-24 mo | 40% | High redundancy (small units) |
| Fuel Cell (SOFC) | $3,000-4,000 | 3-12 mo | 52% | Urban / permitting-constrained |
| CCGT H-Class | $800-1,200 | 24-36 mo | 62% | Permanent GW-scale |
Bloom Energy's Permitting Arbitrage
Bloom's solid oxide fuel cells have no combustion. This changes everything for permitting:
| Permit Type | Gas Turbines | Bloom Fuel Cells |
|---|---|---|
| Air Quality (EPA) | Required | Not required |
| New Source Review | Required | Not required |
| Timeline | 12-36 months | 0 months |
🔍
“Fuel cells cost 50% more per MWh but deploy 70% faster. At permitting-constrained sites, Bloom wins on NPV despite higher LCOE.”
Stock-by-Stock Analysis
GE Vernova (GEV). STRONG BUY
Current Price
$805.02
Market Cap
~$220B
Total Backlog
$135.3B
Slot Reservations
62 GW
Variant Perception: The market treats GEV like traditional power equipment. It's actually the toll road on AI buildout.
- Key Win: 29x LM2500XPRESS units for Crusoe/OpenAI (July 2025)
- Capacity target: 24 GW/year by 2026. 20% above current.
- Capital return: $10B buyback authorized. Dividend doubled to $2/share.
- 62 GW slot reservations: Not in backlog. Pure upside to estimates.
Howmet Aerospace (HWM). STRONG BUY
Current Price
$236.75
Market Cap
~$95B
Backlog
$8.5B
Position
Near-monopoly
Variant Perception: The market sees an aerospace supplier. I see a monopoly bottleneck on AI power.
- Technical moat: Only Western source for high-temp single-crystal nickel superalloy blade castings at scale
- Supply chain logic: AI demand → Turbine orders → Blade casting demand → Howmet revenue
- Critical dependency: GEV's 62 GW slot reservations convert to revenue only as fast as Howmet delivers
- Customers: GE, Siemens, Mitsubishi, Doosan. Every turbine OEM.
Bloom Energy (BE). BUY
Current Price
$154.51
Market Cap
~$34B
Production Capacity
1→2 GW
Key Deal
$5B Brookfield
Variant Perception: High CapEx = bad. Wrong. Permitting speed = NPV positive despite higher LCOE.
- Permitting arbitrage: 3-12 month deployment vs 18-36 months for gas turbines
- Urban advantage: No combustion means no EPA/air quality permits needed
- Key customer: Oracle partnership for urban AI datacenters
- Production expansion: Doubling capacity to 2 GW/year
Caterpillar (CAT). HOLD/BUY
Current Price
$693.99
Market Cap
~$340B
Solar Turbines
1-23 MW
2025 DC Orders
~2,000 MW
Variant Perception: Market sees construction equipment. Solar Turbines optionality is free.
- Diversified exposure: Reduces risk but dilutes AI upside
- Solar Turbines: 12-24 month lead times. Faster than GE aeros at 18-36 months.
- Capacity plan: 2.5x turbine production by 2030
Risk Matrix
| Risk | Severity | Probability | Impact on Thesis | Mitigant |
|---|---|---|---|---|
| Gas >$10/MMBtu sustained | HIGH | 15% | TCO flips negative. BYOG economics break. | Currently ~$3/MMBtu with forward curve flat. Long-term supply contracts hedge. |
| FERC 2023 accelerates queues | MEDIUM | 30% | Grid power arrives faster. BYOG premium shrinks. | 3-5 year implementation. Current backlog doesn't clear. Demand grows 10x faster. |
| Chip efficiency 10,000x | HIGH | 10% | Power demand collapses. Entire thesis dies. | Historically offset by capability expansion. Jevons paradox applies. |
| China rare earth controls | MEDIUM | 25% | Turbine blade production constrained. HWM supply chain disrupted. | Western diversification underway. HWM building inventory positions. |
| GEV execution risk | MEDIUM | 20% | Capacity ramp fails. Backlog conversion slows. | Slot reservations de-risk. Pricing power cushions delays. |
Scenario Analysis
Bear Case (25%)
GEV -20%
Gas spike, demand slows
Base Case (50%)
GEV +50%
Thesis plays out
Bull Case (25%)
GEV +100%
Demand accelerates
Investment Framework
The Core Insight
The new picks and shovels aren't chips. They're turbines.
Portfolio Allocation
| Ticker | Rating | Position Size | Key Thesis |
|---|---|---|---|
| GEV | STRONG BUY | Core (5-8%) | Blue-chip BYOG. $135B backlog. 62 GW slot reservations. |
| HWM | STRONG BUY | Core (4-6%) | Hidden monopoly. Misclassified. Rerating coming. |
| BE | BUY | Meaningful (2-4%) | Permitting arbitrage. Higher vol, higher reward. |
| CAT | HOLD/BUY | Moderate (1-2%) | Diversified. Defensive. Solar optionality is free. |
What I'm Watching
Green Flags (Add)
- GEV backlog keeps expanding
- New hyperscaler BYOG deals announced
- HWM blade orders accelerate
- BE wins more Oracle-style partnerships
Red Flags (Reduce)
- Natural gas spikes above $8/MMBtu
- Grid queue times improve meaningfully
- Hyperscalers cancel BYOG projects
- GEV stumbles on execution
🔍
“The grid is sold out. BYOG is structural. The market prices these as industrial equipment companies, not AI infrastructure bottlenecks. That's the opportunity.”