Bloom Energy: The AI Power Bottleneck Play

AI data center power demand is growing faster than any supply source can deliver, creating a structural shortage that only onsite generation can fill in the near term.

• → AI data centers expected to reach 8–12% of total US electricity by 2030, up from 3–4% today
• → NVIDIA GB200/GB300 racks require 130–140 kW today; next-gen Rubin architecture targets 600 kW–1 MW per rack by 2027, requiring 800-volt DC power delivery to avoid conversion losses
• → Grid interconnection timelines average 4+ years; new transmission takes even longer
• → 35 GW-scale data center projects announced in 2024 alone vs. 3 the prior year
• → 1/3 of data centers expected to be fully off-grid by 2030 (Bloom survey of 152 decision-makers)
• → Hyperscaler capex confirmed and accelerating: Amazon $200B, Google $175–185B for 2026

• ✓ Confirmed: Hyperscaler capex accelerating (Amazon $200B, Google $175–185B for 2026)
• ✓ Confirmed: Grid constraints real — FERC (Federal Energy Regulatory Commission) interconnection queues are multi-year
• ⚠ Assumption: AI capex continues without major temporal lumpiness (a 2–3 quarter pause in data center starts would crush near-term estimates, even if long-term demand is intact)

• ⚡ Bloom delivers onsite power in ~90 days from order to energization (proven in the Oracle deal, pre-approved sites, known utility jurisdictions, repeat customers)
• → GE Vernova has indicated constrained turbine availability through 2028, limited capacity in 2029. For data centers that need power now, there is no turbine option.
• → Nuclear SMRs (Small Modular Reactors) are not commercially viable at scale until the 2030s
• → Bloom SOFCs (Solid Oxide Fuel Cells) have capex on par with turbines but 15–20% lower fuel consumption
• → No combustion = lower emissions = faster local permitting
• → Bloom SOFCs natively output 800-volt DC — exactly matching NVIDIA’s next-gen rack architecture. No AC-to-DC conversion = lower losses, simpler infrastructure, fewer failure points
• → Section 48E ITC: 30% Investment Tax Credit effective January 2026 — makes Bloom fuel cells 30% cheaper to deploy

Our convergence engine flagged BE with multiple confirmed data sources spanning congressional trades, lobbying filings, institutional 13F filings, failure-to-deliver patterns, and federal jobs data. Direction: Bullish. The lobbying data is exceptionally strong for a company of this size.

Our proprietary supply chain graph maps Bloom’s position in the AI power infrastructure ecosystem. Bloom sits at the critical intersection of fuel cell manufacturing, policy tailwinds, and AI data center demand.

Bloom sits between the natural gas supply chain and hyperscaler AI deployments — converting fuel directly into 800V DC power that feeds GPU racks without conversion losses.

Bloom’s manufacturing chain has one critical single-source dependency: MTAR Technologies (India) manufactures the hot box assembly, which is the core SOFC component. This represents both a supply chain risk and a competitive barrier — the manufacturing expertise embedded in this relationship took years to build.

Key risk: MTAR is single-source. Any disruption (geopolitical, quality, capacity) directly impacts Bloom’s production timeline.

• → Product backlog: ~$6B (+140% year-over-year — 3× annual revenue). Record level.
• → Service backlog: ~$14B recurring, ~30% gross margin — every MW deployed = 10+ years of contracted service revenue
• → AEP exercised a $2.65B option for Bloom fuel cells (January 2026)
• → Brookfield $5B strategic partnership (October 2025)
• → 2026 guidance: $3.1–3.3B revenue (+54–63% year-over-year)

Their advantage: Massive installed base, proven at gigawatt scale, lower cost per MW at large scale. Default choice for utility-scale baseload power.

Why Bloom wins now (2026–2028): Supply is the constraint. GE Vernova has indicated constrained availability through 2028 with limited new capacity coming online in 2029. For data centers that need power now, there is no turbine option — Bloom delivers in ~90 days vs 18–24 months.

The 2029 risk: Once turbine supply loosens, Bloom must compete on cost — a fight it currently cannot win at gigawatt scale. The 2026–2028 window is the window to build enough installed base that operational switching costs matter.

Their advantage: Massive distribution network, competitive on speed with 2.5 MW gensets.

Why Bloom wins: Reciprocating internal combustion engines (RICEs) are less efficient, noisier, higher emissions, harder to permit for baseload. No combustion = siting flexibility and ESG (Environmental, Social, and Governance) compliance.

Risk: If Caterpillar captures the “good enough” segment, Bloom’s addressable market shrinks to premium only.

Their advantage: Zero emissions, extremely high energy density, 24/7 baseload for decades. The superior long-term solution if delivered.

Why Bloom wins now: SMRs are a 2030s story at the earliest. No commercially operational SMR exists for data center use today. NRC licensing is multi-year. Bloom is the bridge — and bridges collect tolls for as long as traffic flows.

The coexistence case: Even when SMRs arrive, Bloom and SMRs could coexist for years — Bloom handling rapid-deployment and incremental campus expansions while SMRs serve new greenfield mega-campuses. The real risk is SMR narrative compressing Bloom’s multiple before SMRs deliver a single watt.

Natural gas is Bloom’s primary fuel input. The thesis has an embedded gas price assumption: at current Henry Hub forward prices (~$3.50–4.00/MMBtu), Bloom’s levelized cost of energy (LCOE) beats grid power in key markets. Bloom’s 15–20% efficiency advantage over gas turbines provides a natural hedge — they can tolerate higher gas prices before economics deteriorate.

Current Henry Hub forward curve: ~$3.50–4.00/MMBtu through 2027 — well within the thesis comfort zone. Thesis-breaking threshold: $5.50/MMBtu sustained for 2+ consecutive quarters.