Grid headroom turns negative → datacenters self-generate behind-the-meter → fuel-cell + gas-turbine OEMs (BE, GEV) capture a ~50GW/yr equipment TAM distinct from the genset bridge-power chain

The chain

1. US grid interconnection headroom is collapsing — "approaching zero, turns negative by 2027," with 3-4 year lead times on turbines and transformers blocking conventional interconnection (From 2026-06-25-feed-semianalysis-us-grid-constraints-40gw-behind-the-meter-datacenter)

2. Datacenters respond by securing fast on-site (behind-the-meter) power rather than waiting in the interconnection queue; fuel cells avoid combustion permitting and deploy in quarters rather than years — michael-cembalest in 2026-06-23-podcast-columbia-energy-exchange-michael-cembalest-does-the-math-on-the-energy: operators are "scrambling to get these Bloom fuel cells"

3. Behind-the-meter generation "powers well over half of new US datacenters by 2028," opening a ~50GW/yr equipment TAM by 2029 (From 2026-06-25-feed-semianalysis-us-grid-constraints-40gw-behind-the-meter-datacenter)

4. → Bloom Energy (BE, fuel cells) and GE Vernova (GEV, whose LM2500 turbines dominate the on-site filings) are the primary equipment beneficiaries — dual-named by both SemiAnalysis and Cembalest. BE order-flow is now first-party (2026-08-17). GEV "dominates filings" remains ⚠ unverified — ~20% of GEV's gas GW is data-center; Siemens' DC share is higher; aero is a bridge to HA. See 2026-08-17 update.

5. → Vistra (VST) is an adjacent beneficiary via nuclear co-location as datacenters chase firm primary power. Narrowed 2026-08-17: Comanche Peak is a 20-year 1.2 GW Amazon PPA (energize 2027, full 2032), not a BTM equipment order. Keep VST on the nuclear-PPA chain, not this OEM TAM.

Why it matters

This is a primary on-site generation TAM, structurally distinct from the existing ai-power-gap-to-genset-bridge-power chain (CAT/CMI gensets are the bridge/backup layer; this is the always-on equipment that becomes the datacenter's main power). If BTM genuinely powers more than half of new datacenters by 2028, the equipment OEMs capture a multi-year, ~50GW/yr order cycle that the market may still be pricing as a one-off backup-power story.

• Tradeable: BE (solid-oxide fuel cells, permit-light, fast-deploy), GEV (LM2500 aeroderivative gas turbines dominant in the on-site filings), VST (nuclear co-location adjacent).
• The asymmetry: Two independent sources (SemiAnalysis written analysis + Cembalest first-party) name BE/GEV specifically — and BE was +10.6% / GEV +5.6% on 2026-06-29, suggesting the market is starting to discount this but the TAM scale is still uncertain.

Why it may not work

• OEMs are deliberately under-expanding 10-20% to avoid an overbuild — which means if the demand projections (which Cembalest himself "takes the under" on) disappoint, the capacity that was added becomes stranded and pricing weakens.
• "Surging turbine availability" in secondary markets could soften GEV's near-term pricing power even if unit volume holds.
• Demand-projection risk: the whole chain rests on datacenter buildout continuing at projected pace; an AI-capex pause breaks step 3.
• Permitting / fuel-supply for fuel cells (natural gas hookups, hydrogen logistics) could prove slower than the "deployable in quarters" claim.

What to watch

To graduate this from hypothesis to active thesis:

1. BE / GEV order-backlog disclosures explicitly tied to datacenter behind-the-meter deployments (not generic utility orders)
2. ERCOT "Batch Zero" co-location buildout (effective July 2026) — the first concrete BTM co-location regime to track
3. GEV LM2500 unit shipments and book-to-bill specifically on datacenter accounts
4. Whether BTM share of new datacenters tracks toward the "well over half by 2028" claim in interconnection-queue data

Sources

• 2026-06-25-feed-semianalysis-us-grid-constraints-40gw-behind-the-meter-datacenter — primary (SemiAnalysis written analysis; grid headroom, BTM TAM, OEM naming)
• 2026-06-23-podcast-columbia-energy-exchange-michael-cembalest-does-the-math-on-the-energy — corroborating (Cembalest first-party on the Bloom fuel-cell scramble and demand "under")
• 2026-08-17-autoresearch-btm-self-gen-bloom-gev — weekday-shift order-flow test (BE first-party; GEV 20% DC / Electrification $5B; VST PPA correction)

Related

• ai-capex-to-power-and-materials-cascade — parent cascade (power as the binding constraint)
• ai-power-gap-to-genset-bridge-power — sibling chain (bridge/backup power; this page is the primary on-site counterpart)
• bloom-energy — beneficiary entity
• ge-vernova — beneficiary entity
• vistra — adjacent (nuclear co-location) entity

---

Update (2026-08-10) — a second, independent forcing function: the BTM exemption is now written into state policy, not just implied by physics

From 2026-08-10-autoresearch-datacenter-siting-gate-load-relocation.

This chain's step 1 was a physics/queue argument: grid interconnection headroom "approaching zero, turns negative by 2027." It now has a policy driver alongside it, from a completely different causal direction — and policy drivers move faster than queues.

Two state gates on new grid-connected data centers, twenty days apart, both explicitly exempt purely behind-the-meter on-site generation:

• Texas, 2026-08-03 — Abbott directs PUCT + ERCOT to audit all 1,800+ projects / 474 GW (~90% data centers) in the interconnection queue and halts approvals pending the result. Projects "utilizing purely 'behind the meter' on-site generation and not interconnecting through ERCOT may not be subject to this directive." ERCOT has paused its "batch zero" transmission planning study.
• New York, 2026-07-14 — EO 62 holds discretionary state environmental permits in abeyance for ≥50 MW data centers, with an unbounded end-condition. It does not reach local permits.

And the response is already observed rather than merely predicted: "some data centers are building on-site power generation that allows them to bypass traditional connection to the grid."

Why this is a materially different argument from the original step 1. The queue-headroom driver says BTM is faster; the policy driver says grid connection is unavailable, on an unbounded clock, in the two largest markets. A developer facing a 3–4 year turbine lead time can still choose to wait. A developer facing an open-ended permitting hold with an explicit off-grid carve-out has the choice made for them. That converts BTM from an optimisation into the only route, which is what would lift this from hypothesis toward a real chain.

⚠ Held honest — this does not yet move the un-cited gap. Steps 4–5 (the order-flow-to-ticker link for BE/GEV, and VST co-location) remain ⚠ unverified, exactly as written above. Today's evidence strengthens step 1–3 (why BTM demand exists and accelerates); it says nothing new about who books the orders. Two additional cautions:

• The audit could narrow the exemption rather than widen it. Abbott's audit explicitly verifies "whether data centers provide their own power or depend on ERCOT." That could favour BTM projects — or subject BTM claims to scrutiny. Unresolved; 2026-08-20 (ERCOT good-cause-exemption meeting) is the next dated read.
• The tape disagrees so far. If this channel were being priced, the on-site-generation names should be decoupling upward from grid-interconnection-levered EPC. CAT is currently −8.1% live sector-excess (performance-latest) — i.e. lagging alongside PWR, not separating from it. Either the channel is not yet in the tape or the chain is wrong; see siting-gate-plus-stranded-capacity-to-demand-side-beneficiary-rotation, which today corrected itself on exactly this point.

Added to What to watch

• BE / GEV / CAT / CMI order or backlog commentary explicitly attributing demand to siting or interconnection gates (rather than to generic "AI power demand") — that is the citation that would close this chain's gap, and the state gates make it a question management will now be asked directly.
• 2026-08-20 — ERCOT's good-cause-exemption meeting; first read on whether the BTM carve-out is being widened or narrowed.

Related (added 2026-08-10)

• state-datacenter-siting-moratorium-risk — the gate; the exemption is documented there in full.
• siting-gate-plus-stranded-capacity-to-demand-side-beneficiary-rotation — the sibling chain; note it lists CAT/CMI as disfavoured while this page lists them as beneficiaries. That contradiction is real and is now recorded on both pages — it resolves on whether a name's revenue requires grid interconnection (disfavoured) or supplies the off-grid alternative (favoured).

---

Update (2026-08-11) — the first measured on-site build rate in this wiki, and it cuts against GEV

From 2026-08-07-feed-semianalysis-spacex-10gw-in-2027-microsoft-largest-offtaker.

This page's standing gap is the order-flow-to-ticker link. Today's source supplies the first quantified on-site-generation build rate the wiki has held:

• SpaceX's Southaven power plant expanded from 27 turbines (~495MW) in February 2026 to 69 turbines (1.7GW) in July 2026 — a ~3.4× nameplate increase in five months at a single site.
• SpaceX's 2027 plan "will, naturally, extensively rely on onsite gas generation," and the company chose "to build an onsite generation plant 1km across the border to avoid permitting."
• Second forcing function confirmed: at 10GW of planned build, on-site generation is not a stopgap but the primary power strategy.

Steps 1–3 strengthen. Steps 4–5 do not — and one of them arguably weakens.

⚠ The source names no supplier for the Southaven turbines, and where it does mention GE Vernova it is to route around it: "Gas turbines are 5yr+ backlogged? GEV's are, but there are plenty of other options" — with 30+ alternate manufacturers and a surging secondary market. The largest identified on-site build in the source is therefore evidence that BTM demand is real and accelerating while being, if anything, mild evidence against GEV specifically capturing it. BE is not mentioned at all.

Net: the demand leg of this hypothesis is now the best-evidenced part of it, and the beneficiary leg is no better evidenced than it was on 2026-06-29 — and now has a named competitive leak. See constraint-bypass-by-capitalized-buyers.

---

Update (2026-08-17) — BE order-flow is first-party; GEV is a 20% mix, not an LM2500 monopoly. Do not graduate

From 2026-08-17-autoresearch-btm-self-gen-bloom-gev. Status hypothesis → narrowed. Did not re-litigate ai-power-gap-to-genset-bridge-power.

BE — the watch-item citation landed.

• Q2 2026 (July 28): revenue $1.065B (+165.5% YoY), product $935.4M (+215%), FY26 guide $3.9–$4.2B. kr-sridhar: "Bloom is now a standard for AI onsite power"; "all the major US hyperscalers and over a dozen US neoclouds, AI labs, and colocation data center operators have validated and approved" the platform. Customers "who traditionally defaulted to combustion technologies" are switching — the intended distinction from CAT/CMI.
• Oracle (2026-04-13): MSA up to 2.8 GW; 1.2 GW contracted and deploying in the US. 55-day first delivery vs 90-day schedule.
• Brookfield (2026-06-30): financing shelf $5B → $25B for "islanded" / "onsite" AI-infrastructure power. A shelf, not a 25 GW equipment order.
• AEP (2024-11-14): up to 1 GW, 100 MW initial, co-located at customer sites for AI loads.
• YE2025 backlog ~$20B total / ~$6B product (annual disclosure only). Q2: backlog "growing at a faster pace than revenue"; new customers not in the YE2025 print.

GEV — mix is real; "dominates BTM filings" is not closed.

• Q2 2026: gas equipment + SRAs 100 → 116 GW, YE target ≥125 GW; total backlog $176B. scott-strazik: ~20% of those GW are data-center, 80% traditional utility/IPP.
• The >$5B H1 "data center orders" are Electrification (grid kit: $2.7B in Q2), not Gas Power. Do not cite that print as LM2500 BTM TAM.
• Named aero BTM order is still Crusoe 29× LM2500XPRESS (~1 GW, July 2025) — "power plants alongside AI datacenters." No fetched 2026 primary restates a new LM2500 unit count.
• Strazik: aero is complementary / "providing the bridge" until HA heavy-duty commissions ~18 months later. That leaks this page's "primary vs genset-bridge" distinction on the GEV leg. Bloom's islanded-primary claim is cleaner.
• Competitive leak is now two-sourced. siemens-energy: DC commitments 24 GW of 87 GW (~28%) as of March — a higher DC share than GEV. christian-bruch: customers shuffle sites to wherever air permits clear. Mitsubishi record large-frame intake. The Aug-11 SemiAnalysis 30+ OEM / secondary-market leak is unrefuted.

Texas — Aug-20 has not fired.

• This update is dated 2026-08-17. PUCT good-cause meeting is 2026-08-20. Batch Zero classifications remain paused (ERCOT M-A080326-01).
• The documented BTM path is WLPUN / "bring your own generation." It names natural-gas turbines and solar. Fuel cells are not named in the fetched memo. BTM is not a full ERCOT bypass — WLPUN generators still face generation-interconnection rules.

VST — contradiction recorded, resolved toward the 8-K.

• Step 5's SemiAnalysis "1,200 MW net-metering" is not what the primary says. Vistra 8-K (2025-09-29) + Q2 2026 call: 20-year 1.2 GW PPA with Amazon at Comanche Peak; first energy 2027, full ramp 2032. CEO Jim Burke supports the Texas audit and wants the queue culled. FERC's June co-location order is the tariff path (closer to nuclear-baseload-for-ai-data-centers than to this OEM TAM).
• Do not treat VST as a third name in the BE/GEV equipment book.

2026 answer: Negative grid headroom does accrue to Bloom as a named, converting, islanded-onsite OEM. It does not uniquely accrue to GE Vernova as an LM2500-BTM monopolist, and it does not accrue to Vistra as equipment. Demand steps 1–3 stay the best-evidenced part of the page. Do not graduate the joint BE+GEV+VST chain.

Open questions remaining

• BE: named MW beyond Oracle 1.2 / 2.8; 10-Q product-backlog split data-center vs C&I.
• GEV: 2026 LM2500 (or aero) unit count on datacenter accounts; whether "dominates BTM filings" can be cited from a 2026 primary after Siemens' 28% DC share.
• Texas: 2026-08-20 PUCT meeting — exemption widen vs narrow; does WLPUN ever name fuel cells?
• Already priced in? This pass does not re-underwrite BE / GEV multiples.

Update (2026-09-11) — SemiAnalysis names Google's 900 MW Bloom + 930 MW aero. Demand stronger; GEV monopoly still not closed. Stay narrowed.

From 2026-09-10-what-is-so-hard-about-behind-the-meter-power (SemiAnalysis Energy Model, September 10; partial): 75 GW of firm OEM orders for BTM AI compute; ~3 GW US IT BTM by year-end. Named hyperscaler mix: Microsoft >5 GW YTD BTM (Joulent & Chevron); Google 930 MW aero-derivatives + 900 MW Bloom; Anthropic/Meta 300–500 MW Enchanted Rock 0.5 MW gensets; OpenAI Shackelford 1.4 GW IT on 500+ INNIO Jenbacher J624s. Recip OEM count with multi-hundred-MW off-grid orders 12 → 22. That is first-party-adjacent order-flow color on BE (Google 900 MW) and on non-GEV aero / recip (Jenbacher, Enchanted Rock) — the same competitive leak already on this page from Siemens / 30+ OEMs. Does not close "GEV dominates BTM filings." Steps 1–3 strengthen. Steps 4–5 stay mixed. Do not graduate. Not a net-new AI-infra chain. Partial: OEM/BoP winners-and-losers behind the paywall.

Update (2026-09-19) — PUCT Aug 20 was a process delay, not an exemption rewrite. Stay narrowed.

From 2026-09-19-autoresearch-puct-aug20-btm-gev-bloom: PUCT Project 59142 (2026-08-20) granted ERCOT three timeline good-cause exceptions (delay Batch Zero classification; let conditionally eligible base loads into Aug/Nov QSAs; 24-day dynamic-data cure). That is a process pause for the Abbott audit — not a rewrite that widens or narrows WLPUN/BTM eligibility. First-party HTML still does not name fuel cells as a WLPUN resource. No 2026 first-party GEV LM2500/aero datacenter unit count (last named DC count remains Crusoe 29 / ~1 GW, July 2025). No Bloom-named MW increment past Oracle 1.2 contracted / 2.8 MSA in Bloom Q2 IR. Watch items 1–3 still unmet. Stay narrowed. Do not graduate. Do not re-rate BE/GEV/VST.