AI data centers are waiting on electrons the public grid cannot promise on a GPU procurement calendar. Lawrence Berkeley National Laboratory's Queued Up: 2026 Edition (data through end-2025) still shows roughly 8,200 active U.S. generator interconnection projects totaling about 1,312 GW of generation and 749 GW of storage. For projects that did reach commercial operation in 2025 in regions with timing data, the median span from interconnection request to COD exceeded five years. Only 13% of capacity that entered queues from 2000 through 2020 had reached COD by end-2025, while 75% had withdrawn.
That is the generation-side backlog. Large-load interconnection is a separate queue in many regions, and it is often just as slow. Hyperscalers are answering with behind-the-meter generation: co-located or islanded plants that serve the campus first, and only later (if at all) the public system.
This guide explains what behind-the-meter power means for AI facilities, why interconnection timelines force the choice, which options are shipping or financing in 2026 (gas, fuel cells, bridge generation, batteries and renewables, SMRs later), how the electrical interface meets medium voltage and the path to 800 VDC, and where permits and air-quality rules still break schedules. It sits next to Inside Deep Tech's State of Data Centers 2026, SMR guide, SiC/GaN power electronics guide, and Deep Tech Industry Report: Q3 2026.
Key Takeaways
- U.S. active generator queues held ~8,200 projects / 1,312 GW generation + 749 GW storage at end-2025; median IR→COD >5 years for 2025 CODs; only 13% of 2000–2020 requests reached COD by end-2025 (LBNL Queued Up 2026).
- Microsoft's Pecos campus adds ~2 GW with launch-stage behind-the-meter gas; Chevron's Project Kilby targets ~2.67 GW under a 20-year PPA with first power anticipated in 2028 (Microsoft; Data Center Dynamics).
- Oracle Project Jupiter shifts to up to 2.45 GW of Bloom fuel cells (~92% NOx cut vs prior gas plan); Bloom–Brookfield AI power financing framework expands $5B→$25B (Oracle; Bloom IR).
- Entergy Louisiana + Meta: seven new CC gas plants >5,200 MW, ~240 miles of 500 kV, up to 2,500 MW Meta-funded renewables; ~$2B additional customer savings over 20 years on top of $650M prior (Entergy / PR Newswire).
- BTM is a 2026–2028 bridge for gas and fuel cells; multi-GW COD is still post-FID and permit-gated. SMRs remain a later-decade option (see IDT SMR guide).
What behind-the-meter generation means for an AI campus
In utility language, behind-the-meter generation sits on the customer side of the revenue meter. For AI campuses the phrase is also used for co-located plants that are electrically dedicated to the data hall, even when a future grid tie is planned.
Microsoft's June 22, 2026 Pecos announcement is explicit: at launch the campus "will operate with a co-located natural gas power facility, an arrangement known as "behind the meter,"" serving the campus "directly and independently of the public grid" so the new demand "does not take from the current grid" (Microsoft Blog). Over time Microsoft expects to connect both the plant and the campus into the regional system.
Oracle's Project Jupiter goes further toward an islanded microgrid: Bloom solid-oxide fuel cells replace previously planned gas turbines and diesel generators and consolidate the New Mexico campus into one microgrid sized for up to 2.45 GW of installed fuel-cell capacity, per Oracle's April 27, 2026 announcement.
Here's why that matters for facility design. BTM changes who builds the megawatts, who signs the air permits, how N+1 is sized, and whether your critical power path starts at a utility substation or at on-site generation bus. It does not remove medium-voltage switchgear, transformers, or the rack-level conversion stack that leads toward 800 VDC distribution.
Interconnection queue reality (and what the data actually covers)
Operators often blur two queues: generators waiting to inject, and large loads waiting to take power. LBNL's Queued Up: 2026 Edition is the cleanest public national dataset for the first. It covers transmission-connected generation and storage through end-2025. It does not include large-load interconnection requests or distribution-connected / behind-the-meter projects. Still, its timelines explain why co-located deals are multiplying.
Table 1. U.S. generator interconnection snapshot (end-2025)
| Signal | Hard mark | Window | Source |
|---|---|---|---|
| Active projects | ~8,200 | End-2025 | LBNL Queued Up 2026 |
| Active generation capacity | 1,312 GW | End-2025 | LBNL |
| Active storage capacity | ~749 GW | End-2025 | LBNL |
| YoY change in active queue volume | −10% | 2025 vs prior year | LBNL |
| Active natural gas in queue | 253 GW (+86% YoY) | End-2025 | LBNL |
| Active solar / storage / wind | 773 / 749 / 220 GW (−19% / −16% / −19%) | End-2025 | LBNL |
| Capacity with draft/executed IA, not COD | 549 GW (solar 256 · storage 161 · wind 76 · gas 45) | End-2025 | LBNL |
| Median IR → COD (regions with data) | >5 years | Projects COD in 2025 | LBNL |
| Completion rate, requests 2000–2020 | 13% COD by end-2025; 75% withdrawn; 10% still active | Through 2025 | LBNL |
Source: Lawrence Berkeley National Laboratory, Queued Up (2026 Edition); publication record at ETA Publications. Gas capacity in the active queue jumped 86% year over year to 253 GW while solar, storage, and wind active volumes fell. That mix shift matches the hyperscaler preference for dispatchable on-site or co-located gas and fuel cells when AI load cannot wait on renewable-queue COD.
At first glance a 10% smaller total queue looks like relief. Look closer: withdrawals and fewer new requests drove the decline, while median timelines lengthened. Completing studies is not the same as energizing a hall.
The BTM option stack for AI data centers
No single technology clears every constraint. Operators stack bridge capacity now, firm capacity for the late 2020s, and carbon-free options on a longer clock.
Natural gas: gensets, simple-cycle, and combined-cycle
Natural gas remains the default firm megawatt when interconnection or load studies stretch past GPU deliveries. Microsoft and Chevron's West Texas structure is the clearest 2026 primary-source template: a ~2 GW Microsoft campus paired with Chevron Project Kilby at about 2.67 GW of phased, modular gas capacity under a 20-year power purchase agreement, with first power anticipated in 2028 and FID expected by end-2026 subject to conditions (Microsoft; Data Center Dynamics / Chevron marks).
Chevron says a majority of Kilby generation will come from large GE Vernova turbines plus Solar Turbines (Caterpillar) units, with Selective Catalytic Reduction for NOx, non-potable brackish water for plant operations, and projected regional benefits including more than $10 billion in state and local tax revenue and almost 2,000 jobs (Data Center Dynamics). Microsoft separately cites over 6,000 construction jobs at peak campus build-out and 4.7 GW of renewable electricity already contracted for its Texas electricity use (Microsoft).
Meta's Louisiana Hyperion expansion is a different ownership model: Entergy Louisiana proposes seven new natural gas combined-cycle plants totaling more than 5,200 MW, about 240 miles of new 500 kV transmission, battery storage at three locations, nuclear uprates, and Meta funding for up to 2,500 MW of new renewable resources, with expected additional customer savings of about $2 billion over 20 years on top of $650 million previously announced (combined ~$2.65 billion). Meta's data-center VP framed campus potential "to scale up to 5GW" (Entergy Louisiana via PR Newswire, March 27, 2026). Treat this as customer-funded utility generation plus transmission, not islanded BTM, unless later filings prove otherwise.
Fuel cells: Bloom and the microgrid path
Solid-oxide fuel cells trade combustion for electrochemical conversion. Oracle and BorderPlex said Project Jupiter will use Bloom Energy fuel cells for up to 2.45 GW of installed capacity, cutting NOx by roughly 92% versus the prior gas-turbine plan and using a negligible amount of water for power generation (Oracle, April 27, 2026). Oracle bears the energy costs so residential rates are not the bill-payer for the microgrid.
Capital is following the same thesis. On June 30, 2026, Bloom and Brookfield expanded their AI infrastructure power partnership from a previously announced $5 billion framework to $25 billion (a fivefold increase since October 2025) to finance rapid onsite / islanded power. Brookfield's dedicated AI Infrastructure Fund, launched November 2025, targets $100 billion of deployment; Brookfield also states over $100 billion already invested in digital infrastructure and clean power assets (Bloom Energy IR).
Honest limit: fuel cells prefer steady load. AI training and inference ramps still need batteries or oversizing. That is an engineering constraint, not a press-release footnote.
Temporary and bridge generation
Rental gas turbines, modular reciprocating engines, and temporary substations fill the gap between shell completion and permanent COD. They show up in permitting as short-duration air authorizations and in finance as opex rather than 20-year PPAs. Use them as schedule insurance, not as the long-term energy story you sell to ESG committees.
Batteries plus behind-the-meter renewables
Batteries alone do not create energy. They time-shift it. Paired with on-site or dedicated solar and wind, they can shave peaks and ride through gas-train maintenance. Meta's Entergy package explicitly includes battery energy storage across three locations and up to 2,500 MW of Meta-funded renewables (Entergy). Microsoft's Texas renewables book already stands at 4.7 GW contracted (Microsoft). Renewables still face the same generator interconnection math LBNL documents unless they stay truly behind the meter or on surplus-interconnection products.
SMRs: long-horizon BTM, not a 2026 COD
Small modular reactors remain the long-dated nuclear option for campuses that want carbon-free firm power on site. Licensing, first-of-a-kind construction, and fuel-cycle logistics put most commercial AI-campus SMR MWh after the current GPU cycle. Inside Deep Tech's Small Modular Reactors full guide covers designs, hyperscaler deals, and why most of that electricity will not arrive this decade. Entergy and Meta also signed a memorandum of understanding "to explore the future development and use of nuclear power" (Entergy), which is exploration language, not a COD date.
Table 2. 2026 primary-source deal marks (BTM and adjacent)
| Deal | MW / structure | COD / timeline mark | Source |
|---|---|---|---|
| Microsoft Pecos campus | ~2 GW data center capacity | Announced Jun 22, 2026; BTM gas at launch | Microsoft |
| Chevron Project Kilby | ~2.67 GW phased gas; 20-year PPA | FID targeted end-2026; first power 2028 | Chevron |
| Oracle Project Jupiter + Bloom | Up to 2.45 GW fuel cells; single microgrid | Announced Apr 27, 2026; construction ongoing | Oracle |
| Bloom × Brookfield | $5B → $25B AI power financing framework | Expanded Jun 30, 2026 | Bloom IR |
| Entergy + Meta (LA) | Seven CC plants >5,200 MW; ~240 mi 500 kV; ≤2,500 MW renewables | Filed Mar 27, 2026; LPSC process ongoing | Entergy |
Source lines are in the table. Capacity figures are developer or utility announcements, not Inside Deep Tech measurements. FID, permit, and turbine delivery can still move COD.
Facility electrical interface: MV today, 800 VDC path tomorrow
BTM generation does not erase the facility power train. It relocates the first bus.
Typical path: on-site generators or fuel cells → medium-voltage collection (often 13.8–34.5 kV class, site-specific) → step-down transformers → UPS / battery → rack conversion. As rack loads climb past the copper limits of 48/54 V architectures, operators plan halls around 800 VDC power distribution. Wide-bandgap devices in that conversion stack are covered in the SiC and GaN power electronics guide.
That means your BTM EPC and your data-hall electrical EPC have to share a single one-line diagram early. Islanding controls, black-start, protective relays, and harmonic budgets are not bolt-ons after the white-space layout freezes.
Permitting, air quality, and community constraints
Gas projects live or die on air permits (NOx, CO, VOCs, HAPs), noise, water intake, and land use. Chevron's Kilby materials stress SCR for NOx and brackish rather than freshwater sources (Data Center Dynamics). Microsoft describes SCR and closed-loop cooling that needs only an initial charge in steady state (Microsoft).
Fuel cells win local air-quality arguments when they displace combustion. Oracle cites about a 92% NOx reduction versus its prior gas plan, plus negligible water use for power generation, alongside community packages ($50 million for local water systems, $360 million for schools and services, $6.9 million for workforce and related programs) (Oracle).
On the flip side, state and local politics can still delay FID even for islanded plants. A behind-the-meter label does not automatically exempt a project from data-center permitting pauses or contested air permits. Build schedule contingency for hearings, not just for turbine lead times.
Honest timeline: what ships 2026–2027 versus later
Table 3. Capability versus calendar
| Capability | Realistic window | Evidence mark | Caveat |
|---|---|---|---|
| Bridge gensets / temporary power | 2026–2027 deployments | Industry practice; not a single national COD dataset | Air permits often short-duration |
| Fuel-cell modules at campus scale | Financing and awards in 2026; multi-year install | Oracle up to 2.45 GW Bloom plan (Apr 2026); Bloom–Brookfield $25B (Jun 2026) | Ramp/battery buffering needed |
| Large co-located CC / gas plants | FID 2026–2027; first power often 2028+ | Kilby first power 2028; Entergy/Meta package under LPSC process | Regulatory and turbine supply risk |
| Customer-funded utility buildouts | Multi-year after commission approval | Entergy seven plants >5,200 MW announced Mar 2026 | Not islanded BTM by default |
| SMR commercial MWh for AI campuses | Later decade for most designs | IDT SMR guide; Entergy–Meta nuclear MOU (explore) | Licensing + FOAK construction |
Source: deal rows cite Oracle, Bloom IR, Data Center Dynamics, Entergy, and Inside Deep Tech's SMR guide. Temporary generation is described qualitatively because there is no single public national COD ledger for rental fleets.
Let's start with what you can still change in a 2026 design package: reserve acreage and gas interconnect capacity for BTM, specify MV collection that can accept either turbines or fuel-cell skids, and keep UPS autonomy sized for generation transfer, not only utility recloser events.
Honest limits (read these before you budget)
- LBNL Queued Up measures generator interconnection, not large-load queues. Cite it for supply-side delay, not as a load-queue census (LBNL).
- Announced GW are not energized GW. Kilby's first power is anticipated in 2028; FID remains conditional (Data Center Dynamics).
- Meta/Entergy economics depend on commission certification and cost-allocation rules. Press savings figures are utility projections (Entergy).
- Fuel-cell microgrids still need fuel supply (often natural gas), spare modules, and power-electronics O&M skills on site.
- BTM does not cancel Scope 1 emissions reporting for combustion assets. Fuel cells change the emissions profile; they do not erase feedstock carbon unless paired with certified low-carbon gas or other measures.
Inside Deep Tech's take
Inside Deep Tech's take: behind-the-meter generation is now a first-class AI infrastructure workstream, not an emergency patch. The public evidence is the coincidence of multi-year generator queue times in LBNL Queued Up 2026 with hyperscaler structures that put firm megawatts next to the hall (Microsoft/Chevron), replace combustion with fuel-cell microgrids (Oracle/Bloom), or fund utility fleets under ratepayer-protection pledges (Meta/Entergy).
The team at Inside Deep Tech would rather see operators publish one-line diagrams and air-permit milestones than another vague "we secured power" slide. If your campus story still assumes a greenfield utility feed inside 24 months in a constrained RTO, the schedule is the fiction.
FAQ
What is behind-the-meter generation for AI data centers?
It is on-site or co-located generation that serves the campus electrically without taking that megawatt from the public delivery path at launch. Microsoft's Pecos blog uses the term for a co-located natural gas plant that operates independently of the public grid at startup.
How long are U.S. generator interconnection queues in 2026 data?
Per LBNL Queued Up 2026, active queues held about 8,200 projects and 1,312 GW of generation plus 749 GW of storage at end-2025. Median request-to-COD time exceeded five years for 2025 commercial operations in regions with data, and only 13% of capacity requesting from 2000–2020 had reached COD by end-2025.
What is Microsoft and Chevron building in West Texas?
Why did Oracle choose Bloom fuel cells for Project Jupiter?
Oracle, BorderPlex, and Bloom said up to 2.45 GW of Bloom fuel cells will fully power the Doña Ana County campus as a single microgrid, replacing planned gas turbines and diesel. They cite about a 92% NOx reduction versus the prior gas plan and negligible water use for power generation (Oracle, April 27, 2026).
Is Meta's Louisiana power package behind the meter?
Entergy Louisiana's March 27, 2026 announcement describes Meta-funded utility generation and transmission: seven new combined-cycle plants totaling more than 5,200 MW, roughly 240 miles of 500 kV lines, storage, nuclear uprates, and up to 2,500 MW of Meta-funded renewables. That is customer-funded utility infrastructure with ratepayer-protection framing, not an islanded BTM microgrid by default (Entergy).
How does BTM connect to 800 VDC and SiC/GaN?
BTM usually lands on a medium-voltage collection bus before facility step-down and rack conversion. As rack power rises, halls migrate toward 800 VDC distribution, with wide-bandgap devices in the conversion path covered in Inside Deep Tech's SiC/GaN guide.
When will SMRs power AI data centers behind the meter?
Most commercial SMR megawatt-hours for AI campuses remain a later-decade deliverable. See Inside Deep Tech's SMR full guide for licensing and hyperscaler deal context. Near-term firm power in public 2026 announcements is dominated by gas and fuel cells.
What to do Monday
If you are scoping an AI campus power package this quarter, lock three artifacts before you argue about brand of turbine or fuel cell.
- A dated interconnection and load-study status from the utility or RTO, not a verbal assurance.
- A one-line that shows generation bus, MV collection, UPS autonomy, and the path toward 800 VDC halls.
- An air-permit and water strategy with contingency months, especially if the site sits under a data-center permitting pause.
For capital and infrastructure context beyond power yards, read the Deep Tech Industry Report: Q3 2026. For the wider facility stack, stay with the State of Data Centers 2026.
