US Grid Constraints: Towards 40GW+ of Behind-The-Meter Datacenter by 2028?

Redundant with selected · 13

• claimBehind-the-meter is now the most attractive option for buyersc 0.95 · sim 0.89

  With firm capacity scarce, effective capacity from renewables limited, and headroom vanishing, the decision has shifted to the buyer — and BTM has become the most attractive route, driven by AI labs that dominate demand directly and indirectly through hyperscaler buildouts.

  overlapped with: The grid can no longer keep up with AI datacenter demand

• claimBTM will power more than half of new US datacenters by 2028c 0.90 · sim 0.84

  Behind-the-meter solutions are projected to serve well over half of new US datacenter capacity from 2028 onward, with the TAM for BTM equipment crossing 50GW per year by 2029.

  overlapped with: The grid can no longer keep up with AI datacenter demand

• claimAvailable grid headroom turns negative by 2027c 0.90 · sim 0.90

  Once accredited supply is netted against peak demand and required reserve margins, available headroom is already approaching zero nationally and goes negative by 2027.

  overlapped with: Grid headroom is collapsing across most North American markets

• implicationBTM is now the most attractive path for GW-scale newbuildsc 0.80 · sim 0.82

  Generation and transmission constraints plus weak market incentives make BTM the default for the largest projects, with top developers already planning 5GW+ behind-the-meter facilities in Texas where onsite gas permitting is easier.

  overlapped with: The grid can no longer keep up with AI datacenter demand

• implicationPCLR is the explicit bridge-to-firm pathc 0.80 · sim 0.83

  Because a PCLR's curtailment cap relaxes toward a defined exit date as transmission gets built out, ERCOT positions it as the formal mechanism for moving from flexible to firm interconnection.

  overlapped with: PCLR connects at full size but accepts real-time curtailment

• evidenceOnly ~15GW of net-new firm capacity added per yearc 0.75 · sim 0.85

  The forecast points to barely 15GW of net-new ELCC capacity added annually, rising toward 20GW by decade's end — the only capacity a grid operator can recognize for serving firm datacenter and industrial load.

  overlapped with: Nameplate capacity badly overstates what renewables add to firm supply

• mechanismPower is a small share of AI TCO, so any GW is worth billionsc 0.75 · sim 0.83

  Because power is mostly insignificant as a percent of AI cloud TCO, any amount of power an AI lab secures is worth billions in revenue — implied tens of billions of annual revenue per GW. That economics makes multi-year grid delays unacceptable.

  overlapped with: The grid can no longer keep up with AI datacenter demand

• evidenceDatacenter buildout accelerating from 21GW to 84GW per yearc 0.70 · sim 0.83

  US datacenter additions are forecast to grow from +21GW in 2026 to +84GW by 2030, a record buildout that vastly outpaces what the grid can absorb.

  overlapped with: The grid can no longer keep up with AI datacenter demand

• contextHybrid co-location via ERCOT's Batch Zero is the practical bridgec 0.60 · sim 0.84

  Emerging hybrid structures in ERCOT — codified under the Batch Zero process — blend on-site generation with continued grid access, and are where the early commercial winners are taking shape.

  overlapped with: BYOG lets large loads bypass grid-upgrade queues by bringing their own generation

• mechanismAchievable load is withdrawal limit plus on-site generation, staged as units come onlinec 0.60 · sim 0.84

  A BYOG site can draw up to its withdrawal limit from day one and ramps toward full capacity as each generation unit energizes, subject to system-stability limits.

  overlapped with: BYOG lets large loads bypass grid-upgrade queues by bringing their own generation

• contextPCLR requires no on-site generation, unlike WLPUNc 0.60 · sim 0.83

  PCLR differs from WLPUN by relying purely on dispatchability rather than co-located generation — the load itself absorbs the constraint, not an on-site power source.

  overlapped with: PCLR connects at full size but accepts real-time curtailment

• exampleERCOT now credits some incremental solar at essentially zeroc 0.50 · sim 0.82

  In congested ERCOT zones and segregated transmission corridors, planners leave solar's capacity contribution out of local reliability modeling — a de facto 'no-solar scenario' when sizing firm needs.

  overlapped with: Nameplate capacity badly overstates what renewables add to firm supply

• caveatNew transmission is too slow for the AI buildoutc 0.40 · sim 0.83

  Building bulk transmission could unlock generation and load growth but the market is over-regulated and slow. A major buildout is more likely in the 2030s once large offtakers can underwrite whole projects with parent guarantees.

  overlapped with: The grid can no longer keep up with AI datacenter demand

Below top-k · 35

• mechanismWLPUN trades a hard withdrawal cap for more connected megawattsc 0.85

  A Withdrawal-Limited Private Use Network can connect more load than transmission alone could support in exchange for an enforced cap on grid draw — e.g. 1,000 MW of load that never pulls more than 100 MW from the grid.

• mechanismThe 2026-27 shortfall stacks several distinct bottlenecksc 0.80

  The near-term capacity shortfall isn't one bottleneck but a stack: utility and interconnection-queue friction, a technology mix dominated by slow-to-build CCGTs, and supply-chain lead times that have stretched dramatically.

• mechanismBTM wins on speed and schedule certaintyc 0.80

  Onsite generation can be energized in 2027–28, against grid timelines slipping toward 2030, and the schedule sits in the buyer's hands rather than utilities who routinely revise down promised load with no penalties.

• claimAI tenants have relaxed uptime requirementsc 0.80

  AI labs and some hyperscalers now tolerate lower uptime for both inference and training. Many of Meta's self-built AI datacenters target just two nines and forgo backup generators entirely.

• claimBatch Zero introduced two new self-supply constructs on top of the PUN basec 0.80

  Building on the PUN framework, ERCOT's Batch Zero created two new categories — WLPUN and PCLR — that formalize how large flexible loads can connect under transmission constraints.

• implicationRelaxed uptime removes the historical cost barrier to BTMc 0.75

  Providing four or five nines onsite was prohibitively expensive because a single site can't pool risk the way the grid does. Now that customers accept lower redundancy, BTM economics become competitive with grid power.

• mechanismWhy marginal ELCC declines for correlated renewablesc 0.70

  Solar plants all generate at the same hours, so each incremental GW addresses risk that earlier GWs already covered — the Duck Curve effect — driving marginal capacity value sharply down as penetration rises.

• mechanismBESS suffers the same marginal-value collapse, just by durationc 0.70

  A 4-hour battery only solves outage events shorter than 4 hours; once that risk is saturated, incremental 4-hour BESS adds little ELCC and reliability risk shifts to longer-duration events that need other fuels.

• evidencePJM's 2027/2028 auction cleared 6.5GW short of targetc 0.70

  PJM's Base Residual Auction cleared ~134,478MW UCAP, yielding a 14.4% reserve margin against a 20% target — a physical deficit of about 6,517MW. This is the concrete proof point that headroom has already gone negative in the largest US market.

• claimHybrid BTM-grid structures are where the deal flow isc 0.70

  The market is settling onto a spectrum of hybrid structures, especially in ERCOT, that blend on-site generation with continued grid access. This is where a large share of current deal activity and interest is concentrated.

• mechanismERCOT splits a BYOG project across three parallel review tracksc 0.70

  A Batch Study sets the site's grid withdrawal limit, Generation Interconnection sets its export limit, and Transmission Planning identifies any required network upgrades — each evaluated in parallel.

• contextBYOG sits outside SB6's net-metering review because the generation is newc 0.70

  BYOG is distinguished from an NMA by category and vintage: it describes how a site sources power through newly built generation, which exempts it from SB6's net-metering review.

• caveatWLPUN is not a faster interconnection, just a way to energize more load soonerc 0.70

  ERCOT is explicit that WLPUN does not speed up the interconnection process — it lets a site energize more of its desired load within existing transmission limits by leaning on on-site generation.

• evidenceTurbine and transformer lead times have doubled to 3-4 yearsc 0.65

  Gas turbine and generator step-up transformer lead times have stretched from a historical 18 months to three to four years, pushing total gas-plant development from a 24-month baseline to at least four years.

• evidenceLess than 10GW of gas per year added in 2026-27c 0.60

  Nameplate forecasts show the US will add less than 10GW of gas annually in 2026 and 2027, with material additions only picking up in 2028 and beyond.

• implicationBTM tailwind goes to non-obvious winnersc 0.60

  While BTM is a material tailwind for equipment providers, the key beneficiaries aren't the usual suspects — peak gas turbine orders are a temporary phenomenon and the durable winners sit elsewhere in the stack.

• mechanismUCAP is the view that turns red firstc 0.60

  Unlike ICAP (nameplate proxy), UCAP nets out forced-outage risk on thermal units and the steep ELCC discounts on solar and storage. That makes the UCAP picture materially tighter, and it is the one that crosses into deficit first.

• mechanismERCOT's Batch Zero codifies co-location via NMA and BYOGc 0.60

  ERCOT's NPRR1325/PGRR145 framework added new co-location constructs alongside PUNs, with each large load assigned a maximum withdrawal limit. The structures separate how a site interconnects (PUN, WLPUN, PCLR) from how it sources power (NMA from existing generation, BYOG from new build).

• mechanismA Private Use Network meters only the campus's net exposure to the gridc 0.60

  Under a PUN, an entire campus — load plus its co-located generation — sits behind a single point of interconnection, so ERCOT measures only net flow rather than each individual resource.

• evidenceQueue conversion, not queue size, is now the binding constraint in PJMc 0.55

  In PJM, ~57GW has cleared interconnection studies, yet roughly 24GW with fully executed agreements since 2020 — including 13.5GW of gas — remain unconverted into actual projects.

• exampleUtilities are slipping 2027 load schedules to 2029c 0.55

  Operators routinely see a utility promise to deliver, say, 500MW in 2027 only to push it to 2029 because long-lead grid interconnection equipment like main power transformers isn't available.

• contextFaster-to-build technologies are being grabbed directly by operatorsc 0.55

  Because CCGTs take 4-6 years to build, faster options like fuel cells and reciprocating engines (RICE) are being aggressively secured directly by datacenter operators rather than waiting on utilities.

• mechanismHow headroom is actually calculatedc 0.55

  Headroom equals total accredited supply minus peak demand minus required reserves, computed subregion by subregion. Required reserves are a regulatory artifact set by planners or utilities, typically 15–20% of installed capacity.

• exampleMeta's Ohio campus is designed to never connect to PJMc 0.55

  Meta's Ohio site is built to stay off-grid permanently, with El Paso and Louisiana sites also built around dedicated utility-scale gas rather than near-term interconnection. Cipher's AWS-leased Black Pearl and Fluidstack builds follow similar logic.

• mechanismNet-metering arrangements re-point existing generation to private loadsc 0.55

  Under an NMA, an already-operating generator (pre-September 2025 vintage under SB6) co-locates with a new load and nets output behind a single meter. The PUCT can condition approval, e.g. requiring curtailment during grid emergencies, since the central question is whether re-pointing existing capacity leaves the rest of the system whole.

• evidenceMost announced ERCOT co-location uses existing generationc 0.55

  About 2,885MW of reported projects sit in the existing-generation bucket, including AWS's 1,200MW co-location at Vistra's Comanche Peak nuclear plant on a 20-year PPA ramping to full capacity by 2032.

• contextELCC methodologies diverge sharply across ISOsc 0.50

  PJM uses marginal ELCC with an hourly risk model, MISO accredits seasonally, and ERCOT — an energy-only market — relies on its own effective-capacity measures, so the same solar or storage asset can be credited very differently region to region.

• contextTenants accept worse terms because they have nowhere else to goc 0.50

  The relaxation of uptime and reliability requirements isn't ideological — it reflects tenants adapting to AI's rampant power needs because the grid alternative simply isn't available on the required timeline.

• implicationLoad flexibility could unlock tens of GW if markets allow itc 0.50

  If grid-connected datacenters could curtail a set number of hours per year via workload shifting, backup dispatch, or battery discharge, PJM estimates this could bridge the interconnection gap. But commercial and regulatory frictions slow adoption.

• evidenceFERC and PJM are beginning to write co-location rulesc 0.50

  FERC's December 2025 order directed PJM to create co-location rules and faster interconnection for loads above 20MW. PJM's June 2026 tariff revisions established an Expedited Interconnection Track with a ~10-month accelerated study process.

• contextGensets make sense only for short BTM-to-grid bridgesc 0.45

  Adding gensets is justifiable when grid interconnection is two years out and equipment can be redeployed. For sites staying off-grid for five-plus years, ~$1mm/MW in genset capex delivers only one extra nine and is hard to justify.

• caveatHidden headroom exists for developers who know where to lookc 0.40

  Utilities themselves often struggle to understand their own headroom, and savvy datacenter developers can find pockets of capacity the published numbers miss.

• contextThe traditional five-nines uptime playbookc 0.40

  Historically datacenters reached five nines by combining N+1 substation redundancy (good for three to four nines) with backup generators and batteries sized above 100% of nameplate load. The grid was the critical first leg of that stack.

• mechanismWLPUN sites export their surplus generation through SCEDc 0.40

  When a WLPUN site's on-site generation exceeds its load, the surplus is exported to the grid through ERCOT's Security-Constrained Economic Dispatch.

• caveatLoad flexibility is an alternative but unlikely to winc 0.35

  Alternatives like load flexibility exist and will be covered separately, but BTM is expected to be the dominant solution to the grid constraint over those alternatives.