The silent bottleneck: gensets, transformers, turbines, copper.

In our research desk, we track every structural bottleneck operators name in primary-source conversations — interconnection queues, state PUC rulemakings, community permitting, water withdrawal, equipment lead times. The queue dominates the public discourse because it is the most visible and the most litigated. But when we ran a second pass across the last six months of operator podcasts, investor calls, and industry panels, the story shifted.

Operators who have actually commissioned megawatts in the last eighteen months consistently cite a different top-of-list bottleneck: equipment lead times and the sub-layers underneath them. The gas turbine. The high-voltage transformer. The medium-voltage switchgear. The backup genset. The battery inverter. The copper for the bus work. The LNG tank for fuel-assurance backup. The ride-through model the ISO requires before it will accept the voltage-dip signature of a 1 GW AI training cluster. Each one has its own order book, its own supplier concentration, and its own delivery window — and each one has slipped materially in the last twelve months.

The pattern is consistent across PJM, ERCOT, SPP, and MISO geographies and it replays, compressed, in Europe. Mainova WebHouse Frankfurt reports simultaneous scarcity of experienced data center contractors, gensets, MV/LV gear, BMS/PMS, and commissioning teams — EU permit lead times are up 50–300 percent over the last ten years per a Schneider report cited in the 2026 industry conversations. A project that wins its interconnection filing in month 14 still waits until month 36 or later for turbines. A coal conversion that reuses a 500 kV POI still waits out the transformer queue. A BTM deal that survives FERC review still waits for the LNG tank that backs up the nuclear plant or the gas firm-supply contract. None of this appears on the queue dashboard because the queue dashboard tracks the electrical graph, not the manufacturing graph.

The most striking finding in the journal is not about lead times. It is about the failure mode that AI training workloads introduce into equipment that was never designed for them. Industry sources have described genset crankshafts failing inside twelve months on sites running continuous AI training load profiles — a failure mode that normally appears after five to seven years of rated operation. BESS inverter switches on the same sites were reported to fail inside six months. Multiple sources described harmonic distortion propagating up to fifty miles into adjacent residential feeders, triggering customer complaint cycles at utilities that had no model for the root cause.

The mechanism is load variability. A hyperscale training cluster is not a constant-current load. It is a highly dynamic load that oscillates between idle and peak as gradient descent cycles advance. Those oscillations interact with genset governor loops, BESS inverter switching patterns, and transformer core magnetization in ways that accelerate mechanical and thermal fatigue. The design envelope of most commodity power equipment assumes a load profile closer to industrial or commercial than HPC. Operators running AI workloads are effectively stress-testing equipment past its rated cycle life — and the OEMs are not yet shipping a differentiated SKU for AI operation.

The second-order effect is a shadow demand line the market is not pricing. Equipment failures drive unplanned replacement orders. Replacement orders land in OEM order books that are already backlogged. The backlog extends. Operators with an aggressive AI roadmap face a double-sided shortage: the equipment they need for new sites is delayed, and the equipment they bought for existing sites is failing faster than the amortization schedule assumed.

DG Matrix — a company we now watch closely — raised a $60M Series A from investors including ABB and Mitsubishi Heavy to productize this problem. Their thesis: the gap between nameplate and usable megawatts on a site running AI load is often 40 to 50 percent, because equipment cannot sustain rated output under AI-induced thermal and harmonic stress. DG Matrix claims roughly $10 per watt of revenue for every stranded watt it can recover through solid-state transformer substitution. A 50 MW nameplate site often delivers only 25 to 30 MW under AI surge loads. The new unit of analysis is not nameplate capacity — it is usable megawatts.

We expect this to surface in insurance markets within twelve months. Equipment insurers who underwrote against a traditional-load failure distribution are going to reset premiums once the AI-load failure curve shows up in loss ratios. Operators who treat equipment lifecycle planning as a spreadsheet input will be out-bid by operators who treat it as a dynamic risk with its own hedging strategy.

The gas turbine market for data center peaking and primary generation is concentrated. GE Vernova, Mitsubishi Power, and Hitachi Energy split the majority of the H-class and F-class order book between them. Each has its own factory footprint, its own shift cadence, and its own delivery window.

GE Vernova, the largest of the three in the US footprint, has publicly acknowledged a multi-year order book for its largest frames. Our working estimate is that new orders for H-class turbines placed in Q1 2026 are quoting delivery in 2029 or later. The backlog is not going to clear on its own — adding a line at Greenville would take roughly four years, and GE Vernova's public capital allocation has favored shareholder returns over capacity expansion so far.

Mitsubishi Power and Hitachi Energy both operate on three-shift cadences at their main assembly plants. Three shifts is the maximum sustainable factory cadence for heavy rotating equipment — you cannot compress it further without adding floor space. Lead times at Mitsubishi and Hitachi are shorter than GE for equivalent frames but not by enough to close the bottleneck. The operator advice from the Qatar panels in the journal is consistent: book factory slots and witness tests up front, before the site is fully permitted, because the alternative is writing your COD to the factory calendar and not the other way around.

On transformers, the concentration is even sharper. The US market for large power transformers has four or five active suppliers — Hitachi Energy, Siemens Energy, GE Vernova, WEG, and a short list of Korean and Chinese alternatives. Transformer lead times for 345 kV and 500 kV units have extended from roughly 52 weeks in 2022 to roughly 150-plus weeks in 2026. Medium-voltage switchgear has followed the same pattern. The supply chain response to the IRA and IIJA demand signal is real but delayed — new factory capacity announced today does not produce shipments for 36 to 48 months.

The genset and BESS side is a different shape. More suppliers, more global capacity, but also more acute failure issues under AI load. Caterpillar gensets were reported at a nine-month wait in the 2026 Qatar panel. The genset bottleneck is less about factory capacity and more about the replacement cycle from AI-load-induced failures plus the unprecedented volume of new large-campus orders. The BESS bottleneck is about inverter switch assemblies and specific IGBT components.

Beneath the turbine and transformer layer sit a set of sub-bottlenecks that are less visible but no less binding.

Copper is the first. At roughly $14,000 per ton in early 2026, copper is already expensive enough to change the economics of heavy bus-work and feeder cable. The problem is not the current price — it is the replacement-at-scale problem. New copper mines have 10-to-15 year lead times from discovery to first production. The global copper supply response to data center demand plus EV adoption plus grid electrification is not going to arrive in this decade. Some industry sources are floating 'urban mining' of legacy PSTN copper estates as a bridging supply — the decommissioned copper in central-office and trunk infrastructure is significant at the scale of national telecoms, and the salvage economics work. We do not have a confident supply-response estimate, but operators who are treating copper as a fixed input cost are under-modeling the risk.

LNG fuel assurance is the second. Every BTM or on-site gas generation plan assumes firm fuel supply. In practice, some mid-Atlantic gigawatt campuses have discovered post-permitting that they can only secure 50 percent firm and 50 percent interruptible gas supply from local pipelines. The response is to install on-site LNG storage sized for roughly 120 hours of full-load runthrough. A Cashman Preload tank sized for a gigawatt campus is a multi-million-gallon installation that triggers a full secondary permitting workstream — PHMSA 49 CFR 193, NFPA 59A, state fire marshal, local siting, and in several jurisdictions a setback analysis that looks more like an LNG import facility than a data center accessory. The LNG workstream is a new permitting layer that sits underneath the supply-chain gap and adds 12 to 18 months to projects that discover it late.

Voltage ride-through modeling is the third. ERCOT's dynamic studies group spent six to eight months working with Crusoe on a ride-through model for the 1.2 GW Abilene campus before the ISO would accept the interconnection. The physical problem is that AI training clusters can drop 1 GW of load in milliseconds when a training step finishes, creating a voltage dip signature the existing planning models did not contemplate. Every large AI campus now needs its own ride-through study as a gating condition of ISO acceptance. The dynamic studies groups at ERCOT, PJM, and MISO are the bottleneck — there are a small number of power systems engineers who can run these studies, and they are fully booked.

Harmonic filtering is the fourth — the physical hardware that addresses the distortion problem described above. Filter banks are an additional piece of equipment competing for the same OEM order books that are already backlogged. A large-load harmonic filtering package can add $5M to $20M per site and add a new lead-time dependency that did not exist on the operator's original schedule.

The Crusoe / Oracle / OpenAI Stargate Abilene campus is the cleanest public case study of how these bottlenecks compound. The project was originally sized at roughly 2 GW across multiple buildings; it was capped at 1.2 GW in the final design, with Microsoft taking the remainder of the planned capacity through a separate arrangement. The reported reason is that the campus was designed for NVIDIA GB300 rack power density and had to be retrofitted for Rubin-class equipment, which changed the thermal and electrical envelope in ways the original design could not absorb.

There are multiple bottlenecks stacked here. An OEM bottleneck (the transition from GB300 to Rubin changed what the campus could physically host). A voltage ride-through bottleneck (the six-to-eight month ERCOT dynamic study). A cooling bottleneck (reported winter cooling issues in early operations). A financial bottleneck (the cost model at 2 GW did not pencil under the revised design envelope). Any one of these could have been absorbed. Together they forced a material downsizing.

The Abilene case is instructive because it is a best-case scenario — a well-capitalized project with deep hyperscaler involvement, on a fast-track Texas interconnection, with an experienced developer. If Abilene can lose 40 percent of its planned capacity to a stack of compounding supply-chain and design-envelope issues, the project with less sophisticated management loses more.

We expect to see more Abilene-shaped stories in 2026 and 2027 as the early Nvidia platform transitions interact with campus designs that were locked in against earlier hardware. The operators who will navigate this best are the ones treating the OEM relationship and the ride-through study as first-class project risks from day one.

We give customers five specific pieces of guidance on the equipment stack.

First, treat OEM order books as a first-class siting signal. Before you commit capital to a site, map the equipment delivery windows you need against the realistic commercial operation date. If your site selection logic is interconnect cost plus water access plus zoning and nothing else, you are exposed to an equipment delay that can push your COD by two-plus years after you have already signed for the land.

Second, commission the ride-through study as part of pre-development, not as a study-phase deliverable. For any campus over 500 MW in ERCOT, PJM, or MISO, the ride-through study is on the critical path. Starting it six months earlier saves six months at the back end, and the dynamic studies group will appreciate the lead time because they are fully booked.

Third, map the LNG fuel-assurance layer before you lock in the gas supply contract. If your gas supply is anything less than 100 percent firm, the on-site LNG tank is now a permitting workstream. Scope it, site it, and budget for it alongside the primary generation equipment.

Fourth, build harmonic filtering into the baseline engineering specification. Do not wait for the utility to ask. The filter bank cost is lower if you procure it alongside the primary equipment, and the delivery window is lower if you order it in the first wave rather than the second. Treat it as part of the interconnection package, not a change order.

Fifth, index your site selection to the OEMs you have relationships with, not the OEMs with the shortest quoted lead times. A long-standing supplier relationship with one of the three primary gas turbine OEMs is worth more in 2026 to 2028 than a shorter lead time quote from a supplier you have not worked with before. The OEMs will honor their long-term customers first when the order book gets triaged.

Site Intelligence now flags equipment availability as a structured risk layer on every candidate site in the index. For each site, we record the primary OEM dependencies implied by the likely buildout (gas turbine vendor, transformer vendor, switchgear vendor, LNG tank supplier, harmonic filter vendor), the current public order book status for those OEMs, and a delivery window estimate rolled up to the site COD.

We track ride-through study capacity at the ISO level as a scheduling constraint, so that projects entering pre-development get a realistic lead time on the study rather than a nominal one. We track LNG tank permitting milestones under PHMSA 49 CFR 193 as a parallel workstream on any BTM or on-site gas project. And we track the DG Matrix nameplate-to-usable megawatt ratio as a structural risk parameter on any project that is relying on equipment rated below the AI-load envelope.

Predevelopment Services offers an equipment-stack diligence package — a four-week engagement that maps your specific equipment procurement plan against the current order book, identifies the bottleneck items, and produces a risk-adjusted COD. The engagement output is consistent: equipment diligence shifts the COD by six to eighteen months relative to the developer's initial plan.

The long-horizon bet is that equipment availability becomes the primary site selection driver by 2028. The queue will still matter. The incentives will still matter. But the operators who win the next wave of buildouts will be the ones who treated the OEM relationship, the ride-through study, and the LNG fuel assurance contract as strategic assets in 2025 and 2026 — and we are building the tooling to help the next 200 operators catch up.