Global AI server power demand capacity climbs from 53.8 GW in 2026 to 127.1 GW in 2028 and roughly 264 GW in 2030, close to a fivefold rise in four years, while general-purpose servers see their share of total data center power fall from around 40% historically to 25.6% in 2027 [S1]. The squeeze is not a chip story alone: large power transformers now take roughly two-and-a-half years to procure, gas turbines are booked into the late 2020s, and new transmission lines need years of permitting [S4].
For procurement teams buying general-purpose rack and tower servers, the practical effect is a shared line-item battle with hyperscale AI orders, since both pools pull from the same ODMs, the same 1U/2U chassis, the same industrial router and out-of-band management modules, and increasingly the same behind-the-meter substation capacity. Wholesale AI hardware spend by Microsoft, Google, Amazon and Meta ran $100 to $200 billion in a single six-month window in 2025, a record capex pace that has reordered queue positions across the supply base [S2].
Where the squeeze actually binds in 2026
Three bottleneck layers stack behind every server shipped, and each has a different cycle time, which is the root cause of the general-purpose squeeze [S2]. Chips carry a 6 to 12 month cycle on TSMC leading nodes and advanced packaging, with orders placed in early 2023 not delivering until 2024 or 2025. Data hall buildout adds 12 to 24 months of construction, cooling and network integration. Grid and on-site power run 3 to 5+ years, and several U.S. hubs, including Northern Virginia, Georgia and Ohio, are already hitting grid constraints [S2].
Because general-purpose servers share ODM lines, power conversion stages, fiber converter optics and rack-level PDUs with AI servers, their lead times now move with the AI shipment curve rather than with their own demand. TrendForce puts global data center power demand at 161.0 GW in 2026, up 31.0% year over year, and 211.1 GW in 2027, with AI servers accounting for 33.4% of capacity in 2026 and more than half (53.7%) by 2030 [S1]. General-purpose capacity keeps growing in absolute terms, but at the much slower pace of a minority share.
Power infrastructure, not silicon, is the binding constraint
On 2026-05-12, Wharton summarised the grid-side facts: PJM Interconnection, the nation's largest grid operator, fell nearly 6,600 megawatts short of its reserve target for summer 2027, with ninety-four percent of the projected load growth coming from facilities built to run artificial intelligence [S4]. Capacity prices on PJM hit an all-time high in the same auction, and the political response has been bipartisan, ranging from a proposed moratorium on new data center construction to state-level siting pushback [S4].
Excitement about AI runs at the speed of software; the infrastructure required to sustain it runs at the speed of steel, copper, and concrete, is the exact framing the Wharton operations faculty used to describe the mismatch [S4]. For buyers of general-purpose servers, this matters because any colocation or on-prem build that needs incremental MW now competes directly with AI training campuses on the same interconnect queue, the same substation transformers, and the same behind-the-meter gas turbine slots, and the same racks of protocol gateway and out-of-band management hardware sit on those queues whether the load is AI or not.
Vertical integration is locking out latecomers
The most capital-rich buyers are no longer waiting on the open market. GE Vernova closed a $5.3 billion deal for full ownership of transformer manufacturer Prolec GE, a move that signals how the smartest players are responding: they are buying vertical integration into a chokepoint that their customers cannot reach their AI ambitions without [S4]. The acquisition is a single data point, but it is consistent with the broader pattern of PPAs, equipment reservations and on-site generation that the early hyperscalers used to lock queue position years before general-purpose demand arrived.
The supply base that feeds both AI and general-purpose servers is unusually concentrated. A handful of specialised manufacturers in South Korea and Germany produce the silicon steel inside power transformers, European forging shops supply gas turbine shafts, and a small number of firms build the critical components through which electricity enters a substation [S4]. When these suppliers hit capacity, the effect propagates downstream into rack power, cooling and even serial server management modules, because every data hall expansion or retrofit consumes the same constrained inputs.
What procurement teams can and cannot change
General-purpose server demand is not collapsing, it is being throttled. TrendForce notes that general-purpose servers continue to exhibit growth in absolute power demand capacity, just far behind the much stronger growth shown by AI servers, and that the share decline from around 40% historically to 25.6% in 2027 reflects AI's faster curve, not a shrinking base [S1]. For a buyer, that means long-lead items should be ordered on AI-scale lead times even for non-AI workloads: 18 to 30 months for transformers and turbines, 12 to 24 months for hall buildout, and 6 to 12 months for ODM chassis and PSUs [S2][S4].
Options on the table include behind-the-meter generation contracts, HVDC distribution architectures that improve end-to-end power conversion efficiency, and earlier reservation of substation slots, all of which TrendForce flags as active planning responses as of 2026-09-22 [S1]. The same source notes that supply and demand are expected to begin diverging in 2026, with the gap widening significantly from 2028 onward, and a US-specific shortfall projected to exceed 170 GW by 2030, even before global figures are netted in [S1].
Forecast trajectory and the inference tailwind
Global AI server shipments are forecast to grow over 28% year over year in 2026, with a rising share of ASIC-based systems, and the rapid growth of AI inference services is explicitly boosting demand for general-purpose servers, supporting both replacement and expansion efforts [S5]. That last clause matters for general-purpose buyers: inference traffic, unlike training, runs on standard CPUs, single-GPU nodes and the same industrial modem and management plane as a typical enterprise rack, so the inference tailwind pulls general-purpose unit volumes up even as AI accelerators dominate the watt count.
For more on the upstream side of the same squeeze, the EHV cable and transformer order book situation through 2026 tracks the same bottleneck layer and is covered in EHV Cable Plants Booked Through 2026: Order Books, Lead Times, New Capacity. Track two signals through 2027: PJM and MISO capacity auction clearing prices, which already printed all-time highs in the 2026 auction cycle [S4], and TrendForce's quarterly revision of the AI server power share, which crossed 33.4% in 2026 and is projected above 40% in 2027 [S1].