The great data center delay: Why your AI chips are stuck in 2026

Bruce Bateman
Chief Analyst, Semiconductor

Pillar 1: Assessing risks

Regional semiconductor
market risk assessment

The constraints facing AI infrastructure in 2026 are no longer about demand or design— they are physical. The semiconductor industry is on track to approach $831 billion in revenue in 2025, up roughly 23.3% year-over-year, yet it is entering its most structurally risky phase since the post-COVID correction because supply cannot keep pace with the AI supercycle. The bottlenecks have shifted from transistor density to power delivery, high-bandwidth memory (HBM), advanced packaging, and critical materials, and each represents a single point of failure for AI chip delivery.

Trade policy layers on top of these physical constraints: imported AI accelerators, high-bandwidth memory (HBM) modules, and advanced-packaging substrates remain exposed to shifting US tariff schedules and the United States-Mexico-Canada Agreement (USMCA) rules-of-origin treatment, and chips routed through Mexican assembly or Canadian integration can lose preferential status if regional value-content thresholds are not met—material cost and lead-time risk that buyers should price into sourcing decisions.

Memory sits at the center of the crisis. Samsung, SK Hynix, and Micron's HBM capacity is fully committed through the end of 2026, with no spot market, and HBM consumes three to four times the wafer capacity per gigabyte compared with standard Dynamic Random Access Memory (DRAM). Server DRAM prices surged 60–70% in Q1 2026 as HBM lines displaced conventional supply, and HBM3e pricing is up roughly 20% year-on-year. TSMC's Chip on Wafer on Substrate (CoWoS) advanced packaging line runs at approximately 100% utilization, and shortages of thin glass fiber cloth (T-Glass) and helium threaten to slow packaging further. Consumer and automotive chips are expected to face a new shortage through 2027 as AI absorbs available capacity.

The power wall compounds the silicon shortage. Data center electricity consumption should exceed 500 TWh in 2026—about 2% of global usage—and AI racks now draw 50–100 kW compared with 5–10kW for legacy systems. Grid infrastructure takes 5–10 years to develop while data centers can be built in 12–24 months, creating a structural mismatch. Energy regulations now require hyperscalers to self-generate power, adding 18–24 months to site timelines, and power shortages are projected to constrain 40% of AI data center by 2027.

Pillar 2: Identifying opportunities

The same constraints that are delaying deployment are also unlocking unprecedented capital commitment and a realignment of where AI compute will be built. The Big Five hyperscalers have committed approximately $660–750 billion in capex for 2026, up 67% year-on-year, with roughly 75%—or $500–560 billion—allocated to AI-related investment in GPUs, custom silicon, and data centers. Amazon plans $200 billion, Alphabet $175–185 billion, and Meta $115–135 billion, collectively deploying close to 90% of operating cash flow to capex, up from 65% in 2025. Data center construction alone is projected at $89 billion in 2026, a 782% increase from 2022.

Project Stargate anchors the expansion with a $500 billion blueprint for AI factories across more than six US sites plus the UAE, Norway, and Argentina, targeting approximately 7 GW of capacity. Oracle has signed a 4.5 GW Stargate agreement worth more than $300 billion over five years. On the silicon side, NVIDIA's Rubin platform is in full production with 288 GB of HBM4 and 22 TB/s bandwidth at ten times lower cost per token, with volume shipments beginning in the second half of 2026. AMD's Helios and MI455X platforms deliver 432 GB HBM4 and anchor 6 GW of deals with OpenAI and Meta, while Broadcom's Tomahawk 6 networking silicon is already shipping in volume.

Regional capacity is expanding in parallel. SK Hynix is bringing online its Cheongju M15X fab for 1b-node DRAM and HBM4 and is investing $13 billion in HBM packaging at P&T7. Samsung's Pyeongtaek P4 is scaling to over 60,000 wafers per month by Q2 2026, and Micron's Idaho One fab targets mid-2027 output. For enterprises, the opportunity lies in securing allocation early, diversifying across custom ASICs, and positioning around the rural, gigawatt-scale campuses emerging in Texas and the Midwest to bypass saturated hubs.

Where do you expect the most AI infrastructure investment in 2026?

GPUs & custom silicon
Data centre construction
Cloud & software platforms

Semiconductor market
opportunity identification

Pillar 3: Implementing strategies

Translating this landscape into operational resilience requires concrete actions that decouple financial commitment from physical delivery risk. Leading buyers are moving away from single-vendor, mega-campus strategies toward multi-vendor inference and distributed siting. OpenAI's pivot is instructive: after abandoning the Abilene, Texas expansion from 1.2 to 2.0 GW due to financing disputes and power infrastructure delays, it is now favoring distributed sites in Wisconsin and Detroit and has revised its compute target to $600 billion from $1.4 trillion to align spending with revenue. Organizations sourcing AI compute should adopt the same discipline—spreading workloads across GPUs, TPUs, Trainium, and emerging custom ASICs to avoid exposure to any single supply bottleneck.

Dual-sourcing HBM and packaging allocation is equally critical. With Samsung and SK Hynix as sole HBM4 suppliers for NVIDIA's Vera Rubin and CoWoS-L packaging oversubscribed through 2026, long-term capacity agreements and reserved allocations have become the price of entry. Enterprises should also build optionality into silicon strategy, leveraging the widening GPU-versus-ASIC split, since custom ASICs are roughly twice as cost-effective for inference workloads that are already growing two to three times faster than training.

Finally, site selection now requires political and regulatory risk analysis alongside traditional power and fibre planning. With $98 billion in stalled pipeline projects, 25 data center cancellations in 2025, moratorium bills active in at least 11 states, and PJM capacity prices surging from $30 to $270 per MW-day, only about 10% of planned projects are likely to complete on schedule. The winning strategy pairs diversified silicon sourcing with a portfolio of smaller, politically viable sites—trading the speed of mega-campuses for the certainty of delivery before AI hardware's 3–5 year useful life expires.

Bruce Bateman
Chief Analyst

The insights in this chapter ‘The great data center delay: Why your AI chips are stuck in 2026’ is derived from: