In 2025, SK Hynix overtook Samsung in annual operating profit for the first time in either company’s history — 47.2 trillion won versus Samsung’s 43.6 trillion won. SK Hynix makes memory chips. Samsung makes everything from smartphones to foundry chips to home appliances. That a focused memory specialist outearned the conglomerate comes down to one product: High Bandwidth Memory.
The HBM market is a near-duopoly. Two South Korean companies control approximately 78–92% of all HBM supply. When one of those players holds roughly 70% share and has sold out all 2026 production, the ripple hits every infrastructure budget worldwide. This article maps how SK Hynix built that lead, why Samsung fell behind, and what it means for enterprise buyers planning hardware spend. It’s part of the AI memory crunch series.
SK Hynix and Samsung together hold approximately 78–92% of all HBM revenue. Counterpoint Research‘s Q3 2025 data puts SK Hynix at 57% and Samsung at 22%. TrendForce estimates SK Hynix climbed to roughly 70% by Q1 2026. Micron holds the remaining 21–24% — meaningful, but structurally behind in the AI-grade qualification cycles that actually matter.
It wasn’t always this concentrated. In 2009, ten companies competed in DRAM. The 2011 downcycle triggered terminal consolidation: SK Telecom acquired Hynix in 2012, Micron purchased bankrupt Elpida in 2013. Within five years it was a triopoly.
Re-entry isn’t realistic. A greenfield DRAM fab costs $15–25 billion with a four-to-five year build time, plus 18–36 months of yield ramp. HBM packaging adds another $2–5 billion per node. Total cumulative capex to compete at meaningful scale runs $30–50 billion over seven to ten years. Chinese manufacturers CXMT and YMTC get cited as challengers regularly — they’re three to five nodes behind and blocked from EUV lithography by export controls. This structure holds for at least a decade.
IDC frames it directly: 2026 DRAM supply growth is expected at only 16% year-on-year versus 20–30% historical norms, because wafer capacity has been deliberately reallocated to higher-margin HBM. This isn’t a self-correcting shortage. It’s a permanent structural shift.
SK Hynix’s dominance comes down to a manufacturing process difference that compounds mathematically across every wafer.
HBM is built by stacking 12 DRAM dies vertically using Through-Silicon Vias — copper-filled channels drilled through each die, roughly 5–10 micrometres wide.
💡 A Through-Silicon Via (TSV) is a conductive channel drilled straight through a silicon chip, allowing electrical signals to travel vertically between stacked chips rather than across longer horizontal interconnects.
Here’s why that matters. An HBM3E 12-Hi stack contains 1,200–1,800 TSVs per layer. One defective TSV kills the entire stack. At 99% per-die yield, a 12-die stack yields 88.6% usable. Drop to 95% per-die yield and you get 54%. A 4-point improvement in per-die yield translates to a 34-point improvement in stack yield. Small differences at the die level determine who wins the market.
SK Hynix uses MR-MUF — Mass Reflow with Molded Underfill. All 12 dies bond and seal in a single thermal cycle, reducing voids and raising yield. Their 1bnm DRAM node runs at 75–80% stack yield.
Samsung uses TC-NCF — Thermal Compression with Non-Conductive Film — where adhesive film is applied layer by layer. Historically slower, historically lower yield. Their “Advanced TC-NCF” variant narrowed the gap but didn’t close it. Samsung’s 1cnm node delivers 60–65% stack yield.
That 10–15 point gap, compounded across 12 dies, is the manufacturing basis for SK Hynix’s market position: more usable stacks per wafer, lower defect rates, faster Nvidia qualification. For a full engineering treatment, the technical architecture that explains SK Hynix’s TSV advantage goes deeper.
Samsung’s HBM3E failure wasn’t a single misstep. It was a two-year compounding sequence.
It started with thermal failures during Nvidia qualification in 2023. Samsung’s TC-NCF stacking produced higher thermal resistance than Nvidia’s budget allowed. Under sustained AI training workloads, heat couldn’t dissipate efficiently. Samsung’s parts on the 1anm DRAM node missed yield targets — and the timing couldn’t have been worse. The highest-volume AI GPU production cycle in history was just ramping up.
Nvidia qualifies HBM suppliers for each GPU platform independently. Losing HBM3E qualification for H100 and early Blackwell builds (B200/B300) meant Samsung couldn’t participate in the highest-margin memory window of the AI infrastructure boom. It got pushed into lower-margin segments, pricing HBM3E 12-layer products approximately 30% below SK Hynix just to secure whatever allocation it could.
By Q2 2025, Samsung held only 17% HBM share. Counterpoint’s Q3 2025 data showed Samsung at 22%, SK Hynix at 57%. Samsung eventually received HBM3E 12-Hi qualification from Nvidia in November 2025 — but by then SK Hynix had locked in multi-year hyperscaler supply contracts that qualification alone couldn’t unwind.
Here’s the distinction that gets conflated everywhere: Samsung reclaimed total DRAM revenue leadership in Q4 2025. Real, worth noting. But DRAM revenue leadership and HBM-specific revenue leadership are completely separate measurements. In HBM, Samsung still trails at 22% versus SK Hynix’s 57–70%. Conflating the two gives you a misleading picture of the competitive position that actually matters for AI infrastructure.
SK Hynix reported Q1 2026 operating profit of 37.61 trillion won at a 72% operating margin — a record for any memory company in any quarter. For full year 2025, SK Hynix’s 47.2 trillion won surpassed Samsung’s 43.6 trillion won. Samsung’s memory segment generated approximately 24.9 trillion won in 2025 — meaning SK Hynix earned nearly double Samsung’s memory-only profit.
HBM share sits at approximately 57% per Counterpoint’s Q3 2025 data, climbing toward 70% by TrendForce’s Q1 2026 estimates. All 2026 HBM production is sold out.
SK Hynix has committed ~19 trillion won (~$13 billion) to the new Cheongju M15X mega-fab. SK Group Chairman Chey Tae-won stated publicly in March 2026 that the wafer shortage will likely persist until 2030. M15X is forward-looking — targeting HBM4 and HBM4E ramps. SK Hynix is planning around a multi-year constrained market, not a short cycle.
One thing worth flagging: approximately 42% of SK Hynix’s total revenue comes from a single customer — Nvidia. The dominant position is real, but the financials are heavily exposed to any change in Nvidia’s sourcing strategy. That dependency belongs in any honest assessment of SK Hynix as a long-term supplier.
Samsung shipped its first commercial HBM4 on February 12, 2026. SK Hynix had its HBM4 mass-production system in place in September 2025, delivering large volumes of paid samples to Nvidia well before that. The production readiness gap is roughly five to six months, following about one year of sampling lead time.
That lag is evidence the quality gap from HBM3E is not fully closed. Samsung is recovering while SK Hynix continues to scale — those aren’t the same thing.
Nvidia’s Vera Rubin HBM4 allocation reflects the current position: approximately two-thirds to SK Hynix, one-third to Samsung. An improvement from near-zero during peak HBM3E allocation — but not parity.
Samsung’s most concrete win this cycle is the AMD MI400 AI accelerator, committed primarily to Samsung HBM4. That’s the first major AI GPU socket Samsung has fully won this cycle, giving it a foothold independent of Nvidia. Samsung’s internal projection is that HBM sales will more than triple in 2026 — genuine momentum, though tripling still leaves it well behind.
Samsung is also working to reduce its dependency on TSMC‘s CoWoS-L packaging capacity, sold out through 2027. Samsung’s exploration of alternatives to CoWoS packaging maps this in detail.
The more interesting story is HBM4E, expected in 2027. Samsung is the only memory company with an in-house logic foundry. HBM4 and HBM4E require a logic base die fabricated on an advanced node. Samsung can co-design the DRAM dies and the logic base die within a single organisation — no external foundry dependency, no competing for TSMC allocation. No other memory company can do that. The battleground for parity is HBM4E, not HBM4.
As Micron’s business chief Sumit Sadana put it: “As we increase HBM supply, it leaves less memory left over for the non-HBM portion of the market, because of this three-to-one basis.”
That’s the crux of it. Each wafer allocated to HBM displaces approximately three wafers of conventional DDR5 output. With HBM3E ASPs running six to eight times conventional DDR5 per gigabyte — and HBM4 at roughly ten times — both manufacturers have every incentive to prioritise HBM. DDR5 supply tightens regardless of DDR5 demand.
TrendForce data shows DRAM contract prices rose 90–95% quarter-over-quarter in Q1 2026, with a further 58–63% increase projected in Q2 2026. NAND prices rose 70–75% in Q1 2026. Dell raised hardware prices 17% on March 30, 2026; HP reported memory costs doubled in a single quarter, now representing 35% of PC build materials.
In a normal DRAM cycle, Samsung could flood supply to gain share. In HBM, that’s not possible — qualification takes 6–18 months per product cycle. That lock-in is what makes this shortage structurally different: the duopoly has pricing power in both HBM and conventional DRAM simultaneously, because wafer capacity is coupled. The DRAM budget impact analysis covers the enterprise cost implications in detail.
The one wildcard in this structure is Micron as the US entrant that could disrupt this duopoly. Micron holds 21–24% HBM revenue share today and has its HBM supply sold out through 2026 and 2027 — but as the sole US-headquartered HBM manufacturer, its trajectory matters for both enterprise procurement options and geopolitical supply chain resilience.
The allocation structure has three tiers, and enterprise buyers are at the bottom.
At the top, hyperscalers — Microsoft, Google, Amazon, Meta — have signed multi-year Long-Term Agreements (LTAs) that lock in supply and pricing. SK Hynix is reportedly receiving direct financing from Big Tech in exchange for committed supply. Google has stationed procurement executives in South Korea. These are not negotiations enterprise buyers can participate in.
Below that, buyers without multi-year LTAs compete for residual allocation at spot pricing. DRAM prices have surged 80–90% this quarter. Lead times have stretched beyond 40 weeks, beyond 58 weeks for some configurations.
For HBM: enterprise buyers can’t purchase from SK Hynix or Samsung at all. HBM is sold exclusively to GPU OEMs — Nvidia, AMD — who integrate it into products. Your constraint is GPU availability and server OEM lead times, not direct memory procurement.
For conventional DRAM and NAND, two things to accept:
Multi-year contracts are now a prerequisite, not a negotiating advantage. Without an LTA, you’re on spot pricing with no predictability.
Infrastructure planning has to run over 18–24 month horizons. Spot-market buying for server memory is a losing strategy. Buyers who delay face higher prices and longer waits simultaneously.
The SK Hynix Q1 2026 earnings call framed it well: “customers are prioritising procurement over price.” Paying more doesn’t move you up the queue when production is fully committed. This is a calendar and relationship problem. For a complete view of the structural forces shaping every layer of this market — from hyperscaler contracts through to enterprise and consumer repricing — see the AI memory crunch overview.
SK Hynix’s MR-MUF process produces higher TSV yield than Samsung’s TC-NCF method. At 12 stacked dies, the compounding is significant: 75–80% stack yield (SK Hynix 1bnm) versus 60–65% (Samsung 1cnm). SK Hynix was the sole HBM3 supplier to Nvidia’s H100 and used that position to lock in multi-year contracts. Samsung’s HBM3E qualification delays in 2024–2025 then excluded it from the highest-volume Blackwell GPU run. Samsung reclaimed total DRAM revenue leadership in Q4 2025 — real, but a different measurement from HBM-specific leadership.
Not entirely. Samsung received HBM3E 12-Hi qualification from Nvidia in November 2025, but missed the bulk of Blackwell volume production. Samsung is now an active HBM4 supplier for Vera Rubin, receiving approximately one-third of allocation versus SK Hynix’s two-thirds. The relationship is recovering, not severed.
No. HBM is sold exclusively to GPU OEMs — Nvidia, AMD. Enterprise buyers access it indirectly through AI accelerator hardware. For conventional DRAM and server memory, buyers can negotiate with distributors and OEMs, but without multi-year LTAs they face spot pricing and lead times beyond 40 weeks.
Samsung’s HBM4 programme is live (first shipment February 2026) and HBM sales are projected to more than triple in 2026. TrendForce estimates Samsung’s share could surpass 30% in 2026. Parity with SK Hynix isn’t expected before HBM4E. Samsung’s best path is foundry integration: co-designing DRAM and logic base die in-house is a differentiator no other memory company can replicate.
HBM3E is the current volume product: 12-Hi stacking, ~1.2 TB/s per stack; SK Hynix dominates at 57–70%. HBM4 doubles the interface bus to 2,048 bits and requires a logic base die on an advanced process node. SK Hynix sampled HBM4 roughly one year before Samsung’s first commercial shipment and holds ~70% of Nvidia Vera Rubin allocation. The next battleground for parity is HBM4E in 2027.
MR-MUF (SK Hynix): all 12 dies bond in a single thermal cycle — fewer voids, higher yield; 75–80% stack yield on the 1bnm node. TC-NCF (Samsung): non-conductive film applied layer by layer, historically slower and lower yield; “Advanced TC-NCF” narrowed the gap; 60–65% yield on the 1cnm node. The result: SK Hynix produces more usable stacks per wafer with fewer failures under sustained thermal load — exactly the condition during AI training.
Manufacturers can’t add HBM supply without Nvidia or AMD qualification — 6–18 months per cycle. The three-to-one wafer trade ratio means adding HBM supply simultaneously tightens conventional DRAM supply. IDC characterises this as a permanent strategic reallocation of global wafer capacity, with supply growth at 16% year-on-year versus 20–30% historical norms.
Approximately 42% of SK Hynix’s total revenue comes from Nvidia alone. The dominant position is real, but the financials are disproportionately exposed to any shift in Nvidia’s sourcing strategy, GPU architecture changes, or demand softening. Worth factoring into any assessment of SK Hynix as a durable long-term supplier.
The Cheongju M15X is SK Hynix’s new mega-fab in South Korea — a ~$13 billion investment targeted at HBM4 and HBM4E production ramps. SK Group Chairman Chey Tae-won’s March 2026 statement that the wafer shortage will likely persist until 2030 shows how SK Hynix is planning: around a multi-year constrained market, not a short-term cycle.
HBM the Chip Nobody Planned ForAI GPU prices have roughly doubled in the past 18 months. Cloud AI compute costs keep climbing. And at the centre of all of it is a chip most people outside the semiconductor industry had never heard of three years ago: High Bandwidth Memory, or HBM.
Here is the uncomfortable part. The memory industry came out of a brutal 2022–2023 oversupply bust — manufacturers slashed production to defend prices — then watched AI demand explode in 2024 with almost no spare capacity left. The buffer was gone.
This article is the technical foundation for the AI memory crunch series. It gives you a plain-English explanation of what HBM is, why it is so expensive to make, and why the shortage is structural rather than a temporary blip. It assumes you are technically literate — you have written software, you understand servers — but you are not a semiconductor engineer. Good. Let’s get into it.
High Bandwidth Memory is a type of DRAM — the same base technology as the DDR5 in a server — but physically restructured to deliver radically more bandwidth. Instead of sitting on a PCB slot away from the processor, HBM stacks multiple memory dies vertically and mounts them within approximately 1 mm of the GPU die, connected via a silicon interposer.
The bandwidth gap is the clearest way to understand the difference. HBM3E delivers approximately 1.2 TB/s per stack. DDR5 in dual-channel delivers approximately 102 GB/s. That is a 12-to-1 ratio.
Modern large language models use a transformer architecture. During inference — generating each token — the GPU must load model weights and KV cache values from memory on every single step. The processor can compute far faster than commodity DRAM can supply data. This is the memory wall, and generative AI has weaponised it at commercial scale.
The KV cache holds the “key” and “value” attention representations of every prior token in the context window. For a 128,000-token context window at FP16 precision, that runs to tens of gigabytes per transformer layer per active request. Inference is memory bandwidth-bound, not compute-bound. HBM is not a nice-to-have.
The H100 carries 80 GB of HBM3; the H200 carries 141 GB of HBM3E; the Blackwell B300 carries 288 GB — demand per GPU roughly doubled between generations. HBM now accounts for 50–60% of GPU manufacturing cost. That is why this is the central challenge driving this structural memory shortage.
HBM achieves its bandwidth advantage by stacking up to 16 DRAM dies vertically and connecting them through tiny copper conductors drilled through the silicon itself. Picture a multi-storey car park: each floor is a memory die, and the elevator shafts running floor-to-floor are Through-Silicon Vias (TSVs).
A TSV is a vertical electrical connection drilled through the full thickness of a silicon die and filled with copper. In an HBM3E 12-Hi stack, each die is ground to approximately 30 µm thick — roughly a third of the diameter of a human hair. The dies are joined at each TSV endpoint by microbumps, tiny solder connections completing the circuit.
The result: a 1,024-bit-wide memory bus compared to DDR5’s 64-bit channel. That is why bandwidth is so dramatically higher even though the underlying DRAM cell technology is identical.
Both the HBM stack and the GPU die sit on a silicon interposer — a passive substrate providing the dense wiring to connect HBM’s wide bus to the processor. The assembly is integrated through TSMC’s CoWoS (Chip on Wafer on Substrate) process, which is itself a supply constraint covered in CoWoS advanced packaging — the second bottleneck in the chain.
All of those steps — TSV drilling, wafer thinning, microbump formation — compound to produce the most important economic fact in the shortage.
Micron has stated on record that producing one unit of HBM output requires the wafer capacity that would otherwise have yielded three units of DDR5 output. That 3-to-1 ratio is the single most important economic fact in this shortage.
Three compounding causes explain it.
TSV area overhead. Each TSV and its surrounding keep-out zone consumes silicon area that would otherwise hold memory cells. An HBM die is significantly less bit-dense per unit of wafer area than an equivalent DDR5 die. You are trading storage density for interconnect capability on every square millimetre of silicon.
Multi-step back-end processing. Standard DRAM finishes at the wafer level with relatively straightforward packaging. HBM requires TSV etch, TSV fill, wafer thinning, microbump formation, die-to-die stacking, and CoWoS integration — each step adding its own yield loss.
Wafer thinning yield penalty. At 30 µm thickness, silicon is fragile — closer to glass than rigid chip material. Breakage here is not trivial.
Every wafer diverted to HBM removes roughly three wafers-worth of DDR5 output from the market. DRAM lead times have stretched beyond 40 weeks. Every HBM wafer is a wafer denied to the server RAM in your cloud provider’s next refresh.
Wafer reallocation is the deliberate decision by a DRAM manufacturer to redirect wafer starts — units of production scheduled on fab equipment — from lower-margin products like DDR5 and LPDDR5 to higher-margin HBM. It is a scheduling decision, not a physical constraint.
The cleanroom equipment that runs DDR5 can, with modification, run HBM dies. The shortage is not a physics problem — it is a scheduling decision, and it reverses only when the economics shift.
Right now, the economics are not shifting. HBM3E commands prices approximately 6–8x higher per gigabyte than conventional DDR5. SK Hynix posted record operating profit for full-year 2025, surpassing Samsung for the first time since 1992. Micron’s HBM capacity for both 2025 and 2026 is fully committed.
SK Hynix holds approximately 45% of HBM market share and was first to ship HBM3E at scale. Samsung’s HBM3E modules did not qualify at Nvidia until November 2025. That yield situation is counter-intuitive: capacity that cannot ship qualifying product still consumes wafer starts and removes them from DDR5 supply. Bad yield tightens supply. The full competitive picture is in SK Hynix and Samsung — who makes HBM and why one dominates.
Each generation of HBM delivers higher bandwidth and capacity per stack but adds manufacturing complexity in doing so. The shortage deepens with each generation rather than plateauing — and AI chip roadmaps are already committed to the most complex generation before supply is established.
HBM3 (JEDEC January 2022) was first shipped in Nvidia’s H100, delivering up to 819 GB/s via a 1,024-bit bus across 8–12 die stacks. It established HBM as the default AI accelerator memory interface.
HBM3E (current mainstream, 2024–2025) raises I/O speeds to 1.2 TB/s across the same bus. The H200 and AMD MI300X are its primary consumers. SK Hynix was first to mass production; Samsung qualified at Nvidia only in November 2025.
HBM4 (JEDEC April 2025, shipping 2026) doubles the interface to 2,048 bits and targets 2 TB/s, with up to 16-Hi stacks and 64 GB capacity. The base die is fabricated at TSMC’s N5 node — the first time HBM’s logic layer has required a leading-edge logic process, adding a new dependency on TSMC capacity. SK Hynix shipped its first HBM4 samples to Nvidia in March 2026, six to nine months ahead of Samsung and Micron.
Nvidia’s Vera Rubin platform was designed around HBM4 before HBM4 supply was established. That is exactly why the fab timeline in the next section matters.
New fab capacity will not quickly resolve the shortage. The gap between a groundbreaking announcement and qualified production output is measured in years — and the capacity arriving first is already spoken for.
The construction lag is 3–5 years from groundbreaking to first qualified output. Qualifying a finished fab for HBM production adds another 6–12 months of test runs and customer sign-off. “Fab running” and “qualified HBM shipping at volume” are not the same milestone.
Known timelines: Micron’s Singapore and Boise Idaho fabs come online mid-2027. SK Hynix Cheongju follows in 2027, Indiana in late 2028. Samsung Pyeongtaek is 2028. Intel CEO Lip-Bu Tan was direct: “There’s no relief until 2028.”
When that new capacity arrives, it will not be on the open market. Hyperscalers are signing multi-year supply agreements to lock in capacity before fabs are even operational. SK Hynix has stated that HBM4 demand over the next three years already exceeds its supply capacity.
GPU availability and cloud AI pricing are unlikely to ease materially before 2028. Build that into your infrastructure planning — the budget implications are in what these price surges mean for enterprise hardware budgets.
Training gets most of the attention when people talk about AI memory demand. But agentic inference is structurally more memory-intensive per unit of useful output.
Training is bounded and predictable: data flows through in batches, gradients are computed, weights are updated. You can tune a cluster for it. Agentic inference is a loop — reason, call tools, retrieve data, generate intermediate outputs, arrive at a response. Each step reloads context into the KV cache. SK Hynix said this explicitly in Q1 2026 earnings: “As AI evolves from large-scale model training to agentic AI… demand for memory is expected to continue growing.” That is Nvidia’s primary HBM supplier.
As context windows grow from 8,000 to 1 million tokens, the KV cache grows proportionally — exceeding 600 GB per request at 1 million tokens. Cloud providers serve many simultaneous sessions, each maintaining its own cache. An agentic fleet’s memory requirements scale with adoption and context length in ways no training job ever did. Micron predicts the HBM market will grow from $35 billion in 2025 to $100 billion by 2028 — larger than the entire DRAM market in 2024.
HBM is not a temporary supply blip. It is the simultaneous compounding of three forces: a physics constraint (the memory wall), a deliberate economic choice (wafer reallocation), and a demand acceleration (agentic AI creating memory requirements that scale non-linearly with adoption).
The question to carry out of this article is whether this is structural or cyclical. Cyclical corrections require a period of oversupply. Given advance commitment of new fab capacity and demand forecasts that continue to outpace supply projections, that oversupply window is not visible in the near term.
The manufacturer concentration that makes this shortage so acute — and so resistant to market correction — is the subject of SK Hynix and Samsung — who makes HBM and why one dominates. The packaging bottleneck that constrains GPU production even when HBM dies are available is in CoWoS advanced packaging — the second bottleneck in the chain. What all of this means for enterprise hardware budgets is in what these price surges mean for enterprise hardware budgets. For a complete overview of how all these forces interact, see our comprehensive AI memory crunch resource.
Not exactly. HBM is the memory inside modern AI GPUs — but consumer gaming GPUs use GDDR6X, which is cheaper and lower-bandwidth. HBM is reserved for data-centre AI accelerators (H100, H200, B200, MI300X) where bandwidth is the binding constraint.
No. HBM requires a silicon interposer and CoWoS packaging that is incompatible with standard DIMM slots or laptop form factors, and costs roughly 6–8x more per gigabyte than DDR5. Apple’s M-series unified memory is a different approach — it is not HBM.
Related but distinct. The HBM shortage is driven directly by AI demand. The DDR5 shortage is a secondary effect: every HBM wafer removes three DDR5 wafers-worth of output from the market. DDR5 prices are collateral damage.
Processors can only work on data already in fast memory. If data cannot be loaded fast enough, the processor waits idle — this is the memory wall. For LLMs, model weights and KV cache are enormous relative to on-chip SRAM. Without HBM bandwidth, raw compute is wasted.
Wafer starts per month is the fundamental production unit for semiconductor fabs — fixed in the short term. When HBM consumes 3x the wafer capacity per gigabyte versus DDR5, every HBM wafer start removes three DDR5 wafers-worth of supply from the market. That lands on your cloud and hardware costs.
Two compounding factors: wafer reallocation from DDR5 to HBM cut commodity DRAM output, and production capacity was slashed in 2022–2023, leaving no buffer when AI demand surged in 2024. Samsung and SK Hynix raised server memory prices by up to 70% in Q1 2026, following 50% increases throughout 2025.
HBM3 (2022) delivers up to 819 GB/s via a 1,024-bit bus — first used in the H100. HBM3E (2024) delivers 1.2 TB/s on the same bus, used in the H200 and AMD MI300X. HBM4 (2025) doubles the bus to 2,048 bits and targets 2 TB/s. Each generation adds capability and manufacturing complexity.
HBM stacks and the GPU die both sit on a silicon interposer — a passive substrate providing the dense wiring to connect HBM’s wide bus to the processor. The assembly is integrated through TSMC’s CoWoS process, which is itself a supply constraint covered in CoWoS advanced packaging — the second bottleneck in the chain.
The KV cache stores the “key” and “value” attention representations of every prior token in the context window — so the model does not recompute them on each generation step. At 1 million-token context the cache exceeds 600 GB per active request. Context window size, not model weight size, is the dominant memory constraint.
Partially, but slowly. Micron’s Singapore and Boise fabs arrive in 2027; SK Hynix Indiana and Samsung Pyeongtaek follow in 2028. That capacity is pre-committed to hyperscalers. SK Group Chairman Chey Tae-won (March 2026) put the shortage persisting until 2030, with a projected wafer shortfall exceeding 20%.
SK Hynix (~45% share), Samsung, and Micron. Producing HBM requires mastery of TSV processing, advanced packaging, and a decade of co-development with GPU makers. A greenfield DRAM fab costs $15–25 billion with a 4–5 year lead time; no new entrant is plausible at scale. The full competitive picture is in SK Hynix and Samsung — who makes HBM and why one dominates.
How the AI Memory Crunch Is Reshaping the Global Chip Supply ChainThe AI buildout has consumed the majority of the world’s specialised memory supply — and the knock-on effects are showing up in server quotes, laptop prices, and hardware refresh budgets everywhere. Every major AI accelerator requires vast quantities of high-bandwidth memory (HBM) that draws on the same cleanroom capacity used to make the RAM in your servers, laptops, and phones. Because building one gigabyte of HBM capacity displaces three gigabytes of conventional DRAM production — the 3-to-1 trade ratio — the AI buildout is draining the memory supply for everyone else. The shortage is not a supply blip that will self-correct in a few quarters. It is a structural reallocation of global chip manufacturing capacity that will last through the decade, and it is the background condition for every infrastructure decision your business makes right now. This guide explains the core thesis, answers the ten broadest questions, and links to six in-depth articles covering every angle from chip technology to your hardware budget.
In this guide:
The AI memory crunch is a global shortage of semiconductor memory caused by hyperscalers — Microsoft, Google, Amazon, Meta — pouring hundreds of billions into AI data centres and consuming the majority of the world’s specialised memory supply. Because AI memory is manufactured on the same production lines as ordinary RAM, there is less of everything else to go around. The four largest AI chip designers absorbed roughly 90% of global HBM and advanced packaging capacity in 2025, with hyperscaler capex forecast at $650 billion in 2026 alone. Prices for server and consumer RAM have risen steeply as a result. To orient yourself: “RAM” and “DRAM” refer to the same broad category of memory; HBM is a high-value sub-type; and the shortage in HBM causes a secondary shortage of conventional DRAM that hits every device that uses it. The technical foundation — what HBM is and why it consumes three times the wafer capacity of DDR5 — is the place to start if the chip architecture is unfamiliar.
Deep dive: HBM the Chip Nobody Planned For.
A cyclical shortage self-corrects in 12–18 months as manufacturers ramp production. This shortage is structural because adding memory capacity requires building new semiconductor fabs — a process that takes 3–5 years and costs tens of billions of dollars. The demand driving the shortage is a sustained, multi-year buildout of AI infrastructure, not a temporary spike. Normal boom-bust correction mechanisms cannot operate at the pace AI demand is growing, and every major CEO in the memory and chipmaking industries delivered the same message through 2025 earnings season: demand is rising much faster than capacity can be built. The structural diagnosis is the key planning input: if the shortage were cyclical, waiting it out would be rational — but the evidence is that it is structural, which means adapting procurement and infrastructure strategies now is the more defensible path. For a detailed look at what HBM is and why it consumes three times the wafer capacity of DDR5, the foundational article explains why correction timelines are measured in years, not quarters.
HBM (high-bandwidth memory) is built by stacking 8–16 individual DRAM dies vertically into a single package, delivering far higher data throughput than conventional flat memory chips. AI accelerators require it because large language models have exposed the “memory wall” — the point where processing speed outpaces how quickly data can be fed from memory. Nvidia’s latest GPU carries up to 288 GB of HBM4; a single NVL72 rack contains 13.4 TB, enough memory for a thousand high-end smartphones. HBM cannot be substituted with cheaper memory types for high-end AI training; the architecture of large language models makes high bandwidth non-negotiable. Once HBM is manufactured, it must then be bonded to the GPU die through a process that explains how CoWoS advanced packaging became a second independent bottleneck in its own right.
Full technical breakdown: HBM the Chip Nobody Planned For.
Three manufacturers produce all of the world’s HBM: SK Hynix (South Korea), Samsung Electronics (South Korea), and Micron Technology (United States). SK Hynix leads with roughly 57% revenue share, having capitalised on an early technical advantage to supply the majority of Nvidia’s HBM. Samsung has faced validation delays, while Micron discontinued its consumer Crucial RAM brand in late 2025 to redirect all capacity to HBM. SK Hynix confirmed all 2026 supply is sold out; multi-year hyperscaler contracts effectively remove HBM from the spot market for smaller buyers. The competitive analysis of why SK Hynix holds 70% of the HBM market and what Samsung is doing about it explains why this supply concentration shows no sign of shifting quickly.
Duopoly dynamics: Samsung vs SK Hynix the HBM Duopoly Under Strain.
CoWoS (Chip-on-Wafer-on-Substrate) is TSMC‘s advanced packaging process that physically bonds HBM stacks to a GPU die on a shared silicon interposer — and it is a separate constraint from HBM supply itself. Even if sufficient HBM exists, no AI chip can ship without CoWoS assembly. TSMC is the sole provider of this process at leading-edge scale, it has been sold out through 2026, and Nvidia alone accounts for roughly 60% of that capacity. OSAT subcontractors cannot substitute for leading-edge CoWoS at scale.
Packaging bottleneck deep dive: TSMC CoWoS Packaging the Silent Bottleneck in the AI Chip Supply Chain.
TrendForce forecast a 50–55% quarter-on-quarter increase in Q1 2026, with further increases projected for Q2 2026. Enterprise server DRAM has roughly doubled, and Gartner projects combined DRAM and SSD prices will surge 130% by end of 2026. The mechanism is the 3-to-1 trade ratio: every gigabyte of HBM capacity added removes three gigabytes of conventional DRAM from the global supply pool. Independent analysts confirm the gap persists — IDC has quantified supply-demand gaps of approximately 4% for DRAM and 3% for NAND. No segment of the market is unaffected. Understanding how Micron’s record quarter fits into the US memory chip strategy provides important context on whether new supply is coming fast enough to ease these prices.
Full price analysis: DRAM Up 70 to 110 Percent What It Means for Enterprise Hardware Budgets.
Dell raised hardware prices 17% in March 2026; Lenovo warned customers that all quotations would expire on January 1, 2026, with new pricing reflecting the structural shortage. Server lead times have stretched to 4–8 weeks across major OEMs, and a project budgeted at $500,000 for server procurement can cost an additional $75,000 or more depending on order timing. Cloud vs. on-prem trade-offs have shifted: on-prem locks in today’s elevated prices, while cloud absorbs the increases over time but at a premium to pre-crunch levels.
Procurement analysis: DRAM Up 70 to 110 Percent What It Means for Enterprise Hardware Budgets.
Because AI chips and consumer devices draw on the same wafer manufacturing capacity, a shortage in one creates scarcity in the other. DRAM could account for as much as 30% of low-end smartphones’ bill of materials in 2026, up from around 10% in early 2025, forcing OEMs to absorb costs or pass them to consumers. PC and laptop prices are up 15–20%; Sony is reportedly considering delaying the next PlayStation console to 2028–2029; and Oppo has cut its 2026 shipment forecast by up to 20%. The full causal chain — how AI memory demand cascades from data centres to consumer laptops and phones — runs from hyperscale GPU clusters through enterprise servers all the way to the device in your pocket.
Consumer impact deep dive: From Data Centres to Phones the Consumer Ripple Effect of the AI Memory Crunch.
Micron secured up to $6.1–6.4 billion in CHIPS Act grants and is expanding fabs in Boise, Idaho (initial production mid-2027) and Clay, New York (wafer output H2 2028). US export controls on advanced lithography equipment keep Chinese producer CXMT — which could theoretically add significant DRAM supply — off Western supply chains. The concentration of HBM production in South Korea and CoWoS packaging in Taiwan represents a supply chain risk that US policy is actively working to diversify, with results visible only in the late 2020s.
US strategy and Micron deep dive: Microns Record Quarter and the US Memory Chip Strategy.
Meaningful supply relief is unlikely before 2027 at the earliest, and broad normalisation is a 2028–2029 story. Micron’s Idaho fabs come online mid-2027; SK Hynix’s Yongin campus adds capacity from 2027; Micron’s New York facility is not online until H2 2028. HBM demand is forecast to grow 70% year-on-year in 2026, which will absorb much of the new capacity as it arrives. Intel CEO Lip-Bu Tan said it plainly at the Cisco AI Summit in February 2026: “There’s no relief until 2028.” In the meantime, what a 70-110% DRAM price surge means for enterprise hardware budgets is the most actionable place to focus your procurement planning.
Supply forecasts and timeline analysis: Microns Record Quarter and the US Memory Chip Strategy.
Memory manufacturers are diverting production capacity away from conventional DRAM — used in PCs, servers, and smartphones — and toward HBM, the specialised memory inside AI chips. The mechanism is the 3-to-1 trade ratio explained in the DRAM prices section above: every unit of HBM capacity added removes three units of conventional DRAM from global supply. With hyperscalers spending hundreds of billions on AI infrastructure through 2026, the AI buildout is absorbing the majority of global memory manufacturing output.
The Covid shortage was cyclical — a temporary mismatch caused by factory shutdowns and demand spikes that self-corrected within 18–24 months. The current crunch is structural, driven by a sustained multi-year AI infrastructure buildout that requires new fab construction to resolve. Normal correction mechanisms cannot keep pace with AI demand growth; the planning horizon is years, not quarters.
These are three distinct memory types with different architectures, and they are not interchangeable for their primary use cases. HBM is required for AI training and large-scale inference; GDDR is used in gaming GPUs; DDR5 is the standard for servers and PCs. All three are manufactured on DRAM wafer capacity, making them competitors for the same production resources — but you cannot use DDR5 in an Nvidia H100 or HBM in a laptop.
There is no clearly correct timing decision given the structural diagnosis. If your infrastructure roadmap extends two or more years, locking in hardware at current prices may be preferable to waiting, since meaningful price relief is a 2028 story. If your roadmap is shorter, cloud infrastructure absorbs price volatility more gracefully than on-prem procurement. The enterprise hardware budget article sets out a full framework for this decision.
Building a new semiconductor fab takes 3–5 years and costs $10–20 billion. Existing cleanroom lines can be partially converted to produce more HBM, but conversion is slow, reduces conventional DRAM output, and requires extensive validation. Memory manufacturers are investing aggressively in new capacity — Micron’s Idaho fabs, SK Hynix’s Yongin campus — but these facilities will not meaningfully add to global supply until 2027–2028. There is no short-cut to building semiconductor infrastructure at scale.
Yes, in the near term. HBM demand is forecast to grow 70% year-on-year in 2026; new fab capacity is not online yet; and hyperscaler capex continues to accelerate. The crunch is expected to be tightest through 2026 and into early 2027. From mid-2027 the picture begins to improve as new capacity comes online, but demand growth will absorb much of that relief. Broad normalisation — meaning a return to pre-AI-boom price dynamics — is realistically a 2028–2029 outcome.
Reading Cloud Provider Roadmap Risk from the Permit MapThe organising framework here is the Power-Plus-Permission model, coined by Nixon Peabody in their May 2026 data centre site selection update. Cloud capacity is no longer just a power-grid question — community approval is the second gate, and it is failing at scale. For background, see the data center community revolt and the $710B buildout and its 7 GW shortfall.
Community opposition is now a procurement-level variable. Thirty to fifty percent of planned US data centre capacity for 2026 is expected to be delayed or cancelled, creating a 7 GW shortfall that maps directly to the cloud regions most of us depend on. This article gives you the regional risk tier map, a seven-question vendor due diligence checklist, and a repeatable monitoring workflow using datacentertracker.org.
Community opposition is now a supply chain variable. Full stop.
Nixon Peabody’s May 2026 analysis documented 300+ bills in 30+ states, 140+ local opposition groups, and $60B+ in planned data centre investment blocked or delayed. Project cancellations more than quadrupled — from six in 2024 to 25 in 2025.
Two numbers belong in every internal planning conversation right now.
The 7 GW capacity shortfall: of 12–16 GW of AI data centre capacity announced for 2026, only roughly 5 GW is under active construction per Sightline Climate/Bloomberg. That’s 30–50% of the pipeline delayed or cancelled. The 40% project-level schedule risk: SynMax/IIR Energy satellite analysis identified delay signals at 40% of AI data centre construction sites. The 7 GW is a macro pipeline constraint. The 40% is a per-project delivery input. Use both, source both.
The White House AI Action Plan Pledge does not override county zoning boards. When builds slip, new AZ capacity is delayed, reserved instance lead times stretch, and SLA commitments face strain. This is a 2026 planning problem, not a 2030 scenario.
Northern Virginia and North Carolina are the highest-friction zones for AWS us-east-1, Azure East US and East US 2, and GCP us-east4. All three hyperscalers depend on physical infrastructure in these two states.
The Nixon Peabody Political Durability Index classifies the US landscape across five risk tiers. “High friction” means jurisdictions where community opposition, governance structure, or active litigation create more than six-month permitting delays. Virginia and North Carolina both appear in multiple tiers simultaneously.
Here are the five tiers and the cloud regions they affect:
Moratoria and Bans (Highest Risk) — Maryland, Michigan, Minnesota, NH, NY, Virginia†
Regulated Growth — Virginia†, North Carolina†, California, Florida, Illinois, Massachusetts — affecting us-east-1, Azure East US, GCP us-east4
Tax Incentives Under Fire — Virginia†, North Carolina†, Indiana, Georgia† — affecting us-east-1, Azure East US
Pro-Growth / Tightening — Texas, Arizona, Indiana, Nevada, Ohio, Utah — affecting us-east-2 (partial supply chain)
Neutral / Local Control — Arkansas, Kansas, Kentucky, Mississippi, Missouri, Tennessee — lower-friction alternatives
The dagger (†) indicates a state appears in multiple tiers simultaneously. Virginia appears in three.
Virginia: 61 data centre bills filed; 15 enacted; $1.6B tax exemption under scrutiny; 25+ projects cancelled. The defining case is the QTS Digital Gateway — a 37-data-centre campus in Prince William County where a Virginia appellate court voided the rezoning in March 2026. The Virginia Supreme Court writ panel is scheduled for late May or early June 2026. An adverse ruling sets precedent for every hyperscaler building in the state. Full treatment: the Virginia appellate precedent affecting us-east-1 and Azure East US.
North Carolina: Seven or more municipalities passed moratoriums in 2026. Active case: ECO TIP West LLC v. Chatham County — a developer challenging the county’s 12-month moratorium with an $11 million vested rights claim. See North Carolina’s legal escalation pattern.
Texas is Pro-Growth/Tightening, but Dallas metro opposition is accelerating.
Indiana and Georgia (Home Rule): In Dillon’s Rule states, local governments can only act within state-granted powers — limiting moratoriums but not litigation. In Home Rule states, local governments can impose moratoriums independently. Google withdrew its Project Flo application in Indianapolis minutes before the vote after 17 of 25 councillors publicly opposed it.
Maine vetoed an 18-month moratorium, then signed a bill stripping data centres of state tax incentives. State-level legislative risk signals are in ART003.
Lowest-friction alternative: Midwest states rated Pro-Growth or Neutral — Iowa, Nebraska, Kansas — have no active moratorium legislation and far lower opposition intensity.
Site selection maturity tells you whether a hyperscaler has adapted to the Power-Plus-Permission model. It predicts permitting risk in the capacity serving your region. And it’s something you can actually measure.
The primary measurable signal is Community Benefit Agreements (CBAs). Nixon Peabody’s May 2026 guidance identified CBAs as “quickly becoming a must-have for proposed data center builds to survive scrutiny.” A hyperscaler with a strong CBA record negotiates before filing permits. Ask for the execution record, not the policy statement. There’s a big difference.
NDA-heavy site selection is the opposite signal. A study of 31 Virginia municipalities found 25 — that’s 80% — had NDAs with data centre developers. NDAs suppress the public information that manages opposition. Microsoft announced in 2026 it would stop using NDAs with local governments. That sets the benchmark. See NDA-heavy site selection as a permitting risk signal.
Three maturity signals to look for: proactive CBAs with specific dollar amounts and measurable commitments; transparent pre-permit engagement without NDA requirements for local officials; and a clear no-NDA policy covering local governments in high-friction regions. The operational footprint questions on water, noise, and power are in ART004.
No existing source provides a vendor due diligence checklist in this format for this risk. These seven questions are synthesised from the permit map, the Political Durability Index, and active opposition case data. Use them in your next hyperscaler vendor review or capacity planning conversation.
1. What percentage of your planned 2026–2027 capacity in us-east-1 / Azure East US is currently in active permitting proceedings?
A satisfactory answer names specific projects with percentages. The 40% schedule risk baseline from SynMax/IIR Energy is your calibration point. “All our projects are on track” — set against 25+ Virginia cancellations and an active Supreme Court appeal — is not a credible answer.
2. Do you have active community opposition cases in Virginia or North Carolina that could affect capacity delivery timelines?
A satisfactory answer acknowledges the QTS Digital Gateway Virginia Supreme Court proceedings and the Chatham County moratorium litigation. Denial of active opposition in these two states is implausible.
3. What Community Benefit Agreements have you executed in the last 12 months for facilities serving this region?
A satisfactory answer gives you specific CBAs with municipalities, dollar amounts, and measurable commitments. “We engage with communities” is not sufficient. Absence of CBA data for Virginia and North Carolina is itself a signal.
4. What is your policy on NDAs with local officials during site selection and permitting?
A satisfactory answer confirms that NDAs are not used with local elected officials in ways that prevent disclosure of project scope and community impact. Microsoft’s 2026 policy change sets the benchmark. Any NDA policy covering local officials in high-friction regions predicts the reactive opposition that produces post-approval legal challenges.
5. How does your site selection process account for the Political Durability Index evaluation at candidate sites?
A satisfactory answer shows familiarity with the Nixon Peabody Political Durability Index — or an equivalent risk assessment covering legislative durability, opposition group activity, and governance structure. A hyperscaler using the index flags high-friction counties at the siting stage rather than encountering opposition after capital is committed. No mention of political durability suggests an outdated process.
6. What early warning indicators would you disclose if a major facility serving this region faced a permit challenge?
A satisfactory answer includes a contractual notification threshold — something like “we disclose any permit appeal within 30 days of filing, with a schedule impact assessment.” “We would keep you informed” is not sufficient. The QTS case was before the Virginia Supreme Court before most cloud tenants even knew it existed.
7. What is your delay mitigation strategy if the QTS Digital Gateway Virginia Supreme Court ruling sets adverse precedent?
A satisfactory answer includes specific contingency regions, build pipeline diversification outside Virginia, or alternative AZ capacity plans already in progress. No contingency plan leaves you exposed if you have primary dependence on us-east-1 or Azure East US.
The vendor checklist covers the point-in-time review. The next section gives you the monitoring workflow to track how the landscape shifts between reviews.
datacentertracker.org is the primary tool for monitoring active data centre opposition cases across US geographies. Here’s a repeatable five-step workflow.
Use the county search filter to pull up Prince William County, Virginia, and Chatham County, North Carolina. These are the counties where active legal proceedings most directly affect us-east-1 and Azure East US capacity. Any new filing, appeal, or ruling in these two counties is a procurement-relevant event.
Filter by project status: “opposed,” “appealed,” or “stalled.” Active opposition cases in legal proceedings are the early warning signals.
Monitor the QTS Digital Gateway proceedings as your top priority. The Virginia Supreme Court writ panel is scheduled for late May or early June 2026. An adverse ruling sets precedent for all hyperscaler builds in Virginia — this is the Tier-1 escalation event in the framework below.
Expand to secondary friction zones. Add Dallas County, Texas, and Indiana counties where the Google Project Flo precedent applies.
Set escalation triggers. A new Prince William County appeal filing or Chatham County legal ruling are Tier-2 signals. A Virginia Supreme Court ruling on QTS Digital Gateway is Tier-1 — immediate reassessment of multi-year us-east-1 and Azure East US commitments.
Cadence: Monthly for most geographies. Weekly for Virginia and North Carolina during active QTS litigation. Immediate when a ruling is expected. The tracker bridges the legal dynamics in the data center community revolt and your procurement calendar.
Permitting risk is a plannable variable. The inputs are available and sourced. The task is translating them into scenario assumptions your capacity planning model can actually use.
Delay probability framework by regional friction tier:
Virginia (us-east-1, Azure East US, GCP us-east4) — High friction — 40%+ delay probability baseline. Primary risk mechanism: QTS appellate precedent; 140+ opposition groups; $1.6B tax exemption at risk.
North Carolina (us-east-1 supply chain) — High friction — 40%+ delay probability baseline. Primary risk mechanism: 7+ active moratoriums; active litigation; governor’s tax review.
Texas (Dallas metro) — Growing friction — 15–25% delay probability. Primary risk mechanism: historically Pro-Growth; accelerating opposition.
Indiana, Georgia — Moderate friction — 15–25% delay probability. Primary risk mechanism: Home Rule moratorium authority; Project Flo precedent.
Midwest expansion zones — Lower friction — 10–15% delay probability. Primary risk mechanism: general construction and interconnection risk only.
Use 40% as a scenario input, not a point forecast. The value is in forcing you to model both an on-schedule scenario and a 6–12 month slip scenario before you commit capacity in a high-friction region.
Multi-region architecture is the primary mitigation lever. No enterprise contract required. If us-east-1 or Azure East US is your single-region primary, model a Midwest failover option. Neoclouds like CoreWeave and Lambda Labs fill gaps when reserved instance lead times extend — build 3–6 months additional buffer into reservation timelines for high-friction regions.
For the full treatment of the supply-side figures, see the $710B buildout and its 7 GW shortfall.
Community opposition is the primary risk vector. Two federal developments create independent, additive risk on top of it.
FERC RM26-4-000: A final rule is expected by end of June 2026, setting uniform rules for large electrical loads of 20 MW or more. The key point: a build that clears local permitting can still face FERC-driven delays in grid connection. Ask your hyperscaler whether FERC RM26-4-000 affects any facility in your primary region.
The EIA Mandatory Energy Survey: A mandatory nationwide survey follows the EIA’s voluntary pilot by September 30, 2026, capturing grid-supplied electricity, cooling efficiency, and IT specifications — making previously proprietary operational metrics public record. Before that deadline, requesting this data in vendor reviews may surface build pipeline information hyperscalers have not previously been required to share.
Together, community opposition, FERC interconnection delays, and EIA compliance create a multi-vector risk stack where cumulative delay probability exceeds any single factor. Worth understanding before your next capacity commitment.
Not every opposition development is a procurement event. Here’s a three-tier framework that separates signals requiring immediate action from those that can wait for the next planning cycle.
Tier 1 — Immediate reassessment:
Tier 2 — Review and update risk models:
Tier 3 — Watch, reassess at next planning cycle:
Tier-1 events will get mainstream coverage — but by then you have less time to act. Upstream detail on the Virginia appellate precedent affecting us-east-1 and Azure East US and North Carolina’s legal escalation pattern.
Yes. AWS us-east-1 depends on Northern Virginia — the highest-friction geography in the US. Virginia appears in three Political Durability Index tiers simultaneously; 140+ opposition groups active; at least 25 projects cancelled. An adverse QTS Digital Gateway ruling sets precedent for every hyperscaler building there.
Midwest regions — Azure Central US, GCP us-central1, AWS us-east-2 secondary supply. Nixon Peabody’s May 2026 update identifies the Midwest as the current lowest-risk siting destination. States rated Pro-Growth or Neutral (Iowa, Nebraska, Kansas) have no active moratorium legislation.
Nixon Peabody’s 50-state framework classifying states across five tiers: Moratoria and Bans, Regulated Growth, Tax Incentives Under Fire, Pro-Growth/Tightening, and Neutral/Local Control. Virginia appears in three tiers simultaneously. It is the primary tool for mapping physical data centre geography to named cloud region risk.
A Community Benefit Agreement (CBA) is a legally binding commitment to local governments covering jobs, tax revenue, infrastructure, and environmental mitigation. CBAs are increasingly a prerequisite for permit approval in 2026. A hyperscaler with a strong CBA track record faces lower opposition risk and shorter permitting timelines.
Yes. Opposition groups can challenge approved permits through appeals, litigation, and ballot measures. The QTS Digital Gateway project in Prince William County is currently before the Virginia Supreme Court after being approved and then appealed — approved projects remain at risk until construction is complete.
Neoclouds are smaller GPU-as-a-service providers — CoreWeave and Lambda Labs are the main examples — that fill capacity gaps when hyperscaler builds slip. Consider them when reserved instance lead times extend unexpectedly in high-friction regions, or when you need GPU capacity faster than the hyperscalers can deliver.
FERC RM26-4-000 adds federal utility interconnection requirements to data centre builds. It operates independently of community opposition — a build that clears local permitting can still face FERC-driven delays in grid connection. That is a second independent risk vector compounding community-driven delays in your capacity timeline modelling.
From September 30, 2026, hyperscalers must disclose operational energy data for the first time. Request it now in vendor reviews — before it is mandatory — and you may surface build pipeline information they have not previously been required to share.
In Dillon’s Rule states (Virginia, Texas), local governments can only act within powers granted by state law — limiting moratoriums, but not litigation. In Home Rule states (Indiana, Georgia), local governments can impose moratoriums independently. Google Project Flo in Indiana is the clearest example of what that looks like in practice.
Yes. The risk flows through cloud region availability, not contract scale. If you are on us-east-1 or Azure East US, you face the same capacity constraints when hyperscaler builds slip — you just have fewer levers. Multi-region architecture requires no enterprise contract to implement.
Monthly monitoring of datacentertracker.org covers most geographies. Virginia and North Carolina warrant weekly attention during the active QTS proceedings. Review immediately when a ruling is expected, when a hyperscaler reports a build pause, or when reserved instance lead times extend without explanation.
Not necessarily. Co-location faces the same community opposition dynamics. Hyperscalers have more resources for CBA negotiations, but a larger footprint also creates larger opposition targets. Co-location in a high-friction county does not reduce permitting risk; co-location in a lower-friction region does.
Reading order: This is the seventh article in the cluster, designed as the decision-stage synthesis. For the full context on the revolt’s origins and scale, start with the data center community revolt.
NDA Backlash — The Secrecy Strategy That BackfiredThere’s something bigger than power lines and water consumption driving the broader revolt against hyperscale data centre development. The nondisclosure agreement — a tool the industry treated as standard practice — turned passive neighbourhood scepticism into organised litigation, bipartisan legislation, and voided rezonings across multiple jurisdictions in 2025 and 2026. If you’re evaluating cloud capacity commitments or co-location contracts, this matters. The permitting landscape just fundamentally changed.
Data centre NDAs are legally binding confidentiality agreements. Local officials, economic development administrators, and sometimes landowners are required to keep project identity, scale, power consumption, and tax arrangements under wraps during a project’s pre-announcement phase.
The industry rationale was straightforward: protect competitive intelligence, prevent organised community opposition before entitlements are secured, and preserve negotiating leverage with utilities and local governments. Corporate leaders called it “a normal and necessary part of the economic development process.” That was the mainstream position — before the backlash.
NDAs also gave projects code names. Project Loon (Google, Hermantown MN), Project Delta (Stokes County NC), Project Corn Maze (Wisconsin) — opaque designations that let developers advance through rezonings before anyone outside the room knew what was happening.
What gets concealed goes well beyond trade secrets. Developer identity, project footprint, power estimates, and tax subsidy packages are all typically covered. A survey of 31 Virginia municipalities found 25 — 80% — had NDAs in place. And the strategy consistently produced outcomes that were the exact opposite of what it was designed to achieve.
Here’s how it plays out. Officials sign NDAs before any project details are disclosed. Communities discover a project exists only when rezonings or permits are filed — often after the key decisions are already made.
The objections come down to three things: no informed consent during entitlement votes, no ability to evaluate environmental and infrastructure concerns communities couldn’t evaluate, and a power dynamic where the developer’s interests are protected and the community’s are not.
That last point is the core grievance. When NDAs cover power draw, water consumption, and traffic impact, communities can’t assess how a facility will affect their local grid or aquifer until approval is near-certain. Aubree Derksen of Pine Island, Minnesota testified that city officials had known about a proposed data centre for two years before residents were informed. “The democratic process where my voice is supposed to matter has been hijacked by big tech.”
Port Washington, WI Mayor Ted Neitzke rejected an NDA and fared measurably better. The secrecy around financial terms — tax arrangements negotiated without public input — often proved more inflammatory than the secrecy about the project itself.
In January 2026, Stokes County, NC commissioners voted 3-2 to rezone approximately 1,845 acres of residential-agricultural Dan River land for a hyperscale data centre — “Project Delta” — without publicly naming the developer or disclosing key project details. That vote overruled the county Planning Board’s recommendation to deny.
A six-group lawsuit filed by the Southern Environmental Law Center and Southern Coalition for Social Justice voided the rezoning. Stokes County is now the most documented case of NDA-fuelled secrecy escalating into organised legal action. Stokes County’s opposition was fuelled partly by NDA secrecy.
The community response was particularly organised because the rezoned property includes burial grounds of Native American and enslaved people. Robert Hairston of the National Hairston Clan described “real people, in the ground.” SELC and SCSJ filed suit on behalf of six plaintiff groups. The Board of Commissioners ultimately reversed the zoning decision.
Developer CEO Drew Nations confirmed plans to reapply. Project delay, not cancellation, is the typical outcome — but the cost in legal fees, timeline slippage, and community trust is precisely what the NDA strategy was supposed to prevent.
Yes. Google’s proposed 1.8 million sq ft, $650 million data centre in Hermantown, Minnesota — developed under the code name “Project Loon” by Mortenson Development — required NDAs from the city administrator, his assistant, 22 St Louis County employees, and three county commissioners before any public announcement.
The practice is not confined to smaller developers. Google, Meta, and Microsoft are all documented as hyperscalers that used NDA-first site selection strategies. This was industry-wide.
MCEA (Minnesota Centre for Environmental Advocacy) obtained public records revealing the Project Loon details — including an Xcel Energy representative’s informal disclosure of Google’s identity to a senator the day before the formal announcement. The NDA-shrouded deal leaked anyway. The Stop the Hermantown Data Center Facebook group reached 5,000 members. MCEA filed suit. The AUAR environmental process was restarted, sidelining development for seven months.
Beaver Dam, WI signed an NDA related to a Meta data centre. Microsoft signed NDAs with Kenosha and four other Wisconsin communities. The breadth of the practice across hyperscalers made the backlash inevitable.
Wisconsin Senate Bill 969 would prohibit local governments from signing NDAs with data centre developers. The Senate utilities committee passed it 4-1 on a bipartisan vote — with Republican support — which is a significant signal. NDA prohibition has cross-party backing on community right-to-know grounds.
Minnesota’s House advanced a similar NDA prohibition bill with unanimous bipartisan support. Florida advanced a bill but removed it after industry lobbying — which illustrates how contested this reform is across different state legislative environments.
Wisconsin produced four competing data centre bills in a single session. Husch Blackwell‘s analysis calls it the model for what other states will face, and is explicit that NDA ban legislation “will almost certainly be reintroduced.” Pima County, Arizona acted ahead of state legislatures entirely. At least 10 states have now seen lawmakers propose banning or limiting data centre NDAs.
Microsoft’s March 2026 announcement to stop using NDAs came as Wisconsin legislative pressure mounted. Corporate VP Rima Alaily framed it as a trust question. Bill Lueders of the Wisconsin Freedom of Information Council offered a more empirical verdict: data centre NDA use “did blow up in their faces.”
Nixon Peabody‘s “power-plus-permission” model is the clearest articulation of where the industry is heading. It replaces energy-first site selection by requiring power access, regulatory readiness, and earned community acceptance as co-equal prerequisites for project viability.
💡 Social licence to operate: Nixon Peabody’s term for earned community acceptance — achieved through transparent operations and proactive stakeholder engagement — now required alongside power access and regulatory approvals.
💡 Power-plus-permission model: Nixon Peabody’s coinage for the post-2025 paradigm: power access must be combined with regulatory readiness and earned community acceptance.
In practice, transparent site selection means proactive stakeholder engagement before the project is announced, legislative risk scanning as a due diligence step, and Community Benefit Agreements as the formal mechanism for making public commitments legally binding.
That last part matters. Microsoft’s announcement had no legal mechanism attached — it can be reversed without consequence. A CBA cannot. If you’re evaluating a developer’s transparency commitments, that distinction is the one to focus on.
Here’s the practical concern for technology decision-makers. Hyperscalers that relied on NDA-heavy site selection may have accumulated undisclosed permitting risk in regions where they claim capacity availability — capacity that could be delayed or stranded by community opposition, legislative action, or voided rezonings. This is part of the data center community revolt reshaping how the industry operates at every level — from site selection to infrastructure procurement.
NDA-heavy site selection history is a proxy for opacity-driven permitting risk. That risk does not appear in vendor capacity availability claims.
Twenty-five data centres were scrapped in 2025 — four times the 2024 figure. Google’s Hermantown project remains in limbo while the second AUAR is compiled. That capacity is not available in any meaningful sense, regardless of what vendor materials say.
So when you’re evaluating cloud or co-location providers, here are the questions worth asking:
Microsoft has made a public commitment, however informal. Google and Meta have not. That’s a proxy — imperfect, but available — for the relative permitting risk embedded in their regional capacity claims. Virginia projects where NDA practices contributed to opposition escalation illustrate how this risk materialises at scale. For the full vendor evaluation framework, see the vendor transparency as a procurement evaluation criterion guide.
The NDA strategy was calibrated for an environment that no longer exists. In 2026, the vendors best positioned to deliver on capacity commitments are the ones whose site selection practices can withstand public scrutiny.
A nondisclosure agreement requiring local officials, administrators, or landowners to keep project details — developer identity, scale, power consumption, tax arrangements — confidential during the pre-announcement phase. Typically persists until entitlements are nearly secured.
The rationale was threefold: protect competitive intelligence, prevent organised community opposition before entitlements were secured, and preserve negotiating leverage with utilities and local governments. It was characterised as standard economic development practice before the backlash.
Nixon Peabody’s term for earned community acceptance through transparent operations and proactive stakeholder engagement — now required alongside power access and regulatory approvals. It’s the alternative to NDA-first site selection.
A legally binding agreement specifying public commitments — jobs, tax contributions, environmental standards — as a condition of project approval. Brookings Institution identifies CBAs as the replacement for NDA-first approaches. The “legally binding” characteristic is what distinguishes them from informal announcements with no enforcement mechanism.
Nixon Peabody’s term for the post-2025 paradigm: power access must be combined with regulatory readiness and earned community acceptance to be viable. It replaces the energy-first model where power access alone was sufficient.
Wisconsin (SB-969 passed committee 4-1 with bipartisan support), Minnesota (House NDA prohibition bill advanced unanimously), and Florida (advanced then removed after industry lobbying) are the primary jurisdictions. Pima County, AZ was the first to enact limits. At least 10 states have seen lawmakers propose banning or limiting data centre NDAs.
Yes. The Senate utilities committee voted 4-1 with Republican sponsorship from Sen. André Jacque. Community right-to-know concerns transcend partisan divides.
Microsoft announced in March 2026 it would cease using NDAs with local governments. Google and Meta have not made equivalent public commitments.
The January 2026 rezoning of 1,845 acres for Project Delta was voided following a lawsuit filed by SELC and SCSJ. The developer intends to reapply — project delay, not cancellation, is the typical outcome.
Project Loon was the code name for a proposed Google data centre in Hermantown, Minnesota. NDAs were signed with the city administrator, 22 county employees, and three county commissioners. Public records obtained by MCEA triggered a 5,000-member opposition group, a lawsuit, and a seven-month environmental review restart.
The evidence favours transparency. Menomonie, WI passed an ordinance blocking a proposed $1.6 billion data centre. Stokes County’s rezoning was voided. Hermantown’s AUAR was restarted for seven months. Port Washington, WI — which rejected an NDA — fared measurably better. Nixon Peabody frames early community engagement as a risk reduction strategy, not merely an ethical preference.
Generally no. Microsoft’s March 2026 announcement had no legal mechanism attached. Brookings Institution notes that Community Benefit Agreements should be “legally binding” — distinguishing them from voluntary commitments that can be reversed without consequence. Treat CBAs as a substantially stronger signal than announcements alone.
Virginia Cancellation Wave — Data Center Alley Under PressureVirginia’s Data Center Alley runs more than 600 operational facilities and holds the title of the world’s most data-centre-dense market. Even so, it is not immune to the data center community revolt that is reshaping US cloud infrastructure.
On 31 March 2026, the Virginia Court of Appeals voided the rezoning approvals for the QTS Digital Gateway project — a proposed 2,100-acre campus in Prince William County that would have been the largest data centre complex ever built. The court’s reason was procedural: the county failed to legally notify the public before its December 2023 rezoning hearing. The project is now before the Virginia Supreme Court, with QTS proceeding alone after its co-developer withdrew and Prince William County dropped its own defence.
Meanwhile, Loudoun County eliminated by-right data centre approval in March 2025, replacing it with a Special Exception process that requires public hearings and a board vote on every new application. That zoning change is baked into the ordinance. No Supreme Court ruling touches it.
So you have two parallel tracks — a court precedent and a durable zoning change — running at the same time. That makes Virginia the highest-stakes test case for cloud infrastructure supply in the US. This article breaks down both tracks, what they mean for AWS us-east-1 and Azure East US capacity planning, and why the legal template created here can be copied by opposition groups in any other US jurisdiction.
Virginia’s status as the world’s largest data centre market means permitting friction here affects cloud capacity at a national scale.
Data Center Alley — concentrated in Loudoun County and extending into Prince William, Fairfax, and Fauquier counties — hosts more than 600 operational facilities. AWS us-east-1, the default region for the majority of US workloads, is anchored here. So is Azure East US. Sustained permitting friction here ripples nationally.
Here is the interesting wrinkle. Loudoun County derives nearly 50% of its property tax revenue from data centres. The local government most financially dependent on the industry has also moved furthest to constrain new approvals. That tells you something about how much pressure residents have been applying.
Sixty-one data centre-related bills were introduced in the 2026 Virginia General Assembly, with 15 enacted — the highest state-level legislative churn in any US jurisdiction. Nearly half of 2026 pipeline projects face delays nationally. Virginia is the leading indicator.
And here is something worth noting: the opposition in Virginia is explicitly bipartisan. The Coalition to Protect Prince William County spans Democratic and Republican elected officials. The Board of Supervisors withdrew from the Digital Gateway litigation unanimously, under a Democrat-led board. This is not a political trend that reverses with an election cycle.
💡 Hyperscaler: A company — such as AWS, Google, or Microsoft — that owns and operates data centre campuses at massive scale, typically hundreds of megawatts to multiple gigawatts per campus, to power its cloud services.
The PW Digital Gateway was proposed as a 37-building data centre campus on approximately 2,100 acres near Gainesville in Prince William County. Full buildout would have meant over 22 million square feet and a projected 3 to 6 gigawatts of power demand — roughly equivalent to all the data centre power currently consumed across Northern Virginia. The largest data centre campus in the world.
QTS Realty Trust was the primary developer. Compass Datacenters was the co-developer.
Prince William County approved the rezonings in December 2023 after a 27-hour public hearing. The Oak Valley Homeowners Association and the American Battlefield Trust filed separate legal challenges — both arguing that the county had not properly notified the public, making the entire approval process legally defective.
A Circuit Court judge agreed in August 2025. The Virginia Court of Appeals issued a unanimous ruling on 31 March 2026 declaring all three rezoning decisions void from the outset.
The site sits adjacent to Manassas National Battlefield Park, which draws 700,000+ visitors annually. The Occoquan watershed — drinking water for approximately 800,000 Northern Virginians — provided a second environmental argument.
Prince William County had spent $1.72 million defending the rezonings through two court levels. Then in April 2026, the Board voted unanimously to walk away. Approved it, spent $1.7 million defending it, then reversed course. QTS filed a Supreme Court petition on 30 April. Two days earlier, Compass had ended its appeal.
“Procedural deficiency” sounds like dry legal language. In practice it means this: the county failed to legally notify residents about the December 2023 rezoning hearing in the manner required by Virginia zoning law.
Virginia law sets specific public notice requirements before rezoning hearings. The county did not follow them correctly. The Court of Appeals found that failure sufficient to void not just the hearing, but all three rezoning decisions that came out of it — declared void from the outset, as though the hearings had never legally occurred.
The court did not rule on whether the project was a good or bad idea. It ruled the process was legally defective. That is exactly what makes this ruling significant for anyone watching from outside Virginia.
💡 By-right zoning: Automatic approval of a land use application at staff level, without requiring a public hearing or board vote, provided the application meets the applicable zoning specifications.
The Oak Valley Homeowners Association, under Mac Haddow, drove the procedural challenge through both courts. The Piedmont Environmental Council filed an amicus brief at the appellate level — showing how environmental advocacy organisations can amplify HOA-driven challenges through formal legal participation.
AFS Law puts it plainly: “even minor procedural deficiencies have been held sufficient to void project approvals.” Public notice compliance is now a litigation risk priority, not just an administrative formality.
Compass withdrew on 28 April 2026 — two days before QTS filed its Supreme Court petition — making an independent decision not to pursue the appeal. Prince William County had already reached the same conclusion after two court levels. QTS is proceeding alone.
Capstone DC frames the broader dynamic bluntly: even well-capitalised developers face a calculation where “projects will either eventually circumvent these challenges by relenting to public demands or relocating to a more amenable area.” Compass reached that conclusion. The county reached that conclusion.
The result: QTS faces the writ panel without a co-developer and without the county that originally approved and defended the rezonings. No co-petitioner, no government defender — a Virginia Supreme Court reversal is less probable, and the Court of Appeals ruling more likely to stand as precedent.
QTS’s petition asks the Virginia Supreme Court to exercise discretionary review. The court is not obligated to hear the case. It chooses.
A three-justice writ panel will hear a 20-minute session from QTS in late May or early June 2026.
💡 Writ panel: In Virginia, a panel of three Supreme Court justices that reviews a petition for appeal and decides whether the full court will hear the case. Only the petitioner argues before the panel; the opposing parties do not appear.
Two outcomes. If the panel declines, the Court of Appeals ruling becomes final confirmed precedent. If it accepts, a full hearing follows — potentially in late 2026 or 2027.
Mac Haddow of the Oak Valley HOA confirmed the uncertainty: the decision is “entirely discretionary” and “there is no set timeline.” Infrastructure planning that depends on this outcome cannot assume a fixed date.
WTOP (wtop.com) and InsideNoVa.com are tracking the writ panel schedule in real time.
In March 2025, the Loudoun County Board of Supervisors eliminated by-right approval for new data centre applications. Every new application now requires a Special Exception process: a public hearing and an official board vote.
This is a structural policy change, not a court ruling. It is embedded in the zoning ordinance. It applies regardless of how the QTS appeal resolves. Even if QTS wins at the Virginia Supreme Court, Loudoun’s by-right elimination stays in place.
💡 Dillon’s Rule: A legal doctrine under which local governments can only exercise powers explicitly granted to them by state law. Virginia is a Dillon’s Rule state — counties cannot impose blanket moratoriums, but they can rewrite their zoning ordinances within state-authorised land use powers.
Loudoun acknowledged it “does not have the legal authority to implement a moratorium” under Dillon’s Rule. So it rewrote its zoning ordinance instead — same practical effect, different mechanism. After the change, the Board approved a subsequent application only after extracting concessions from the developer to reduce the project’s square footage.
Adjacent counties are watching. Fairfax, Prince William, and Fauquier are all considering similar changes. Stafford County has already enacted enhanced setback and buffer requirements: a 750-foot residential setback and a 55-decibel noise cap. AFS Law documents at least 25 project cancellations across Virginia from this combined pressure.
Two parallel tracks — court rulings that can be appealed and zoning ordinance changes that cannot — mean the permitting environment is tightening regardless of how the QTS case resolves.
The QTS Court of Appeals ruling is exportable. The procedural deficiency argument works in Texas, Georgia, Indiana, or any other state with a public notice requirement in its rezoning process. AFS Law: “even minor procedural deficiencies have been held sufficient to void project approvals.” Virginia is one chapter in a broader $710B buildout collision playing out across the country.
The American Battlefield Trust’s involvement adds a second tool. Any project near a national battlefield, historic landmark, or federal heritage site now faces potential litigation from preservation organisations with standing and resources. The Wilderness Battlefield in Orange County, Virginia has already been named one of the most endangered historic places in the US due to data centre proximity.
AWS us-east-1 and Azure East US are both anchored in Northern Virginia. The 25+ cancellations reduce the pipeline of future capacity in the region serving the largest share of US cloud workloads. As direct implications for AWS us-east-1 and Azure East US capacity tighten, lead times will extend.
Hyperscalers build at scales that attract political attention. Smaller co-location operators such as Equinix and Iron Mountain typically operate below the threshold of organised opposition — lower permitting risk.
Virginia is the most legally complex data centre permitting environment in the US — not despite being the largest market, but partly because of it. Maine is tracking in the same direction, with a bill blocking builds over 20 megawatts until late 2027 on Governor Mills’ desk.
The Coalition to Protect Prince William County put it plainly: “Never before has such a small group of people made such a difference to their community and to their state.” That is the playbook that opposition groups in every other US jurisdiction will study.
The corridor of data centre facilities concentrated in Loudoun County, Virginia, extending into Prince William, Fairfax, and Fauquier counties. More than 600 operational facilities — the world’s largest data centre market by capacity.
Prince William County failed to comply with required public notification procedures before its December 2023 rezoning hearing. That procedural deficiency voided all three rezoning approvals. The court did not rule on the merits of the project itself.
The county defended the rezonings through two court levels, then reversed course after losing at the appellate level — a political recalculation about the cost and probability of success before the Virginia Supreme Court.
A 37-building data centre campus on approximately 2,100 acres near Gainesville, Prince William County. At full buildout: over 22 million square feet and 3 to 6 gigawatts of power demand — the largest data centre campus in the world, if it had been built.
Dillon’s Rule limits Virginia counties to powers explicitly granted by state law — it prevents blanket moratoriums. But it does not prevent counties from rewriting their zoning ordinances, which is exactly what Loudoun County did in March 2025.
No. Review is entirely discretionary. A three-justice writ panel will hold a 20-minute hearing in late May or early June 2026. There is no obligation to accept the petition and no set timeline for the decision.
The Digital Gateway site is adjacent to Manassas National Battlefield Park, a federally designated heritage site drawing 700,000+ visitors annually. Any proposed site near a national battlefield or federal heritage site now faces this same litigation vector.
Under by-right zoning, a data centre meeting specifications was approved at staff level — no public hearing, no board vote. Under Loudoun’s Special Exception process, every application must go through a public hearing and receive an official Board of Supervisors vote.
Yes. AWS us-east-1 and Azure East US are both anchored in Northern Virginia. The 25+ project cancellations and new permitting friction reduce the pipeline of future capacity in the region serving the largest share of US cloud workloads.
Indiana and Georgia operate under Home Rule, giving municipalities broad authority to enact blanket moratoriums — at least eight Georgia jurisdictions and four Indiana counties have banned new data centre construction outright. Virginia’s Dillon’s Rule prevents that direct path. But Loudoun achieved the same practical outcome through its zoning ordinance anyway.
Compass withdrew before QTS filed the Supreme Court petition. QTS now proceeds without a co-petitioner, without the county government that approved the rezonings, and with reduced legal resources — weakening the position before the writ panel on every dimension.
WTOP (wtop.com) and InsideNoVa.com are the primary sources tracking the writ panel schedule and any subsequent developments. WTOP’s 8 May 2026 report is the most current primary reference at time of publication.
Water, Noise, Power — The Real Community Costs of Data CentersA large hyperscale data centre draws up to five million gallons of water per day — the same daily consumption as a town of 50,000 people. Its cooling systems run non-stop and emit up to 70 dB at 400 feet. When it hooks up to the regional grid, the transmission upgrades it needs get spread across every household in the same zone — whether those households see any benefit or not.
Water draw volumes, decibel readings, and grid load calculations. Those three things are the quantified grievances behind Nixon Peabody‘s documented count of 140+ community opposition groups and $60 billion in blocked investment since early 2025. This article is part of the data center community revolt reshaping infrastructure planning, and it documents those impacts with source-verified numbers.
Water depletion, continuous noise, and rising electricity bills are what turn passive concern into organised legal action. They’re measurable, they’re continuous, and they affect residents who get nothing in return. They also generate bipartisan opposition — water depletion alarms rural conservatives, noise disrupts suburban homeowners, and ratepayer cost-shifting has triggered Republican- and Democrat-sponsored legislation in the same state sessions.
Between April and June 2025 alone, Data Center Watch counted 20 proposals worth $98 billion blocked or delayed by local opposition — two-thirds of all projects tracked in that period. That marks the end of the “energy-first” era across the data center community revolt. The Power-Plus-Permission model (Nixon Peabody, 2026) replaces it: power access is necessary, but it’s no longer enough.
💡 Social licence to operate is the combination of community trust, transparent operations, and legislative durability that a facility must maintain alongside its formal permits — without it, approved projects face injunctions, moratoriums, and ballot measures even after construction begins.
A typical mid-sized data centre uses around 300,000 gallons of water per day — roughly equivalent to 1,000 households. A large hyperscale facility can draw up to five million gallons per day. That gap comes down to cooling technology choice more than anything else.
In Northern Virginia alone, data centre water consumption hit close to two billion gallons in 2023 — a 63% increase from 2019 — with Loudoun County accounting for around 900 million of those gallons. Approximately 80% of the water withdrawn evaporates through cooling towers rather than being recycled. The standard efficiency metric is Water Usage Effectiveness (WUE) — litres of water consumed per kilowatt-hour of IT energy. The industry average sits at 1.9 L/kWh.
Two-thirds of all US data centres under development since 2022 are in water-stressed areas (World Resources Institute). In western states under prior appropriation water law — Colorado, Nevada, and Arizona — the newest users face curtailment during drought. Florida has moved most directly: demonstrated water access is now a gating prerequisite for new projects there.
The cooling system a facility uses is the single most important variable in determining its water footprint. It should be the first question you ask any vendor.
Evaporative (open-loop) cooling is the most common type. Around 80% of withdrawn water evaporates, and cooling towers emit up to 70 dBA within 400 feet — the highest-water and highest-noise option of the lot. Closed-loop cooling recirculates water internally, cutting freshwater use by up to 70%. Immersion cooling submerges servers in synthetic dielectric fluid, reducing water consumption by 90% or more and eliminating cooling towers entirely. Air cooling uses near-zero water but only works in cooler, drier climates.
The Lancaster, Pennsylvania CBA caps water use at 20,000 gallons per day. The El Paso, Texas Meta CBA specifies closed-loop cooling starting at 750,000 gallons per day. When you’re evaluating co-location providers, ask for the WUE figure upfront — it’s the most direct proxy for cooling technology choice.
Cooling systems generate up to 70 dBA within 400 feet, continuously, every day of the year. Backup diesel generators hit up to 105 dB during testing or emergencies. Residents in Prince William County, Virginia routinely report noise exceeding 60 dB.
Three sources explain that range:
1. HVAC and cooling systems produce a continuous 24/7 hum — consistently the primary community complaint, not because it’s the loudest source, but because it never stops.
2. Backup diesel generators are tested monthly and can run without a federally enforced time limit during emergencies. Larger industrial units approach 105 dB.
3. Behind-the-meter (BTM) gas turbines are a growing category that deserves its own attention.
💡 Behind-the-meter (BTM) generation means a data centre generates its own power on-site — through diesel generators, gas turbines, or batteries — independently of the utility grid. Unlike backup generators, BTM turbines run 24 hours a day.
Forty-six planned or permitted US data centres will use off-grid gas turbines. xAI‘s Memphis Colossus operates 30+ natural gas turbines daily; local residents and the NAACP filed a Clean Air Act notice of intent to sue. Virginia JLARC’s 2024 report found nearly one-third of Virginia’s data centres sit within 200 feet of residentially zoned properties. And because the EPA defunded its Office of Noise Abatement and Control in 1981, federal noise standards for data centres simply don’t exist.
Yes — in regions where large loads connect without bearing the full cost of the infrastructure upgrades they require. In 2025, Americans paid approximately $60 billion more in electricity costs nationwide compared to 2024 (WRI), with data centre load growth a documented contributor in the Mid-Atlantic region.
When a facility drawing 20 megawatts or more connects to the grid, transmission lines, substations, and capacity reserves need upgrading. By default, those costs get distributed across all existing ratepayers.
💡 PJM Interconnection is the regional transmission organisation that coordinates wholesale electricity across 13 Mid-Atlantic and Midwest states — it determines whether a new large load pays for its own infrastructure upgrades or distributes those costs across existing customers.
US summer peak demand is expected to rise 24% in the next 10 years, with data centres accounting for the majority of that increase (NERC Level 3 Alert, May 4, 2026). By 2030, data centre demand could hit 90 GW — nine times New York City’s peak summer demand.
AEP Ohio now requires data centres to pay for at least 85% of subscribed energy capacity regardless of actual use. FERC‘s docket RM26-4-000 is expected to finalise uniform rules for large electrical loads (≥20 MW) by end of June 2026. The White House Ratepayer Protection Pledge (March 4, 2026) commits Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI to build or buy their own generation and pay for delivery upgrades — voluntary today, but the same language is now showing up in 18+ state bills.
Social licence to operate is Nixon Peabody’s term for the community approval now required alongside grid access as a practical prerequisite for permitting. It’s not a legal document — it’s a combination of five commitments:
The consequences of getting this wrong: 300+ bills in 30+ states; 12+ moratorium proposals; 140+ local groups blocking approximately $60 billion. In Menomonie, Wisconsin, community opposition led the city council to block a $1.6 billion project outright. As Nixon Peabody puts it — “Don’t mistake silence for safety.”
A Community Benefit Agreement (CBA) is a formal, legally binding contract between a data centre developer and a host community. Unlike NDAs, CBAs must be publicly available. Nixon Peabody describes them as “permitting prerequisites” in jurisdictions with active opposition — not optional goodwill gestures.
Standard CBA coverage includes job creation targets with wage floors, water usage caps, noise limits, clean energy commitments, a public dashboard, and exit fees. Proposed fines run up to $50,000 per day for wilful violations.
Four verified examples: Lancaster, Pennsylvania (Chirisa / CoreWeave) caps water at 20,000 gallons per day with $20 million in community commitments. Cedar Rapids, Iowa (Google + QTS) delivers a 70% tax exemption tied to job thresholds. West Des Moines, Iowa (Microsoft) commits to 100% renewable energy and $2 billion-plus in projected tax revenues. El Paso, Texas (Meta) specifies closed-loop cooling and 80% property tax abatement over 35 years.
Some CBAs are, as Brookings puts it, “heavily redacted such that key sections are hidden.” That changes after September 30, 2026, when the EIA mandatory survey makes energy, cooling, and BTM generation data public for the first time.
The three footprints documented in this article — water, noise, and grid cost-shifting — determine whether a hyperscaler’s regional capacity plans will actually proceed on schedule. Ask about each of them directly before signing any long-term contracts.
Vendor Evaluation Checklist (10 Questions)
Nixon Peabody maps all 50 states into four risk tiers — moratoria/bans, regulated growth, tax incentives under fire, and pro-growth — with Virginia, Michigan, Minnesota, and Wisconsin in the highest-risk tier.
Until September 30, 2026, direct questions are the only proxy for facility-level data. After that date, the EIA mandatory survey makes water consumption, BTM generation, and cooling efficiency metrics matters of public record. For the due diligence questions to ask cloud providers about site operations, continue to the CTO Guide.
A typical mid-sized facility uses around 300,000 gallons per day. A large hyperscale facility can draw up to five million gallons per day. The industry average WUE is 1.9 litres per kWh; facilities using closed-loop or immersion cooling score significantly lower.
Data centres could represent up to 12% of all US electricity consumption by 2028 (Lawrence Berkeley National Laboratory). By 2030, data centre demand may reach 90 GW — nine times New York City’s peak summer demand.
The EPA defunded its Office of Noise Abatement and Control in 1981, and federal noise standards for data centres don’t exist. Local zoning designed for occasional industrial activity is poorly suited to continuous 24/7 low-frequency hum, and multi-frequency noise makes standard decibel-meter enforcement difficult.
WUE measures litres of water consumed per kilowatt-hour of IT energy. The industry average is 1.9 L/kWh; the ideal score is 0 (fully air-cooled systems). Request the facility WUE from co-location providers as standard due diligence before committing to a contract.
BTM generation means a data centre generates its own power on-site, independently of the utility grid. Generators and turbines that run continuously produce 24/7 noise rather than occasional emergency operation. The EIA mandatory survey will capture BTM generation data for the first time after September 2026.
Yes. Two-thirds of US data centres under development since 2022 are in water-stressed areas (WRI). In western states under prior appropriation water law — Colorado, Nevada, and Arizona — the newest users face curtailment during drought and cannot guarantee long-term access.
Nearly one-third of Virginia’s data centres sit within 200 feet of residentially zoned properties (Virginia JLARC 2024 report); some are within 50 feet of homes. Prince William County residents routinely report noise above 60 dB, and Amazon has been retrofitting acoustical shrouds at some Northern Virginia facilities in response.
NERC issued a Level 3 Essential Action Alert on May 4, 2026 — its highest designation — in response to data centres dropping load or oscillating demand rapidly, creating grid reliability risks. Seven mandatory actions are required by August 3, 2026, with NERC CLE Compliance expected to become enforceable for in-scope data centres by 2027.
After September 30, 2026, the EIA’s mandatory nationwide survey will require data centres to report grid-supplied electricity, behind-the-meter generation, cooling efficiency metrics, and IT specifications — making previously proprietary operational data matters of public record for the first time.
Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI signed the voluntary pledge on March 4, 2026, committing to build or buy their own generation, pay for delivery upgrades, and negotiate separate rate structures. Voluntary today — but its vocabulary now appears in 18+ state bills.
Yes: through local zoning denials (upheld by the Virginia Court of Appeals in the QTS Digital Gateway case); moratoriums (adopted by 12+ jurisdictions); ballot measures (Ohio residents are pursuing a permanent ban); and community litigation. In Menomonie, Wisconsin, the city council blocked a $1.6 billion project outright.
Rural opposition is often more intense — loss of farmland, groundwater depletion, a fundamental change in community character, and less organisational capacity to negotiate CBA terms. Suburban communities have more resources to mount legal challenges but face greater pressure to accept economic development arguments.
Maine Freezes — The First State-Level Data Center MoratoriumIn April 2026, the Maine Legislature became the first state legislature in the United States to pass a bill freezing approvals for new data centres. That matters for infrastructure planning, regardless of what happened next.
What happened next: Governor Janet Mills vetoed LD 307 on 24 April 2026.
A vetoed bill is not a failed signal. Both chambers passed it. That means a state-level moratorium is not fringe advocacy — it is a credible policy tool that cleared the full institutional machinery of American state lawmaking. The veto does not erase that.
Maine sits at the leading edge of a 300+ bill legislative wave across 30+ states, documented by Nixon Peabody in their May 2026 50-state framework. For the full picture of the national buildout driving this backlash, read the $710B buildout driving state action. For the broader revolt, see the data centre community revolt.
Maine Legislative Document 307 proposed an 18-month moratorium on new state approvals for data centres requiring more than 20 megawatts of grid power — a pause to study the impacts on the electric grid, electricity rates, water supply, and environment.
Two definitional points matter here.
This was a moratorium, not a ban. A moratorium is a temporary pause with a defined end date. Some media coverage used “ban” in headlines — that is an overstatement. No US state has enacted a permanent ban on data centres as of May 2026.
The 20 MW threshold targets hyperscale, not enterprise. We are talking about the kind of infrastructure AWS, Google, or Microsoft build at scale — not the server room in your office. FERC uses the same 20 MW figure in its large-load interconnection rulemaking (Docket RM26-4-000). When state legislatures and federal regulators independently land on the same number, it becomes the de facto industry standard. Pay attention to that number.
The bill was sponsored by State Representative Melanie Sachs (D). It passed the Maine House 79–62 and the Senate 21–13 — simple majority in both chambers, short of the two-thirds threshold required to override a veto.
The Maine Legislature passed LD 307 for the same reasons community opposition has emerged across dozens of states: grid strain, rising electricity rates, water consumption, and NDA-enforced secrecy. By the Senate vote, 4,900 Mainers had written letters of support — the largest response organisers had seen for state energy legislation.
Governor Mills’s veto on 24 April created a political paradox. Her veto message included this line: “A moratorium is appropriate given the impacts of massive data centers in other states on the environment and on electricity rates.” She agreed with the concept. She vetoed the implementation.
Her stated reason was a $550 million Sentinel Data Centers development on the site of the former Androscoggin paper mill in Jay, Maine — a mill that closed in 2023, taking hundreds of jobs with it. The Sentinel project would create over 800 construction jobs in an economically distressed community, and would draw up to 25 megawatts from the grid, squarely above the threshold. As drafted, LD 307 would have stopped it cold.
Mills framed the veto as a drafting problem, not opposition to moratoriums. She simultaneously signed a companion bill blocking data centres from Maine’s business development tax incentive programmes, and issued an executive order for an advisory council on data centre impacts.
Representative Sachs called the veto “simply wrong.” The override was mathematically unavailable. The politics are instructive: this was not a partisan split — the conflict occurred largely within Democratic ranks. Data centre regulation cuts across conventional political alignments.
The legislative vote is the milestone. When a state legislature passes a moratorium bill, it signals that the tool has cleared the full machinery of American state lawmaking. The veto is a footnote about exemptions — it does not erase the precedent.
The companion measures persist regardless of the moratorium’s fate. The tax incentive ban is law. The advisory council is running. Councils that study an issue tend to recommend regulation. Maine has also provided a template: the next drafter gets Maine’s drafting, vote tallies, and a footnote about exemptions.
For compliance planning, what counts is the 3–5 year infrastructure horizon. A state at “legislative proposal” stage can reach “passed” stage within a single planning window. A March 2026 Quinnipiac poll found 65% of Americans oppose building AI data centres in their communities — a political majority, not a fringe position.
Nixon Peabody’s conclusion on all of this: “The ‘energy-first’ data center siting era in the US is over.” Grid access is no longer enough. Projects now require a “social licence to operate” — earned community acceptance that must be built, not assumed.
For parallel governance pressure through the courts, see the Virginia Cancellation Wave.
Maine is the leading edge of a legislative wave. Nixon Peabody documents more than 300 state bills across 30+ states — at least 12 proposing moratoriums or bans, the rest addressing ratepayer protections, NDA prohibitions, environmental review, and tax incentive reform. The scale of the national buildout driving state action explains why so many states are reaching for the same tool at the same time.
Nixon Peabody’s 50-state framework classifies every US state into five tiers by legislative posture:
Tier 1 — Moratoria and Bans (highest compliance risk): Maine, Virginia, Michigan, Minnesota, Maryland, New Hampshire, New York, Oklahoma, Pennsylvania, South Carolina, South Dakota, Vermont.
Tier 2 — Regulated Growth (complex permitting; elevated but manageable): California, Connecticut, Florida, Illinois, Massachusetts, North Carolina, New Jersey, Oregon, Washington.
Tier 3 — Tax Incentives Under Fire (indirect cost risk): Georgia, Indiana, North Carolina, Oklahoma, Virginia, Washington.
Tier 4 — Pro-Growth / Tightening (favourable with emerging restrictions): Alabama, Arizona, Iowa, Nebraska, Nevada, Ohio, Texas, Utah.
Tier 5 — Neutral / Local Control (minimal state-level activity; local moratoriums still possible): Arkansas, Kansas, Kentucky, Mississippi, Missouri, Montana, Tennessee.
Some states appear in multiple tiers. Virginia simultaneously faces moratorium proposals, complex permitting, and tax incentive pressure, with 61 bills filed and $1.6 billion at stake from potential exemption loss.
Bipartisanship is a feature of this landscape, not a bug. The federal moratorium bill is led by progressives. Wisconsin and Florida NDA opposition reflects conservative concerns about local control. Missouri voters replaced half a city council over a data centre dispute. The grievances — electricity costs, water use, noise, opacity — do not respect party lines.
On 25 March 2026, Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez announced the Artificial Intelligence Data Center Moratorium Act (S.4214): a nationwide freeze on data centres requiring 20 megawatts or more. The political prognosis is poor — Senator Mark Warner called a federal moratorium “idiocy.” Treat it as a political signal, not a near-term outcome.
Three federal certainties are advancing regardless of S.4214’s fate.
FERC RM26-4-000 (April 2026): Uniform rules for large electrical loads of 20 MW or more, governing interconnection and grid upgrade costs. Action expected by end of June 2026.
EIA Mandatory Energy Survey: Mandatory nationwide reporting on data centre energy use — grid-supplied electricity, behind-the-meter generation, cooling efficiency — becomes compulsory after 30 September 2026. Previously proprietary metrics will become public record.
NERC Level 3 Essential Action Alert (4 May 2026): Seven actions addressing reliability risks from computational loads, with enforceable obligations expected in 2027.
The federal moratorium bill is a signal. The three items above are compliance certainties. For the broader story, see the data centre community revolt.
Tier 1 states require immediate attention. Any data centre project at or above 20 MW in Maine, Virginia, Michigan, Minnesota, Maryland, New Hampshire, New York, Oklahoma, Pennsylvania, South Carolina, South Dakota, or Vermont should carry moratorium-termination rights in acquisition contracts, and should budget for 12–24 month approval delays as a base-case scenario.
The structural factor that determines where risk sits is the Home Rule versus Dillon’s Rule distinction.
Home Rule grants local governments broad independent authority over land use. A county or municipality can enact a moratorium without state-level legislation — risk is decentralised and hard to track. Eight Georgia towns and four Indiana counties have already banned new data centre construction independently.
Dillon’s Rule constrains local governments to powers explicitly granted by state law. Virginia’s Loudoun County Board has stated it lacks the legal authority to implement a moratorium. In Dillon’s Rule states like Virginia and Texas, the state legislature is the only viable channel for a legally enforceable freeze.
Maine is a Home Rule state — LD 307 was a choice, not a necessity, which is why it is a template.
Nixon Peabody’s practical guidance here is blunt: “Don’t mistake silence for safety.” Here is what you need to do:
Map your cloud exposure to the tier framework. Work out which tier applies to the regions where your primary cloud providers — AWS, Google, Microsoft, Azure — operate significant capacity.
Ask your cloud providers directly. Request permitting pipeline status in Tier 1 states and ask whether any moratorium-triggered delays are projected.
Review your contracts. Multi-year cloud and co-location agreements may not cover legislative delays in force majeure clauses. Check now, not when a moratorium lands.
Monitor datacentertracker.org. Near-real-time updates on community opposition by county — the early warning signal for where pressure is building before it becomes legislation.
Build delay assumptions into capacity planning. For Tier 1 states, model your infrastructure plan with and without a 12–24 month permitting delay. If the numbers only work with on-time approvals in high-risk states, that is a problem you need to address now.
What is Maine LD 307 and what would it have done? An 18-month moratorium bill that passed the Maine Legislature in April 2026 — the first such bill passed by any US state legislature. It would have frozen approvals for new data centres requiring more than 20 megawatts of grid power until approximately late 2027. Governor Janet Mills vetoed it on 24 April 2026.
Why did Governor Mills veto a bill she said she agreed with in principle? The bill lacked an exemption for a $550M Sentinel Data Centers project in Jay, Maine — expected to bring over 800 construction jobs to a community hit by mill closures. Mills described it as a drafting problem, not opposition to moratoriums. She signed a companion tax incentive ban and issued an executive order for an advisory council as alternative measures.
Could the Maine Legislature override the veto? No. A veto override requires two-thirds in both chambers. LD 307 passed 79–62 in the House and 21–13 in the Senate — both below two-thirds. The override path was mathematically closed.
What is the 20-megawatt threshold in data centre legislation? The power consumption level used in LD 307, the Sanders/AOC federal bill, and FERC’s large-load interconnection rules (RM26-4-000) to define a “large” data centre. Facilities at or above 20 MW are the target of most moratorium proposals; below that threshold, facilities typically fall outside scope.
What is the difference between a data centre moratorium and a permanent ban? A moratorium is a temporary pause with a defined end date. A ban is indefinite. LD 307 was an 18-month moratorium; no US state has enacted a permanent ban as of May 2026. Ohio’s proposed ballot initiative would come closer to a permanent restriction.
What is the Home Rule vs. Dillon’s Rule distinction and why does it matter for data centres? Home Rule states let local governments enact moratoriums independently. Dillon’s Rule states allow only powers explicitly granted by state law — local moratoriums require state authorisation. In Dillon’s Rule states like Virginia and Texas, the state legislature is the only viable channel for legally imposing a moratorium.
What is the Sanders/Ocasio-Cortez AI Data Center Moratorium Act? Announced 25 March 2026 as S.4214, the Act proposes a nationwide moratorium on new data centres at or above 20 MW. It faces very low prospects of passage. Its significance is as a political signal — data centre opposition has now reached the federal legislative level.
What is Nixon Peabody’s 50-state data centre framework? Nixon Peabody’s 7 May 2026 analysis classifies all 50 US states into five tiers based on legislative posture toward data centres. Tier 1 (Moratoria and Bans) carries the highest compliance risk. It translates directly into a tool for assessing cloud provider regional risk.
What should you do given 300+ pending state data centre bills? Map your cloud providers to the Nixon Peabody tier framework. Ask providers directly about permitting pipeline status in Tier 1 states. Review multi-year contracts for force majeure clauses that may not cover legislative delays. Build 12–24 month permitting delay assumptions into capacity planning for Tier 1 states. Monitor datacentertracker.org for near-real-time community opposition by county.
What happened to Maine’s data centre regulatory effort after the veto? Governor Mills issued an executive order for an advisory council to study data centre impacts on Maine’s grid, rates, environment, and water supply. She signed a companion bill blocking data centres from Maine’s tax incentive programmes. The moratorium concept has not been abandoned — the advisory council’s findings will inform future legislation.
NC Rebellion — Lawsuits, Moratoriums, and Burial GroundsWhen residents of Walnut Cove, North Carolina found out that ancestral graves sat inside the footprint of a proposed 1,844-acre hyperscale data centre, what had started as a local zoning dispute became a potential federal heritage law case overnight. The burial sites included graves of Hairston family descendants — people formerly enslaved on the very plantation now targeted for heavy industrial development — and sacred cultural landscape sites of the Saura, a Siouan-speaking Indigenous nation with deep roots in the Dan River Piedmont.
That escalation tells you something about where North Carolina is right now. An April 2026 Elon University poll of 800 residents found 44% oppose data centres in their community versus only 24% in support. A resident-compiled map documented 21 active projects facing organised opposition across the state — not a NIMBY cluster, but a statewide pattern. North Carolina is the sharpest edge of the broader data centre community revolt now unfolding nationally, where more than 140 community groups are pushing back against a $710 billion buildout that developers did not anticipate would meet this kind of resistance.
In this article we’re going to look at three NC case studies — Stokes County (six-organisation lawsuit, federal heritage law), Natelli Apex (full project cancellation), and Chatham County (a moratorium that generated its own developer lawsuit) — and what they mean for data centre development nationally.
North Carolina offered developers everything they wanted. Cheap Piedmont land, coal-dependent power infrastructure, and proximity to Research Triangle hyperscaler demand made it an obvious target. The state is on track to approve over 100 data centre projects — volume that is placing measurable strain on Duke Energy‘s grid and drawing the state’s 2050 decarbonisation goals into serious question.
The 21-project opposition map makes the NIMBY framing impossible to sustain. Rania Masri, co-director of the NC Environmental Justice Network, put it plainly: “This is across economic lines, it’s across racial lines, it’s across county lines. Everyone that I have talked to is upset about data centres.” Democrats oppose at 55%, independents at 48%, Republicans are split.
The structural legal feature shaping what communities can actually do here is Dillon’s Rule.
💡 Dillon’s Rule limits local government authority to powers expressly granted by state law — which means county-level moratoriums face a threshold question of whether counties have statutory authority to adopt them at all.
Political pressure has reached the state level. Governor Josh Stein has called for a review of data centre tax breaks. House Bill 1063 would mandate consumption reporting, require 25% clean power sourcing, and repeal existing tax exemptions. There is also a complication tucked into a 2024 Hurricane Helene relief bill — a provision that already prevents municipalities from downzoning to exclude data centre use, which makes any blanket county response considerably harder.
In January 2026, the Stokes County Board of Commissioners voted 3-2 to approve Engineered Land Solutions (ELS, CEO Drew Nations) rezoning 1,844 acres of residential-agricultural land along the Dan River for heavy industrial use. This enabled “Project Delta,” a proposed hyperscale data centre. The vote overrode the Planning Board’s recommendation to deny. ELS had incorporated only in December 2024 and had not yet named a tenant or substantiated the promised 250–500 jobs and 20–40 million in annual tax revenue.
The commissioners also amended the county-wide zoning ordinance to add data centres as a permitted use throughout all heavy manufacturing districts — meaning future projects could proceed by right.
💡 By-right development means a project can proceed without public notice or a hearing as long as the use is already listed as permitted in that zone.
In March 2026, six advocacy organisations filed Hairston Clan et al. v. Stokes County — the National Hairston Clan, CleanAIRE NC, Dan River Basin Association, 7 Directions of Service, SCSJ, and SELC. Their case: the county had failed to follow statutorily mandated notification timing, had not conducted a meaningful environmental impact assessment, and had approved the rezoning on “false or materially incomplete premises.”
In April 2026, the board voided its own rezoning. County Manager Jeff Sanborn acknowledged: “We did not follow statutorily mandated timing of our notifications to residents. It was an oversight.”
ELS plans to refile. The by-right county-wide amendment remains in place. The takeaway is stark. A procedural defect in notification timing — not a substantive environmental ruling — was sufficient to void an approved rezoning and unwind months of developer work.
The procedural void only suspended the project. It did not touch the heritage law questions the January approval left entirely unresolved.
Within the Project Delta footprint lie ancestral burial sites connected to two distinct communities. Robert Hairston, Board Chairperson of the National Hairston Clan, was direct about it: “Our foreparents are buried on the data centre site. Not history in a book — real people, in the ground.” Crystal Cavalier-Keck, co-founder of 7 Directions of Service, put the Indigenous dimension in context: “For Indigenous people of this region, the Dan River corridor is not just land — it is a living cultural landscape tied to the Saura and other Siouan-speaking nations who have lived along these rivers for centuries.” Add the adjacent historically Black neighbourhood of Baileytown and you have an environmental justice dimension sitting alongside the heritage law questions.
Two federal statutes become relevant when projects with a federal permit nexus disturb culturally significant sites. NHPA Section 106 requires mandatory consultation with the State Historic Preservation Officer and potentially affected tribal nations before ground disturbance — a process the developer cannot waive. NAGPRA (Native American Graves Protection and Repatriation Act) separately governs treatment of Native American human remains discovered during excavation, with federal penalties triggered independently of local approvals.
Here is the mechanism that matters. A FERC power interconnection or an Army Corps environmental discharge permit is sufficient to activate Section 106 review. Even if ELS refiles and Stokes County approves Project Delta, any federal permit nexus activates federal consultation that survives local zoning decisions entirely. The NDA secrecy that helped organise community opposition at Stokes County was a documented catalyst — the Project Delta codename and NDA requirements left communities with little information before the January 2026 vote.
In March 2026, Natelli Investments withdrew all applications for its 250 MW New Hill Digital Campus in Apex, Wake County, adjacent to Duke Energy’s Harris Nuclear Plant.
Community opposition focused on water consumption of up to one million gallons per day, energy costs, diesel generator health effects, and nuclear plant proximity. Melissa Ripper of the volunteer-led Protect Wake County Coalition described the withdrawal timing as a surprise: “We thought we had, probably, another six to eight months of trying to push back.” The developer exited rather than fight.
This is the clearest documented case in North Carolina of community pressure achieving a full project cancellation. Natelli subsequently pursued a rezoning in Vance County, generating fresh opposition there. The pattern is site-specific cancellation, not regional exit.
Electoral accountability is also now a documented mechanism. In Edgecombe County’s Kingsboro, David Batts warned commissioners at a December 2025 hearing: “We will primary you.” In the March 2026 primary, Batts received 337 votes against 162 for four-term incumbent Donald Boswell — defeating a 16-year commissioner on an anti-data centre platform.
On February 11, 2026, Chatham County adopted a 12-month moratorium on new data centre development. By April 2026, roughly a dozen NC localities had adopted moratoriums or were positioned to do so.
On April 27, 2026, ECO TIP West LLC filed suit in Chatham County Superior Court. The developer plans the 750 MW Triangle Innovation West project at the Moncure Megasite and claims to have invested more than $11 million in pre-moratorium development activity, and holds a county-issued October 2025 zoning permit confirming data centre use was by-right. Represented by Robin Tatum and Mitchell Armbruster of Smith Anderson, ECO TIP West is seeking a declaration that the moratorium is invalid, confirmation of vested rights, project exemption, damages, and attorney’s fees — and is pushing for expedited judicial review.
💡 Vested rights is a legal doctrine under which a developer who has made substantial investment in good-faith reliance on existing zoning acquires rights that cannot be stripped away by a subsequent moratorium or zoning change. NC’s exact statutory vesting threshold is not yet resolved in case law.
A moratorium shifts rather than eliminates legal risk. Chatham County adopted one — and now faces a developer claiming $11 million in vested rights.
Both approving and blocking a data centre project now carry litigation risk for local governments. That is bidirectional legal exposure. Stokes County: approved rezoning → community lawsuit → voided decision. Chatham County: adopted moratorium → developer lawsuit claiming $11M in vested rights. There is no procedurally safe path.
The Vorys legal analysis identifies five grounds on which a moratorium may be challenged: lack of authority to adopt it, vested rights interference, regulatory takings, undue burden on interstate commerce, and constitutional violations. Under Dillon’s Rule, the authority question alone is a threshold risk for NC counties.
Nixon Peabody‘s May 2026 analysis offers the counternarrative. Community Benefit Agreements (CBAs) are now functioning as permitting prerequisites. The Saline Township, Michigan settlement secured approximately $14 million in community benefits and let the project proceed. Nixon Peabody’s bottom line: “Don’t mistake silence for safety.” Run a political durability assessment at every candidate site.
Moving planned capacity to a county that “seems friendlier” doesn’t eliminate the risk. The whole permitting environment has become litigation-prone regardless of local political lean.
North Carolina is functioning as a template. At least 19 Michigan municipalities have enacted data centre moratoriums. New York’s Senate Bill S.9144 proposes a three-year statewide moratorium on facilities exceeding 20 MW. More than 230 advocacy groups have petitioned Congress.
Three developer-side precedents from NC cases carry national relevance.
North Carolina hosts significant hyperscaler infrastructure for AWS, Azure, and GCP in the Research Triangle. Sustained opposition here affects regional capacity supply for cloud customers. Nixon Peabody’s power-plus-permission model — replacing energy-first site selection with a five-factor assessment of grid access, regulatory clearance, community acceptance, legislative durability, and ratepayer protection — is the direct strategic output of the NC experience. It is one of the central frameworks examined in our data centre community revolt overview.
How NC regulatory patterns spread to other states where hyperscalers have facilities is examined in our guide for technology leaders. The NDA secrecy that helped organise community opposition is moving into legislation — NDA prohibition bills are advancing in Wisconsin, Minnesota, and elsewhere.
Miquel Vila of Data Center Watch summed it up: “There’s no safe space for data centres. Opposition is happening in very different communities.” The full picture of how the $710B buildout collision is unfolding across all vectors — geographic, legal, legislative, and environmental — is in our comprehensive data centre community revolt resource.
Project Delta is the codename for a proposed hyperscale data centre complex on 1,844 acres along the Dan River, developed by Engineered Land Solutions (ELS). The Stokes County Board approved the rezoning 3-2 in January 2026, overriding its Planning Board. Six organisations sued in March 2026 over statutory notification timing. The board voided its own rezoning in April 2026 citing that procedural failure. ELS plans to refile.
Six plaintiff organisations filed Hairston Clan et al. v. Stokes County: the National Hairston Clan, CleanAIRE NC, Dan River Basin Association, 7 Directions of Service, SELC, and SCSJ. Their core claim was that the county failed to follow statutory notification timing requirements. Secondary concerns included the absence of a meaningful environmental impact assessment, ancestral burial grounds within the project footprint, and air, water, and noise impacts.
Vested rights is a doctrine under which substantial investment made in good-faith reliance on existing zoning generates rights that a subsequent moratorium cannot extinguish. ECO TIP West’s claim rests on $11 million invested pre-moratorium plus a county-issued October 2025 zoning permit confirming by-right use. NC’s exact statutory vesting threshold is not yet resolved in case law.
NHPA Section 106 requires mandatory federal consultation with the State Historic Preservation Officer and potentially affected tribal nations before ground disturbance at any project requiring a federal permit — such as a FERC power interconnection or Army Corps environmental discharge permit. This process cannot be waived regardless of local zoning approvals. Where Native American remains may be present, NAGPRA adds further federal obligations triggered independently of local decisions.
Yes — Natelli Apex in Wake County is the NC proof point. Organised community opposition drove a complete withdrawal of the 250 MW New Hill Digital Campus application. Natelli subsequently pursued a site in Vance County. Opposition achieves cancellation at a specific site; the developer typically redirects rather than exits the region.
A moratorium is a time-limited freeze on new permits — typically six to twelve months — while a local government develops regulatory standards. A zoning ban restricts a use indefinitely. Both delay projects, but moratoriums have stronger legal footing when purpose-stated and time-limited. Neither is litigation-free: the Chatham County case demonstrates that moratoriums generate their own developer lawsuit exposure through vested rights and regulatory takings claims.
Dillon’s Rule limits local government authority to powers expressly granted by state law. In North Carolina, county-level moratoriums face a threshold challenge: does the county actually have statutory authority to adopt one? This is the first of five challenge vectors in the Vorys developer challenge framework. Loudoun County, Virginia — also Dillon’s Rule — stated it lacks legal authority to implement a data centre moratorium. Home Rule jurisdictions can act on any matter not prohibited by state law.
Both approving and blocking a data centre project now carry litigation risk for local governments — that is bidirectional legal exposure. Stokes County: approved rezoning → community lawsuit → voided decision. Chatham County: adopted moratorium → developer lawsuit claiming $11M vested rights. No procedurally safe path exists.
By-right zoning allows a use to proceed without public notice or a hearing. The Stokes County rezoning added data centres as a by-right use in all heavy manufacturing districts county-wide. Even after the Project Delta site rezoning was voided, that amendment remains — future data centre projects in those districts can proceed without community input.
Nixon Peabody’s term for the post-2025 standard. Grid access is necessary but not sufficient; projects positioned for success pair credible speed-to-power plans with earned community buy-in, transparent operations, reliable utility relationships, and ratepayer-protective cost allocations. The model is a direct output of the NC experience, where projects with power access still faced lawsuits, moratoriums, and full cancellations.
datacentertracker.org provides near-real-time county-level tracking of opposition actions, moratoriums, lawsuits, and project status across NC and nationally. The WUNC/NC Newsroom “Hyperscale, Hyperspeed” investigative series provides case-by-case reporting.
Natelli Investments withdrew all applications for the 250 MW New Hill Digital Campus in Apex, Wake County, adjacent to Duke Energy’s Harris Nuclear Plant, citing “ongoing deliberations over zoning ordinance changes.” Opposition focused on water consumption of up to one million gallons per day, energy costs, diesel generator health effects, and nuclear plant proximity. Natelli subsequently pursued a different site in Vance County, NC.