The Physical Stack of AI

Established Between a model someone wants to train and the electrons that train it stands a physical stack of roughly eight layers, and every layer is now on a public waiting list. From the top: the accelerator die; the high-bandwidth memory stacked beside it; the advanced package that marries logic to memory; the substrate under the package; the server, rack and liquid-cooling loop; the building; the electricity system that feeds it — generation, transformers, switchgear, transmission, an interconnection agreement; and beneath everything, land, water and a permit. This brief treats the stack as a single object because its failures are joint. The 2023 shortage was accelerator dies. The 2024 shortage was packaging and memory. The 2025–2026 shortage is electricity delivery equipment and grid connections. The binding constraint migrates downward, toward the layers with the slowest clocks.

Established The clock speeds are the argument. A training run can be commissioned in weeks. A gigawatt-scale campus can go from construction start to a gigawatt of live power in one to 3.6 years at the measured frontier. A high-bandwidth-memory part takes roughly nine to fifteen months from first sample to qualified volume production when nothing goes wrong, and more than two years when something does. A grid interconnection in a US organised market has a median of more than five years from request to commercial operation for the projects that make it at all; a heavy-frame gas turbine ordered in late 2025 was being scheduled into the next decade; and the wait for large transformers and cable roughly doubled in three years. The slowest permitted layer sets the pace of the whole industry, and nearly every strategic move in the sector — on-site turbines, nuclear power purchase agreements, federal land offers, panel-level packaging programmes — is an attempt to route around whichever layer currently binds.

Established Three boundaries keep this synthesis honest. The demand side of the energy ledger — joules per operation, facility overhead ratios, the terawatt-hour totals read as claims about need, and every architecture offered to reduce them — is owned by Ultra Efficient Computing Energy Systems, and this brief adopts its arbitration rule: if the question is how many joules a computation costs, it belongs there; if the question is what has to be built, qualified, connected and permitted so the joules arrive, it belongs here. Water at national scale — desalination, inter-basin transfer, aquifer storage — is owned by Water Infrastructure Megaprojects; water appears in this brief at facility and basin scale, as a cooling input, a siting variable and a permit condition. And the automation of goods movement is Autonomous Supply Chains; the supply chain appears here as capacity, qualification and lead time, not as logistics.

Frontier The joint question none of those briefs owns, and this one does: when demand for computation outruns every physical layer at once, which layer binds, for how long, and who decides — the market, the interconnection queue, or the permit? The measured record below says the answer has already changed twice in three years, and that the decisive evidence is scheduled: a cohort of announced gigawatt campuses carries public delivery dates clustered on late 2028, and their arrival or slippage will localise the constraint better than any forecast.

Established Start with the electricity, globally measured. The International Energy Agency’s Energy and AI report (2025) puts data-centre electricity consumption at about 415 TWh in 2024 — around 1.5% of world electricity — with the United States accounting for 45% of that consumption, China about 25% and Europe about 15%. Its base case more than doubles the total to around 945 TWh by 2030, with data centres taking roughly one tenth of global electricity demand growth to 2030 but more than 20% of demand growth in advanced economies, and nearly half of US demand growth. The same report carries the stack’s two quiet warnings: about 20% of planned data-centre projects could be at risk of delay for want of grid connection, and wait times for transformers and cables had doubled in three years.

Established The US record is congressionally mandated and unusually honest about its own uncertainty. Lawrence Berkeley National Laboratory’s 2024 United States Data Center Energy Usage Report (December 2024) measures US data-centre consumption at 176 TWh in 2023 — 4.4% of US electricity, up from 58 TWh in 2014 — and projects 325 to 580 TWh by 2028, or 6.7% to 12% of US electricity. The width of that range is the finding: it is dominated by assumptions about accelerator shipments and utilisation, which is to say by decisions not yet taken. Any single-number citation of this report is misquoting it.

Established National averages conceal the concentration, which is where the stack actually breaks. In Ireland, data centres took 21% of metered national electricity in 2023, up from 5% in 2015, and the state grid operator has a pause on new Dublin-area data-centre connections running to 2028. In Virginia, the legislature’s audit agency (JLARC, December 2024) found that unconstrained data-centre demand would double the state’s power requirement within ten years and would need roughly 150% more in-state generation and 40% more transmission; a single modern campus can exceed 1,000 MW, more than the output of a large nuclear unit, and an 18 MW building draws what roughly 4,500 homes draw. These are the two best-documented jurisdictions; they bracket the phenomenon rather than exhaust it.

Established The interconnection queue is the stack’s longest-measured layer, and its base rates are brutal. Berkeley Lab’s Queued Up: 2026 Edition counts roughly 1,312 GW of generation and 749 GW of storage active in US interconnection queues at the end of 2025 across about 8,200 projects. The median duration from interconnection request to commercial operation exceeded five years for projects completed in 2025. Of all capacity that requested interconnection between 2000 and 2020, only 13% had reached commercial operation by the end of 2025; 75% had withdrawn. And the queue’s composition pivoted with the AI buildout: natural gas requests rose 86% in one year while solar, wind and storage each fell 16–19%. On the load side the numbers are stranger still: ERCOT faced an estimated 474 GW of large-load interconnection requests by August 2026 — described by the state’s governor as more than five times the Texas record peak — with data centres making up roughly 90% of new requests; Texas ordered an audit and paused its transmission planning study, and New York had imposed a data-centre approval pause of up to a year the month before. Load queues, unlike generation queues, have no two-decade completion statistics; nobody yet knows their materialisation rate, and that ignorance is itself a planning input.

Frontier The money is measured; its attribution is contested. In PJM, the largest US market, the independent market monitor attributed the recent capacity-price escalation squarely to data-centre load: across two auctions it counted about $16.6 billion of data-centre-related revenue, roughly half of the $30.8 billion total, with the most recent auction up $7.3 billion (82%) on the prior year, and stated that current conditions “are not the result of organic load growth”. Utilities and some developers dispute how much of the price is properly assigned to data centres versus retirements and supply stagnation; the double-digit retail bill increases in parts of PJM’s footprint are not disputed.

Established The machinery layer has an order book, and the order book is the forecast. GE Vernova expected to end 2025 with an 80 GW gas-turbine backlog stretching into 2029, booked 18 GW of orders in the fourth quarter of 2025 alone, and its chief executive projected turbine reservations sold out through 2030 by the end of 2026 (company figures). Its production capacity moves from about 20 GW per year in mid-2026 to 24 GW per year by mid-2028 — a two-year lag between demand signal and metal. On the firm-power side, Microsoft signed a 20-year power purchase agreement for the restart of Three Mile Island Unit 1 (renamed the Crane Clean Energy Center): 837 MW, a $1.6 billion restart bill carried by Constellation, targeting 2028 subject to regulatory approval. A hyperscaler paying to reopen a mothballed nuclear plant is the clearest single price signal in the record: firm electrons, not chips, were the scarce good in 2024–2026.

Frontier Water is measured where companies choose to measure it, which is the problem. Google reported 8.1 billion gallons of water consumed across its data centres and offices in 2024, against 4.5 billion gallons replenished (about 64% of freshwater consumption, versus a 120% target for 2030), with electricity consumption up 27% year on year (company’s own figures). Industry-average water usage effectiveness is commonly cited near 1.8 litres per kilowatt-hour, while AWS reports 0.19 and Meta’s newest builds about 0.20 — a spread of nearly 10× that mixes genuinely different cooling choices with genuinely different accounting boundaries, and the indirect water consumed at the power plant serving the facility is routinely excluded altogether. A typical hyperscale campus is estimated around 550,000 gallons per day. Against that, Virginia’s JLARC found data centres took less than 0.5% of state water withdrawals. Both facts are true; the first is a basin-scale fact and the second a state-scale one, and almost all public argument consists of quoting them at each other.

Established Inside the building, the power density jumped an order of magnitude in two product generations. An NVIDIA GB200 NVL72 rack draws about 120 kW and its GB300 successor up to about 142 kW, against 3–10 kW for a conventional air-cooled server rack; direct-to-chip liquid cooling is mandatory, with rack-level coolant flows of roughly 59 to 177 litres per minute depending on facility water temperature. This is why the building layer, almost alone in the stack, got faster: purpose-built halls with busbars, coolant distribution units and warm-water loops are erected in months. Epoch AI’s satellite-tracked build census finds construction start to a full gigawatt of operational power taking 1 to 3.6 years, expected first gigawatt-scale sites online in early 2026, and xAI projecting twelve months for its second Memphis-area campus. OpenAI’s flagship Abilene site had about 0.3 GW live in April 2026 — four of eight buildings — targeting 1.2 GW by the fourth quarter of 2026.

Established The semiconductor module, sharply bounded: six facts about how silicon becomes deliverable compute. This brief does not cover nodes, lithography or fab economics; it covers the five technologies — high-bandwidth memory, chiplets, advanced packaging, substrates, optical I/O — and the one institution, qualification, through which chip supply becomes a delivery schedule. First, HBM: the JEDEC JESD270-4 HBM4 standard (April 2025) specifies a 2048-bit interface at up to 8 Gb/s per pin — up to 2 TB/s per stack — in stacks up to 16 die high and 64 GB, with vendor-selectable supply voltages for power. Memory bandwidth per package, not logic transistors, is the currency of the accelerator generation, and each HBM generation doubles the interface rather than shrinking a device.

Frontier Second, qualification is the gate that actually schedules the industry. The trade record assembled from 2022–2026 reads: SK hynix qualified HBM3 for volume in June 2022; Micron and SK hynix reached HBM3E volume in early 2024; Samsung — the largest memory maker by capacity — failed the lead customer’s heat and power qualification in April–May 2024 and did not pass until September 2025, roughly eighteen months of the strongest demand in memory history spent outside the market. HBM4 samples shipped from June 2025 and all three vendors were reported qualified and in production by mid-2026. A clean sample-to-volume cycle runs nine to fifteen months; a dirty one exceeds two years; and no capital expenditure compresses it, because the test is thermal and electrical behaviour of a 12-to-16-high bonded stack inside a customer’s package, which can only be measured by building and running it.

Frontier Third, packaging: chiplet architectures moved the integration point from the die to the package, and the package became the product. TSMC’s CoWoS-class capacity — the packaging family carrying nearly all frontier accelerators — ran about 75,000 wafers per month entering 2025 (roughly double 2024) and was reported heading for 120,000–140,000 by late 2026 with another 50,000–60,000 at outsourced assembly partners; TSMC has described a compound growth rate above 80% per year across 2022–2027, and analysts put the supply-demand gap at about 20% in mid-2026, narrowing to perhaps 10% by year-end (trade estimates, not audited figures). Fourth, substrates: the layer below the package has its own shortage — Ajinomoto build-up film and low-thermal-expansion T-glass cloth — with trade forecasts of a 10% ABF supply gap in late 2026 widening toward 20% in 2027 before announced capacity lands, an AI processor consuming roughly ten times the substrate material of a PC chip, and Ibiden committing about 500 billion yen through FY2028. Fifth, optical I/O: co-packaged optics entered the product cycle in March 2025 with NVIDIA’s photonic switches claiming 3.5× better power efficiency and 4× fewer lasers than pluggable transceivers (vendor claims; fleet-scale measurements do not yet exist). Sixth, the policy overlay: the US NAPMP programme finalised $1.4 billion in advanced-packaging awards in January 2025 — $100 million each for glass-core substrates (Absolics), silicon-core substrates (Applied Materials) and fan-out wafer-level processing (Arizona State), plus $1.1 billion for a national packaging piloting facility — on the stated logic that chips fabricated in the United States should also be packageable there. That is the whole module; everything else about semiconductors belongs elsewhere.

Frontier How much of the queued demand is real? This is the stack’s governing open question. Generation queues have a measured base rate — 13% of 2000–2020 requests completed — but large-load queues have none, and the incentives run toward duplication: a developer courts five utilities for one campus, and each utility books the request. ERCOT’s own preliminary forecast that peak demand could more than quadruple by 2032 came with the operator’s caution that the projection “may be inflated” by speculative projects. Until somebody publishes a load-queue completion statistic with the rigour Berkeley Lab applies to generation, every capacity plan, capacity auction and state audit is pricing a number nobody has measured.

Frontier Does the constraint keep migrating, and where does it settle? The record shows dies binding in 2023, packaging and memory in 2024–2025, and electricity delivery in 2025–2026. If CoWoS-class capacity and HBM supply normalise on the announced curves while interconnection medians stay above five years and turbine order books stay sold out, the constraint settles in the electricity layer for a decade, and compute geography becomes energy geography. The alternative — that packaging or substrates re-bind because package sizes and layer counts grow faster than capacity — is live: the substrate trade’s own forecast has the ABF gap widening through 2027, not closing.

Frontier Does behind-the-meter generation become the default, and does it survive contact with air-quality law? The 2028 cohort of announced campuses is predominantly powered by on-site natural gas microgrids rather than grid interconnections — a direct purchase of schedule. But the Memphis-area record shows the routing-around has its own regulator: dozens of turbines operated without air permits drew a federal Clean Air Act suit in April 2026. Whether on-site gas is a bridge, a permanent fixture, or a liability that consent decrees and community litigation gradually reprice is unresolved and decisive for both emissions and delivery schedules.

Frontier Is water a physical constraint or a consent constraint? The measured on-site volumes are small against state withdrawals and against agriculture; the basin-level and disclosure records are poor; and closed-loop and dry cooling exist at a known energy penalty (roughly 10% efficiency loss moving away from evaporative designs, per industry guides). The frontier question is not whether the water exists but whether communities will permit its use without facility-level disclosure — and whether the industry will accept disclosure before being compelled.

Frontier Who pays for the grid the stack needs? The PJM monitor’s attribution — half of $30.8 billion in capacity-auction revenue traced to data-centre load — is one accounting; utilities offer others; and the emerging instruments (special large-load tariffs, take-or-pay contracts, co-location behind existing plants, curtailment obligations) are all attempts to move the incidence. The allocation question is genuinely open in every US state and in Ireland, and its answer determines whether the buildout is politically durable.

Frontier Can qualification cycles compress? The nine-to-fifteen-month sample-to-volume cycle for HBM, and the analogous cycles for substrates and co-packaged optics, are the semiconductor stack’s hard clock. Standardised known-good-die testing, better thermal modelling of tall stacks, and the NAPMP metrology programmes all aim at it. There is no published demonstration yet of a materially shorter cycle at frontier stack heights; sixteen-high hybrid-bonded HBM4E qualification, expected in the 2026–2028 window, is the next natural reading.

Frontier And the demand curve itself — whether efficiency gains bend it or feed it — is the one frontier question this brief refuses to own, because Ultra Efficient Computing Energy Systems owns it, names compute demand growth as the most decisive variable on its own map, and this brief’s entire subject is the consequence of that variable taking high values.

Established Ranked by measured lead time, the bottlenecks are not where the public argument is. The slowest layer is electricity delivery: a median above five years from interconnection request to commercial operation for completed generation projects, a 13% historical completion rate, transformer and cable waits that doubled in three years, and heavy-frame gas turbines whose reservations were projected sold out through 2030. Nothing in silicon is that slow.

Established Second: the qualified-capacity bottlenecks of the semiconductor module. Advanced packaging capacity roughly doubling year on year and still carrying a double-digit supply gap; HBM supply concentrated in three firms, one of which spent eighteen months failing a single customer’s thermal qualification; ABF substrate and T-glass gaps forecast to widen through 2027. The distinguishing feature of this tier is that capacity is not fungible: wafers of packaging capacity are useless until a specific stack, in a specific package, passes a specific customer’s qualification, so headline capacity numbers systematically overstate deliverable supply.

Frontier Third: people. The trade press converges on electrical trades as the binding construction input — substation electricians, high-voltage linemen, and the specialised winding labour behind transformer production — and vendors report multi-year backlogs attributed to labour rather than materials. The claim is plausible and repeated by manufacturers, but no rigorous labour-market study of the transformer and switchgear workforce comparable to the Berkeley queue statistics exists in the public record; this bottleneck is asserted more than it is measured.

Established Fourth: minerals, concentrated rather than absent. The IEA’s critical-minerals series puts the average top-supplier refining share at 72% in 2025 and rising; China refines about 99% of gallium, with data-centre demand alone plausibly exceeding 10% of today’s supply by 2030; European gallium prices ran about five times Chinese domestic prices under export controls; and copper — the stack’s bulk electrical material — carries a projected supply deficit of 25% by 2035. None of this stops a buildout this decade; all of it prices and politicises one.

Established Fifth: consent. Data Center Watch, a research outfit tracking local opposition, counted $18 billion in blocked and $46 billion in delayed US projects over roughly two years to March 2025, with 142 opposition groups across 24 states and an opposition that is bipartisan — 55% of opposing elected officials Republican, 45% Democratic. Two statewide pauses (Texas audit, New York approval pause) arrived in a single summer. Consent is a bottleneck with no engineering workaround, only an institutional one.

Frontier What is conspicuously not a bottleneck: the building. Purpose-built halls reached gigawatt scale in one to 3.6 years, with the fastest builders projecting twelve months. Concrete, steel and shell construction respond to money at the speed money arrives. The stack’s popular image — that the hard part is building the data centre — inverts the measured record: the hall is the fast layer wrapped in slow ones. Water, similarly, binds at specific basins and specific permits rather than in aggregate; and the die itself, the layer that named the shortage in 2023, was by 2026 the layer with the most committed capacity behind it.

Established The stack’s research dependencies are mostly measurement institutions that do not yet exist. First: a load-queue statistic. Berkeley Lab’s generation-queue series is the only reason anyone knows the 13% completion base rate; there is no equivalent census of large-load requests, no standard for de-duplicating multi-utility shopping, and therefore no empirical basis for the capacity plans now being financed against 474 GW-scale request books. Producing one requires utility disclosure rules, not laboratory work.

Established Second: facility-level energy and water disclosure. The measured water record is a patchwork of voluntary corporate totals with incompatible boundaries; the power-usage and water-usage effectiveness figures in circulation are self-reported and unaudited — a gap Ultra Efficient Computing Energy Systems flags on the energy side and which is strictly worse on the water side, where the 10× spread between reported WUE values cannot be decomposed into cooling physics versus accounting choices from public data.

Frontier Third: packaging science with public metrology. The NAPMP research stack — glass-core and silicon-core substrates, fan-out wafer-level processing, and a piloting facility — is an attempt to give advanced packaging what lithography has had for decades: shared roadmaps, shared test vehicles, and metrology that lets a second source qualify without repeating the first source’s entire learning curve. Whether a $1.4 billion federal programme can seed that against an incumbent ecosystem doing multiples of that privately per year is the open institutional experiment.

Frontier Fourth: thermal and reliability models for tall bonded stacks. The Samsung HBM3E record demonstrates that heat and power behaviour of 12-high stacks inside a customer package was not predictable from vendor-side test data; sixteen-high HBM4 stacks and co-packaged optics push further into regimes where qualification is the only trusted model. A validated public thermal-mechanical simulation capability for advanced packages would shorten the industry’s slowest silicon clock; nothing published yet does it at qualification grade.

Established Fifth: grid engineering imports. The stack inherits, rather than owns, the dependencies of transmission expansion, storage, and firm low-carbon generation — the subjects of Energy Corridors, Energy Storage Revolutions, Small Modular Reactors and Advanced Fission. Every year those fields run late, the on-site gas share of the stack rises by default.

Frontier The decisive test is already running. The announced 2028 cohort — five Stargate-class sites carrying roughly 8 GW of stated capacity against fourth-quarter-2028 targets, the Crane nuclear restart, and the giant load queues behind them — will either reach powered operation on schedule or slip, and the pattern of the slips will localise the binding layer of the stack in a way no forecast can. Sites on on-site gas microgrids test the turbine order book and the air permit; grid-connected sites in Wisconsin and Michigan test the interconnection and consent layers; the restart tests nuclear regulatory throughput against a hyperscaler’s clock. The cohort is, in effect, a controlled comparison of routing-around strategies run at hundred-billion-dollar scale with published dates, and it needs no new instrument beyond satellite imagery, permit dockets and utility filings — the census Epoch AI already runs.

Frontier Second experiment: the load-queue materialisation measurement. Take the 474 GW of ERCOT large-load requests logged by August 2026 and the equivalent books in PJM and the Southeast, and measure delivered megawatts against them in 2030. The generation-side base rate is 13%; if load queues complete at anything like that rate, most current state-level capacity planning is provisioning for phantoms, and the capacity-market price signals now flowing into retail bills are partly the cost of unmeasured duplication. Texas’s August 2026 audit is the first institutional attempt at this measurement; its methodology will matter more than its headline.

Frontier Third: the qualification-cycle experiment, which the memory industry reruns every generation. HBM4E and sixteen-high hybrid-bonded stacks in the 2026–2028 window will show whether the nine-to-fifteen-month sample-to-volume cycle is compressible or whether taller stacks stretch it; a second Samsung-style eighteen-month failure at a moment of peak demand would establish qualification, not capacity, as the semiconductor module’s permanent constraint.

Frontier Fourth: the policy-result experiments. Ireland’s December 2025 connection rules — requiring new data centres to bring 80% new renewable supply and dispatchable on-site generation or storage matched to import capacity on a six-year glide path — constitute the strictest connection regime in the OECD applied to the most data-centre-dependent grid in it; whether it produces compliant connections, a construction halt, or emigration of the load to other member states will be visible by 2028–2030. The US federal permitting acceleration of July 2025 runs the opposite treatment; the pair is as close to an A/B test of permitting philosophy as the field will get.

Established What would count as disconfirmation matters here. If the 2028 cohort lands broadly on time, on on-site gas, with queues quietly shrinking as duplicates withdraw, then the stack thesis weakens into a procurement story: money bought schedule, and the binding constraint was price all along. This brief’s assessment — that the electricity layer binds for a decade — is falsifiable on a published calendar, which is more than most infrastructure theses can say.

Established The engineering programme inside the building is thermal. Racks at 120–142 kW require direct-to-chip liquid cooling as a design baseline: coolant distribution units separating a technology loop from facility water, rack-level flows of 59–177 litres per minute scaling steeply with facility water temperature, 50–51 V DC busbar distribution, and residual airflow for the components the cold plates miss. The retrofit question is largely settled in the negative: conventional air-cooled halls cannot take these racks without rebuilding power distribution and cooling from the slab up, which is why the fleet bifurcates into legacy cloud and purpose-built AI factories.

Frontier Water engineering is a three-way trade the sector has stopped pretending away. Evaporative cooling buys the lowest energy overhead at the highest water consumption; closed-loop and dry systems cut water toward zero at an energy penalty commonly put near 10%; warm-water liquid cooling shifts the optimum by raising the temperature at which heat leaves the rack, widening the envelope where dry heat rejection works. The engineering choice is now made basin by basin, and increasingly written into permits rather than left to the operator.

Frontier Power engineering is bifurcating into grid-side and island-side schools. The island school builds gas microgrids, batteries and eventually small reactors behind the fence, buying schedule at the price of fuel logistics, air permits and stranded-asset risk. The grid school signs structured tariffs with curtailment obligations — Texas’s SB6 regime, with its provisions for curtailing large flexible loads in emergencies, is the template — betting that a campus that can shed a quarter of its load for tens of hours a year is a grid asset rather than a burden. The unresolved engineering question underneath both: how much curtailment a frontier training run can actually absorb, on which no operator has published data.

Frontier In the semiconductor module, the engineering frontier is the package as a system. Co-packaged optics moves the electro-optical conversion inside the switch package for a claimed 3.5× power saving and 4× laser reduction (vendor figures, first products 2025–2026); glass-core and silicon-core substrates aim at the warpage and flatness limits of organic ABF at reticle-multiple package sizes; panel-level fan-out processing attacks the cost side; and 16-high stacks push hybrid bonding from roadmap to requirement. Each of these must pass the same institution — customer qualification — that the module’s history says is the true clock.

Established Construction engineering is the stack’s solved problem, and the solution is speed itself. Prefabricated electrical rooms, modular cooling plant, and design-for-repetition let the measured frontier hit a gigawatt in a year from a standing start where power exists. The engineering requirement that remains is not the hall but its interfaces: substations, transformer procurement, and the milestone-payment discipline that keeps a 1,500-worker site synchronised with a turbine delivery schedule set three years earlier — the classic megaproject interface problem, treated at length in Megaproject Governance.

Established The stack’s adjacencies are its seams, and each seam has an owner on this map. The demand curve and every efficiency claim made against it live in Ultra Efficient Computing Energy Systems, which also documents why the sector’s self-reported efficiency series cannot currently be audited. The supply of firm electrons is contested among Small Modular Reactors, Advanced Fission, Geothermal Megaprojects and Energy Storage Revolutions — each of which now aims a large share of its commercial case at this load — while Energy Corridors owns the transmission that would let siting decouple from generation.

Established Water’s national-scale machinery — desalination as an electricity purchase, storage as a property-rights problem — is Water Infrastructure Megaprojects, and its finding that the levelized cost of desalinated water is dominated by the energy term closes a loop with this brief: a data-centre buildout that raises regional power prices raises the cost of the water infrastructure its own siting disputes demand.

Frontier The construction seam runs to Automated Construction Systems and Robotics in Infrastructure — the data-centre hall, repetitive and schedule-priced, is the most natural early market either names — and the delivery-record seam to Megaproject Governance, because a $500 billion multi-site programme with published quarterly targets is a megaproject portfolio whether or not its sponsors use the word. Component movement, port throughput and the logistics of getting a 400-tonne transformer to a Texas county seat belong to Autonomous Supply Chains and Future Ports and Shipping.

Frontier Finally, the end-of-life seam: a fleet refreshing accelerators on two-to-three-year cycles at million-GPU scale is about to become one of the largest high-value waste streams in electronics, a problem Circular Infrastructure Systems owns and this brief merely schedules.

Established Permitting is where the stack meets the state, and 2025–2026 produced natural experiments in both directions at once. A July 2025 US executive action, Accelerating Federal Permitting of Data Center Infrastructure, directed agencies to speed environmental review and to open federal sites for qualifying data-centre and grid projects. In the same eighteen months: Texas ordered an audit of data-centre grid, water and subsidy claims and paused its large-load transmission study (August 2026); New York paused data-centre approvals for up to a year (July 2026); Ireland’s regulator imposed the 80%-new-renewables connection regime (December 2025); and dozens of US counties enacted moratoria or rejected rezonings. Federal acceleration and subnational braking are now running simultaneously on the same projects.

Established The consent record is quantified and bipartisan. The Data Center Watch census counted $18 billion in blocked and $46 billion in delayed projects across roughly two years to March 2025, 142 opposition groups in 24 states, and opposing elected officials split 55% Republican to 45% Democratic; in Warrenton, Virginia, every council member who had supported a proposed Amazon facility was voted out in November 2024. Opposition motives are local — bills, water, noise, land value — not ideological, which is why it does not map onto partisan control and cannot be legislated away from one capital.

Frontier The Memphis-area litigation is the institutional test case for the island strategy. By the plaintiffs’ account — the NAACP and Mississippi State Conference, represented by the Southern Environmental Law Center and Earthjustice — xAI operated some 35 turbines without air permits at its first campus before obtaining permits for 15, and by April 2026 was running 27 unpermitted turbines totalling up to 495 MW at its second, with claimed emissions above 1,700 tons of NOx per year; the Clean Air Act suit was filed on 14 April 2026 and is pending. The numbers are contested filings, not adjudicated facts. The institutional point is not: behind-the-meter generation does not exit the permitting state, it only changes which permit binds, and the enforcement clock — sixty-day notices, emergency motions — runs far slower than the twelve-month construction clock it is chasing.

Frontier Cost allocation is becoming the central institutional fight in the electricity layer. The PJM market monitor’s attribution of roughly half of $30.8 billion in capacity-auction revenue to data-centre load, and the double-digit bill increases that followed, have produced a wave of special large-load tariff dockets whose common design question is whether a hyperscaler’s contract can be made to carry the grid risk its load creates — minimum-take obligations, exit fees, curtailment duties. Ireland’s regime answers by making the load bring its own supply; Texas’s SB6 answers with emergency curtailment authority; no jurisdiction has yet run its answer through a full investment cycle.

Established On the semiconductor side, the institution being built is packaging sovereignty. The NAPMP’s $1.4 billion in final awards (January 2025) — substrates, fan-out processing, and a $1.1 billion national piloting facility operated by Natcast — is the first US industrial-policy programme aimed below the die, at the layer where 2024’s shortage actually lived. Its stated goal, advanced-node chips both manufactured and packaged domestically, is a decade-scale institutional bet whose first checkpoint is simply whether the piloting facility runs qualification-grade lots on a schedule customers use.

Frontier The distributional questions are concrete, current and mostly about incidence. Who breathes the exhaust: the unpermitted-turbine litigation sits in a corridor of majority-Black South Memphis neighbourhoods with a long industrial-burden history, and whatever the adjudicated emissions turn out to be, the pattern — fastest-built campuses seeking jurisdictions with the least enforcement friction — is an environmental-justice claim with a factual record attached. Who pays for the grid: capacity-market escalations traced by the market monitor to data-centre load land on retail bills years before any local benefit, a regressive transfer unless tariff design deliberately reverses it.

Frontier Who gets the water is less a volume question than a knowledge question. Aggregate withdrawals are small; specific basins are stressed; and the public record is voluntary corporate disclosure with incompatible boundaries, frequently shielded further by non-disclosure agreements between developers and localities during siting. A community asked to approve a campus without knowing its water draw, its cooling design or even its operator’s name is not consenting in any meaningful sense; the NDA-mediated siting process is the single practice this literature most consistently identifies as corrosive.

Established The labour ethics are boom-shaped. Construction peaks near 1,500 workers per site and falls to a few hundred permanent staff; the tax-revenue and jobs cases that carry local approvals are dominated by the construction phase; and the JLARC review — the most thorough public accounting — found genuine fiscal benefit alongside demand growth that the same jurisdiction’s ratepayers must finance. Honest siting arithmetic exists; it is simply rarely presented at the rezoning hearing.

Speculative The intergenerational question is stranded carbon and stranded steel. A buildout that resolves its schedule problem with behind-the-meter gas installs combustion assets on twenty-to-thirty-year lives against a demand curve nobody can forecast five years out; if demand disappoints or efficiency surprises, communities inherit the turbines, the emissions and the rate base. No jurisdiction currently requires decommissioning assurance for data-centre generation the way many do for wind and solar — an asymmetry with no principled defence.

Frontier The stack territorialises a technology that markets itself as placeless. If the electricity layer binds for a decade, frontier-model capability becomes a function of grid geography: gas basins, nuclear restarts, hydro provinces and permissive jurisdictions acquire the relationship to AI that bauxite and cheap power once had to aluminium. The 2025–2026 record already shows the migration — campuses following stranded gas and friendly permits rather than fibre routes — and states have begun treating interconnection queues as instruments of industrial strategy.

Established The concentration numbers are the systemic-risk numbers. Frontier packaging capacity sits overwhelmingly in Taiwan; HBM in three firms in two countries; refined gallium 99% in one country; top-supplier refining shares across critical minerals averaging 72% and rising. A stack whose every layer is queued is also a stack whose every layer is a single point of failure, and the NAPMP, European and Japanese packaging programmes are best read as civilizational hedging against exactly that concentration.

Frontier The capital scale has crossed into macroeconomic territory. A bank estimate put the five largest US hyperscalers’ 2026 capital expenditure near $697 billion — a private infrastructure programme in the range of national defence budgets — financed increasingly through project debt, SPVs and take-or-pay structures that will transmit any demand disappointment to credit markets rather than merely to equity holders. The honest historical comparisons are the railway manias and the 2000 fibre buildout: transformative infrastructure, genuinely built, with investor outcomes largely uncorrelated with the infrastructure’s eventual usefulness.

Speculative Two long-run branches, stated as branches. If demand holds, the stack becomes the organising customer of the mid-century electricity system — the load around which nuclear restarts, transmission corridors and water policy arrange themselves — and compute-rich regions compound advantages in science and defence. If demand breaks, the residue is a fleet of liquid-cooled halls, substations and gas turbines whose salvage uses (industrial heat, grid services, conventional cloud) are real but worth a fraction of cost. Either branch leaves the permanent gains where infrastructure booms always leave them: in the queues reformed, the workforces trained and the manufacturing capacity built while the money was easy.

These horizons track the stack’s delivery record against its announced schedule, not AI capability.

• 10 yr: Frontier The 2028 cohort resolves and localises the binding layer; load queues acquire their first honest completion statistics; turbine and transformer capacity expansions (24 GW per year of turbines by 2028) meet the backlog; CoWoS-class packaging and HBM normalise or re-bind at 16-high; the NAPMP piloting facility either runs customer-grade qualification lots or quietly does not; two or three US states and Ireland settle workable large-load tariff and connection regimes that others copy.
• 25 yr: Speculative Compute siting completes its migration to energy geography; small modular reactors and restarts supply a material share of campus power if their own briefs’ timelines hold; optical I/O and glass substrates are commodity layers; closed-loop cooling is the permitted default in stressed basins; the first full accelerator-fleet recycling industry operates at scale.
• 50 yr: Speculative The stack is either the electricity system’s anchor tenant — with grids planned around compute the way twentieth-century grids were planned around smelters — or a case study in overbuild whose halls found second lives; which branch obtained will have been decided by the demand curve, not by any layer this brief covers.
• 100 / 250+ yr: Handwave Projections at this range — orbital compute, planetary compute budgets, thermodynamic-limit computing economies — assert more about their authors than about infrastructure; nothing in the measured record constrains them.

• Depends on The demand curve, which is the decisive variable of Ultra Efficient Computing Energy Systems and is imported here unmodified. Firm supply and delivery run through Small Modular Reactors, Advanced Fission, Energy Storage Revolutions and Energy Corridors; basin-scale water through Water Infrastructure Megaprojects; and the delivery record of very large, schedule-driven programmes through Megaproject Governance.
• Requires (not on this map) Six conditions no brief on this map produces. Qualified high-bandwidth-memory and advanced-packaging capacity, where headline wafer counts overstate deliverable supply until each stack passes each customer’s qualification. Large-power-transformer and gas-turbine production, the machinery layer whose order books currently extend past 2030 and whose expansion lags demand by about two years. Large-load interconnection rules with real commitment tests, because a 474 GW request book with no deposit discipline and no completion statistics cannot be planned against. Facility-level energy and water disclosure, without which the efficiency and consumption record remains self-reported and unauditable. Credible long-horizon compute demand contracts — take-or-pay instruments that let generators, utilities and substrate makers finance capacity against something firmer than announcements. And diversification of gallium and copper refining, the two mineral concentrations the electricity and packaging layers share.
• Enables Delivered at schedule, the stack is the substrate for every brief on this map that assumes abundant computation — which is most of the artificial-minds category and much of the rest — but the honest typed claim is narrower: it enables frontier training runs to keep scaling on their announced cadence, and it enables the electricity system to acquire, for the first time since electrified rail, an anchor customer willing to finance generation. Failure enables something too: the queues, disclosure rules and manufacturing capacity built for this load outlast any demand disappointment.
• Adjacent Autonomous Supply Chains and Future Ports and Shipping move the stack’s components; Automated Construction Systems and Robotics in Infrastructure court its halls as an early market; Circular Infrastructure Systems inherits its refresh cycles; Geothermal Megaprojects and Planetary-Scale Energy Systems compete to supply it.

Established “AI will consume half the world’s electricity.” The measured record: data centres of all kinds took about 1.5% of world electricity in 2024, heading toward roughly 3% by 2030 on the IEA base case. The claim survives because it borrows local truths — 21% of Ireland’s metered electricity, a projected doubling of Virginia’s demand — and promotes them to planetary scale. The correct statement is that the load is globally modest and locally extreme, and that every real crisis in the record is local.

Frontier “Every chatbot query costs a bottle of water.” Per-query water figures in circulation span orders of magnitude because they divide contested numerators by arbitrary denominators; the measured quantities are facility-level — reported water-usage effectiveness from 0.19 to about 1.8 litres per kilowatt-hour depending on cooling design and accounting boundary, and corporate totals like Google’s 8.1 billion gallons in 2024. The honest criticism of the industry is not a per-query factoid; it is that disclosure is voluntary, boundaries are incompatible, and the indirect water at the power plant is usually omitted — by both sides.

Established “Chips are the bottleneck.” True in 2023; false as stated since. The die gave way to packaging and memory in 2024 — a supply gap near 20% at the packaging layer, a leading vendor spending eighteen months failing thermal qualification — and by 2025–2026 the binding layer was electricity delivery: five-year interconnection medians, turbine order books sold out toward 2030, transformer waits doubled. Anyone still arguing export controls and die supply is arguing about the layer with the most committed capacity behind it.

Frontier “The queue is the demand.” Utilities, regulators and journalists routinely quote interconnection request books — 474 GW in ERCOT alone — as if they were orders. The only measured completion base rate, from the generation side, is 13%; the load side has no statistic at all, and the grid operator publishing the scariest forecast attached its own warning that the number may be inflated by speculative duplicates. Planning as if the queue were real and planning as if it were phantom are both unjustified; the missing artefact is a measurement, and it is measurable.

Established “Permitting reform will fix the timeline.” Permitting is one of two slow clocks, and the July 2025 federal acceleration touches only it. The other clock is machinery: a turbine reservation book extending past 2030 and a transformer queue measured in years are indifferent to environmental-review streamlining. A campus with a categorical exclusion and no transformer is a field.

Established “On-site generation escapes the grid’s politics.” The Memphis-area record is the counterexample: behind-the-meter turbines exchanged the interconnection queue for the Clean Air Act, sixty-day notices and a federal lawsuit, plus fuel logistics and community opposition with sharper moral standing than any rate case. The island strategy buys schedule; it does not buy exit from the state.

Frontier “Data centres barely use water” / “data centres are draining the aquifers.” Both circulate; both misread scale. Statewide, the best audit found under 0.5% of withdrawals; basin-level, a single campus at 550,000 gallons a day sited on a stressed aquifer is a genuine allocation problem. The resolution is boring: water is a siting variable, decided per basin, and the fight should be over disclosure and permit conditions rather than over a single national number.

Speculative “The buildout pays for itself in jobs and taxes.” The fiscal record is genuinely positive where audited — but construction-phase-dominated, and booked against ratepayer and infrastructure costs that arrive on a different ledger and a different timetable. Whether a given locality nets out ahead is an empirical question with a known method and, in almost every jurisdiction except Virginia, no published answer.