Scenarios for 2030

Every chapter before this one decides something about a building you can draw. This one decides which world the building lives in. The same 200 MW campus is a brilliant asset in one 2030 and a stranded liability in another, depending on which of a few distinct futures arrives. The scenario is the largest uncontrolled variable in any AI-infrastructure thesis, and it usually enters as an unexamined assumption.

Three axes frame 2030, and they move semi-independently. The demand-and-capital regime is supercycle, digestion or bubble. Chapter 1.8 covers firm-level economics; here the regime is the sector-wide lens. The architectural shape is centralized mega-campus versus distributed inference, which decides where the megawatts go. The industry structure and geographic re-map decides who owns the capacity and where it lands. The chapter then turns to the power-supply endgame (CFE, SMR, fusion, and the demand-side wildcards that could invalidate the whole demand curve) and to the stranded-asset and systemic risk each scenario hides.

Start with the demand-and-capital regime, because it sets the boundary conditions for everything else. The three scenarios are not moods of optimism or pessimism. They are different beliefs about one mechanism: does AI revenue grow fast enough to service the capital being deployed before the assets depreciate? McKinsey's own framing spans a $3.7T-to-$7.9T range of AI-capable capex to 2030 against a ~$5.2T midline, wide enough that the constrained and accelerated cases describe different industries (McKinsey, 2025). Bain's September 2026 report has the buildout costing $5–6.5T and bringing about 150 GW or more online by 2030. In Bain's view power, chips, skilled labor and permitting all hold it back at the same time. In Q1 2026 alone, local opposition blocked or delayed at least 75 projects worth $130B, by Bain's count (29 September 2026). The two houses measure different scopes, so their ranges sit side by side and do not form one series.

Supercycle is the case where demand keeps outrunning supply. Reasoning and agentic workloads inflate tokens per task, inference compute compounds, and power, not chips or capital, stays the limit through 2030. The scarce asset is an energized megawatt, time-to-power outranks every other lever, and the operators who locked firm power and long-lead gear early win. Digestion is the soft-landing case. The extraordinary base is absorbed and capex growth decelerates sharply after 2026 without a crash. That makes 2026 the peak it falls from: the top four US hyperscalers' capex guidance for the year is up more than 60% over 2025 (→ Chapter 16.4). The operators who win are the disciplined ones that held back at the top of the cycle. Bubble is the case where the revenue gap wins. Spend behaves like a utility build-out while revenue still behaves like software subscriptions, and the mismatch forces write-downs, cancelled announcements and a depreciation reckoning.

Three estimates size the revenue the capex needs, or the shortfall against it, on different bases and horizons. David Cahn’s July 8, 2026 heuristic puts lifetime end-customer revenue at ~$1.5T for one annual capex cohort. Bain sized a ~$800B annual shortfall by 2030 even after AI-driven savings (Bain, 2025). Its September 2026 model, on a different horizon, puts the new revenue still needed in 2031 at ~$4.2T a year to reach a ~$6T annual AI market.

Nobody knows which scenario will arrive, so the decision is which one your balance sheet is structured to survive. A supercycle bet is contracted, levered and over-built on power. A bubble hedge is merchant-light, opex-flexible and short on irreversible commitments. You cannot be optimized for both, and pretending otherwise is how operators get caught.

The second axis is architectural, and it is orthogonal to the demand regime. Even a supercycle spans a workload continuum that pulls siting and design in different directions. A centralized mega-campus is the gigawatt-scale training factory: one tightly coupled supercomputer chasing the cheapest firm power and the coldest climate, indifferent to user proximity (→ Chapter 16.1). Inference splits by workload. Frontier, model-parallel and batch inference fit the same dense liquid-cooled campus as training. Latency- or residency-bound inference fits regional sites, and strict-locality inference fits edge or on-premises capacity.

The inference transition makes this fork decisive. Bain forecasts that by the end of the decade the majority of AI compute is inference, not training, even as frontier training concentrates into 1 GW-plus campuses that become standard (Bain, 2025). The crossover is not distant: Deloitte's November 2025 outlook already forecasts inference at about two-thirds of 2026 AI compute (Chapter 1.3). If you build only for centralized training, you cannot serve every latency or residency need. If you build only a small distributed fleet, you cannot host frontier training or model-parallel inference. Density, redundancy, fabric blocking and siting driver all vary across the continuum, so this is a fork to decide at scoping, not a dial to tune later. Reasoning and test-time compute raise the decode-heavy share but do not by themselves decide whether serving is centralized or distributed (→ Chapter 16.3).

Who owns the 2030 capacity is a third axis. Three operator archetypes are consolidating, each with a different cost of capital and obsolescence exposure. Hyperscalers self-fund the largest share from cash flow and carry the depreciation debate on their own books. The top four US hyperscalers have guided to roughly $735–760B of capex in 2026 (company guidance through the Q2 2026 prints). Amazon raised its guide to ~$220B on 2026-07-30, with CEO Jassy naming higher memory cost, not more megawatts, as a driver. The Oracle-inclusive top-five cluster is tracked at ~$750B+ (CreditSights). Neoclouds are the pure-play GPU landlords: capital-intensive, debt-financed and margin-pressured, and the most exposed to a utilization or residual-value shock because their entire asset is depreciating silicon. Sovereign and enterprise buyers are the fastest-broadening demand source and the one least governed by pure unit economics. They include nation-states pursuing compute independence and regulated industries pursuing data residency. SK Telecom's planned up-to-2 GW Vera Rubin DSX cloud in Korea (letters of intent with NVIDIA, July 2026; first AI factory targeted for 2027) is the APAC instance.

A fourth archetype is emerging in 2026: the energy/launch integrator selling time-to-megawatt. SemiAnalysis reads SpaceX's earnings-call plans as a path to ~10 GW of ground compute by year-end 2027, with Microsoft the logical largest offtaker. That structure is neither hyperscaler nor neocloud. Apply the announced-vs-under-construction test to every such figure before you underwrite it as energized.

The geographic re-map follows firm power and permissive policy. Power-first training campuses migrate to stranded-generation and cold-climate regions, while latency-first inference clusters stay in the metros (→ grid integration in Chapter 15.8). Sovereign demand and export controls redraw the map again: the (later rescinded) US AI-diffusion tiering, and Gulf-state build-outs converting energy wealth into compute. A review of 775 non-US data-center projects found US companies operate about 48% of them by investment value, so residency is not control: a local site does not guarantee sovereignty when a foreign operator runs it (arXiv 2508.00932, 2025). For a 2026 siting decision, the cheapest-power, fastest-permit jurisdiction may carry a geopolitical or export-control tail risk that a spreadsheet does not price. That tail is no longer hypothetical. In March 2026 drone strikes damaged AWS facilities in the UAE and Bahrain, and in September AWS said its Bahrain region cannot be restored (Chapter 11.2). Price physical-security and regional-concentration risk into Gulf and other contested-region capacity, and spread critical workloads across regions, not only availability zones.

Every 2030 scenario eventually hits the same wall: where do the firm, clean megawatts come from, and when? The IEA projects global data-center electricity demand to roughly double from ~485 TWh in 2025 to ~950 TWh by 2030 (~3% of global electricity), with AI-specific load tripling (IEA, 2026). In the US, EPRI's scenarios move data centers from ~4-5% of electricity today toward 9-17% by 2030; LBNL's June 2026 update (192 TWh in 2024, a 649 TWh reference case for 2030) is in Chapter 16.1. The endgame is a ranking of supply options by the variable that actually gates a build: time-to-megawatt, not headline LCOE.

Ranked that way, the clean-and-firm options that look best on a carbon spreadsheet are the slowest to energize. Firm, clean supply available now includes operating nuclear, hydro and geothermal, where resource geography and uncontracted capacity permit. Restarts are project-specific future supply: Constellation targets a 2027 restart of the 835 MW Crane Clean Energy Center, which serves Microsoft. SMRs are real but late. Google–Kairos targets first deployment in 2030 and a fleet by 2035 (Kairos Power, 14 October 2024). Amazon’s Cascade targets operations in the 2030s and the wider X-energy ambition runs to 2039 (Amazon, 16 October 2025). The 2030 dates are widely judged optimistic.

Power-reactor construction permits have started to land. In March 2026 the NRC voted to award TerraPower a construction permit for its Natrium sodium-cooled fast reactor at Kemmerer, Wyoming, which TerraPower expects to complete in 2030 (TerraPower, 4 March 2026). The NRC issued TVA a construction permit for the Clinch River BWRX-300, a light-water SMR, on 28 September 2026, at the end of a 14-month review. TVA has not set a construction start and is still negotiating cost sharing. A construction permit is not an operating licence, and it is not delivered power.

Fusion is an option, not a plan. SPARC is the demonstration machine and ARC the generating plant, and Commonwealth Fusion targets grid electricity from its first ARC plant in Virginia in the early 2030s (CFS, Aug 2026). No one should underwrite a 2030 build on a fusion megawatt. That leaves gas-and-grid as the default bridge that carries the load this decade. It combines behind-the-meter generation (Cleanview tracks ~90 GW announced at 59 US data centers by mid-2026, largely gas turbines and engines, with about 2 GW operating and roughly 1 GW under construction) and host-utility grid service (project-specific tariff, studies, upgrades, agreements and milestones). Fuel emissions stay in the inventory: CFE procurement and any separate offset claim do not erase the gas combustion. → speed-to-power mechanics in Chapter 3.2.

Carry the combined shock through the power portfolio. Use the 100 MW announced / 40 MW financed / 20 MW energized phase example in Chapter 16.1, not another forecast of delivered capacity. Suppose power slips six months while useful-service demand drops to 50% and the assumed residual falls 25%. Then suspend the next fit-out release, re-price equipment commitments, and keep only the power rights whose holding cost clears the downside case. For each grid, gas, nuclear or storage option, record contracted MW, earliest deliverable date, duration, fuel or energy dependency, permits, counterparty, termination payment and carbon boundary. Announced generation is not firm supply, storage needs recharge, and a clean-energy certificate is not a dispatch right. Release the next phase only when the power milestone, funding and accepted customer workload all return together. The case’s 60% useful-service hurdle is one explicit management choice. An SMR or fusion target moving closer on a slide does not meet those gates. → operating flexibility offers in Chapter 15.8; financing consequences in Chapter 2.5.

The entire demand curve rests on assumptions that a single shock could invalidate, and the wildcards cut both ways. On the downside sit three. One is an algorithmic-efficiency step (a DeepSeek-style collapse in training or inference cost) that deflates the compute needed per unit of value. Another is an enterprise-adoption air-pocket: the MIT NANDA study found ~95% of enterprise GenAI pilots produced no measurable P&L impact on $30-40B of spend. The third is a regulatory or energy backlash that caps siting. That one is no longer hypothetical. New York's Executive Order 62 (14 July 2026) holds DEC discretionary permits for new or expanded data centers able to draw 50 MW or more until the final statewide GEIS, with no fixed end date. Texas has told TCEQ to halt data-center permits pending ERCOT's audit (→ Chapter 3.9). On the upside, the same efficiency gains can increase total compute under the Jevons rebound condition: cheaper inference begets more inference. Reasoning and agentic workloads are a structural demand multiplier that could keep the supercycle alive longer than the bears expect (→ Chapter 16.3).

The decision consequence is asymmetric. A pure-play, single-generation, debt-financed asset is fragile to every downside wildcard and benefits from the upside ones only if it is already full. A flexible substrate is the cheapest insurance against the wildcard you cannot predict: over-provisioned floor loading and water, modal procurement, contracted rather than merchant revenue, and a power deal that flexes. What it buys is the right to be wrong about the forecast.

Each scenario hides a different stranded-asset failure mode, and naming them is the point of the exercise. Stranded silicon is the obsolescence risk, and GPU life is disputed: a contested 2–3-year obsolescence bear case vs published 4–6-year useful-life estimates and 5–6-year book policies. If the bear case materializes, a generational density step can leave current-generation fleets uneconomic against newer parts and uncompetitive against grid power before they are depreciated. The step runs from ~132 kW NVL72 racks (HPE’s nominal for GB200 and GB300) to Vera Rubin NVL72, shipping since June 2026, where NVIDIA’s DSX reference design plans 330 kW per cabinet. Kyber-class racks at ~600 kW come next: NVIDIA targets H2 2027 and SemiAnalysis reports 2028 (→ Chapter 16.2). Stranded substations are the power-delivery mirror image: an interconnection slot energized for a campus whose demand never materializes, or a behind-the-meter gas plant carrying carbon liability for a load that digested. Stranded balance sheets are the financial endgame: long-lived debt underwritten against short-lived silicon, and an under-depreciation question that turns reported earnings into borrowed future write-downs (→ Chapter 1.8).

Systemic and societal constraints form the outer boundary on all three scenarios. Grid reliability has become a planning constraint. On 4 May 2026 NERC issued a rare Level 3 alert, after large-load loss events that included Virginia in July 2024. In that event about 1,500 MW of data-center demand left the grid over a six-fault, 82-second sequence, as facilities transferred to backup. Then came a larger one. When a single 230 kV fault near Ashburn cleared normally on 22 July 2026, about 3.8 GW still moved to backup. That is roughly 2.5 times the 2024 scale. The alert does not make its Essential Actions mandatory. One of them asks transmission owners to work with their customers, where possible, so facilities ride through as much as they can. So far, binding ride-through rules are regional. SPP applies large-load requirements to requests from January 2026, and ERCOT’s NOGRR282 covers new large computational loads from August 2026. PJM has proposed its own requirements (September 2026), and NERC’s computational-load standards passed final ballot in October 2026 but still need Board adoption and FERC approval (→ Chapter 4.10).

Water, community social license, and the concentration of load in a few jurisdictions are real ceilings that a demand curve cannot wish away. Virginia alone is projected at 39-57% of state electricity by 2030 (EPRI, Powering Intelligence 2026). The deepest systemic risk is correlation. Every scenario assumes the constraints relax independently, but a power slip, a residual-value shock and a demand air-pocket are correlated. They tend to arrive together in a downturn, which is exactly when a levered, single-generation, merchant-exposed asset has the least room to survive them. An attacker can also induce the failure modes that random faults trigger (→ Chapter 11.10).