The $7 Trillion AI Build

Any analytical framework for the AI capital cycle must begin with the numerical baseline. The headline figure on the cover is sourced and verifiable. Eight separate lines of evidence converge on the same picture: the AI infrastructure build represents the largest concentrated capital deployment in modern industrial history. This section establishes that picture from primary sources, then qualifies the nature of the cycle.

Evidence Ledger · how to read load-bearing claims

This report distinguishes between four classes of factual claim. Every load-bearing assertion in the body falls into one of these categories. The reader can audit any specific claim against the corresponding source class.

The strongest claims in this report are disclosed facts and analyst forecasts. The most useful claims are the framework-level inferences. The most contestable claims are the author estimates that bridge the two. Where the report makes a numeric claim that does not fit one of the first two classes, it is by definition author estimate or inference; the body labels these explicitly. The bibliography at the end of the document maps every numbered reference to its source class.

The 2026 number

Microsoft, Alphabet, Amazon, and Meta will spend roughly $690 to $725 billion on capital expenditures in 2026 based on guidance issued at Q1 2026 earnings, with the high end representing an approximately 77% increase over 2025's record of $410 billion^[1][2]. Add Oracle's roughly $50 billion target and the figure clears $740 billion across the five US cloud and AI infrastructure providers^[3]. The vast majority of that spending is directly tied to AI compute, data centers, and networking infrastructure.

The breakdown by company indicates which firms are deploying capital most aggressively. Amazon held its number at $200 billion. Alphabet raised twice during Q1 reporting, landing at $175 to $190 billion. Microsoft set its calendar 2026 capex at $190 billion, with the company explicitly disclosing that approximately $25 billion of that figure represents component price inflation rather than additional capacity^[1][4]. Meta raised its 2026 guidance at Q1 2026 earnings (April 29, 2026) from $115 to $135 billion to $125 to $145 billion, citing higher component pricing and additional data center costs to support future-year capacity. The Hyperion campus in Richland Parish, Louisiana represents a $27 billion total facility cost destined to scale to 5 GW long-term^[5][6].

The 2030 number

The five-year cumulative figure is where the $7 trillion headline comes from. Three independent research desks have built bottom-up data center capacity models with broadly converging conclusions. McKinsey's base scenario projects $6.7 trillion in cumulative global data center capital expenditure through 2030, including $5.2 trillion AI-specific and $1.5 trillion non-AI^[7]. McKinsey's three AI-only scenarios bracket the AI portion at $3.7 trillion (constrained), $5.2 trillion (base), and $7.9 trillion (accelerated)^[7]. Goldman Sachs forecasts $7.6 trillion in cumulative AI capex globally between 2026 and 2031, with annual spending rising from $765 billion in 2026 to roughly $1.6 trillion by 2031^[5]. Omdia's data center capex forecast lands at approximately $1.6 trillion in annualized investment by 2030^[8].

The composition of that AI-specific spend matters as much as the number. McKinsey's bottom-up model breaks the $5.2 trillion AI-specific figure into roughly $3.1 trillion at the technology developer and chip layer, $1.3 trillion at the energizers (power generation, transmission, distribution, equipment), and $0.8 trillion at the builders (data center construction, mechanical and electrical systems, real estate)^[7][9]. The implied capacity demand is substantial: global data center capacity rises from 82 GW in 2025 to 219 GW in 2030, of which AI workloads account for 156 GW (70% of the total)^[7]. Approximately 124 GW of incremental capacity must therefore come online over five years, a buildout pace without modern precedent.

How this compares to history

Capital cycles of this magnitude have occurred before, but rarely. The three useful analogs are the late-1990s telecommunications buildout, the 1880s American railway expansion, and the 1920s electrification of the United States. The order of magnitude is informative.

The US telecom industry raised approximately $1.6 trillion in equity and an additional $600 billion in debt during the 1990s buildout, of which roughly $500 billion was deployed during the 1998-2000 peak^[11]. Adjusted for inflation, the peak years of telecom capex represented around 1.5% of US GDP. The American railway expansion of the 1880s peaked at approximately 6% of GDP in 1882, the highest concentrated capital cycle in modern US economic history^[12]. The electrification of the United States peaked at roughly 3% of GDP in the early 1920s^[13].

What's actually different this time

Three structural features distinguish the AI capital cycle from each of its historical analogs.

First, the cycle is being financed primarily by mature, cash-flow-generative incumbents rather than by speculative new entrants. The 1990s telecom buildout was financed largely by debt issued by new long-distance carriers (Worldcom, Global Crossing, Qwest) that had no operating history at the scale they were attempting. The 1880s railway expansion involved heavy public subsidies and frequent bondholder defaults. The current AI buildout is being funded predominantly from the operating cash flows and balance sheets of Microsoft, Alphabet, Amazon, and Meta, all of which entered 2026 with combined cash and equivalents of over $420 billion^[14]. The credit profile of the buyers is structurally different from prior cycles.

Second, the financing is starting to shift away from cash flow. The Bank for International Settlements documented in its January 2026 Bulletin No 120 that AI infrastructure investment is moving from being predominantly cash-financed in 2023-2024 to relying increasingly on debt issuance, private credit, and off-balance-sheet vendor financing in 2025-2026^[15]. Moody's reported in February 2026 that hyperscalers had approximately $662 billion in data center lease commitments signed but not yet commenced, an off-balance-sheet figure that exceeds the on-balance-sheet debt of the same companies^[16]. The cycle is starting from a stronger position than telecom did, but it is moving into the same financial structure.

Third, the binding constraint is physical, not capital. Capital availability has not been the limiting factor for any major AI infrastructure project in 2025-2026. Power availability has. Microsoft's Q3 FY2026 earnings call described Azure capacity as remaining capacity-constrained through fiscal 2026, with CFO Amy Hood noting that demand continues to outpace deliverable capacity^[17]. Subsequent industry coverage and CEO Satya Nadella's BG2 podcast commentary placed the binding constraint specifically at power and grid interconnection rather than GPU inventory^[18][29]. Section 03 develops this point in detail. It represents the most consequential structural feature of the current cycle and the one most under-appreciated in retail-facing analysis.

Calibrating the cycle

The supercycle question deserves a calibrated answer rather than a slogan. Capital cycles meeting the historical magnitude threshold (3% of GDP or higher at peak) have run for 8 to 18 years from inflection to mature decline. Railroads ran from roughly 1865 to 1893. Electrification ran from roughly 1908 to the late 1920s. Telecom ran from roughly 1996 to 2002. None of these cycles ended on the schedule predicted at their start. None were timeable in advance.

The defensible position is that the AI capital cycle has all the structural features of a multi-year supercycle and is currently in its early-to-middle phase, having inflected in roughly 2024. Whether the cycle sustains for ten years, fifteen, or breaks at year six depends on three variables mapped in detail across the next eight sections: whether application-layer revenue catches up with infrastructure capex by 2028, whether power infrastructure can be expanded fast enough to clear the demand backlog, and whether algorithmic efficiency gains compress demand for compute faster than capacity expands. Which variable dominates is not knowable in advance, and no published source resolves the question. The defensible operational response is to position now and treat layer-level allocation as more consequential than the timing call.

Most retail-facing analysis of the AI buildout reduces to a single dimension: chips. The argument is that exposure to compute infrastructure subsumes the rest of the cycle. This view is incomplete in three respects. It under-weights power infrastructure as the actual bottleneck. It misreads the capital structure that is making the cycle financeable in the first place. It also leaves builders, traders, and long-horizon investors without an evaluation framework for names that fall outside the chip narrative. This report uses a three-force framework instead.

Why three forces

The largest and most durable industrial capital cycles in history have been governed by the interaction of three forces in roughly the same proportions: the productive technology itself, the energy required to operate it at scale, and the financing structure that bridges the gap between capital deployment and revenue realization. Railways had locomotives, coal, and bond markets. Electrification had generators, fuel, and utility-scale equity. Computing through the 1990s had silicon, electricity, and venture capital. Each of these cycles can be analyzed by tracking the bottleneck rotation between the three forces over time.

The AI capital cycle has the same structure. Compute is the productive technology. Power is the energy required to operate it. Capital is the financing structure that connects the two and absorbs the timing mismatch between $7 trillion of infrastructure spending and the eventual revenue from AI applications. Every name in our atlas, every risk in Section 07, and every rotation call in Section 06 maps cleanly to one or more of these three.

Compute

Compute is the most-discussed force and the most-priced. NVIDIA's data center revenue reached $193.7 billion in fiscal 2026, up 68% year over year, with Q4 alone delivering $62.3 billion (up 75% YoY)^[19]. AMD and Meta announced a multi-year strategic compute agreement in February 2026 covering up to 6 GW of deployed capacity and AMD performance-based warrants for up to 160 million shares; Reuters reported the value at up to $60 billion over five years, with shipments of custom MI450-architecture accelerators commencing 2H 2026^[20]. Marvell's FY26 revenue reached $8.2 billion (up 42% YoY) with a $1.5 billion run rate in custom silicon, and NVIDIA invested $2 billion in Marvell as part of the NVLink Fusion partnership^[21]. Micron's HBM capacity is sold out through 2026 with HBM4 entering volume production one quarter ahead of schedule^[22].

The compute story has three sub-currents that matter for positioning. Custom silicon (Google TPU, Amazon Trainium, Microsoft Maia, Meta MTIA) is taking accelerator share from merchant GPUs, with NVIDIA's inference share projected to fall from approximately 90% in 2025 to 20-30% by 2028 according to Counterpoint Research^[23]. Photonics and optical networking are hitting their inflection point: NVIDIA's $4 billion combined investment in Lumentum and Coherent on March 2, 2026 signaled that optical interconnect is now strategic^[24][25]. And the foundry layer has consolidated to TSMC for advanced nodes, making Taiwan a single-point-of-failure risk that we treat in Section 07.

Power

Power is the most-bottlenecked and least-priced force in the cycle. The IEA's Energy and AI report (April 2025) projects global data center electricity consumption rising from 415 TWh in 2024 to 945 TWh in 2030, slightly more than Japan's total current electricity consumption^[18]. The United States accounts for the largest share of that increase, with US data centers projected to consume more electricity for processing data than for the production of aluminum, steel, cement, and chemicals combined by 2030^[18].

The constraint is not the quantity of available power generation. It is the speed at which that generation can be brought online and connected to data center sites. Power transformer lead times have stretched into the multi-year range; Hitachi Energy and other major manufacturers have publicly cited 80-130 week lead times for high-voltage transformers, with secondary industry coverage placing some grades at 128 weeks^[17]. Grid interconnection queues in major US data center markets stretch five to seven years. The Department of Energy's SPARK program is targeted at reconductoring upgrades on existing transmission, not aggregate transmission expansion, and even at $1.9 billion will not materially relieve the multi-year backlog on its own. The result is a structural power-capacity backlog that is forcing hyperscalers into long-duration power purchase agreements at prices significantly above merchant rates: Vistra's combined PPAs with AWS at Comanche Peak (1,200 MW) and with Meta at PJM nuclear facilities (~2,600 MW) total approximately 3,800 MW under contract^[26].

The investing implication is that power generation, transmission, and distribution names are systematically under-priced relative to compute names. The trading implication is that the catalysts for power names (PPA announcements, regulatory approvals, capacity additions) are predictable and operate on monthly-to-quarterly time horizons rather than the annual time horizons of chip cycles. The building implication is that power constraints determine inference economics: where you can put compute is now constrained by where you can put electrons.

Capital

Capital is the connective tissue of the cycle and the source of most of the systemic risk. Three structural features matter.

The hyperscaler balance sheet is funding most of the build. Microsoft, Alphabet, Amazon, and Meta entered 2026 with combined cash and equivalents over $420 billion^[14]. The big four operating cash flows ran at roughly $480 billion combined in 2025. This is the strongest financing position any capital cycle in modern history has started from. It is also being depleted faster than free cash flow can replenish it: Pivotal Research projects Alphabet's free cash flow to fall almost 90% in 2026, from $73.3 billion in 2025 to $8.2 billion^[14]. Barclays estimates Microsoft's free cash flow will decline 28% in 2026 before recovering in 2027.

Vendor financing is taking an increasing share. NVIDIA invested $100 billion in OpenAI as part of the September 2025 multi-year compute partnership. NVIDIA invested $2 billion in CoreWeave as Class A common stock in Q1 2026^[27]. NVIDIA invested $2 billion each in Lumentum and Coherent in March 2026^[24][25]. NVIDIA holds warrants in IREN exercisable into 30 million shares at $70 strike, plus a $3.4 billion AI Cloud commitment^[28]. AMD issued Meta a performance-based warrant for up to 160 million AMD shares (approximately 10% of AMD's diluted equity if fully exercised) as part of the multi-year strategic compute agreement; Reuters reported the implied contract value at up to $60 billion over five years^[20]. The vendor at the top of the supply chain is increasingly taking equity stakes in customers and partners further down it. We discuss the systemic risk implications in Section 07.

Off-balance-sheet leverage is rising. The $662 billion in signed-but-uncommenced data center lease commitments documented by Moody's exceed the on-balance-sheet debt of the four hyperscalers combined^[16]. The accounting rule that delays disclosure of these commitments until lease commencement makes the true debt-equivalent obligation of the cycle larger than headline balance sheets indicate. The BIS bulletin specifically flagged this as the primary structural concern for the financial stability implications of the AI buildout^[15].

How the three forces interact

The three forces are not independent. They are a triangle, and the binding constraint rotates between them over the course of any capital cycle.

In 2022-2024 the binding constraint was compute supply: not enough HBM, not enough advanced-node foundry capacity, not enough rack-scale systems to absorb the demand. The names that won during that window were the chip-tier suppliers (NVIDIA, SK Hynix, TSMC, ASML).

In 2025-2027 the binding constraint is power: chip supply has caught up at the high end, but the GW of electricity and the transformers required to turn those chips into operating data centers cannot be built fast enough. The names that win during this window are the power tier (Vistra, Constellation, Talen, GE Vernova, Vertiv, Eaton). This is the rotation we are currently inside.

In 2027-2029 we expect the binding constraint to rotate to capital, specifically the question of whether application-layer revenue is materializing fast enough to support the cumulative debt and lease obligations of the previous three years. This rotation determines whether the cycle continues smoothly into a long maturity phase (the application-layer compounders, the cybersecurity layer, the agentic SaaS players win) or whether it triggers a meaningful drawdown in infrastructure names. We will lay out both scenarios in Section 07.

The framework is straightforward to apply. For any AI-related name, ask: which force does this company sit on? What is the binding constraint of that force right now? Where in the rotation cycle are we? The answers are not always immediately obvious, but the framework is consistent. It is the same lens we apply across the six layers in Section 04.

If the report has a single thesis statement, it is this section. The argument is structural rather than cyclical. The AI buildout cannot scale at the pace its sponsors are guiding because the power infrastructure required to support it cannot be built at that pace. This creates a multi-year structural premium for any company that owns, generates, transmits, distributes, or supplies equipment to the power grid in regions where data center demand is concentrated. The implication for investors and traders is direct: power-tier exposure is materially under-weighted in most retail AI portfolios.

The hard data on the constraint

Microsoft's Q3 FY2026 earnings release (April 29, 2026) reported a commercial Remaining Performance Obligation (RPO) of $627 billion and reaffirmed that Azure capacity will remain demand-constrained through fiscal 2026, with CFO Amy Hood describing capacity as "short for many quarters running" in prepared remarks^[17]. CEO Satya Nadella subsequently clarified the binding constraint in his BG2 podcast appearance: GPUs were sitting in inventory because Microsoft "could not find electricity to power them," with grid interconnection and transformer availability cited as the structural bottleneck^[29]. The earnings disclosure documents the demand-supply gap; the podcast commentary identifies the binding mechanism.

Microsoft's commercial Remaining Performance Obligation (RPO), a standard accounting disclosure of contracted but undelivered cloud and AI services, reached $627 billion at Q3 FY2026, nearly doubling year over year^[30]. Microsoft management noted on the Q3 earnings call that this 99% YoY growth includes the renegotiated OpenAI commercial commitments; excluding OpenAI, RPO grew approximately 26% YoY, in line with historical enterprise cloud seasonality^[17]. The ex-OpenAI growth rate is the cleaner enterprise demand signal. The 99% figure is significantly influenced by a single ecosystem partner. Both numbers matter: the headline confirms the absolute scale of contracted commitments, and the ex-OpenAI baseline confirms that broad-based enterprise AI demand is also growing meaningfully. Secondary analyst estimates of the portion specifically attributable to power-constrained Azure capacity cluster around $80 billion^[3][31]. Whether the precise figure is $80 billion or somewhat above or below, the structural conclusion is clear: Microsoft holds more contracted demand than it can deliver, and the constraint is power availability, not compute supply.

Google Cloud disclosed at the Q1 2026 Alphabet earnings call that backlog reached $462 billion, more than doubling sequentially. Alphabet management noted that just over 50% of the backlog is expected to convert to revenue within 24 months, with conversion timing constrained primarily by data center build pace^[32]. The conversion-rate disclosure is critical: it confirms that the gating factor on revenue recognition is delivery, not demand.

Why power is hard to scale

Three sub-constraints compound. Each operates on a different time horizon, and all three must be resolved for capacity to come online.

Generation buildout takes years, not months. A new combined-cycle natural gas plant takes 24-36 months from groundbreaking to commercial operation. A new nuclear unit, even an SMR, takes 5-7 years from regulatory filing to first criticality, with current SMR commercial deployment timelines slipping to 2027-2030^[33]. Renewable buildout is faster (12-18 months for utility-scale solar) but is intermittent and requires storage to serve 24/7 data center loads. The aggregate result is that even with capital available and demand obvious, generation capacity expansion is governed by physical, regulatory, and supply chain timelines that cannot be compressed below 18 months at the absolute minimum.

Transformers are the chokepoint inside the chokepoint. Industry coverage has placed lead times for high-voltage power transformers in a range of 80 to 130 weeks, with secondary reporting citing 128 weeks (about 2.5 years) at the upper end^[17]. Distribution transformers required for behind-the-fence data center substations have similar lead times. The global transformer manufacturing industry has approximately a dozen major manufacturers, and capacity expansion for transformers itself takes 3-5 years from investment decision to first delivery. Without transformers, generation capacity exists but cannot be connected to load. The transformer bottleneck is the actual binding constraint within the broader power constraint, and it is the single highest-leverage piece of the supply chain to track.

Grid interconnection takes longer than generation. Lawrence Berkeley National Laboratory data on US grid interconnection queues shows that 2024 interconnection requests faced average wait times of 5-7 years before clearance, with a clearance rate of approximately 15-20% (most projects in the queue never get built)^[34]. The Department of Energy in March 2026 launched SPARK (Speed to Power through Accelerated Reconductoring and other Key Advanced Transmission Technology Upgrades), a $1.9 billion funding opportunity targeting reconductoring of existing transmission lines and deployment of advanced transmission technologies such as dynamic line rating^[35]. SPARK is the third tranche of the broader Grid Resilience and Innovation Partnerships (GRIP) program. The program is meaningful at the project level (eligible reconductoring projects must demonstrate a 50% capacity gain on the upgraded line) but the aggregate $1.9 billion represents incremental capacity, not a structural expansion of the US grid. In specific data center hubs (Northern Virginia, Texas, certain parts of the Pacific Northwest), interconnection moratoria have been imposed as utilities triage demand.

Where the demand is going

The aggregate demand picture from the IEA is the foundation. Global data center electricity consumption is projected to more than double, from 415 TWh in 2024 to 945 TWh in 2030, accounting for just under 3% of global electricity consumption by that date^[18]. The growth concentrates: the United States and China account for approximately 80% of the projected global growth, with the US alone adding up to 240 TWh of incremental demand (a 130% increase from 2024 levels)^[36]. Data centers are forecast to drive nearly half of US electricity demand growth between 2024 and 2030^[18].

Where the supply is coming from

The IEA's base-case forecast for how the additional ~530 TWh of incremental data center demand gets supplied is informative both for what is included and what is excluded. Renewables (solar plus wind plus hydro plus storage) cover approximately 50% of the additional demand by 2030, requiring more than a further 450 TWh of incremental renewable generation tied to data center loads^[18]. Natural gas covers a further 175 TWh, with most of that growth concentrated in the United States behind-the-meter or via dedicated PPAs^[18]. Nuclear contributes a smaller absolute increment, primarily through the restart of mothballed units (Crane Clean Energy Center, formerly Three Mile Island Unit 1, etc.) and limited SMR deployment.

The economic implication of the supply mix is that gas turbine manufacturers (GE Vernova), nuclear operators with restart-able assets (Constellation, Vistra, Talen), grid equipment manufacturers (Eaton, Vertiv), and traditional regulated utilities in data center hubs all sit at the intersection of two trends: the secular AI demand and the structural inability of the grid to expand at the pace required. Each of these names trades at single-digit to mid-teens P/E multiples appropriate to defensive utilities, despite carrying contracted cash flows that tie them to the largest growth story in the global economy.

The most validated power deals to date

The five power agreements summarized below are public, primary-source-documented, and structural. They establish the pattern that we expect to repeat across dozens of similar deals through 2027.

The pattern across these five deals is consistent: 15-20 year fixed-price agreements with investment-grade counterparties, capacity dedicated to specific data center load, and prices materially above merchant rates in exchange for delivery certainty. The hyperscalers are paying a premium for power that they would not have paid for compute. This is what binding constraints look like in financial form.

The trade in plainest terms

If power binds the cycle, the names that sit closest to the binding constraint capture disproportionate value. Three categories matter most.

Power generation operators with re-rateable assets. Vistra, Constellation, and Talen each entered the AI cycle with large nuclear and gas portfolios trading at utility multiples appropriate to a slow-growth regulated industry. Each is now signing 15-20 year PPAs at prices significantly above merchant rates. Each will see contracted cash flows expand 30-50% by 2030 if their announced deals execute. The multiple expansion that follows the cash flow expansion is the trade.

Grid and power equipment manufacturers with order books expanding. GE Vernova (gas turbines, transmission), Eaton (electrical distribution), Vertiv (data center power and cooling) each have backlogs growing 50-100% YoY across their AI infrastructure exposed segments. The order book itself is the leading indicator. Watch quarterly disclosures of book-to-bill ratios.

Specialist nuclear and SMR developers. Oklo, NuScale (SMR), and Centrus (uranium fuel) are higher-risk, longer-duration plays on the same thesis. The investment case rests on commercial deployment of SMRs by 2027-2030, which is plausible but not yet proven. Size accordingly.

We expand each of these through the Section 04 layer-by-layer treatment. The point of this section is to establish the structural argument: power binds, the constraint is multi-year, and the names sitting at the constraint trade at multiples that have not yet caught up with the demand picture.

Sovereign capital extends the demand curve

The McKinsey, Goldman, and Omdia forecasts referenced in Section 01 model commercial hyperscaler and enterprise demand. They under-count a category that is now material: sovereign-funded AI infrastructure. National wealth funds and government-aligned entities are deploying capital outside the commercial-return logic that bounds the hyperscaler models. This adds a second demand layer to the cycle that does not appear in the base-case forecasts and that shifts the supply-demand picture further toward the structural undersupply scenario.

Stargate UAE. Announced May 2025, partnered between G42, OpenAI, Oracle, NVIDIA, Cisco, and SoftBank within a 5 GW UAE-US AI campus in Abu Dhabi. The first 200 MW phase is scheduled for completion in Q3 2026, with the full 1 GW Stargate UAE cluster targeted within three years. The US authorized export of up to approximately 35,000 NVIDIA Blackwell GB300 systems to G42 in early 2026, materially expanding sovereign access to frontier hardware^[51].

Saudi HUMAIN and PIF programs. Saudi Arabia's Public Investment Fund committed approximately $40 billion to AI infrastructure beginning in 2024, with HUMAIN positioned as the sovereign AI data center platform. HUMAIN awarded MIS a $501 million contract in February 2026 to design and build a private AI data center. A 250 MW AirTrunk-HUMAIN co-developed campus in Riyadh is in tender. Qualcomm-HUMAIN agreements cover advanced AI data center deployment plus edge inference. Total Saudi commitments are difficult to triangulate precisely but exceed $40 billion in announced or contracted capital across data center construction, GPU procurement, and software partnerships^[52].

European AI Continent Action Plan. The European Union targets €200 billion in AI investment through 2030 across compute infrastructure, model development, and applications. The AI Gigafactory program is the largest single component, funding multi-GW data center construction in member states. Less liquid in capital deployment terms than UAE or Saudi, but adds a third sovereign demand layer that is contractually committed and politically supported.

India and Japan. India's national AI infrastructure programme targets 10,000+ GPUs deployed via the IndiaAI Mission. Japan has multi-billion-dollar sovereign AI commitments through METI. Korea is a smaller but growing buyer.

Combining the announced sovereign programs, conservatively, adds roughly $80-150 billion of incremental AI infrastructure capex through 2030 that does not appear in the McKinsey base case. This is not enough to change the order of magnitude of the cycle, but it does shift the supply-demand balance toward the constrained-supply scenario, particularly for GPUs, HBM, and advanced packaging where the global supply chain is shared between commercial and sovereign buyers.

The procurement bottleneck stack

Power transformers are the most discussed supply-side bottleneck, but they are not alone. The full procurement stack contains at least nine constrained inputs, each with its own lead time, supplier concentration, and capacity-expansion timeline. The trade in each constraint is different.

The investment implication is that not every name in the AI infrastructure stack benefits equally. The names with the longest lead times and most concentrated suppliers (HBM, advanced packaging, transformers) capture the most pricing power. The names with shorter lead times and more competitive supply (fiber, cooling) capture less. Position sizing should track procurement leverage, not just AI exposure.

Regional power grid asymmetry

"Power binds the cycle" is a national-level statement. The reality on the ground is that different regional grids face different binding constraints, with materially different implications for which Layer 2 names benefit. The five major US grid regions break down as follows:

PJM (Mid-Atlantic, Northern Virginia, Ohio Valley). The most demand-constrained region in the United States. Northern Virginia's "Data Center Alley" has interconnection moratoria in some sub-areas. Capacity prices in the 2025-26 PJM auction cleared at record highs. Vistra (~2,600 MW Meta PPA), Talen (1,920 MW Amazon), and Constellation (Crane restart) all have direct PJM exposure. PJM also has the highest concentration of contracted nuclear-restart upside.

ERCOT (Texas). The largest absolute load growth, and the most flexible regulatory environment. West Texas Permian Basin gas plants (Vistra Permian expansion, GE Vernova turbines) supply behind-the-meter loads with limited interconnection bottleneck. Comanche Peak (1,200 MW VST/AWS) is the major nuclear PPA. ERCOT's biggest constraint is gas turbine availability; capacity is buildable but lead times stretch into 2027-2028.

MISO (Midwest). Mixed picture. Significant new data center activity in Iowa, Illinois, Wisconsin. Less concentrated than PJM. Wind-heavy supply mix creates intermittency challenges that hyperscalers solve via natural gas backup. GE Vernova and Vertiv equipment exposure.

SPP (Plains). Smaller absolute scale, but growing. Wyoming and Oklahoma data center campuses are emerging behind-the-meter solutions to PJM/ERCOT congestion. Less direct equity exposure for public Layer 2 names.

WECC (West). California is supply-constrained and politically resistant to new gas. Pacific Northwest hydro is fully allocated. Arizona is the active growth corridor, with NV Energy, APS, and SRP increasing data center service offerings. Limited public-equity Layer 2 exposure beyond GE Vernova and the IPP names.

The implication is that VST, CEG, and TLN have asymmetric concentration in PJM, the most-binding region. Their multi-year cash flow durability is therefore tied to the most structural single regional constraint in the US grid system. This is good for the long thesis and bad for the diversification argument. Layer 2 exposure is regionally concentrated whether the position-sizing recognizes that explicitly or not.

Behind-the-meter vs front-of-the-meter

Two power-delivery structures dominate the new wave of AI data center development, and they have very different implications for which Layer 2 names benefit and which utility regulators have jurisdiction.

Front-of-the-meter (FOTM). Power flows through the public grid via standard utility interconnection. The customer pays retail rates plus transmission charges. Hyperscaler PPAs with utilities or with merchant generators that wheel through the grid (Vistra-Meta PJM nuclear, Talen-Amazon Susquehanna restructured) are FOTM. Subject to FERC and state PUC oversight. Slower to deploy because of interconnection queues, but legally durable.

Behind-the-meter (BTM). Power is generated on-site or co-located with the data center, typically natural gas plants built specifically to serve a single hyperscaler customer. The Hyperion campus in Louisiana (Meta-Entergy) is partially BTM. West Texas gas plant developments serving Stargate sites are BTM. Subject to less regulatory oversight, faster to deploy, but exposes the customer to fuel price risk and emissions liability.

The thesis implication: if interconnection delays force more capacity BTM, the gas turbine OEMs (GE Vernova specifically) capture upside while the contracted-utility nuclear thesis (CEG specifically) sees its scarcity premium erode. The Risk 6 inverse case in Section 07 explores this dynamic in detail. For now, the deployed mix is roughly 70% FOTM, 30% BTM in the announced 2026-2028 pipeline; that ratio is plausibly the variable that determines whether the structural Layer 2 multiple expansion holds.

This is the analytical core of the report. We map the AI build-out into six functional layers, evaluate each layer's economics, name the dominant players, and assess each through three lenses simultaneously. The investing lens addresses long-horizon capital allocation. The trading lens addresses position sizing, catalysts, and rotation logic. The builder lens addresses what the layer means for AI product economics. The same analytical framework applies to all six layers, which makes them directly comparable.

Layer 1 · Foundation: Compute Silicon

The foundation layer is the silicon that does the actual computation. NVIDIA dominates merchant GPU. AMD is the credible alternative. Custom ASICs from Broadcom-designed parts (Google TPU, Meta MTIA, Microsoft Maia, OpenAI Titan) are taking inference share. Memory (HBM) is supplied by SK Hynix, Samsung, and Micron. Foundry capacity is consolidated at TSMC. ASML holds the EUV lithography monopoly that gates every leading-edge node and is therefore an unavoidable bottleneck name; the High-NA EUV ramp through 2026-2028 is the supply-side constraint nobody can route around. Photonics is split between Lumentum and Coherent.

Investing view: NVIDIA at fiscal 2026 revenue of $215.9 billion, data center segment $193.7 billion (up 68%), remains the highest-growth megacap in the public markets^[19]. The bear case is custom ASIC erosion: Counterpoint Research projects NVIDIA's inference share falling from approximately 90% in 2025 to 20-30% by 2028^[23]. The bull case is that the total accelerator market grows faster than NVIDIA's share erodes. Both arguments can hold simultaneously. AMD's Meta agreement covers up to 6 GW of deployed capacity and AMD performance-based warrants for up to 160 million shares per company filings; Reuters reported the implied contract value at up to $60 billion over five years^[20]. AMD therefore has a structural growth path independent of NVIDIA. Broadcom and Marvell are positioned as primary beneficiaries of the custom-ASIC trend.

The NVIDIA quantum platform extension is an under-priced structural moat. NVIDIA's Ising launch (April 14, 2026) combined with CUDA-Q (2022) and NVQLink (2025) positions NVIDIA as the control plane for hybrid quantum-classical computing. The strategy is identical to the CUDA playbook applied to a new compute paradigm before the ecosystem consolidates. Every quantum hardware company that adopts the stack creates lock-in independent of which qubit modality eventually wins. Twelve major quantum institutions (IonQ, Atom Computing, Academia Sinica, EeroQ, Conductor Quantum, Fermilab, Harvard, Infleqtion, IQM, Lawrence Berkeley National Lab, Q-CTRL, UK National Physical Laboratory) signed on at launch. The market does not price NVIDIA's quantum platform position because near-term revenue contribution is small; if quantum commercialization continues on the accelerated 2026-2030 timeline and NVIDIA's control-plane position holds, the platform lock-in becomes another decade-scale structural moat. Section 08 develops this thesis in detail. Intel's silicon spin-qubit program is the other Layer 1 name with quantum optionality, leveraging existing semiconductor fab infrastructure as a long-term scaling advantage. The market currently prices INTC on classical-CPU turnaround context; spin-qubit re-rating is unpriced optionality.

Trading view: Layer 1 has the highest signal-to-noise ratio of any tier. NVDA has weekly options with sub-$0.10 spreads at the core strikes. AMD, AVGO, and MRVL all have deep options chains. The relevant catalysts are well-defined: quarterly earnings, hyperscaler capex guides, and custom ASIC ship-rate disclosures. The principal tail risk is a DeepSeek-class shock; the January 27, 2025 event removed $589 billion from NVIDIA's market capitalization in a single session, the largest single-day loss in US market history^[39]. Position sizing should account for the possibility of a similar event.

Building view: Foundation-layer choice is increasingly multi-vendor. Open-weight models (Llama 4, Qwen 3.5, Mistral Large 3) running on AMD MI300/MI450 hardware deliver 60-80% of frontier-model quality at 30-50% of the cost of NVIDIA reference deployments. For builders running inference at scale, multi-silicon support is now a baseline architectural requirement.

Layer 2 · Power & Grid

Covered in detail in Section 03. Layer summary: the most under-priced tier of the supercycle relative to contracted cash flows. Vistra, Constellation, and Talen are the three publicly traded nuclear-and-gas operators with material AI exposure. GE Vernova is the dominant gas turbine and transmission supplier. Eaton and Vertiv supply electrical distribution and data center power systems. Oklo, NuScale, and Cameco round out the SMR and uranium plays.

Building view for Layer 2: Power constraints determine where you can host inference. The cheapest tokens in 2026-2028 will not be in Northern Virginia or Silicon Valley but in regions with newly built behind-the-meter gas plants (West Texas, Louisiana, parts of Oklahoma and Wyoming). For builders running large-scale inference, pricing geography matters as much as model choice.

Layer 3 · Infrastructure: Data Centers, Neoclouds, REITs

The infrastructure layer is the physical hosting tier between silicon and software. Three sub-segments: data center REITs (Equinix, Digital Realty), neoclouds (CoreWeave, Nebius, IREN), and dedicated colocation. CoreWeave's Q1 2026 disclosure crystallizes the segment economics: revenue backlog $99.4 billion (up nearly 50% sequentially), Q1 revenue $2.078 billion (up 112% YoY), GAAP net loss $740 million, total debt $24.859 billion plus roughly $10 billion in lease liabilities, quarterly interest expense $536 million^[27].

Investing view: Two distinct opportunities. The REITs (EQIX, DLR) trade at utility-style multiples appropriate to their dividend profiles, but the AI demand picture is asymmetric upside that the multiple has not yet captured. The neoclouds (CRWV, NBIS, IREN) are pure-play AI infrastructure with leverage. CoreWeave's 10 customers each committed to spending at least $1 billion^[27]; Nebius is signing similar large customer commitments at higher gross margin; IREN secured $3.4B AI Cloud commitment plus $2.1B NVIDIA warrant for 30M shares at $70 strike^[28].

Trading view: CoreWeave moves on backlog and debt-issuance news. The next quarterly RPO disclosure is the highest-information-content event in the entire AI complex. EQIX and DLR are slower, lower-volatility names appropriate to credit-spread structures. The neoclouds carry meaningful binary risk: any major customer cancellation or financing tightening hits the equity hard.

Building view: Multi-cloud + neocloud architecture is now standard. CoreWeave for training (highest performance, NVIDIA reference racks). Nebius for inference (better cost-per-token). IREN for raw GPU rental at sub-hyperscaler prices. Hyperscaler reserved capacity for guaranteed availability and credit packages. The economics have flipped: a year ago, hyperscalers paid neoclouds; now neoclouds pay hyperscalers for power and customers pay neoclouds for compute, with all parties in the chain margin-stacking on the supply gap.

Layer 4 · Networks: Optical Interconnect, Switching, ASICs

The networking layer connects accelerators inside data centers and between them. The economics matter more than most retail analysis recognizes: rack-scale AI systems require terabytes per second of bandwidth, not gigabytes; this is achievable only through optical interconnect at the chip-to-chip level. NVIDIA's $4 billion combined investment in Lumentum and Coherent on March 2, 2026 (with multi-billion dollar purchase commitments and capacity rights) signaled that optical interconnect is now strategic, not commodity^[24][25].

Investing view: Lumentum (LITE) shares closed nearly 12% higher on the NVIDIA investment day; Coherent (COHR) jumped 15%. Both companies' equity has approximately quadrupled in 12 months, but the structural story is durable: co-packaged optics (CPO) is the network architecture for the next generation of AI systems, and Lumentum and Coherent are the two largest US-based suppliers. Arista Networks (ANET) dominates Ethernet networking switches at hyperscale. Marvell (MRVL) supplies the custom networking silicon and active electrical cable interconnect that ties racks together; NVIDIA's $2 billion investment in Marvell as part of NVLink Fusion confirmed the strategic importance.

Trading view: Higher beta than Layer 1 due to smaller market caps. LITE and COHR move on photonics-specific catalysts (qualification milestones, capacity announcements). ANET has cleaner technicals and trades closer to large-cap chip names.

Building view: Network choice typically follows silicon choice. NVIDIA-based clusters use NVLink + InfiniBand. AMD-based and custom-ASIC clusters increasingly use Ethernet at 800G or 1.6T speeds (Arista, Broadcom). For builders, network architecture choice cascades downstream into latency profiles for distributed training and multi-region inference.

Layer 5 · Application Layer

This is where the AI cycle eventually monetizes, and where the largest gap currently exists between capex and revenue. The 2025 State of AI survey published by McKinsey found that only 39% of firms attribute any EBIT impact to AI deployment, with most reporting less than 5% impact^[40]. MIT's Project NANDA, in its GenAI Divide study, concluded that 95% of generative AI pilots fail to deliver measurable P&L impact, with only 5% achieving meaningful financial returns^[41]. Two large independent studies pointing at the same picture: the application layer is the gating factor on whether the supercycle's revenue catches up with its capex by 2028.

Investing view: The application layer is split between incumbent enterprise software (PLTR, NOW, CRM, SNOW, MDB, ADBE) and AI-native cybersecurity (CRWD, PANW, S, ZS). The incumbent software names are re-bundling AI capabilities into existing distribution channels (Microsoft Copilot through Office, Salesforce Einstein, ServiceNow Now Assist) and capturing premium pricing. Palantir reported FY2025 revenue of $4.475 billion, up 56% YoY, with Q4 2025 alone up 70% YoY; US commercial revenue grew 109% YoY full-year (Q4 alone +137% YoY) and US government grew 55% YoY full-year, demonstrating the pure-play AI-application opportunity. The cybersecurity tier is leveraging the AI buildout to position AI defense as requiring AI-native tooling: CrowdStrike's Q4 FY2026 ending ARR reached $5.25 billion, up 24% YoY, with $1.01 billion of net new ARR for the year (the first year exceeding $1 billion).

Trading view: Application-layer earnings inflections are the highest-information-content trades in the next 12-24 months. If AI-attributable revenue at PLTR, NOW, CRM, SNOW exceeds 10% of revenue by 2026 year-end, the entire tier re-rates upward. If it stalls, the multiple compression is severe. Watch quarterly disclosures of "AI revenue" or "Copilot revenue" line items.

Building view: The application layer is where most AI builders should compete, not at the model layer (commodity) or the infrastructure layer (impossible to compete at scale). The 5% of AI pilots that succeed (per MIT NANDA) capture outsized share of the productivity gains. Common patterns in successful deployments: deep workflow integration, line-manager-driven adoption (not centralized AI labs), MCP-based tool integration, and willingness to redesign processes around AI rather than bolt AI onto existing processes.

Layer 6 · Frontier: Quantum, Rare Earths, Robotics, Defense, Space

The frontier layer covers the speculative or longer-duration adjacencies to the AI core: quantum computing (IonQ, Rigetti, D-Wave), rare earth processing (MP Materials, USA Rare Earth), humanoid robotics (Tesla Optimus, Figure AI, Apptronik), defense AI (Anduril, Palantir Defense, Kratos), and space launch and communications (Rocket Lab, AST SpaceMobile). These tiers are real but trade with much higher beta and longer time horizons than Layers 1-5.

Quantum: Quantum computing has moved from "decade-out science project" to "category with first commercial revenues and a credible 2026 inflection point" in the last twelve months. IBM publicly committed to 2026 as the year quantum first outperforms classical computing on specific problem classes. IonQ Q1 2026 revenue of $64.7 million (up 755% YoY) with FY26 guidance raised to $260-270M and Remaining Performance Obligations of $470M (+554% YoY) makes IonQ the first public quantum company past meaningful commercial scale^[40]. NVIDIA launched Ising on April 14, 2026, the first family of open-source AI models for quantum, integrating with CUDA-Q and NVQLink. The architectural consensus is hybrid quantum-classical: quantum as coprocessor inside classical AI infrastructure, not as a separate computing paradigm. Quantum does not draw from the same power demand pool as AI; cooling-dominated quantum energy budgets are roughly three orders of magnitude below AI workload demand at equivalent commercial scale. Position quantum names as portfolio exposure rather than single-name bets; see Section 08 for a full treatment of the modality landscape, public proxies, and the strategic positioning of NVIDIA's platform play.

Rare earths and magnetics: China controls approximately 60% of mined production of magnet rare earths (per IEA 2026 critical minerals data), 91% of global refining and processing capacity, and 94% of permanent magnet manufacturing^[43]. As of late 2025, China imposed strict export controls on seven rare earth elements (scandium, yttrium, samarium, gadolinium, terbium, dysprosium, lutetium), with new dual-use restrictions targeting Japan in early 2026. MP Materials (Q1 2026 NdPr production +63% YoY to 917 MT, $90.6M consolidated revenue +49% YoY, $72.2M Materials segment revenue / $36.7M segment EBITDA, $21.1M Magnetics segment revenue / $9.6M segment EBITDA) and USA Rare Earth (announced a non-binding CHIPS Program LOI in early 2026 covering proposed federal support of up to $1.6B, including up to $277M direct funding and up to $1.3B in a senior secured loan, plus a separately closed $1.5B PIPE in January 2026) are the two principal US-listed plays^[44][45].

Humanoid robotics and Tesla Optimus: Tesla's Q4 2025 and Q1 2026 disclosures confirmed that Model S and Model X production at Fremont ends in early May 2026, with the production lines being converted to Optimus humanoid robots; first-generation Optimus production targeting 1 million units per year is scheduled to begin in summer 2026 (Q2-Q3 startup, full ramp 2027)^[46]. Tesla is also constructing a 5.2 million sq ft Optimus factory expansion at Gigafactory Texas, with a long-term target of 10 million units per year^[47]. The Optimus thesis has substantive technical and operational backing, though the commercial monetization path remains uncertain; position size should reflect this uncertainty.

Defense AI and space: Anduril (private) is the dominant defense-AI primary contractor. Palantir (PLTR) generates more revenue from defense than from commercial. Among space names, Rocket Lab (RKLB) Neutron timeline slipped further into 2026; AST SpaceMobile (ASTS) lost BlueBird 7 on Blue Origin New Glenn NG-3 on April 19, 2026, when the upper stage BE-3U engine underperformed and the satellite reached an unsustainable orbit; FAA grounded New Glenn pending investigation^[48][49]. AST is falling back on SpaceX Falcon 9; year-end 2026 target of 45 satellites in orbit at risk per William Blair commentary.

Building view for Layer 6: Most builders should not be building in Layer 6. The exception is defense and physical-AI builders, where the application surface is starting to materialize (drone defense, autonomous logistics, embodied agents). Quantum, rare earths, and space are infrastructure plays for institutional capital, not opportunities for individual AI builders.

The thesis-map sections of this report identify Layer 2 as the highest-conviction tier of the supercycle. Identifying a tier is not the same as proving the names within it are mispriced. The Layer 2 trade has been widely discussed since 2024, and the easy alpha from being early to "power binds" is gone. The question this section addresses is whether the public-equity Layer 2 names are still mispriced relative to their contracted-cash-flow trajectory, after the market has already moved. The short answer is: partially. Two of the six core names look fully priced. Three look partially priced with asymmetric upside if specific catalysts hit. One looks structurally mispriced. The differentiation matters for position sizing.

The valuation framework

The right framework for evaluating Layer 2 names is not a static multiple comparison. These are utility-adjacent businesses transitioning from regulated-growth utility multiples to AI-infrastructure-adjacent growth multiples, and the appropriate forward multiple depends on which framing the market eventually adopts. The framework used here is straightforward: estimate contracted incremental EBITDA from announced PPAs and AI-infrastructure deals through 2030, apply a defensible forward multiple range, compare to current market capitalization. The result indicates whether the contracted growth is already priced or whether asymmetric upside remains.

This is a thesis-level valuation framework, not a position-grade DCF. It surfaces the bigger picture: which names have most of their AI-tailwind cash flows already in the multiple and which still have unfilled valuation gap. Position-grade work requires per-name DCF modeling, debt-adjusted enterprise value, contract-by-contract pricing analysis, and regulatory risk assessment. That work is a research starting point for an active trader, not a substitute for it.

Vistra (VST): partially priced, regional concentration risk

Current position: ~$160-166 stock price as of early May 2026, ~$50-52B market cap. FY2025 adjusted EBITDA of $5.91B, EV/EBITDA of approximately 12.8x as of Q4 2025. 52-week range $134-220. Median analyst price target $228, range $97-323. The 52-week range is itself diagnostic; this is a name that has been volatile around the AI infrastructure narrative.

Contracted growth picture: AWS Comanche Peak PPA (1,200 MW, 20 years, ramping to full 2032); Meta PJM nuclear PPAs (~2,600 MW, 20 years, executed Q1 2026); Lotus Infrastructure 2,600 MW gas portfolio acquired 2025; pending Cogentrix acquisition (5,500 MW gas portfolio, announced Q1 2026). Combined nuclear PPA portfolio totals ~3,800 MW under 20-year contracts at premium-to-merchant pricing.

Implied incremental EBITDA: Conservatively, the AWS + Meta nuclear PPAs add $1.0-1.5B of incremental EBITDA at full ramp by 2032, depending on contract pricing assumptions. The Cogentrix gas portfolio adds another $700M-1.2B EBITDA depending on capacity factor and merchant exposure. Combined, FY2030E EBITDA could reach $8-9B versus FY25 actual of $5.9B, a 40-50% increase over 4-5 years.

Valuation read: At the median analyst $228 target, VST trades at approximately 9-10x FY2030E EBITDA. That is not screaming cheap, but it is reasonable for a utility transitioning into AI-infrastructure growth multiples. The asymmetric upside is in the multiple compression toward 12-14x EV/EBITDA if the market adopts AI-infrastructure framing, which would imply $250-300/share. The downside is regional concentration: roughly 60-70% of VST's AI-related EBITDA contribution is PJM-exposed, and PJM regulatory or political action against capacity prices is the single biggest risk.

Verdict: Partially priced. Hold or add on PJM-related dips. Do not chase strength above $200 without a specific catalyst.

Constellation Energy (CEG): fully priced, watch the Calpine integration

Current position: ~$300-311 stock price as of early May 2026, $107-110B market cap. Trailing P/E approximately 40x, revenue $25.5B TTM. Median analyst target $367-386, range $272-481. CEG closed the Calpine acquisition in January 2026, materially expanding the gas-fleet portfolio.

Contracted growth picture: Microsoft Crane Clean Energy Center (formerly TMI Unit 1) restart, 835 MW dedicated, 20-year PPA. Calpine acquisition adds ~26,000 MW of mostly gas capacity, of which a meaningful fraction will become available for AI infrastructure PPAs. Constellation has the largest US nuclear fleet (~22 GW) and is positioned as the dominant nuclear-restart and SMR-deployment counterparty.

Implied incremental EBITDA: The Microsoft Crane PPA adds approximately $400-600M EBITDA at full ramp (mid-2028+). Calpine integration is more complex; gross EBITDA contribution of $2-3B annually, but offset by financing and integration costs. Net incremental EBITDA likely $1-1.5B by 2028.

Valuation read: At ~40x P/E, CEG is priced as if substantial AI-infrastructure growth is already in the multiple. The bull case requires either much faster than expected commercial SMR deployment or additional large hyperscaler PPAs at premium pricing. The bear case is multiple compression as the Calpine integration creates short-term margin pressure. Median analyst target implies 25-30% upside, but the path to that target is gated on execution that has not yet been demonstrated.

Verdict: Fully priced at current levels. Hold existing positions. New entry below $280 only.

Talen Energy (TLN): the most-mispriced PJM nuclear pure-play

Current position: ~$369 stock price as of late April 2026, ~$16.5-17.4B market cap. 52-week range $158-451 (high volatility). Q1 2026 saw a strong profit recovery and updated outlook.

Contracted growth picture: Susquehanna nuclear expansion ISA with Amazon (1,920 MW through 2042). The deal was restructured in 2025 from co-located BTM to PJM grid integration with revised wholesale pricing. The Susquehanna asset is among the most under-monetized large nuclear plants in the US, and the AWS deal demonstrates the structural premium available for firm baseload power dedicated to AI loads.

Implied incremental EBITDA: The AWS Susquehanna deal at full economics adds roughly $1.0-1.4B annual EBITDA versus prior merchant economics. Total Talen EBITDA could double from FY24 levels by 2028, depending on PPA pricing and capacity factors. This is a smaller absolute-dollar story than VST or CEG but a much larger percentage growth story.

Valuation read: TLN trades at approximately 10-12x forward EBITDA on consensus estimates, which reflects the AWS deal but does not yet fully price the multi-year tail of similar nuclear-AI premiums. The pure-play nature (single major asset, single major customer, single structural thesis) creates concentrated upside if the thesis plays out and concentrated downside if the AWS contract is renegotiated unfavorably.

Verdict: The most-mispriced of the public Layer 2 nuclear names on a pure-play basis. Highest beta to the structural thesis. Position size accordingly given the single-asset concentration.

GE Vernova (GEV): structurally rich, gas turbine cycle exposure

Current position: ~$1,040 stock price, $279.5B market cap. Trailing P/E 30.4x, forward P/E 37.2x. EV/EBITDA over 100x on trailing (the recent profitability inflection makes trailing metrics misleading). Revenue $39.4B TTM.

Contracted growth picture: Gas turbine backlog expanding rapidly (50-100% YoY in AI-exposed segments). Transmission equipment book-to-bill above 1.5x. The core thesis is that gas turbines and transmission equipment are necessary components of any data center buildout, and the OEM position is structurally advantaged. GE Vernova has the largest gas turbine market share globally and a duopoly position in transmission equipment with Siemens Energy.

Implied incremental EBITDA: The order book is the leading indicator. Backlog growth implies materially higher revenue and EBITDA through 2027-2028. But execution risk is real: gas turbine manufacturing has historical margin volatility, and the AI-infrastructure backlog is a relatively new segment without a long execution track record.

Valuation read: The market has fully priced the gas turbine cycle thesis. The forward P/E of 37x is rich for an industrial. The bull case requires the gas turbine backlog to translate into sustained 20%+ EBITDA growth through 2028, which is plausible but not yet proven. The bear case is execution miss + cycle peak by 2028 leaving a multiple-compression event in 2029-2030.

Verdict: Structurally rich. Market has anticipated the AI-infrastructure thesis. New entry only on 15-20% pullbacks. Existing holders should consider trimming above $1,200.

Vertiv (VRT): backlog inflection, multi-year visibility

Current position: Q4 2025 backlog of $15.0B (+109% YoY per company filings), FY2026 organic sales growth guidance of 29-31% per Q1 2026 earnings, FY26 revenue range $13.5-14B.

Contracted growth picture: Vertiv supplies data center power systems (UPS, switchgear, busbar), liquid cooling (CDUs, manifolds, in-row cooling), and thermal management. Every hyperscaler AI campus and every Stargate-class deployment requires multiple Vertiv-supplied components. The backlog growth indicates that Vertiv is capturing share in the highest-growth segment of data center capex.

Implied incremental EBITDA: 29-31% organic growth at FY26 mid-teens EBITDA margins implies ~$2.0-2.3B FY26 EBITDA versus ~$1.3-1.5B in FY24. The backlog visibility extends growth into 2027-2028.

Valuation read: Vertiv has been one of the better-performing AI-infrastructure adjacencies. The forward multiple has expanded but the backlog visibility supports the multiple. The asymmetric risk is gross margin compression as competition intensifies (Schneider Electric in particular). The asymmetric upside is liquid cooling becoming the dominant thermal solution at scale, which would expand Vertiv's content per data center materially.

Verdict: Fairly priced at current levels with backlog-supported growth visibility. Reasonable core position. New entry on 10-15% pullbacks.

Eaton (ETN): the lowest-beta Layer 2 exposure

Current position: ~$401 stock price, $156B market cap. Eaton is the most diversified of the Layer 2 names, with electrical distribution, automotive, aerospace, and industrial businesses.

Contracted growth picture: The Electrical Americas segment is the AI-infrastructure pure-play within Eaton. Backlog expanded materially in 2024-2025. Eaton announced transformer plant capacity expansion and is building dedicated US electrical distribution capacity for data center customers.

Valuation read: The diversification dilutes the AI thesis. Roughly 30-40% of Eaton's earnings are AI-infrastructure exposed. The other 60-70% is industrial cycle exposure. This makes Eaton a lower-beta Layer 2 trade with less upside in a strong cycle but more downside protection in a sector drawdown.

Verdict: Fairly priced. Best for investors who want Layer 2 exposure without the volatility of pure-plays. Not the asymmetric trade.

Putting it together

The Layer 2 pure-plays (VST, CEG, TLN) capture the asymmetric upside of the structural thesis but have specific name-by-name risks (VST regional concentration, CEG fully priced, TLN single-asset concentration). The equipment names (GEV, VRT, ETN) capture more diversified upside but trade closer to fair value, with GEV especially looking rich. The market has done meaningful work pricing the Layer 2 thesis since 2024.

Where structural mispricing remains: TLN as the cleanest PJM nuclear pure-play, VST on PJM-related dips, VRT as the highest-conviction equipment name with backlog visibility. Where the trade is largely complete: CEG above $300, GEV at current multiples.

The takeaway is the framework, not the picks. Position sizing in Layer 2 should reflect that the thesis is still structurally correct (power binds, multi-year duration) but that the easy multiple-expansion alpha is gone, and the next round of returns has to come from cash flow execution rather than narrative re-rating.

Capital cycles do not run linearly. Different layers of the stack peak at different times, and the trade that worked in the previous phase often does not work in the next. The historical analogs are instructive. The chip-equivalent of the 1990s telecom buildout (long-haul fiber laid by Worldcom, Global Crossing, et al) peaked in 1999-2000 and lost 90% of its value over the next two years. The application-layer equivalent (Google, Amazon, eBay) peaked five to fifteen years later and compounded for the next two decades. Same cycle. Different layers. Different timing. Different outcomes.

The rotation framework

We map the AI cycle into three rotation phases, each lasting roughly two to three years. The phases overlap; this is not a clean handoff. But the relative weight of each layer in a sensible portfolio shifts measurably across the phases.

How each layer's catalysts time

The forecast below maps each layer's expected peak intensity. We are not forecasting precise drawdown timing; we are mapping where contracted growth, primary catalyst density, and multiple expansion are concentrated. Position weighting should track these phases on a rolling basis.

Foundation (Layer 1) peaks 2026-2027. NVIDIA's data center growth rate decelerates from +68% YoY in fiscal 2026 to a forecast +30-40% in fiscal 2027 as the merchant GPU base case rises. Custom ASIC ship rates inflect in 2027 (Google TPU v8, AWS Trainium 3, Meta MTIA Gen 2). The accelerator market continues to expand but NVIDIA's share compresses. The trade evolves from "buy NVIDIA" to "buy the broader silicon basket" (NVDA, AMD, AVGO, MRVL, MU, LITE, COHR, TSM).

Power and grid (Layer 2) inflects 2026-2028. The contracted PPA pipeline is largely set through 2028. The multiple expansion has not happened yet; this is the largest gap between contracted growth and trading multiple in the entire AI complex. The catalysts are well-defined: each major hyperscaler will sign 1-3 additional PPAs annually through 2028, and each announcement materially re-rates the supplier. Watch quarterly disclosure cycles: Vistra, Constellation, and Talen typically announce PPAs at quarterly earnings, with high catalyst density.

Infrastructure (Layer 3) tracks Phase B. Neocloud RPO and revenue growth are tied directly to data center buildout and customer concentration. CoreWeave's growth rate decelerates as the customer base saturates with hyperscaler-equivalent counterparties; Nebius and IREN have steeper growth curves but smaller base. Data center REITs (EQIX, DLR) are slower but durable, with multiple expansion expected as Phase C application-layer revenue catches up.

Application (Layer 5) inflects 2026-2028. This is the most uncertain timing. If application-layer AI revenue grows from current levels (low single-digit percent of enterprise software revenue) to mid-teens by 2028, the entire tier re-rates. PLTR, NOW, CRM, SNOW, MDB, ADBE all have meaningful AI-revenue line items now. Watch quarterly disclosures of "AI revenue" or equivalent. The acceleration in the cybersecurity tier (CRWD, PANW, S, ZS) is already underway with CRWD's $5.25B ending ARR and Palantir's 109% FY2025 US commercial growth signaling the inflection has begun; the broader enterprise software acceleration follows by 12-18 months.

Frontier (Layer 6) is event-driven. Quantum, rare earths, robotics, defense, and space each have idiosyncratic catalysts that don't track the broader cycle. Tesla Optimus production launch in late July or August 2026 is a binary event. AST SpaceMobile's recovery from the New Glenn loss depends on Falcon 9 manifest availability. Rocket Lab Neutron's commercial debut. IonQ's revenue trajectory. These are individual-name trades, not basket trades.

Sequencing the trade

For an investor building exposure in mid-2026, the practical sequencing is: maintain Layer 1 core position (NVDA, AVGO, MU, TSM as core) but reduce the rate of additions; add to Layer 2 power tier (VST, CEG, GEV, VRT) as the highest-conviction addition through 2027; size Layer 3 neocloud exposure according to individual risk tolerance; build Layer 5 application exposure gradually through 2026-2027 to position ahead of the inflection; maintain Layer 6 at small individual position sizes.

For traders running active positions, the rotation logic is: when hyperscaler capex guides increase at quarterly earnings, add to Layers 1 and 2; when capex guides plateau or decline, trim Layer 1 (chip multiples compress) and increase Layer 5 (the monetization narrative replaces the capex narrative); when an efficiency shock occurs (DeepSeek-class), trim Layer 1 substantially, preserve Layer 2 (the power constraint remains unchanged), and build Layer 5; when a major application-layer earnings beat occurs, the entire monetization-phase basket re-rates simultaneously.

What we expect to see in the next 12 months

Three calls we are willing to make explicitly. We expect to see them validated or invalidated in real time, which makes them tradeable.

One: at least one major application-layer software company (most likely Palantir, ServiceNow, or Microsoft commercial Copilot revenue) will report greater than 15% of revenue attributable to AI in the next four quarters. This will trigger a multiple expansion across the application layer. The setup is consistent with the McKinsey State of AI 2025 finding that 6% of organizations are already capturing greater than 5% EBIT impact from AI^[40]; these are the customers buying these products.

Two: at least one hyperscaler will announce a 12-month or longer delay to a major data center commissioning due to power constraints. This is the visible event that drives the Layer 2 multiple expansion. The disclosure is most likely from Microsoft (most explicit about constraints) or Meta (largest single project at Hyperion).

Three: at least one custom-silicon ramp will surprise to the upside, with Google TPU v8 or AWS Trainium 3 disclosed shipment volumes exceeding consensus by 20%+. This compresses NVDA inference multiples. The setup is in place: hyperscaler custom silicon ship rates are the single most-watched datapoint in the chip-tier coverage and the disclosure cadence is quarterly.

A responsible research piece on a multi-year capital cycle requires a candid assessment of what could break it. The risks identified below are not tail concerns. Each is documented, sourced, and actively monitored by major institutional research desks. The thesis is directionally correct. The execution risks are material. Position sizing, defined stop levels, and diversification across layers represent the appropriate response framework. The intent of this section is to surface these risks ahead of any drawdown that might force them into focus.

1. Circular financing

The single most-discussed structural risk in the AI complex is the closed-loop financing pattern between NVIDIA, OpenAI, Oracle, CoreWeave, and the hyperscalers. The pattern: NVIDIA invests in OpenAI ($100 billion announced commitment as part of the September 2025 multi-year compute partnership). OpenAI commits $300 billion to Oracle for cloud capacity (validated by OpenAI press release September 23, 2025)^[50]. Oracle uses those commitments to backstop debt issuance for data center construction. CoreWeave buys NVIDIA GPUs financed by NVIDIA-backed debt facilities. NVIDIA takes equity in CoreWeave ($2 billion Class A common stock, closed Q1 2026)^[27], warrants in IREN ($2.1 billion over five years for 30 million shares at $70 strike)^[28], and similar arrangements with other neoclouds. The same dollars circulate through multiple companies' revenue lines.

The Bank for International Settlements bulletin in January 2026 specifically flagged this structure as the primary concern for AI infrastructure financial stability^[15]. The structure works as long as ultimate end-customer demand grows fast enough to support the cumulative obligations being layered on top of it. If end-customer demand decelerates, the entire chain unwinds simultaneously. This is the most plausible mechanism for a major AI infrastructure drawdown.

2. The capex-revenue gap

Hyperscalers and neoclouds are spending $690 to $725 billion in 2026 against AI revenue (defined narrowly as Microsoft AI segment + Google AI segment + AWS AI/inference + Meta AI infrastructure-attributable revenue) of perhaps $80 to $120 billion. The math does not work indefinitely. Either application-layer revenue accelerates dramatically through 2028, or the multiples on infrastructure names compress hard.

Bulls argue this is a feature, not a bug: the same gap existed during the early years of cloud computing and the early years of mobile. Revenue catches up. Bears argue the gap is unsustainable and the inflection point at which capex must be backed by revenue arrives sooner than the bulls expect. The data supporting either view is partial. McKinsey's State of AI 2025 found that 39% of firms now attribute any level of EBIT impact to AI deployment, with most reporting less than 5% impact^[40]. The MIT NANDA study found 95% of generative AI pilots fail to deliver measurable P&L impact^[41]. Neither datapoint resolves the gap question definitively.

3. Algorithmic efficiency shocks

The January 27, 2025 DeepSeek R1 release triggered the market shock: NVIDIA lost $589 billion of market cap in a single session, the largest single-day loss in US stock-market history. R1 was a reasoning model distilled from DeepSeek V3, the underlying base model released in December 2024 that demonstrated frontier-quality models could be trained at a fraction of the cost Western labs assumed. McKinsey's analysis attributes approximately 18x training efficiency and 36x inference efficiency to V3 versus GPT-4o-class economics, driven by V3's MoE architecture activating 37 billion parameters per token from a 671 billion total parameter base^[39]. Most of the price action that day was reversed within weeks but the structural concern persisted: if algorithmic efficiency improves faster than infrastructure expansion, demand for compute compresses.

The Jevons Paradox holds that efficiency gains expand total demand rather than contract it. In the long run this likely applies (cheaper compute drives more compute consumption). In the short run, a major efficiency shock can compress multiples on infrastructure names within days. The ongoing risk is that another DeepSeek-equivalent release could emerge at any time. The relevant indicators to monitor are quarterly model-quality benchmarks (HELM, MMLU, ARC-AGI) and the open-source release cadence from Chinese frontier labs.

4. Geopolitical: Taiwan and rare earths

Two single points of failure dominate the AI supply chain: Taiwan-based TSMC for advanced chip manufacturing, and China-based rare earth processing for magnetics. Either becoming inaccessible would cause severe supply disruption. We do not think either is likely in the near term, but the magnitude of impact deserves explicit mention.

TSMC manufactures essentially all leading-edge AI accelerators (NVIDIA H100, B200, Vera Rubin; AMD MI300, MI450; AWS Trainium; Google TPU; Apple Silicon). A Taiwan supply disruption (military, natural disaster, regulatory) would compress global AI infrastructure capacity expansion to roughly zero for 18-36 months while production migrated to TSMC Arizona, Samsung, and other foundries. The TSMC Arizona facility is operational but its capacity is a small fraction of the Taiwan production base.

Rare earth processing concentration in China is structural. China controls approximately 91% of global rare earth refining and 94% of permanent magnet manufacturing^[43]. The seven-element export ban announced in late 2025 (scandium, yttrium, samarium, gadolinium, terbium, dysprosium, lutetium), with new dual-use restrictions targeting Japan in early 2026, demonstrated the leverage. US response (DoD MP Materials investment, USA Rare Earth's non-binding $1.6B CHIPS Program LOI plus $1.5B closed PIPE, Apple-MP $500M partnership) is decisive but multi-year. Investors with significant AI infrastructure exposure should size positions with this tail risk in mind.

5. The free cash flow inversion

Pivotal Research projects Alphabet's free cash flow to fall almost 90% in 2026 to $8.2 billion from $73.3 billion in 2025^[14]. Barclays estimates Microsoft's free cash flow will decline 28% in 2026 before recovering in 2027^[14]. The four big hyperscalers had $420 billion combined cash and equivalents at the end of Q1 2026, but that buffer is being drawn down faster than operating cash flow can replenish it. This is a one-to-two-year stress, not a multi-year one, but it has implications for capital allocation: dividends and buybacks at the big four will be re-prioritized below capex through 2027 at minimum.

The credit market is starting to price this. Hyperscaler bond spreads widened modestly in Q1 2026 as the market began pricing in higher leverage. Moody's flagged the $662 billion in off-balance-sheet lease commitments as the principal concern^[16]. None of this is yet a credit event, but it establishes the boundary conditions: if anyone of (capex acceleration, revenue deceleration, credit market tightening) breaks adversely, the financial stability of the cycle compresses fast.

6. The inverse case: power stops binding

The bull thesis depends on power remaining the binding constraint through Phase B. The counter-thesis is to identify what breaks if it does not. Three mechanisms could relieve the power constraint faster than the base case assumes, and each would compress the Layer 2 multiple expansion that anchors a meaningful share of the report's recommended exposure.

Grid scales faster than the queue suggests. The 5-7 year average interconnection wait is an average across legacy queue management. If FERC accelerates queue reform, transmission operators implement dynamic line rating at scale, and reconductoring under the SPARK program plus state-level grid investments deliver capacity 2-3 years ahead of the consensus schedule, the contracted-PPA premium compresses. Vistra, Constellation, and Talen would still earn their contracted cash flows, but the multiple expansion thesis that assumes structural scarcity becomes a re-rating in the wrong direction.

Distributed inference reduces centralized power demand. Section 08 makes the case that local inference on consumer-grade hardware is now viable for a meaningful fraction of workloads. If that share grows faster than the application layer monetizes, the marginal data center build slows. The compute requirement persists but is distributed across edge devices and behind-the-meter end-user installations, rather than concentrated in centralized AI campuses requiring 5 GW PPAs. Model efficiency improvements on the DeepSeek V3 trajectory accelerate this dynamic.

Behind-the-meter gas displaces utility-scale procurement. If hyperscalers respond to interconnection delays by building dedicated behind-the-meter combined-cycle gas plants on-site (the West Texas, Louisiana, Oklahoma trajectory in Section 08), the structural premium for utility-grade nuclear PPAs erodes. The gas turbine OEMs (GE Vernova specifically) capture upside; the contracted nuclear operators see their pricing power normalize toward merchant-plus rather than the structural premium pricing currently embedded in some 2027-2028 forecasts.

None of these scenarios represents our base case. None, however, is implausible enough to discount. The position-sizing implication is that Layer 2 exposure should be concentrated in names whose contracted cash flows remain durable regardless of which scenario plays out (the 15-20 year nuclear PPAs already signed) rather than names whose multiples assume the constraint persists indefinitely. The asymmetric risk argues for owning the contracted growth without paying for the scarcity premium.

7. Application-layer revenue stalls

The most plausible cycle-breaking outcome is that the application layer simply does not monetize fast enough. If the MIT NANDA finding (95% of pilots fail) extends through 2027 without meaningful improvement, the gap between $7 trillion of cumulative infrastructure capex and the application-layer revenue required to support it widens to a level that cannot be closed.

The early evidence is mixed. Cybersecurity AI (CRWD, PANW) is growing fast; horizontal enterprise software AI (NOW Now Assist, CRM Einstein, MSFT Copilot) is in early commercial ramp; vertical AI applications (legal, healthcare, finance) are scaling slowly. The trajectory matters. Watch the application-layer revenue disclosures every quarter through 2027.

8. Depreciation and utilization risk

The capex headlines obscure a critical accounting reality: the $7 trillion build is composed of assets with very different economic lives. GPUs depreciate over 3-5 years (NVIDIA's H100/H200 generations are already showing performance gaps versus B200/Vera Rubin generations). Power equipment depreciates over 15-25 years. Data center shells depreciate over 25-40 years. Land does not depreciate. The blended useful-life assumption embedded in current hyperscaler capex guidance is being aggressively tested.

Microsoft, Alphabet, Amazon, and Meta have all extended useful-life assumptions on data center assets at various points since 2022 (typically from 4 years to 6 years for servers, from 30 to 40 years for buildings). This is accounting tailwind that supports current GAAP earnings but does not change the underlying economic reality. If GPU obsolescence accelerates faster than the extended depreciation schedule assumes, the gap eventually shows up in capacity write-downs.

Utilization is the second hidden variable. Not every deployed GPU is producing tokens. Microsoft management acknowledged on the BG2 podcast that GPUs were sitting idle in inventory because of grid interconnection delays. Industry analysts estimate that 15-25% of deployed AI accelerator capacity in 2025-2026 is operating below 60% utilization due to power, network, or workload-allocation friction. That is a material drag on the apparent return on the $7T capex picture.

The investment implication is that headline capex growth may overstate productive capacity growth by 20-30% during the buildout phase. The names that benefit are those exposed to capex itself (NVDA, AVGO, the Layer 1 silicon names) rather than those exposed to monetization (Layer 5 application names). The names most at risk are neoclouds with debt-financed GPU inventories that need full utilization to service the financing structure.

9. Sensitivity to the headline forecast

The $7 trillion cumulative figure is best understood as a midpoint within a wide scenario range, not a confirmed forecast. McKinsey explicitly publishes constrained ($3.7T), base ($5.2T), and accelerated ($7.9T) scenarios for AI-specific capex through 2030. The factors that move the cycle between scenarios are well-defined; the timing of any individual factor is not.

The honest reading is that the cycle is plausibly anywhere from $3.7T to $7.9T cumulative through 2030. The midpoint is the right anchor for thesis construction; the wings are the right anchors for risk management. Any position sized as if $7.9T is certain is over-sized.

10. China beyond rare earths

China is treated in this report primarily as a rare-earth supply concentration risk. The fuller picture matters and is more complex. Three additional dimensions of Chinese AI infrastructure activity are material to the global thesis.

Huawei Ascend and the parallel accelerator stack. Huawei's Ascend 910 series (910B, 910C) is the leading Chinese alternative to NVIDIA H100/H200. Reported performance is roughly 70-80% of equivalent NVIDIA on training, with significant software-stack maturity gaps closing through 2026. Huawei is positioned to absorb most of the displaced Chinese demand from US export controls. Domestic Chinese hyperscalers (Alibaba, Tencent, Baidu, ByteDance) are already deploying Ascend at scale.

SMIC and domestic foundry progress. SMIC achieved 7nm production in 2023 and is reportedly progressing toward 5nm capability through 2026. The technology gap to TSMC remains 2-3 nodes, but it is closing. The implication for the global thesis is that a Taiwan supply disruption (Risk 4) does not necessarily compress AI capex globally to zero; it compresses Western AI capex while potentially accelerating Chinese AI capex on a domestic-only supply chain.

Chinese hyperscaler capex. Alibaba, Tencent, Baidu, and ByteDance combined capex is meaningful and undercounted in McKinsey/Goldman/Omdia models that focus on US hyperscalers. Alibaba's capex doubled in fiscal 2024-2025. ByteDance is reported to be among the largest single buyers of NVIDIA GPUs globally despite export restrictions, with demand routed through Singapore and Malaysia distributors. Conservatively, Chinese hyperscaler AI capex is $80-120 billion annually by 2026, a meaningful fraction of the $725B Big Four US figure.

The investment implication is that the global AI infrastructure cycle is not as US-concentrated as the Layer 1-4 narrative suggests. Demand exists for accelerators, networking, and data center equipment from a parallel Chinese ecosystem that is not directly accessible to US public-equity investors but does affect the demand-supply balance for components shared between the two ecosystems (HBM, advanced packaging, optical components).

11. Hyperscaler bond market signal

Risk 5 introduced the free cash flow inversion. The credit market is starting to price the implications. Hyperscaler bond spreads widened modestly through Q1 2026, with Microsoft 30-year notes trading at approximately 95 bps over Treasuries (up from 75 bps in early 2025) and similar pattern at Alphabet, Amazon, and Meta. The widening is small in absolute terms but directionally consistent with the leverage trajectory.

Moody's flagged the $662 billion in off-balance-sheet operating lease commitments at the four major hyperscalers as the principal forward concern^[16]. These leases are not currently on-balance-sheet under GAAP but represent real future cash obligations that compete with shareholder returns. Moody's has not downgraded any of the four hyperscalers, but the analytical commentary represents a credit-market shift from "AI capex is funded by free cash flow" to "AI capex is increasingly funded by debt and lease obligations."

The signal to monitor is whether spreads continue widening through 2026-2027. Spread compression back to early-2025 levels would indicate the credit market views the capex as self-funding. Continued widening (or a downgrade) would indicate the cycle is moving toward a financing-constrained phase, which materially compresses the multiple expansion thesis on Layer 2 and Layer 3 names that depend on hyperscaler customer credit quality.

12. The quantum substitution question

A sophisticated reader will ask whether quantum computing represents a substitution risk to the AI capex thesis. The framing is: if quantum commercializes on the accelerated 2026-2028 timeline that IBM, NVIDIA, and IonQ are now indicating, does it compress demand for classical AI compute fast enough to break the capex cycle the report describes? The answer is no, and Section 08 develops the reasoning in detail, but the framework belongs in the risk section because the question will be asked.

Three reasons quantum does not threaten the AI capex thesis. First, quantum's energy footprint is approximately three orders of magnitude below AI workload demand at any plausible commercialization timeline through 2030. A quantum data center campus does not appear in the IEA 945 TWh 2030 forecast at any material scale. Second, the architectural consensus is hybrid quantum-classical: quantum processors sit as coprocessors INSIDE existing AI data center infrastructure, accessed through the same NVIDIA CUDA-Q stack as GPUs. This adds marginal demand for Layer 3 (data center capacity) and Layer 4 (interconnect) rather than displacing it. Third, where quantum actually applies for AI workloads (hyperparameter tuning, architecture search, attention mechanism optimization) is a narrow slice of the training pipeline; the bulk of LLM pretraining remains a classical GPU workload through 2030 and beyond.

The asymmetric framing: quantum is more likely to extend the AI capex cycle by adding QPU racks to existing campuses than to compress it by substituting AI workloads. The risk to manage is hype-cycle compression in the pure-play quantum names (Section 08), not substitution risk to the underlying AI infrastructure thesis.

Any institutional research piece on the AI capital cycle in 2026 has to address quantum computing directly. The technology has shifted in the last twelve months from "decade-out science project" to "category with first commercial revenues and a credible 2026 inflection point." The question for this report is whether quantum changes the AI infrastructure thesis. The short answer is no, but the longer answer is more interesting: quantum reinforces the thesis in several specific ways, complicates it in one, and creates a separate investment category that overlaps with but does not displace the AI capex cycle.

The 2026 inflection

Three developments in the last twelve months collectively constitute a regime change in how quantum computing should be treated in an AI infrastructure thesis. None of them existed in a credible form when the prior consensus timeline of 2029-2030 commercial advantage was set.

IBM publicly committed to 2026 as the year quantum first outperforms classical computing on specific problem classes. IBM's positioning is not a marketing claim from a quantum startup. It is an explicit corporate commitment from one of the dominant players, supported by their Nighthawk processor roadmap and the surface-code error correction breakthroughs of 2024-2025. The framing has moved from "we are working toward quantum advantage" to "we are using current-generation hardware on real use cases including drug development, materials discovery, financial optimization, and logistics"^[53].

Google Willow demonstrated a 13,000x speedup over Frontier (the leading classical supercomputer) using 65 qubits for physics simulations in October 2025. This was a scientific-advantage demonstration rather than commercial advantage, but it was the most credible single result since the 2019 supremacy claim. The Willow result depended on error correction breakthroughs that materially extended quantum coherence times. Multiple independent teams (IBM, Google, IonQ, Quantinuum, Pasqal, QuEra) hit complementary milestones in the same window, and the convergence is what made 2026 an inflection year rather than another announcement cycle^[54].

NVIDIA Ising launched April 14, 2026 as the first family of open AI models purpose-built for quantum computing. Ising integrates with CUDA-Q (the NVIDIA programming model for hybrid quantum-classical workloads, introduced 2022) and NVQLink (the low-latency hardware interconnect between GPUs and quantum processing units, introduced 2025). Twelve named launch adopters span the field including IonQ, Atom Computing, Academia Sinica, EeroQ, Conductor Quantum, Fermilab, Harvard, Infleqtion, IQM, Lawrence Berkeley National Lab, Q-CTRL, and the UK National Physical Laboratory^[42]. The market reaction was immediate: IonQ traded up 21%, Rigetti up 13.3%, D-Wave up 22.6% on the announcement day. KB Securities analyst commentary captured the institutional shift: "The industry had expected commercialization in 2029-2030, but technological advances are accelerating that timeline."

The power consumption story is the opposite of AI's

This is the part that matters most for the report's structural thesis, and it is counterintuitive enough that most readers will not have processed it correctly. Quantum computers consume vastly less power than AI workloads at equivalent commercial scale.

A typical dilution refrigerator for a superconducting quantum system (the modality used by IBM, Google, Rigetti) consumes up to 25 kW per system at full operation. PASQAL's neutral atom processor (a competing modality used by Pasqal, QuEra, Atom Computing) draws only 2.6 kW per processor and is structurally power-independent of qubit count, with half the power going to lasers and the rest to electronics and environmental control. Even the largest research-scale quantum installations consume kilowatts to low megawatts in total^[55].

Compare these figures to AI workloads. Training a GPT-4-class model burns through 51,000-62,000 megawatt-hours over a 90-100 day training cycle. A single hyperscaler AI campus consumes hundreds of megawatts to gigawatts continuously. The 5 GW Hyperion campus alone exceeds the total power draw of every quantum research facility in the world combined by approximately three orders of magnitude.

The architectural reason for the gap is structural. In classical computing, computation drives the energy budget and cooling is 2-20% of total power. In quantum, that ratio inverts: cooling dominates the energy budget, but the absolute numbers stay small because quantum compute is doing structurally efficient work on the right problem classes. Each watt of quantum compute is solving exponential-complexity problems in polynomial time, work that would otherwise require massive classical compute infrastructure to replicate.

Implication for the Layer 2 power thesis: quantum does not draw from the same demand pool as AI. The 945 TWh data center demand projection for 2030 is not threatened by quantum substitution. A future quantum data center campus does not appear in the IEA's AI-attributable power forecast because the total power required for quantum is rounding error against AI demand at any plausible quantum commercialization timeline. If anything, the integration of quantum coprocessors INTO existing AI data centers (the NVIDIA Ising / CUDA-Q / NVQLink architecture) marginally increases AI data center demand by adding QPU racks to GPU campuses.

The architectural consensus: hybrid quantum-classical

The most important framing shift of 2025-2026 is that the industry has converged on hybrid quantum-classical architectures rather than standalone quantum systems. The implications for the report's thesis are significant.

Quantum processors are now positioned and engineered as coprocessors operating inside classical AI infrastructure, accessed through APIs the same way GPUs are accessed. NVIDIA's positioning is the clearest signal: Ising sits on CUDA-Q sits on NVQLink. Jensen Huang's positioning statement on the Ising launch: "AI becomes the control plane, the operating system of quantum machines." This is not corporate communications. It is an architectural statement about which layer of the stack owns the integration logic.

The architectural consequence is that quantum is structurally constrained to enter the data center as an additional accelerator class alongside GPUs and custom ASICs, not as a separate parallel computing paradigm. The same hyperscaler campuses that house H100/B200/Vera Rubin GPUs will house QPUs running on the same network fabric, with the same cooling utilities, in the same operations footprint. Layer 3 (data center / neoclouds) benefits at the margin from this trajectory. Layer 4 (networks and optics) benefits from the additional interconnect requirements. Layer 1 (foundation silicon) benefits structurally from NVIDIA's platform-lock-in via the CUDA-Q ecosystem.

Where quantum actually applies for AI workloads is a narrower question than the marketing suggests. The honest answer is: hyperparameter tuning across millions of model parameters, neural architecture search, attention mechanism optimization, and small-scale chemistry/materials/finance simulations. The October 2025 Zurich/IBM result trained a language model's attention mechanisms with quantum circuits using approximately one-tenth the compute resources of an equivalent classical approach^[54]. That result is real and significant for specific training phases. It does not apply to the bulk of LLM pretraining, which remains a classical GPU workload through 2030 and beyond.

The investment landscape by modality

The quantum sector has not converged on a single architectural approach. Five distinct modalities are in active commercial development, each with different timelines, supplier concentration, and public-equity exposure profiles. The right framing is portfolio exposure across modalities, not single-name bets.

IonQ as the leading public proxy

IonQ deserves specific treatment as the most credible public-equity quantum exposure available in mid-2026. The Q1 2026 results were a structural inflection rather than a single quarter beat.

Revenue of $64.7 million for Q1 2026 represented 755% year-over-year growth and exceeded the midpoint of management guidance by 30%. Full-year 2026 guidance was raised to $260-270 million, implying continued growth above 100% organic. Remaining Performance Obligations grew 554% year-over-year to $470 million, providing multi-year revenue visibility unusual for a pre-profitability quantum hardware company. IonQ became the first public quantum company to exceed $100 million in annual GAAP revenue in 2025 ($130 million)^[40]. IonQ shipped its first sixth-generation, chip-based 256-qubit system in Q1 2026, anchored by a secure quantum network and broad IP-generation partnership spanning computing, networking, sensing, and security. Market capitalization stood at $17.6 billion as of Q1 2026 results^[40].

The bear case is straightforward and worth holding. Adjusted EBITDA loss of $96.8 million for Q1 2026, free cash flow of negative $159 million, full-year adjusted EBITDA loss guidance of $310-330 million. IonQ requires sustained capital markets access to fund operations through profitability, which the market is pricing at $17.6B market cap with limited near-term path to GAAP profitability. The structural revenue growth is real; the path to operating profitability is not yet demonstrated.

The bull case is the platform position. IonQ is one of the most credible commercial quantum companies globally, the only named launch adopter of NVIDIA Ising among public-equity quantum names, and has demonstrated a 10,000x time-to-solution advantage on specific algorithms versus classical approaches. The 256-qubit system shipment in Q1 2026 is materially ahead of consensus expectations. Government and defense contract flow has been a meaningful contributor, with QuantumBasel ($60M four-year), KISTI delivery, and a growing national-scale initiative pipeline.

NVIDIA's strategic positioning is the most overlooked angle

The quantum opportunity for NVIDIA is not in selling quantum hardware. NVIDIA is not building qubits. The opportunity is in owning the software and AI control plane for quantum hardware operated by everyone else.

The playbook is the CUDA playbook, applied to a new compute paradigm before the ecosystem consolidates. CUDA-Q (2022) established the programming model. NVQLink (2025) established the hardware interconnect. Ising (April 2026) established the AI models for the hardest engineering problems in quantum (calibration, error correction decoding) and made them open-source on top of NVIDIA infrastructure. Every quantum hardware company that adopts the stack creates lock-in. Twelve major quantum institutions named at Ising launch represent the dominant share of credible commercial quantum effort globally.

The investment implication is that NVIDIA's Layer 1 thesis is structurally extended by quantum. The market does not yet price NVIDIA's quantum-platform position because the absolute revenue contribution is small in the near term. If quantum commercialization continues on the accelerated timeline and NVIDIA's control-plane position holds, the platform lock-in becomes another decade-scale structural moat similar to CUDA itself. This is not in current NVDA consensus forecasts.

Bottom line on quantum and the AI thesis

Quantum computing does not change the structural conclusions of this report. The Layer 2 power thesis remains intact because quantum does not draw from the same demand pool. The Layer 1 thesis is structurally extended because NVIDIA's quantum platform position adds an undervalued component to the existing CUDA moat. The Layer 6 frontier tier is meaningfully more interesting in 2026 than it was in 2024, with IonQ as the leading credible public-equity exposure and the broader quantum complex as a portfolio allocation rather than single-name bet.

The risk to be aware of is hype cycle compression. The 2024-2025 quantum stock rallies were partially driven by ETF flows, momentum trading, and headline catalysts rather than fundamentals. Q1 2026 results across the sector have begun to validate revenue trajectories at IonQ and selectively elsewhere, but the public-equity quantum names trade at multiples that assume continued execution. Position sizing should reflect that this is a category where the operational gap between leaders and laggards will widen materially through 2026-2027, and where the wrong name held in concentration creates real drawdown risk even within a sector that is broadly directionally correct.

The financial story is one part of the picture. The other part is what the cycle means for the people building AI products on top of all this infrastructure. Building economics determine which application-layer companies actually capture the productivity gains the rest of the report describes. This section is the operating manual for the AI build, written from the perspective of someone running infrastructure decisions for an AI company today. It is the lens Ferrox Labs applies to its own R&D.

The cost-of-tokens problem

The model layer is rapidly commoditizing on inference economics, though not uniformly. Closed-frontier API pricing remains substantial (GPT-5.5 at roughly $5 per million input / $30 per million output as of mid-2026), but open-weight self-hosted inference on optimized stacks (quantized open-weight models on AMD MI300/MI450 or NVIDIA H100/H200, with batched serving) has fallen to roughly $0.20 to $1.00 per million tokens for selected production workloads, depending on model size, hardware, batch size, context length, and quality target. The directional implication holds even if the specific number does not: builders should not anchor product economics or pricing to today's frontier API tier. Anything you build on $0.20 per million tokens for self-hosted inference needs to plausibly work at $0.02 per million in 2028; anything you build on closed-frontier pricing needs to survive a 50-90% cost compression as open-weight quality closes the gap.

This commoditization has a second-order implication that most builders are not yet adjusting to: the value migrates up the stack. The model layer is not where margin lives. The integration layer (where AI meets actual workflows, tools, and data) is. The agent layer (where AI does multi-step work autonomously) is. The verification and trust layer (where AI output gets validated against ground truth) is. Builders who put their effort at the model layer are competing in a commodity. Builders who put their effort at the layers above are competing in a rapidly expanding margin pool.

Multi-provider routing as default

Single-provider AI architecture is technical debt now. Frontier-quality models exist from at least four independent providers, each with multiple SKUs at different cost-quality tradeoffs: OpenAI (GPT-5.5 and GPT-5.5 Pro for frontier reasoning and coding, GPT-5.4 mini and GPT-5.4 nano for lower-latency lower-cost workloads), Anthropic (Claude Opus 4.7, Sonnet 4.6, Opus 4.6), Google (Gemini 3 Pro, Gemini 3 Flash, Gemini 3 Deep Think), and open-weight models from Alibaba (Qwen3.5 family, with Qwen3.6 in early rollout) and Meta (Llama 4 Scout, Maverick, Behemoth). Each has different cost-quality tradeoffs at different request types. The right architecture is a model router that selects providers based on request characteristics, with fallback hierarchies for capacity and cost optimization. Note that frontier model lineups iterate on weeks-to-months timescales; OpenAI alone has shipped GPT-5, GPT-5.4, and GPT-5.5 within a four-month window. Any specific model named here may be superseded by the time this report is read.

The pattern Ferrox Labs uses internally for its own infrastructure is a Thompson-sampling-based router that dynamically allocates traffic across providers based on observed quality and cost: highest-quality requests go to frontier closed-weight models when the value-per-token justifies the price; bulk inference goes to optimized open-weight inference on lower-cost compute; latency-sensitive requests go to the lowest-latency provider regardless of cost. The architecture itself is straightforward; the discipline of building it correctly from day one is what most teams skip.

Local inference is now viable

Open-weight models running on consumer-grade hardware are good enough for a meaningful portion of AI workloads. Llama 4 Scout at 4-bit quantization runs on a single high-end consumer GPU. Qwen3.5-4B runs natively on commodity laptops with native multimodal capability. The implication for builders is that some workloads should not be sent to cloud inference at all: privacy-sensitive workloads, low-latency interaction, offline-capable applications, and any workload where the marginal cost of cloud inference matters.

The constraint is architectural. Building applications that work transparently across local and cloud inference (with capability detection and graceful fallback) is more complex than cloud-only architecture. But the cost differential favors local-capable architecture by 10-50x for the workloads where local inference is viable. For builders running consumer applications at scale, this is the difference between unit economics that work and unit economics that don't.

MCP and the agent interoperability layer

The Model Context Protocol, introduced by Anthropic in late 2024, has become the de facto standard for connecting AI agents to tools, data, and other agents. The standardization matters because it is the layer at which multi-provider AI infrastructure actually composes: an MCP-compatible tool works equally with Claude, GPT-5, Gemini, and open-weight model deployments without rewriting integrations. Builders investing in MCP-compatible architecture today get future-proof multi-provider compatibility for free.

The implication for product economics is that AI features that depend on tool use, data access, or external system integration should be built MCP-first. The category of "AI products that automate workflows" is shifting from custom integrations per AI provider to MCP-standardized integrations that work everywhere. The economic rent shifts from integration vendors to the workflow layer above.

Power and the geography of inference

Section 03 established that power binds the cycle. The implication for builders is that the geography of cheap inference is shifting. Inference clusters that operate behind-the-meter at new gas plants in West Texas, Louisiana, parts of Oklahoma and Wyoming will deliver substantially lower cost-per-token than equivalent capacity in Northern Virginia or Silicon Valley. Some of the lowest-cost inference in 2027 will be in regions that don't currently host meaningful tech infrastructure.

Builders running at scale should track which neoclouds and AI cloud providers are siting capacity in cheap-power regions and route inference accordingly. The 30-50% cost differential between premium-region inference and cheap-region inference will widen, not narrow, through 2028.

The desktop agent layer

The next architectural shift is from cloud-only AI assistants to local-cloud hybrid agents that operate on the user's actual desktop, with access to local files, applications, and workflows. This is the layer Ferrox Labs targets with the Wayland project (June 2026 launch), and it is the layer where we expect the most economically interesting AI builds of the next two years to be concentrated. The thesis: most knowledge-work productivity gain requires AI to operate on the user's actual data and tools, not on cloud abstractions of them. Local-first agents with cloud capability for heavy reasoning are the right architecture.

This is a builder's view, not an investor's view. We highlight it because the financial implications cascade: if the desktop agent layer takes off, the value migrates further up the stack from the cloud APIs that the rest of the report covers, and the application layer becomes the dominant tier of the AI economy by 2028. If it doesn't, the cloud-API status quo persists and infrastructure remains the dominant tier.

This section is written for the active trader: someone running positions on a daily-to-quarterly time horizon with stops, position-sizing discipline, and a defined edge. The framework below is the same analytical structure as the rest of the report, applied to active position management. It is offered as a thought experiment and starting framework, not as a model portfolio to be copied. Three components: a baseline allocation framework by phase, a watch list of indicators that drive rotation decisions, and a catalyst calendar that maps the high-information-content events through the next 12 months. The intended output is a repeatable analytical process, not a list of stock picks.

Allocation framework: $100K reference book

For a $100K notional book actively trading the AI complex in mid-2026 (Phase B, power-binding phase), the reference allocation below is provided as an analytical starting point rather than a recommendation. An experienced trader would calibrate the framework to individual conviction, time horizon, and risk profile.

The indicator dashboard

Six indicators drive most of the rotation decisions. They are checked weekly. Each maps to a specific layer or to the cycle as a whole.

Catalyst calendar, next 12 months

Specific high-information-content events scheduled or expected through May 2027. Each is tradeable in size; each requires positioning before the event, not after.

Rotation logic in practice

The rotation between phases happens at specific trigger events, not on calendar dates. Three patterns to recognize:

The capex-decel rotation: when hyperscaler capex guides plateau or decline (typical Phase A to Phase B handoff signal), trim Layer 1 (NVDA, AVGO, MRVL multiples compress as growth normalizes), maintain Layer 2 (the power constraint is independent of capex deceleration), and add to Layer 5 (application monetization story replaces infrastructure capex story). Position size for chip names should compress 30-50% on this trigger.

The efficiency-shock rotation: when a DeepSeek-class open-weight release demonstrates 5-10x cost compression, trim Layer 1 hard (the immediate price action), maintain Layer 2 (power constraint unchanged), defensive trim Layer 3 (neoclouds with debt), preserve Layer 4 (networking layer benefits from total compute growth even at lower per-token margin), opportunistic add Layer 5 (application layer benefits from cheaper inputs). The window between event and position adjustment is hours, not days.

The application-revenue inflection: when a major application-layer name (PLTR, NOW, CRM, MSFT) reports AI revenue exceeding 15% of total revenue, accelerate rotation from Layer 1 (where chip multiples are largely priced in) to Layer 5 (where application multiples re-rate). This represents the Phase B to Phase C handoff. Expected timing: late 2026 to mid-2027. The recommended approach is to position pre-emptively beginning in late 2026.

Position sizing rules

Three rules we apply to every position in the AI complex.

One: no single name greater than 10% of the book. The DeepSeek event demonstrated that even the highest-conviction megacap can lose 17% in a single session on idiosyncratic risk. Concentration above 10% means a single bad event can take the book down materially.

Two: Layer 6 (frontier) names individually capped at 1% of book, total Layer 6 exposure capped at 5%. The five recommended Layer 6 names (IONQ, MP, USAR, RKLB, ASTS) at 1% each fits the cap exactly. These are venture-style positions with binary outcomes; size accordingly. A full loss on a 1% position is recoverable; a full loss on a 10% position is not. TSLA, when included in Layer 6 exposure for the Optimus thesis, sizes as a separate sleeve outside the venture allocation given its market-cap profile.

Three: maintain the 2% Cash/Hedge sleeve as a baseline reserve and expand to 5% under defined trigger conditions (sustained drawdown of 10%+ in either Layer 1 or Layer 2, DeepSeek-class efficiency event, or hyperscaler capex guide cut of 15%+). Major drawdowns in the AI complex typically last 3-7 sessions before high-conviction names recover. The trader who can deploy fresh capital into a panicked tape captures the recovery; the trader fully invested into the panic does not.

Quantum complex: portfolio exposure across modalities. IONQ is the leading pure-play (also Layer 6); RGTI superconducting; QBTS annealing; QUBT photonic speculative. NVDA captures quantum upside via platform lock-in without binary risk. Adjacencies: TSLA is a large-cap physical-AI thesis, sized independently. INTC is turnaround context plus silicon spin-qubit optionality.

Position-sizing reminder: no single name >10% of book; Layer 6 names capped at 1% individual / 5% aggregate; quantum complex sized as portfolio exposure across modalities; adjacencies sized separately.

The rotation timeline

READ WITH · Section 04 (layer details) · Section 05 (Layer 2 valuation) · Section 06 (rotation logic) · Section 07 (counter-thesis) · Section 08 (quantum) · Section 10 (allocation framework)

This is original research by Ferrox Labs and Trade Canyon, conducted during March, April, and early May 2026. We synthesized over 50 primary and secondary sources, validated every numeric claim against company filings or named research desks, and developed the Compute / Power / Capital framework and the six-layer atlas as our analytical lens. The framework is our own. The data underneath is sourced and cited inline. The argument is ours. No single external source provided the structure of this report.

Source hierarchy

We organized sources into three tiers with descending weight on contested claims.

Tier 1: Primary documents. SEC filings (10-K, 10-Q, 8-K), earnings call transcripts, company press releases, government data (US Department of Energy SPARK program disclosures, FERC interconnection queue data, Lawrence Berkeley National Lab studies, US Geological Survey rare earths data), and intergovernmental research (International Energy Agency, Bank for International Settlements). Anything cited from these sources is verbatim or paraphrased without interpretation.

Tier 2: Named research desk reports. McKinsey & Company, Goldman Sachs, Morgan Stanley, JPMorgan, Bank of America, UBS, KKR, Bernstein, Counterpoint Research, TrendForce, Omdia, MIT NANDA, Stanford HAI. We use these for forecast figures and analytical framework inputs. Where multiple desks have published convergent forecasts, we cite the convergence.

Tier 3: News of record. Bloomberg, Financial Times, Reuters, Wall Street Journal, CNBC, Tom's Hardware, IT Brew, DCD Magazine. Used for events of public record (announcement timing, market cap movements, executive quotes verifying primary disclosures) but never as the sole source for a numeric claim.

Validation discipline

Every numeric claim in this report was validated against at least one Tier 1 source. Where Tier 2 forecasts are cited, we name the desk and date. Where claims are estimates rather than disclosures (e.g., Microsoft's "$80 billion power-constrained backlog" figure), we explicitly label them as analyst estimates and give the analyst source. Quotes from executives are sourced to specific earnings calls or press events and verified against the published transcript when available.

Two specific validation patterns warrant emphasis. First, McKinsey's $6.7T continued-momentum scenario was cross-checked against Goldman Sachs's $7.6T cumulative 2026-2031 figure and Omdia's $1.6T 2030 annualized figure; the three independently-built models converge on the same order of magnitude, which represents significantly stronger validation than any single source. Second, specific company disclosures (CoreWeave's $99.4B RPO, Microsoft's $627B RPO, Google's $462B backlog) have been treated as primary facts rather than analyst interpretations.

What the framework owes (and doesn't owe) to others

The Compute / Power / Capital triangle is our own framing, though the three-force pattern parallels analysis of historical capital cycles by Carlota Perez (2002) and others. The six-layer atlas (Foundation, Power & Grid, Infrastructure, Networks, Application, Frontier) is our segmentation. The Phase A / B / C rotation framework is our timeline construction. McKinsey's investor-archetype framework (Builders, Energizers, Technology Developers, Operators, AI Architects) is a different but compatible lens; we cite it where relevant but our six-layer structure is more granular and oriented to position-sizing decisions.

Disclosures and conflicts

Sean Donahoe, lead author, holds positions in several names referenced in this report through Trade Canyon, Inc. and personal accounts. Specific positions are not disclosed for confidentiality but the reader should assume conflicts of interest exist. Trade Canyon is an active trading and education business with revenue streams that benefit from both sides of trades discussed in this report. Ferrox Labs is the R&D entity that publishes Wayland (June 2026 launch), which is a builder-side product positioned in the desktop-agent layer covered in Section 09. Neither company has commercial relationships with any of the publicly-traded names referenced here.

This report is research and education, not investment advice. No recommendation to buy, sell, hold, or trade any security is intended or implied. Readers should consult appropriate professional advisors before making investment decisions. Past performance does not guarantee future results. Capital cycles are uncertain by nature; the framework presented here is a tool for thinking, not a guarantee of outcomes.

1. [1]Tom's Hardware, "Hyperscaler 2026 capex tally crosses $725B," May 2026. Compilation of Q1 2026 earnings disclosures from Microsoft, Alphabet, Amazon, and Meta.
2. [2]Statista, Chart 35046, "Capital Expenditure of Leading Hyperscaler Operators," April 2026 update.
3. [3]Futurum Group / Introl, "AI Capex 2026," February 2026 analyst estimate.
4. [4]Microsoft Corporation, Q3 FY2026 Earnings Release and Conference Call Transcript, April 29, 2026.
5. [5]Goldman Sachs Global Investment Research, "Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out," May 1, 2026, supporting $765B 2026 annual AI capex, $1.6T 2031 annual capex, and $7.6T 2026-2031 cumulative baseline.
6. [6]Engineering News-Record / Entergy Louisiana regulatory filings on Meta Hyperion campus, 2025-2026.
7. [7]McKinsey & Company (Patel, Sachdeva, Noffsinger, Bhan, Chang, Goodpaster), "The cost of compute: A $7 trillion race to scale data centers," McKinsey Quarterly, April 28, 2025.
8. [8]Omdia, "Data Center Capital Expenditure Forecast," 2025-2026.
9. [9]McKinsey & Company, "Beyond compute: Infrastructure that powers and cools AI data centers," August 2025; "The data center dividend," October 7, 2025.
10. [10]KKR Insights, "Beyond the Bubble: Calibrating the AI Capital Cycle," February 2026.
11. [11]Behind the Balance Sheet (Steve Clapham et al.), "The Telecom Build of the 1990s: A Comparable History," April 2026.
12. [12]Investment Research Partners, "Comparing Capital Cycles: AI vs Railways, Electrification, and Telecom," August 2025.
13. [13]Newbery, D.M. (1999), "Privatization, Restructuring, and Regulation of Network Utilities," MIT Press; cited for railway capex historical data.
14. [14]CNBC / Pivotal Research / Barclays compiled hyperscaler cash and FCF analysis, Q1 2026 earnings cycle.
15. [15]Bank for International Settlements (Aldasoro, Doerr, Rees), "Financing the AI Boom: From Cash Flows to Debt," BIS Bulletin No. 120, January 7, 2026.
16. [16]Moody's Investors Service, "Hyperscaler Off-Balance-Sheet Lease Commitments Reach $662 Billion," February 2026.
17. [17]Microsoft Corporation Q3 FY2026 Earnings Conference Call, April 29, 2026, prepared remarks by CFO Amy Hood. Transformer lead time of 80-130 weeks, with 128-week upper end, sourced from secondary industry coverage including Data Center Dynamics, McKinsey 2026 data center supply chain analysis, and Hitachi Energy disclosures.
18. [18]International Energy Agency, "Energy and AI," April 2025 special report.
19. [19]NVIDIA Corporation Fiscal Year 2026 Annual Results, Form 10-K and Q4 FY2026 Earnings Release.
20. [20]AMD and Meta press releases on multi-year strategic compute agreement, February 2026, covering up to 6 GW deployment and AMD performance-based warrants for up to 160M shares; contract value of "up to $60 billion over five years" reported by Reuters and corroborated by Bloomberg.
21. [21]Marvell Technology Inc. FY2026 Annual Results and NVIDIA NVLink Fusion Partnership disclosures.
22. [22]Micron Technology Inc. Q1, Q2, Q3 FY2026 Earnings Disclosures on HBM4 production.
23. [23]Counterpoint Research, "AI Inference Accelerator Market: Custom ASIC vs Merchant GPU Forecast," 2025-2026 update; corroborating data from TrendForce.
24. [24]Lumentum Holdings Inc. Form 8-K, March 2, 2026, NVIDIA Investment Disclosure.
25. [25]Coherent Corp. Form 8-K, March 2, 2026, NVIDIA Investment Disclosure.
26. [26]Vistra Corp. Form 8-K filings on Comanche Peak Nuclear AWS PPA and PJM nuclear Meta agreements; Lotus Infrastructure 2,600 MW gas portfolio acquisition.
27. [27]CoreWeave Inc. Q1 2026 Form 8-K and Earnings Release, including $99.4B RPO disclosure, $24.859B debt position, and NVIDIA $2B Class A common stock investment.
28. [28]IREN Limited Q3 FY26 Earnings Release and 8-K disclosures on $3.4B AI Cloud commitment, $2.1B NVIDIA warrant for 30M shares at $70 strike, 5GW Sweetwater pipeline.
29. [29]Satya Nadella, BG2 Podcast appearance (Brad Gerstner, Bill Gurley) covering Microsoft AI infrastructure constraints, including the "could not find electricity to power them" comment on idle GPU inventory. Coverage by Data Center Dynamics, The Verge, and other secondary outlets in early November 2025.
30. [30]Microsoft Corporation Q3 FY2026 Earnings Release commercial RPO disclosure.
31. [31]Introl analysis of Microsoft Azure backlog composition, Q1 2026.
32. [32]Alphabet Inc. Q1 2026 Earnings Release and Investor Relations disclosures, Google Cloud backlog.
33. [33]NuScale, Oklo, X-Energy, TerraPower regulatory filings and DOE deployment timeline disclosures.
34. [34]Lawrence Berkeley National Laboratory, "Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection," 2024 annual update.
35. [35]US Department of Energy, Office of Electricity, "Speed to Power through Accelerated Reconductoring and other Key Advanced Transmission Technology Upgrades (SPARK)," $1.9B Notice of Funding Opportunity DE-FOA-0003580, announced March 12, 2026; third tranche of the Grid Resilience and Innovation Partnerships (GRIP) program.
36. [36]International Energy Agency, "Key Questions on Energy and AI," April 2026 update to the 2025 Energy and AI report.
37. [37]Constellation Energy Corporation 8-K and press release on Crane Clean Energy Center (formerly Three Mile Island Unit 1) nuclear restart agreement with Microsoft, September 2024.
38. [38]Talen Energy Corporation 8-K disclosures on Susquehanna Steam Electric Station expansion ISA with Amazon Web Services.
39. [39]Bloomberg / FT / Reuters market data on NVIDIA $589B single-day market cap loss, January 27, 2025, triggered by DeepSeek R1 release. DeepSeek V3 technical report (arXiv:2412.19437) for the 671B-parameter MoE architecture with 37B active per token. McKinsey "DeepSeek and the new economics of AI" 2025 analysis for the 18x training and 36x inference efficiency comparison versus GPT-4o-class economics, attributed to V3 (the underlying base model), not R1 (the distilled reasoning model).
40. [40]McKinsey & Company, "The State of AI in 2025: Agents, Innovation, and Transformation," November 2025; survey of 1,993 respondents in 105 countries, June 25-July 29, 2025.
41. [41]Massachusetts Institute of Technology, NANDA initiative (Aditya Challapally et al.), "GenAI Divide: Why 95% of Pilots Fail," 2025.
42. [42]IonQ Inc. Q1 2026 Earnings Release and full-year 2026 revenue guidance ($260-270M range).
43. [43]USGS Mineral Commodity Summaries 2025 and 2026 for broad rare earth mine production. IEA "The Role of Critical Minerals in Clean Energy Transitions" 2026 update for category-specific shares: ~60% mined production of magnet rare earths, 91% refined output, 94% sintered permanent magnet production (2024 figures).
44. [44]MP Materials Corp. Q1 2026 Form 10-Q and Earnings Release, NdPr production and Magnetics segment disclosures.
45. [45]USA Rare Earth Inc. Press Release and Form 8-K on non-binding Department of Commerce CHIPS Program LOI (proposed up to $277M direct funding plus up to $1.3B senior secured loan, subject to diligence and final agreements) and separately closed $1.5B PIPE financing, January-April 2026.
46. [46]Tesla Inc. Q4 2025 Earnings Conference Call Transcript on Optimus production timing and Fremont Model S/X production end.
47. [47]Tesla Inc. Q1 2026 Form 8-K disclosures on Gigafactory Texas Optimus expansion.
48. [48]AST SpaceMobile Inc. Form 8-K, April 2026, BlueBird 7 satellite loss disclosure.
49. [49]Blue Origin / Federal Aviation Administration disclosures on New Glenn NG-3 mission anomaly, April 19, 2026.
50. [50]OpenAI Press Release, "OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites," September 23, 2025, which states the OpenAI/Oracle agreement entered in July 2025 "exceeds $300 billion between the two companies over the next five years" for 4.5 GW of additional Stargate capacity. Initial WSJ reporting September 10, 2025.
51. [51]Stargate UAE program announcements: G42 Press Release "Global Tech Alliance Launches Stargate UAE" May 22, 2025; OpenAI "Introducing Stargate UAE" May 2025; National (UAE) coverage of Mubadala progress update Q4 2025. NVIDIA Blackwell GB300 export approval to G42 reported by US Commerce Department, January 2026.
52. [52]Saudi Arabia AI infrastructure programs: Public Investment Fund $40B AI fund commitments 2024-2026; HUMAIN-MIS contract $501M February 2026; LEAP 2025 announcements totaling $15B including PIF-Google Cloud $10B partnership; Qualcomm-HUMAIN MOU May 2025. EU AI Continent Action Plan: European Commission announcements February 2025, €200B target through 2030.
53. [53]IBM "The trends that will shape AI and tech in 2026," IBM Think, March 2026, including the corporate commitment that "2026 will mark the first time a quantum computer will be able to outperform a classical computer" and the Quantum Roadmap for utility-scale advantage. IBM Nighthawk processor specifications and Quantum Network commercial usage data.
54. [54]Google Quantum AI publication on Willow processor 13,000x speedup over Frontier supercomputer on physics simulation, October 2025. IBM Quantum + ETH Zurich research on quantum-enhanced attention mechanism training, October 2025, demonstrating ~10x compute resource reduction for specific training phases. Multiple independent quantum advantage demonstrations from IBM, Google, IonQ, Quantinuum, Pasqal, and QuEra during 2025-2026.
55. [55]Energy consumption data for quantum computing systems: Pasqal corporate disclosure on neutral atom processor power draw of approximately 2.6 kW independent of qubit count. IEEE/NREL research on superconducting dilution refrigerator power requirements (~25 kW per system). M. Martin et al., "Energy use in quantum data centers: Scaling the impact of computer architecture, qubit performance, size, and thermal parameters," arXiv:2103.16726 and subsequent IEEE publication, establishing first-principles energy modeling for quantum data centers.