The hyperscalers have guided to close to $700 billion in combined capital expenditure for calendar year 2026. That is not money they have spent yet; it is money they have committed to spend, backed by land purchases, construction contracts, and equipment orders with lead times measured in years. Yet the consensus narrative still treats this as a winner-take-all story between three or four well-known names. That framing appears incomplete, and the incompleteness may be creating a sourcing dislocation across the supply chain.
When the market tells a simple story about a complex system, the second-order and third-order effects tend to be underweighted. Those overlooked effects are where the analytical value lives. What follows is a breakdown of what the available filings suggest the market is mispricing, and where the asymmetric data points cluster.
The capex number is real, and it is much bigger than the public narrative suggests
The four largest hyperscalers have each guided to individual 2026 capex budgets that would have sounded fictitious two years ago:
- Microsoft: approximately $190 billion for calendar year 2026, per the FY26 Q3 earnings call on April 29, 2026. That figure includes roughly $25 billion from the impact of higher component pricing. In Q3 FY26 alone, Microsoft spent $31.9 billion on capex, roughly two-thirds of which went to short-lived assets like GPUs and CPUs. (Source: Microsoft FY26 Q3 earnings call, microsoft.com/investor/events/fy-2026/earnings-fy-2026-q3, April 29, 2026.)
- Alphabet: $175 billion to $185 billion in initial 2026 guidance, subsequently raised to $180 billion to $190 billion on the Q1 2026 earnings call. CFO Anat Ashkenazi stated that 2027 capex will “significantly increase compared to 2026.” Q1 2026 capex was $35.7 billion. (Source: Alphabet Q1 2026 earnings, CNBC, April 29, 2026; cnbc.com/2026/04/29/alphabet-googl-q1-2026-earnings.html; abc.xyz/investor.)
- Amazon: $200 billion for 2026, announced in February 2026 and reaffirmed on the Q1 2026 call. Q1 2026 capex was $43.2 billion. For context, Amazon spent $131.8 billion in all of 2025. (Source: Amazon Q1 2026 earnings, Business Insider, April 29, 2026; businessinsider.com/amazon-q1-earnings-amzn-stock-price-aws-ai-capex-2026-4.)
- Meta: $125 billion to $145 billion, raised from a prior range of $115 billion to $135 billion on the Q1 2026 call. Meta spent $72.2 billion in 2025, so the 2026 guidance is nearly double. (Source: Meta Q1 2026 earnings, CNBC, April 29, 2026; cnbc.com/2026/04/29/meta-q1-earnings-report-2026.html; investor.atmeta.com.)
Figure 1 — Hyperscaler 2026 capex guidance by company. At the midpoints, roughly $710B across four companies in a single calendar year. Source: company Q1/Q3 2026 earnings calls.
At the midpoints, that is roughly $710 billion across four companies in a single calendar year. These are not aspirational numbers. Microsoft added one gigawatt of data center capacity in Q3 FY26 alone, with the Fairwater, Wisconsin campus coming online six weeks ahead of schedule (Microsoft FY26 Q3 earnings call, April 29, 2026). A gigawatt-scale data center costs roughly $30 billion and takes approximately 2.1 years to build (Epoch AI trends page, epoch.ai/trends, retrieved June 2026).
The analytical mistake is not disputing the number. It is treating $710B as a single trade instead of a supply chain. Most of that chain is still priced like a cyclical peak rather than a structural buildout.
Three observations from the filings
1. Power, not chips, appears to be the binding constraint.
The popular narrative fixates on GPU supply. NVIDIA’s H100 shortage in 2023 trained an entire generation of analysts to think about AI infrastructure as a semiconductor story. That lens needs updating.
GPU availability has improved materially since the 2023 pinch. NVIDIA’s market capitalization sits at approximately $5.2 trillion as of June 2, 2026 (companiesmarketcap.com), and the company remains the default supplier for frontier training clusters. Demand patterns have shifted from pure training toward inference, but the total compute appetite has not plateaued.
What is clearly bottlenecked is power delivery. The numbers from the filings and industry reports:
- Power transformers average a 128-week lead time in the United States as of Q2 2025, with prices up 77% since 2019. Generator step-up transformers average 144 weeks. (Source: Wood Mackenzie analysis, reported in POWER Magazine, powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/, 2026.)
- High-voltage switchgear averaged 44 weeks in Q2 2025, with prices up 50% since 2021. (Source: Wood Mackenzie via POWER Magazine, same source.)
- Pre-pandemic, large power transformer lead times ran 7 to 14 months. They have roughly doubled to tripled. (Source: CWIEME industry analysis, berlin.cwiemeevents.com.)
- Demand for power transformers is up 119% since 2019. Demand for generator step-up transformers is up 274%. Projected supply shortfall in 2025 is 30% for power transformers and 10% for distribution transformers. (Source: Wood Mackenzie, via POWER Magazine, same source.)
- PJM, the largest regional transmission organization in the United States, now takes eight years to bring new generation online through its interconnection queue. (Source: Rocky Mountain Institute, rmi.org/pjms-speed-to-power-problem-and-how-to-fix-it/, 2025.)
- Nationally, over 2,060 gigawatts of generation and storage capacity were actively seeking grid interconnection as of the end of 2025. The median time from interconnection request to commercial operation has doubled from under two years (projects built 2000-2007) to over four years (projects built 2018-2024). (Source: Lawrence Berkeley National Laboratory, “Queued Up” report, emp.lbl.gov/queues, 2026 edition.)
Every company in the power delivery chain is running a backlog that did not exist in this form three years ago. Manufacturers have committed nearly $2 billion to new or expanded North American transformer production capacity since 2023, including Hitachi Energy (over $1 billion), Eaton ($340 million), Prolec GE (over $300 million), and Siemens Energy ($150 million). New factories take years to come online, and the supply gap will not close quickly even with these investments. (Source: POWER Magazine, same source.)
2. The second-order beneficiaries may be larger than the first-order ones.
NVIDIA is a $5.2 trillion company as of June 2, 2026 (companiesmarketcap.com). The re-rating for the AI cycle is already reflected in the price. The analytical question is whether the same re-rating has reached the companies sitting one or two steps removed from the GPU.
The electrical infrastructure manufacturers that supply data centers are seeing revenue transformations, but the market has re-rated them unevenly. That unevenness is where the analytical interest lives.
Vertiv Holdings (NYSE: VRT) is the closest thing to a data center infrastructure pure-play. In Q1 2026, Vertiv reported $2.65 billion in net sales, up 30% year-over-year. The Americas segment, which carries the bulk of the data center business, grew 53.1% YoY to $1.81 billion, with organic growth of 44.3%. Adjusted operating margin expanded 430 basis points to 20.8%. Management raised full-year organic growth guidance to 29-31% and adjusted operating margin guidance to 22.8-23.8%. As of April 22, 2026, backlog stands at $15.0 billion with a book-to-bill ratio of approximately 2.9x. Vertiv trades at roughly 48x forward earnings as of early June 2026 (valueinvesting.io/VRT/metric/forward-pe, June 2, 2026; Yahoo Finance forward P/E data). (Source: Vertiv Q1 2026 earnings release, investors.vertiv.com, April 22, 2026.)
Eaton Corporation (NYSE: ETN) is the large-cap power management company. Q1 2026 sales were a record $7.45 billion, up 17% YoY. The Electrical Americas segment grew 20% YoY to $3.60 billion, with 12-month rolling orders up 42% organically and backlog up 44%. Electrical Global backlog grew 73%. Eaton closed $11 billion in acquisitions during the quarter, including Boyd Thermal ($9.55 billion), a data center thermal solutions business. Management raised 2026 organic growth guidance to 9-11% from 8%. Eaton trades at roughly 35x forward earnings (Trefis, trefis.com/data/companies/ETN, Q1 2026 data). Eaton also carries aerospace and defense exposure, including its $1.14 billion Aerospace segment where Q1 backlog grew 28% YoY. (Source: Eaton Q1 2026 earnings release, businesswire.com, May 5, 2026; eaton.com/investor-relations.)
Hubbell (NYSE: HUBB) is where the valuation gap appears widest. Q1 2026 net sales rose 11% to $1.52 billion. The Electrical Solutions segment grew 12% with organic growth of 10.6%, driven specifically by “strong datacenter and light industrial markets” per CEO Gerben Bakker. The Utility Solutions segment grew 11%, with Grid Infrastructure products up 18%. Adjusted operating margin expanded 110 basis points to 19.8%. Management raised full-year guidance to 8-11% total sales growth with $19.30 to $19.85 adjusted EPS. Hubbell trades at approximately 24x forward earnings as of late May 2026 (GuruFocus, gurufocus.com/term/forward-pe-ratio/HUBB, May 28, 2026). That is a traditional industrial multiple for a company with double-digit organic growth driven by data center demand and expanding margins. (Source: Hubbell Q1 2026 earnings release, hubbell.gcs-web.com, April 30, 2026.)
A note on the multiples cited above: these are point-in-time aggregator figures (Trefis, GuruFocus, valueinvesting.io, Yahoo Finance) as of the dates shown. They are not pulled from primary filings and should be treated as indicative. The relative ordering — Hubbell cheapest, Eaton in the middle, Vertiv most expensive — is what carries the argument, not the absolute levels.
The pattern extends across the second-order chain. Modine Manufacturing (NYSE: MOD) reported data center revenue up 78% year-over-year in Q3 FY2026. The Climate Solutions segment grew 51% YoY. Management guided data center revenue to increase by more than 70% YoY for the full fiscal year 2026, and then grow at 50% to 70% annually for the next two years, putting the company on track to exceed its target of more than $2 billion in data center revenue by FY2028. Modine is also spinning off its Performance Technologies segment via a Reverse Morris Trust transaction with Gentherm, transforming into a pure-play climate solutions company focused on data center cooling and commercial HVAC. (Source: Modine Manufacturing Q3 FY2026 earnings release, investors.modine.com, February 4, 2026; Bizjournals, bizjournals.com/milwaukee/news/2025/08/01/manufacturer-modine-2b-data-center-revenue-2028.html, August 2025.)
Cooling has moved from a niche specialty to a baseline requirement for high-density AI racks. Liquid cooling, direct-to-chip cooling, and immersion cooling are no longer exotic; they are table stakes for any rack running at the power densities that modern AI accelerators demand.
Switchgear, bus ducts, power distribution units, generators, fuel cells, on-site nuclear microreactors. Each of these is a real product category with real companies behind it, selling into real purchase orders with real delivery dates. Every watt that reaches a GPU has to pass through a dozen pieces of equipment, each manufactured by a different company.
3. The duration of capex may matter more than the magnitude.
This is where the analytical tension is sharpest. The market appears to be pricing AI infrastructure spending as though it may be a two-to-three-year cycle that peaks and reverts, similar to the 5G buildout or the shale drilling boom. The multiples assigned to AI-adjacent industrials suggest the market still treats these as cyclicals despite having backlog visibility that should otherwise command compounder multiples.
The filings point toward a longer cycle. Consider:
- Training compute requirements for frontier models grow at 4 to 5 times per year (Epoch AI, “Power Demands of Frontier AI Training,” epoch.ai/publications/power-demands-of-frontier-ai-training, May 28, 2024; epoch.ai/trends confirms 5x per year since 2020). Each new generation of models demands materially more hardware than the last.
- Inference demand is scaling alongside training. As AI agents, copilots, and embedded AI features reach billions of end users, the inference compute load compounds.
- Sovereign AI is a new demand layer. Governments in the Middle East, Europe, and Asia are building national AI compute infrastructure. This adds a buyer class that did not exist in previous technology cycles.
Perhaps the most telling data point: Alphabet’s CFO explicitly stated on April 29, 2026 that 2027 capex will “significantly increase compared to 2026” (Alphabet Q1 2026 earnings call, CNBC, April 29, 2026). That is not language consistent with a cycle about to peak. It is language consistent with a cycle where demand outruns supply for years.
If the buildout runs for five to seven years rather than two to three, then every company in the supply chain warrants a fundamentally different multiple. A company with six years of visible, contracted backlog growth is a structurally different investment than one with eighteen months of backlog and a hope. The market does not appear to have made this distinction yet.
What the data does not yet show
No framing is complete without its own stress test. The available filings and industry reports point to several unresolved risks:
- Hyperscaler capex pullback. If AI monetization disappoints, the hyperscalers could collectively reduce spending guidance and the chain would reprice. This is the tail risk, but current evidence does not support it as the base case. These companies have made public, multi-year commitments backed by board-level approval and contracted spend. Microsoft guided to more than $40 billion in Q4 FY26 capex alone, meaning their near-term quarterly spend exceeds what most of these companies spent in an entire year a decade ago (Microsoft FY26 Q3 earnings call, April 29, 2026). A 10-15% trim is plausible. A 50% cut appears unlikely absent a macro shock.
- Policy and permitting intervention. The US government could accelerate or restrict data center construction through executive action, environmental regulation, or export controls on AI hardware. The tariff landscape, including copper tariffs up to 50% and expanded Section 232 steel and aluminum duties, already creates cost uncertainty for equipment sourcing (POWER Magazine, powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/, 2026). Any policy that materially slows construction timelines compresses the backlog duration thesis.
- Technology substitution. If a breakthrough in chip architecture dramatically reduces the power-per-token of AI inference, the power bottleneck loosens faster than expected. This is a long-tail risk. Software-level efficiency gains are real: Epoch AI tracks pre-training compute efficiency improvements at 3.0x per year (epoch.ai/trends, retrieved June 2026). Historically, efficiency gains have not reduced total compute demand; they have expanded what is possible within the same power envelope, which creates more demand.
- Concentration risk in the second-order chain. Several of the names cited here have customer concentrations tied to the top three hyperscalers. A company that derives an outsized share of its AI-exposed revenue from a single customer carries single-name credit risk beneath a growth multiple. Breadth of demand is more durable than depth.
Timelines and catalysts
The infrastructure buildout is most analytically interesting during the construction and commissioning phase. Once the data centers are built, powered, and operational, the spending shifts from capex to opex and the beneficiary set changes. Cooling and power companies benefit during the build. Software and services companies benefit during the operate phase.
The current analytical sweet spot runs through 2028, with the most notable valuation gaps in names where the market does not appear to fully believe the backlog. After 2028, the investable question is expected to rotate toward the software layer and the inference economics story.
Near-term catalysts worth tracking: subsequent hyperscaler earnings for updated capex guidance, interconnection queue data from regional transmission organizations, transformer shipment data from the major electrical manufacturers, and any updates on the domestic manufacturing buildout (Hitachi’s Virginia plant targeting 2028, Eaton’s South Carolina facility targeting 2027, Siemens’s Charlotte plant targeting early 2027, per POWER Magazine).
Glossary
- Book-to-Bill Ratio: A metric comparing the amount of new orders received (booked) to the amount of products shipped and billed. A ratio greater than 1.0x indicates a growing backlog.
- Generator Step-Up (GSU) Transformer: A specialized heavy-duty transformer used to step up voltage from a power generation facility to transmission-line levels.
- Interconnection Queue: The backlog of proposed power generation and storage projects waiting for grid impact studies and necessary transmission upgrades before they can reach commercial operation.
- Reverse Morris Trust: A tax-efficient corporate restructuring transaction used to spin off a business unit by merging it with a third-party company, allowing the parent to divest without immediate tax liability.
This is an AI Briefing — AI-generated analysis published under TLCapital.AI. It is not personal research or positions, and it is not investment advice. Figures are sourced to primary filings with dates noted throughout. Do your own diligence.