Hyperscaler Capex Wars 2026: AI Compute Reshapes Cloud Economics

Last Updated: 2026-05-16

Architecture at a glance

The five companies most responsible for the world’s AI compute — Microsoft, Amazon, Alphabet, Meta, and Oracle — will spend somewhere between $330 billion and $355 billion on capital expenditure in calendar 2026, based on their own public guidance and first-quarter actuals. That single number, larger than the annual GDP of Finland, is the most consequential variable in enterprise technology this decade. It funds the GPUs that train frontier models, the substations that keep them online, and the long-dated leases that have quietly converted three of the five companies into something closer to industrial-scale utilities than software businesses.

This analysis treats the 2026 capex cycle as a coherent industrial phenomenon rather than a set of company press releases. The thesis is simple: the binding constraint on AI compute is no longer silicon — it is power, grid interconnects, and the multi-year duration of the assets the hyperscalers are now buying. That single shift reshapes unit economics, depreciation, the GPU-versus-ASIC question, and the bear case if model efficiency keeps improving. We will work through each in turn, using only what the companies and reputable trackers have published, and flagging uncertainty where it is real.

The 2026 Capex Numbers: $300B+ Annual AI Buildout

The headline figures, drawn from each company’s most recent earnings disclosures and analyst consensus as of Q1 2026:

• Amazon has guided to capex of roughly $100-110 billion for FY2026, the largest absolute spend among hyperscalers, with the company telling analysts that “the vast majority” is going to AI infrastructure for AWS. The starting point is the ~$83 billion run-rate Amazon disclosed in 2024 commentary; the 2025 step-up was confirmed in the Q4 2024 results release on aboutamazon.com, and FY26 commentary on the Q1 2026 call extended the trajectory.
• Microsoft guided FY26 (ending June 2026) capex of approximately $80 billion, communicated on the July 2025 Q4 FY25 earnings call, with management framing the spend as “front-loaded” against committed Azure AI demand. The October 2025 Q1 FY26 call reaffirmed the trajectory; consensus analyst notes place the range at roughly $75-90 billion depending on how lease assumption is treated.
• Alphabet guided 2026 capex to roughly $75-80 billion during its Q4 2025 and Q1 2026 calls (abc.xyz/investor), a step up from ~$52 billion in 2024, with Google framing the increase as primarily AI infrastructure for both Google Cloud and Search.
• Meta raised its 2026 capex range to $60-72 billion on its Q1 2026 results (investor.atmeta.com), a substantial increase from the ~$39-40 billion 2024 actual. Meta is unusual in that the spend funds first-party AI workloads (Llama, GenAI ad ranking, Reality Labs) rather than a public cloud business.
• Oracle has guided capex of roughly $25-30 billion for FY26 (ending May 2026), per its Q3 FY26 earnings release on oracle.com/investor. Oracle is small in absolute terms but has the highest share of capex flowing to AI infrastructure — its remaining performance obligations (RPO) crossed half a trillion dollars in the most recent quarter, driven almost entirely by AI contracts.

These are guidance ranges, not forecasts; the actual numbers will move with FX, lease accounting, and how the companies classify shells they take on operating leases versus build directly. Two qualifications matter. First, a meaningful share of “capex” is now finance leases on third-party data center shells — particularly true for Microsoft and Oracle, which lease capacity from CoreWeave, Crusoe, Stargate-affiliated entities, and traditional operators like Digital Realty and Equinix. The economic spend is higher than the GAAP capex line for those companies. Second, third-party tracking firms — Dell’Oro, Synergy Research, Omdia — publish their own estimates that aggregate slightly differently; Dell’Oro’s most recent AI infrastructure forecast (dellgroup.com) puts total 2026 AI capex from the top six globally at roughly $400 billion when you include ByteDance, Tencent, Alibaba, and the GPU clouds.

The directional message holds regardless: 2026 capex is roughly double 2023 levels and the curve has not visibly flattened. Two and a half years into the GenAI cycle, the spend is still going up.

Where the Money Goes: GPU, ASIC, Network, Power, Land

The most common misconception in casual hyperscaler coverage is that capex equals NVIDIA orders. That is wrong in both directions — GPUs are less than the headline share suggests, and the non-silicon spend is far stickier than the silicon spend.

A reasonable illustrative breakdown of a marginal dollar of 2026 AI capex, synthesized from earnings commentary, Dell’Oro tracking, and disclosed PPA values:

• Compute silicon: 50-60 percent. This bucket is GPUs (still the majority — NVIDIA Blackwell and Rubin generations), in-house ASIC (Maia, Trainium, TPU, MTIA), CPU hosts, HBM memory, and SSDs. The split between NVIDIA and in-house silicon varies dramatically by hyperscaler — Google leans hardest on its own TPU, Oracle leans almost entirely on NVIDIA — but in aggregate NVIDIA still receives the largest single share of compute dollars.
• Networking: 8-12 percent. Scale-out AI fabrics (NVIDIA Spectrum-X Ethernet, Arista 7800R, Broadcom Tomahawk 5/6), NVLink switches and trays for scale-up, optics (800G transitioning to 1.6T), and structured cabling. As cluster sizes pass 100,000 GPUs the fabric becomes a meaningful percentage of cost. We covered the engineering trade-offs in our NVIDIA Spectrum-X analysis.
• Power and cooling: 12-18 percent. Substations, transformers, switchgear, UPS, backup gensets, busways, and increasingly direct-to-chip liquid cooling distribution. Liquid cooling is the line item that has surprised cost engineers the most — a fully liquid-cooled rack can cost 30-50 percent more in mechanical/electrical infrastructure than air-cooled, but it is mandatory at the 130+ kW rack densities Blackwell-class systems require.
• Shell and fit-out: 10-15 percent. The physical building, racks, fire suppression, security, and structured site work. Modular shells of 30-50 MW per hall have become the dominant build pattern because they map cleanly onto utility power-block delivery cadences.
• Land and utilities: 3-6 percent. Parcel acquisition, water rights, fiber bring-in, and the soft costs of utility coordination.

Two patterns inside these numbers are worth flagging.

First, the in-house ASIC share of compute spend is rising every quarter but is still a minority position across the industry. Google is the closest to majority in-house — TPU has been a multi-generation effort and v7 (“Trillium successor”) is now in volume — but even Google has continued buying NVIDIA for its third-party customers who request CUDA-native workloads. Microsoft Maia v2 is a real chip but the FY26 capacity additions still skew heavily NVIDIA. The custom-silicon story is real and gradual, not a step change.

Second, the power and cooling share is rising faster than any other category as rack densities climb from ~30 kW (2022 norm) toward 130-250 kW (Blackwell-class racks today) and projected 400+ kW (Rubin Ultra and successors). The mechanical infrastructure to handle that heat — coolant distribution units, primary and secondary loops, dry coolers or evaporative towers — is real construction work, not commoditized hardware. It scales sub-linearly with effort, and that is a piece of why the hyperscalers are increasingly building their own shells rather than leasing from colocation providers.

The GPU-vs-ASIC Bet: Each Hyperscaler’s Strategy

Every hyperscaler has now articulated a public position on custom silicon. The strategies differ in a way that maps cleanly to each company’s customer mix and revenue model.

Microsoft: Maia plus heavy NVIDIA

Microsoft’s silicon strategy has been pragmatic. Maia 100 shipped in limited Azure regions in 2024; Maia v2 is in qualification with disclosed targets that put it competitive with H100/B200-class workloads for specific inference patterns. But the overwhelming majority of FY26 capacity is NVIDIA — Microsoft has been one of NVIDIA’s largest customers since the OpenAI partnership scaled. The strategic logic is straightforward: Microsoft sells Azure to enterprise customers, most of whom want CUDA-native deployment paths, and Microsoft has a separate massive obligation to OpenAI that runs on NVIDIA infrastructure. Maia is a hedge and a margin lever on internal Microsoft inference (Copilot, Bing), not a wholesale replacement.

Amazon AWS: Trainium 3 and Inferentia at scale

AWS has been the most committed to in-house silicon over the longest timeframe. Trainium 3 is in volume production in 2026, with Anthropic disclosed as the anchor training customer for Project Rainier and several others; Inferentia continues to serve inference workloads in production. AWS publishes price-performance comparisons in its launch materials that claim 30-40 percent better $/training-hour than equivalent NVIDIA configurations — claims that are difficult to independently verify but consistent with what other ASIC programs show on narrow workloads. AWS also remains an enormous NVIDIA customer; the in-house chips serve specific workload patterns and customers, not the full market.

Google Cloud: TPU v7 generation, NVIDIA for hybrid customers

Google is unique in that TPU has been a research-and-production effort since 2015 and is now on its seventh generation. TPU v7 (the Trillium-successor generation) is in production for Gemini training and for Google Cloud customers who consume TPU directly via Vertex AI. Google has been more public than other hyperscalers about the cost-per-token advantage TPU offers on workloads engineered for it. But Google Cloud’s third-party customer base is more heterogeneous than Google Search internal workloads, and Google still buys NVIDIA in volume for those customers — including the major H100/B200 and Blackwell deployments announced through 2025-2026.

Meta: MTIA for ads and ranking, NVIDIA for everything else

Meta’s silicon strategy is the most internally-focused. MTIA v2 is in production running ads-ranking and recommendation workloads that have specific structural characteristics — sparse, embedding-heavy, latency-sensitive — that map well to a custom ASIC. For LLM training (Llama family) and general-purpose GenAI, Meta has been one of NVIDIA’s largest 2024-2026 customers; Mark Zuckerberg disclosed targets of 350,000+ H100-equivalent units multiple times across earnings calls. MTIA solves a specific Meta problem; it is not positioned as a product.

Oracle: heavy NVIDIA, narrow positioning

Oracle has taken the opposite position: its OCI strategy is to be the cleanest NVIDIA reference deployment, with high RDMA performance, predictable pricing, and large committed contracts. Oracle does not have a custom AI ASIC and has not signaled one. The RPO growth — half a trillion dollars in remaining performance obligations as of Q3 FY26 — reflects a small number of very large customers (including OpenAI under the Stargate umbrella, by reported terms) buying NVIDIA-based capacity on multi-year commits. It is the highest-conviction NVIDIA bet among hyperscalers.

The pattern across all five: NVIDIA remains the default for customer-facing workloads, custom silicon serves first-party or narrowly-defined cases, and the in-house chips are best understood as margin levers and supply-chain hedges rather than NVIDIA replacements at this point in the cycle.

Power as the Real Constraint: Why Capex Stalls on the Grid

If there is one thesis worth keeping when the noise of quarterly earnings fades, it is this: the binding constraint on hyperscaler AI buildouts in 2026 is not GPU supply. It is electrical power, transmission, and the multi-year permitting cycles that gate them.

NVIDIA’s supply has been catching up since H2 2024; CoWoS packaging capacity at TSMC has expanded; HBM supply from SK Hynix, Samsung, and Micron is no longer the binding lane on most builds. What has not caught up is power. An AI campus today is 500 MW to 2+ GW of essentially flat 24×7 demand — a load profile closer to an aluminum smelter than to a traditional data center. Three things follow from that.

First, firm power is the planning unit. Variable renewables, however abundant on paper, do not match a flat AI training profile without massive over-build and storage. The companies have responded with three firming patterns:

• Nuclear PPAs. Microsoft signed a 20-year offtake for Three Mile Island Unit 1 (now Crane Clean Energy Center), Constellation has multiple hyperscaler-related agreements, and Amazon acquired the Talen Cumulus campus and committed to SMR development with X-Energy. Restart timelines run 2027-2028; greenfield SMRs are 2030+. Useful, but slow.
• Natural gas. Combined-cycle and peaker plants are being permitted at unprecedented rates near new campuses — particularly in Texas, Louisiana, Oklahoma, Wyoming, and the Mid-Atlantic. Gas is the only firm source that can be commissioned on a 24-36 month horizon for new builds, and it is doing the bulk of the actual work.
• Geothermal and SMR forwards. Google’s Fervo deal in Nevada, Microsoft’s TerraPower-adjacent commitments, and Oracle’s stated interest in modular nuclear all matter for 2028+ but contribute little to the 2026 numbers.

Second, transmission interconnect queues are the actual scheduling bottleneck. PJM, MISO, ERCOT, and SPP have all reported interconnect queues stretching 4-7 years for new large-load interconnections in 2025-2026 commentary. Several hyperscalers have shifted site selection away from traditional Northern Virginia and the Pacific Northwest to where new power can actually be delivered — Wisconsin, Indiana, Texas, the Mountain West, and increasingly the Middle East (UAE) and parts of Asia. Site selection is now downstream of power, not the other way around.

Third, water and dry-cooling trade-offs are getting more attention. Water-cooled or evaporative-cooled designs are cheaper to operate but draw heavily on local water supply, an increasingly fraught permitting conversation in arid regions. The shift toward direct-to-chip liquid cooling with dry external rejection — engineered to a stricter WUE (water usage effectiveness) target — adds capex but de-risks the social and regulatory side. The IEA’s most recent data center electricity report (iea.org) places data center electricity demand at roughly 415 TWh globally in 2024 with AI-driven growth as a primary 2025-2030 driver.

The implication for capex modeling: hyperscalers cannot spend faster than they can energize, and that gates the 2027-2028 trajectory as much as any demand-side question. When you see a capex guide raised by $10 billion, the relevant follow-up is whether the power for that capacity has been signed, not whether NVIDIA can deliver the chips.

The New Cloud Unit Economics

The shift from CPU-dominant cloud workloads to GPU-dominant ones has changed unit economics at every step from balance sheet to API price. Three structural changes deserve careful framing.

First, depreciation accounting is now a material lever, and the hyperscalers have used it. A GPU costs roughly $40,000 in marginal terms today (Blackwell HGX-class, all-in); with network, power infrastructure, and shell allocations the effective capex per useful GPU is meaningfully higher. That asset is depreciated on a straight-line basis over a useful life the company chooses. Microsoft and Meta both extended useful-life assumptions for server hardware in 2022 (from 4 to 6 years) and Microsoft extended again in 2023 — each extension flowed straight to operating margin because annual depreciation dropped. AWS made similar extensions over the same window. The lever is real and has been pulled. It cannot be pulled indefinitely — at some point the chips genuinely retire — and the question for 2027-2028 modeling is whether useful life can be extended further or whether reality catches up.

Second, AI tokens-per-dollar is improving rapidly, but unevenly across the stack. Token output costs have fallen roughly 80-95 percent for equivalent-quality outputs over the past 18 months across the major providers (OpenAI, Anthropic, Google), driven by model efficiency (MoE architectures, distillation, smaller dense models matching larger ones), inference stack improvements (we covered the engine landscape in our LLM inference benchmark), and hardware generation improvements. But the cost the hyperscaler experiences per GPU-hour is falling much more slowly than the price per token — because utilization and software improvements compound multiplicatively while the underlying capex bill is sticky. That gap is where hyperscaler AI gross margins are made or lost.

Third, long-term compute commitments from OpenAI, Anthropic, and similar labs are now a significant share of forward hyperscaler revenue. Microsoft’s OpenAI relationship anchors a meaningful share of Azure AI revenue; Amazon’s Anthropic stake and Project Rainier anchor a meaningful share of AWS AI revenue; Oracle’s Stargate-related commitments are documented in its RPO disclosures and form the largest single concentration in its book. These are not normal cloud contracts — they are multi-year, take-or-pay-flavored commits backed by labs whose own revenue is growing fast but is not yet matching the capex they have committed to. This is the most important risk factor in the entire stack: if any one of those labs slows materially, the hyperscaler underwriting the capacity has limited optionality, because the assets have been built.

The honest summary of unit economics: hyperscaler AI margins are positive and improving for inference workloads, are still under pressure for training workloads, and depend on continuing token-volume growth to absorb a depreciation curve that gets steeper each year. The 2026 number is fine; the 2028 number requires demand to keep compounding.

Bear-Case Math: What If Demand Slows

A disciplined analysis needs an honest bear case. There are three orthogonal risks worth treating individually.

Oversupply risk

The hyperscalers and the GPU clouds (CoreWeave, Crusoe, Lambda, and the new entrants) are building capacity against an aggregate demand forecast that no single party owns. If the forecast is wrong by 15-25 percent, the resulting overhang could be substantial. The historical analogy that comes up — and that should be treated carefully — is the 2000-2001 telecom long-haul fiber buildout, where committed capex against an extrapolated demand curve collided with reality and left an oversupply that took years to absorb. The key difference is that AI inference demand has visible enterprise pull today that long-haul fiber largely did not, and inference scales with usage rather than provisioned capacity. The similarity is the multi-year asset life against an unstable demand curve.

Depreciation hangover

If utilization drops below 55-60 percent across a 2026-vintage GPU fleet, the depreciation bill becomes harder to absorb. The 6-7 year useful-life assumption that helped margins on the way up becomes a drag on the way down, because the cost stays on the income statement whether the chips are running or not. A meaningful write-down or impairment is possible in a hard-bear scenario; it is not a base case, but it is not far-fetched either. We discussed the gap between AI capability claims and actual deployment in our fact-check on AI replacing engineers, which is relevant context for how enterprise demand actually translates to GPU hours.

Model-efficiency improvements eroding need

This is the most subtle risk. Every 6-12 months, the frontier of useful model size for a given quality target shrinks — distillation, mixture-of-experts, quantization, and architectural innovation all reduce the FLOPs required for a given task. If efficiency outpaces demand growth, capacity needs flatten even as usage rises. That is not a hypothetical; it has been the trajectory for 18+ months on inference. The bear case is that this trajectory continues, on-device inference (Apple, Qualcomm) absorbs more workload, and 2026-vintage data center GPU capacity ends up partially underutilized for general-purpose inference even as agent workloads grow.

The bull-case rebuttal to all three: agents, multimodal workloads, and enterprise pull are still very early in adoption, and each new modality (video, robotics, scientific computing) historically expands the addressable compute envelope faster than efficiency can compress it. That is a defensible position. The honest analyst’s stance is that the range of plausible 2027-2028 outcomes is wide, and the asymmetry favors the bear case more than the bull case rhetoric of 2024-2025 implied.

What This Means for Enterprise Customers

For enterprise IT and platform leaders, the practical implications of the capex cycle are concrete and increasingly negotiable.

Pricing is now multi-dimensional. On-demand GPU pricing has fallen on most hyperscalers through 2025-2026 as supply caught up; reserved and committed-use pricing has fallen further. For customers with predictable annual consumption above ~$2-5 million on AI workloads, 1-3 year committed-capacity contracts now carry meaningful discounts — frequently 30-50 percent off on-demand list. The hyperscalers want the committed revenue to underwrite their own capex; you can use that.

Workload portability matters more than it did. The GPU-vs-ASIC fragmentation across hyperscalers means that a workload running on AWS Trainium is not trivially portable to Google TPU or to NVIDIA elsewhere. Frameworks that abstract this (PyTorch with XLA, OpenXLA, custom inference engines) reduce switching cost, but the work is real. For new workloads, the question of “what runtime do we commit to” is now a strategic procurement choice. For data-format choices that affect portability, our lakehouse table format comparison covers a parallel architectural debate.

Power-aware geography is back. For customers building private AI capacity or negotiating colocation, the question “can this site actually deliver 50 MW in 24 months” matters again. Hyperscaler regions have implicit power answers; private builds do not.

Sustainability claims need engineering scrutiny. The 24×7 carbon-free energy claims being made by hyperscalers vary substantially in rigor — some are time-matched, some are annual-volume-matched, some are PPA-based without delivery matching. For customers with their own Scope 3 reporting obligations, the difference matters and is increasingly auditable.

The strategic point: enterprises are no longer price-takers on AI infrastructure. The buildout has created enough capacity, and enough competition between hyperscalers and the GPU clouds, that committed-volume customers have negotiating leverage that did not exist in 2023-2024.

Trade-offs and Open Questions

Honest analysis requires naming what we do not know.

The actual NVIDIA-versus-in-house split is not publicly disclosed at the granularity that would let us model margin precisely. Each hyperscaler discloses capex but not the silicon mix; analyst estimates vary by 10-15 percentage points in either direction.

The depreciation lever has limits that are not yet visible. Extending useful life from 4 to 6 to 7 years has flowed through margins. Extending to 8 or 9 years would require the chips to actually be useful that long — and Blackwell-generation chips depreciating into a Rubin Ultra world might or might not retain that economic life.

The OpenAI/Anthropic dependency is more concentrated than the headline numbers suggest. A meaningful share of forward-committed hyperscaler revenue rests on the financial trajectory of a small number of AI labs. That is a real risk and not one the hyperscalers can fully insure against.

Power constraints are knowable but timeline-sensitive. Whether transmission queues clear in 3 years or 6 years is the difference between continued capex growth and a forced plateau. We do not know which.

We have not predicted stock prices, and this article should not be read as investment advice on any of the named companies. The aim is to give engineering and platform leaders an honest map of the capex landscape they are operating in.

FAQ

How much will hyperscalers spend on AI capex in 2026?
Public guidance and Q1 2026 actuals put the combined 2026 capex for Microsoft, Amazon, Alphabet, Meta, and Oracle in the range of $330-355 billion, with the AI infrastructure share of that ranging from roughly 55 percent (Alphabet) to 80+ percent (Oracle). Total global AI infrastructure capex including Chinese hyperscalers and the GPU clouds is tracked by Dell’Oro at roughly $400+ billion for the same period.

Is NVIDIA still the dominant supplier to hyperscalers in 2026?
Yes. Custom ASIC programs (Maia, Trainium, TPU, MTIA) have grown every year and serve real production workloads, but NVIDIA remains the largest single recipient of hyperscaler compute spend in 2026. The custom-silicon share is rising gradually, not in step changes.

What is the biggest constraint on hyperscaler AI buildout?
Electrical power, transmission interconnect queues, and the multi-year permitting cycles that gate them. GPU supply has caught up; power has not. Several hyperscalers have shifted site selection geography to follow new firm power capacity rather than legacy network hubs.

How do hyperscalers account for depreciation on AI hardware?
Most have extended useful-life assumptions for server hardware to 6 years (Microsoft and Meta in 2022, with Microsoft extending again in 2023; AWS and Alphabet over the same window). Each extension reduces annual depreciation expense and supports operating margins; the lever has been pulled and has limits.

What is the bear case for AI capex?
Three risks combine in a hard-bear scenario: oversupply if demand forecasts overshoot by 15-25 percent, a depreciation hangover if utilization drops below 55-60 percent, and model-efficiency improvements compressing FLOPs-per-task faster than demand expands. The base case still has demand absorbing capacity, but the bear case is more credible than 2024 narratives implied.

Further Reading

• LLM Inference Benchmark Q2 2026: vLLM, TGI, SGLang, Triton
• NVIDIA Spectrum-X AI Ethernet Fabric for 100k GPU Clusters
• Fact-Check: Did AI Replace Software Engineers in 2025?
• Apache Iceberg vs Paimon: Lakehouse Table Formats 2026
• OpenUSD for Industrial Digital Twins: Architecture Guide 2026

References

1. Microsoft Investor Relations, Q4 FY25 earnings call (July 2025) and Q1 FY26 earnings call (October 2025) — microsoft.com/en-us/Investor
2. Amazon Q4 2024 earnings release and Q1 2026 results — aboutamazon.com/news/company-news
3. Alphabet Q1 2026 earnings — abc.xyz/investor
4. Meta Q1 2026 results — investor.atmeta.com
5. Oracle Q3 FY26 earnings release — investor.oracle.com
6. Dell’Oro Group, AI Infrastructure Capex tracking and 2026 forecast — delloro.com
7. IEA, Electricity 2024 / Data Centres and Data Transmission Networks report — iea.org/reports
8. Synergy Research Group, hyperscale capex tracking commentary
9. Omdia, AI server and accelerator forecasts
10. Bloomberg and Financial Times reporting on hyperscaler capex through Q1 2026