Hyperscalers are committing $700 billion in 2026 and over $1 trillion in 2027. Here's what the math looks like across three scenarios — and why the sustaining capex embedded in the buildout makes the bear case worse than the market is pricing.
The headline number is staggering. Microsoft, Alphabet, Meta, and Amazon have collectively guided to roughly $725 billion in 2026 capex, with analysts at Bank of America and Evercore now penciling in $1 trillion-plus for 2027. Roughly three-quarters of that is AI infrastructure — data centers, GPUs, networking, power.
It's the largest infrastructure bet in corporate history. And the return math is much tighter than the share prices suggest.
Hyperscalers historically earn 15–20% return on invested capital. Their cost of capital sits around 8–10%. To clear that hurdle on $700 billion of new capex, you need to recover depreciation, cost of capital, and operating costs every year on top of the asset base.
Run the math on $700B with five-year depreciation, 10% cost of capital, and operating costs around 25% of capex:
| Component | Annual cost |
|---|---|
| Depreciation ($700B / 5 yrs) | $140B |
| Cost of capital (10%) | $70B |
| Operating costs (25% of capex) | $175B |
| Total annual recovery needed | ~$385B |
To generate $385B in incremental revenue at $25/month blended pricing across consumer and enterprise AI subscriptions, you need roughly 1.28 billion paying subscribers. There are about 1.3 billion knowledge workers on the planet. The subscription math says you need essentially all of them. Every month. No churn.
The actual path runs through cloud consumption, API spend, and enterprise licensing — Microsoft's AI business is already at a $37B annualized run rate growing 123% year-over-year. The bet is real. But the spread between "real" and "enough" is where the next two years live.
Before walking through scenarios, you need to see why this system can spiral in either direction. The flywheel:
Hyperscalers commit capex → NVIDIA and AMD orders surge → HBM memory demand explodes → Samsung, SK Hynix, and Micron raise prices → GPU costs go up 25–33% per generation → hyperscalers need more capex to buy the same compute → hyperscalers commit more capex.
TrendForce's Q1 2026 data shows conventional DRAM prices up 90–95% quarter-over-quarter, server DRAM up 88–93%, and DRAM prices on track to rise more than 70% across 2026. Bernstein's Mark Li described memory pricing as going "parabolic." Micron's HBM is sold out through 2026. SK Hynix has committed $74.8B through 2028, with 80% earmarked for HBM. None of that capacity comes online before 2027.
Then layer in power. Goldman projects a 15% CAGR in U.S. data center power demand through 2030, hitting 8% of all U.S. electricity. Gartner expects power shortages to operationally constrain 40% of AI data centers by 2027. PJM — the grid operator serving 65 million people across the Mid-Atlantic — will be 6 gigawatts short of its reliability requirement by 2027. High-voltage transformer lead times have stretched from 24–30 months pre-2020 to roughly 5 years today. Of the 12 GW of U.S. data center capacity scheduled for 2026, only about 5 GW is actually under construction.
You have a system where AI investment drives up chip prices, chip prices drive up power needs, power needs drive up grid costs, grid costs drive up operating expenses — and all of it rolls into the justification for charging more, which requires enterprise adoption to close the loop.
Enterprise AI adoption accelerates faster than compute costs rise. The 2026–2027 bottlenecks turn out to be temporary — Micron's Idaho fab ramps in 2027, SK Hynix's M15X starts mass production in November 2026, HBM supply normalizes. Cost-per-token resumes its downward trajectory.
On the demand side, every Fortune 500 runs AI agents across legal, finance, procurement, customer service. API consumption — not subscriptions — becomes the dominant revenue model. AWS's 28% growth on a $150B base sustains. Google Cloud's $460B backlog converts to revenue at scale. Model efficiency improves enough that 10x agent adoption only requires 3–4x compute.
The underappreciated piece: the same GPU clusters running AI also serve traditional cloud, HPC, genomics, climate modeling, rendering. If AI takes 70% of the cluster, the other 30% subsidizes the fixed costs. Hyperscalers hit 15–18% ROIC on AI infrastructure by 2029. The capex looks, in retrospect, like Amazon building fulfillment centers in 2012 — early, expensive, completely right.
AI adoption is real but slower and more concentrated than the capex bets assumed. Large companies deploy AI in 3–5 workflows, not 30–50. Knowledge workers absorb AI as a tool rather than being replaced — which is fine for workers but kills the "AI as labor substitute" thesis that justified the investment scale.
The transformer and grid bottleneck persists. The $150–200B in deferred 2026 construction pushes into 2027–2028, creating compressed construction surges and new bottlenecks. Cost overruns on stalled projects push actual capex well above guidance. The $700B 2026 figure becomes $900B+ when change orders, rebid material prices, and labor inflation are counted.
Consumer AI subscription growth hits a ceiling around 500–600 million paid users globally — well short of the 1.3 billion needed to close the math on subscriptions alone. Free cash flow takes a real hit. Alphabet's projected FCF drops nearly 90% in 2026. Amazon turns FCF-negative. Hyperscalers earn 8–10% ROIC on AI infrastructure — covering cost of capital but not generating spread. The market reprices these names from "growth infrastructure" to "capital-intensive utilities." Multiples compress for 2–3 years.
Two or three constraint categories hit simultaneously. PJM goes short on power in a hot summer — rolling blackouts in Virginia, the world's largest data center market. SLA violations propagate. Customer trust cracks. "It's at a crisis stage right now. PJM has never been this short." — the independent market monitor said exactly that in January.
Memory and chip costs spiral beyond model. The next GPU generation (Rubin, late 2026) commands 20–30% higher prices, and if memory costs stay high, that premium grows. BOM for a 2027 cluster runs 40–60% higher than 2025. Every return model built on 2025 hardware economics breaks.
Political risk turns into regulatory risk. The Sanders/AOC AI Data Center Moratorium Act, introduced in March, would halt construction of any data center over 20 MW. Of the 38 states offering data center tax incentives, 28 are considering rollbacks. Local opposition blocked or delayed $156B in projects across 48 sites in 2025 alone.
Enterprise adoption disappoints. Agents run. Tokens flow. But the value is diffuse — a bit faster, a bit cheaper, not transformatively better. Utilization on the most expensive infrastructure ever built sits at 30–40%. Two or more hyperscalers impair $100–200B in AI assets between 2027–2030. NVIDIA's P/E compresses from 35x to 15x. Memory manufacturers face the classic semiconductor bust — they just spent $100B building into a demand cliff.
Here's the part of the analysis that doesn't show up in the headlines, and it's the most important piece for understanding where this actually leads.
The headline capex number ignores the recurring obligation embedded in it. AI data centers aren't built once — they need a continuous refresh cycle to stay competitive. The useful economic life of the equipment is shorter than most people assume.
| Component | Useful life | What drives refresh |
|---|---|---|
| Training GPUs (H100, B200, B300, Rubin) | 3–4 years | Performance-per-watt vs. new generation. Physical hardware lasts longer; economics force replacement. |
| HBM memory | Tied to GPU | Stacked on the GPU package. Replaced when GPU is replaced. |
| CPUs (Grace, EPYC, Xeon) | 5–7 years | Slower performance curve; less aggressive refresh pressure. |
| Networking (InfiniBand, NVLink, switches) | 5–7 years | Bandwidth ceilings vs. cluster requirements. |
| Liquid cooling systems | 7–10 years | Direct-to-chip systems wear from thermal cycling. |
| Power distribution / electrical | 15–20 years | Long life, but expensive when due. |
| Building shell | 30+ years | Effectively permanent. |
The GPU cycle is what kills the math. H100s already dropped from $40,000 to $12,000–18,000 on the secondary market — that's the depreciation curve in action, and the hardware is barely three years old. By year four, training GPUs get pushed to inference duty or sold off entirely.
Roughly 70% of every dollar of data center capex is IT equipment (GPUs, memory, servers, networking). The other 30% is facility (power, cooling, shell). In steady state:
On a $700B installed base, that's $154B per year just to stay where you are. Not to grow. And it compounds — every year of growth capex builds the base that must be sustained 3–4 years later.
It gets worse: sustaining capex isn't immune to cost inflation. If the next GPU generation costs 25% more than the one being retired — which it does — then sustaining capex inflates faster than the underlying installed base. You're paying more to replace the same compute capability. NVIDIA held gross margins at ~84% through the Hopper-to-Blackwell transition rather than absorbing the cost increase. There's no margin compression cycle coming to rescue buyers.
The chart below shows total annual capex across the three scenarios from 2026 through 2035, with sustaining capex broken out separately. The third panel — sustaining as a percentage of total — is where the maintenance treadmill becomes visible.
The 10-year spread is roughly $1.5 trillion per year by 2035. Bull case ends at $2.1T annual, base at $800B, reckoning at $600B. That's the dispersion you're underwriting when you buy NVDA, the hyperscalers, or the memory names today.
The maintenance burden lags growth by 3–4 years and then hits hard. Sustaining capex barely registers through 2027 — pre-2024 installed base was small. By 2030, it's $140–175B across all scenarios because the 2026 vintage is hitting end of life simultaneously. Companies don't choose whether to refresh. Competitive pressure forces it.
In the reckoning case, sustaining capex eats more than half of all spending in 2030. The red line in Panel 3 spikes to 52% as growth capex collapses but maintenance obligations don't. This is the trap. You can't fully fund the refresh, your performance-per-watt slips behind, your enterprise customers churn to whoever still has Rubin-generation hardware, and you spiral down. The companies that overbuilt in 2026–2027 face the worst version of this.
The base case is more painful than it looks. Sustaining peaks at 24% of total spend in 2031. Nearly a quarter of all capex is just keeping the lights on at current capability. That's a profile that looks more like a utility than a growth tech company. The market is currently pricing these names like growth tech.
In the bull case, sustaining stays manageable as a percentage because growth capex grows alongside. In the reckoning, sustaining stays large in absolute terms while growth craters. There is no "soft landing" version of the reckoning — once you commit to the installed base, you're committed to the refresh cycle that comes with it.
| Scenario | 10-year cumulative capex | Sustaining portion | % sustaining |
|---|---|---|---|
| Bull | ~$15.2T | ~$2.0T | 13% |
| Base | ~$7.9T | ~$1.2T | 15% |
| Reckoning | ~$5.0T | ~$0.8T | 16% |
The bear case isn't where the numbers shrink most — it's where the concentration shifts most. In the reckoning case, $785B in maintenance hits in 2029–2032, exactly when revenue is also disappointing. FCF goes deeply negative for 3–4 years. Two or three hyperscalers exit the AI infrastructure race or get acquired by stronger balance sheets. The semiconductor supply chain — which built capacity for the bull case — faces a brutal multi-year inventory unwind.
The current Wall Street model treats AI capex as growth investment that produces a return curve. The lifecycle reality says a meaningful portion of "growth capex" is actually disguised maintenance — replacement of equipment whose economic life has expired.
Once the buildout matures, even sustained $700B annual capex doesn't mean $700B of new capability. It means roughly $550B of new capability plus $150B of replacement. For hyperscaler equity holders, the question isn't "will AI capex stay high." It's "will the net growth capex — total minus sustaining — be enough to justify the multiples." That's a much harder question to answer optimistically.
The bull case probably gets the headlines through 2027. The base case probably gets the FCF numbers through 2028. And the reckoning case — the 25% probability that the circular dependencies become circular failures — is the one no one wants to model.
$700B in 2026. $1T+ in 2027. And a transformer supply chain that needs five years to catch up.
That's the bet.
Sources include Bank of America capex tracking (April 2026), Bloomberg / Statista capex aggregations, Goldman Sachs hyperscaler capex projections, TrendForce memory pricing data (Q1 2026), Wood Mackenzie transformer lead time research, Sightline Climate data center capacity tracking, PJM Interconnection reliability filings (January 2026), Pivotal Research and Barclays FCF projections, IDC and Deloitte inference demand forecasts, and CNBC reporting on hyperscaler earnings (April–May 2026). Scenario probabilities and 10-year capex trajectories are illustrative and reflect Laverton Advisory's analytical judgment based on cited inputs, not point forecasts. Sustaining capex math assumes 3.5-year average IT equipment life, 15-year facility life, and a 70/30 IT-to-facility capex mix.