AIFEATURE

AI Capex Nears $725 Billion in 2026: Rational Arms Race or Bubble?

E
EffectStory 編輯部Editorial Team
Published · Updated
Microsoft, Alphabet, Amazon, and Meta are on pace for roughly $725 billion in combined 2026 capital spending, up 77% from 2025, and NVIDIA's data center revenue confirms real near-term chip demand. But Bain projects an $800 billion annual revenue gap by 2030, investor Michael Burry alleges depreciation-driven profit inflation at Oracle and Meta, and a NVIDIA-OpenAI-Oracle financing loop raises structural questions even as bulls cite self-funded cash flow and scarce data center capacity.

How large is the 2026 capex plan across the four hyperscalers, and who is scaling fastest?

Microsoft, Alphabet, Amazon, and Meta are on track to spend roughly $725 billion combined on capital expenditure in 2026, a 77% increase from about $410 billion in 2025CITE:E1. Individual 2026 guidance ranges are wide but directionally consistent across all four companiesCITE:E1.

Company2026 Capex Guidance
Microsoft$175B–$190B
Alphabet$195B–$205B
Amazon$200B–$220B
Meta$130B–$145B
Combined (four hyperscalers)~$725B, +77% YoY vs. ~$410B in 2025

Among the four, Meta's guidance represents the steepest increase: its $130 billion to $145 billion plan for 2026 is nearly double its $72.2 billion actual spend in 2025CITE:E2.

Is the spending converting into real revenue? What does NVIDIA's data center growth show?

NVIDIA's data center segment generated $89 billion in revenue for its fiscal 2027 second quarter, ended July 26, 2026, up 117% year over yearCITE:E3. NVIDIA guided the following quarter's revenue to $108 billion, indicating that hyperscaler capex is, for now, showing up directly as chip salesCITE:E3.

Can this spending generate enough revenue? How large is the projected 2030 revenue gap?

Bain estimates the AI industry will need about $2 trillion in new annual revenue by 2030 to support the current scale of investment, leaving a projected shortfall of roughly $800 billion per year even after factoring in AI-driven cost savingsCITE:E4. This figure is a consulting-firm forecast, not a realized resultCITE:E4.

Are there accounting disputes in earnings? How much do extended depreciation schedules inflate profit?

Investor Michael Burry has accused hyperscalers of extending the useful-life assumptions on AI chips and servers, which understates depreciation expense and inflates reported profitCITE:E5. Burry estimates the cumulative underestimation at roughly $176 billion across 2026–2028, with 2028 profit overstated by about 27% at Oracle and about 21% at MetaCITE:E5.

Burry's EstimateFigure
Cumulative underestimated depreciation, 2026–2028~$176B
Oracle 2028 profit overstatement~27%
Meta 2028 profit overstatement~21%

These figures are Burry's personal calculations rather than restated or audited resultsCITE:E5.

Is capital flowing in circular financing? How do NVIDIA, OpenAI, and Oracle form a funding loop?

NVIDIA, OpenAI, and Oracle have entered a chain of deals that critics describe as circular financingCITE:E6. NVIDIA announced an investment cap of up to $100 billion in OpenAI but actually participated at roughly $30 billion in one funding round, a commitment later reported to have been paused in January 2026CITE:E6. Separately, OpenAI signed a data-center contract with Oracle worth up to $300 billion, and Oracle in turn purchases chips from NVIDIA, closing the loopCITE:E6.

DealAmount
NVIDIA's announced investment cap in OpenAIup to $100B
NVIDIA's actual participation in one funding round~$30B
OpenAI's data-center contract with Oracleup to $300B

Critics have compared this structure to "vendor financing" arrangements from the dot-com era, though deal terms and status continue to shiftCITE:E6.

How do bulls rebut the bubble thesis? Do self-funded cash flow and data center scarcity hold up?

NVIDIA CEO Jensen Huang says "compute is revenue" and that demand is accelerating, a position JPMorgan Asset Management supports by noting that current AI investment is funded primarily through hyperscalers' own cash flow rather than external debtCITE:E7. JPMorgan Asset Management also points to a data center vacancy rate of about 1.6%, arguing that capacity remains undersupplied relative to demand — a structural contrast with the dot-com era's externally financed overbuildCITE:E7.

What do industry leaders say about return on AI investment? What does an early-stage adoption curve imply?

OpenAI CEO Sam Altman has said AI adoption remains at an early stage, describing the current moment as still the "Palm Pilot" phase rather than having reached an "iPhone moment"CITE:E8. Coming from the leader of one of the largest AI labs, the framing suggests that even insiders are not yet claiming the investment payoff is provenCITE:E8.

What this means

The $725 billion in combined 2026 capex and NVIDIA's $89 billion in realized data center revenue (with $108 billion guided for the following quarter) show that spending is, in the near term, converting into chip salesCITE:E1CITE:E3. But Bain's projected $800 billion annual revenue gap through 2030 implies the industry-wide payoff required is far larger than what current results demonstrateCITE:E4. Burry's depreciation critique and the NVIDIA-OpenAI-Oracle financing loop both raise questions about how much of the reported revenue and profit is durable versus circularly generated within a small set of counterpartiesCITE:E5CITE:E6. Bulls counter with self-funded cash flow and a 1.6% data center vacancy rate as evidence of genuine scarcity rather than speculative overbuildCITE:E7, while Altman's own "early stage" framing indicates that even the industry's most prominent builder has not yet declared the return on this investment settledCITE:E8.

📊 Evidence

FAQ

How large is the 2026 capex plan across the four hyperscalers, and who is scaling fastest?

Microsoft, Alphabet, Amazon, and Meta are on track to spend roughly $725 billion combined on capital expenditure in 2026, a 77% increase from about $410 billion…

Is the spending converting into real revenue? What does NVIDIA's data center growth show?

NVIDIA's data center segment generated $89 billion in revenue for its fiscal 2027 second quarter, ended July 26, 2026, up 117% year over yearCITE:E3.

Can this spending generate enough revenue? How large is the projected 2030 revenue gap?

Bain estimates the AI industry will need about $2 trillion in new annual revenue by 2030 to support the current scale of investment, leaving a projected shortfa…

Are there accounting disputes in earnings? How much do extended depreciation schedules inflate profit?

Investor Michael Burry has accused hyperscalers of extending the useful-life assumptions on AI chips and servers, which understates depreciation expense and inf…

📎 Sources

  1. valueaddvc.com
  2. datacenterdynamics.com
  3. nvidianews.nvidia.com
  4. bain.com
  5. cnbc.com
  6. bloomberg.com
  7. finance.yahoo.com

Related data

Author's TakeEffectStory 編輯部

The number worth watching isn't the $725 billion aggregate — it's whether NVIDIA's own guidance keeps climbing. Data center revenue hit $89 billion last quarter and the next quarter's guidance already points to $108 billion, a run-rate that keeps pace with, but does not close, the $800 billion annual gap Bain projects for 2030. The depreciation dispute sharpens the stakes: if Burry's estimate that Oracle's and Meta's 2028 profits are overstated by 27% and 21% respectively holds up, then the cash flow JPMorgan Asset Management cites as proof this buildout is self-funded is less solid than it looks, since profit and cash generation both roll through the same depreciation assumptions. The concrete indicator to track next earnings cycle: whether NVIDIA's actual quarterly revenue lands at or above its $108 billion guidance. A miss there — not the size of the capex headline — would be the first hard evidence that hyperscaler spending has outrun demand rather than tracked it.

E
EffectStory 編輯部Editorial Team

Related

FEATURE

DRAM Prices Nearly Doubled in a Quarter: Why This Memory Supercycle Is Different

General-purpose DRAM contract prices jumped roughly 93–98% quarter-on-quarter in Q1 2026, with NAND Flash close behind. The surge stems from memory makers diverting capacity to HBM and AI servers under long-term agreements, leaving standard DRAM supply unable to catch up. Micron's 2026 HBM output is already sold out through 2027, and record revenue at SK Hynix and Micron shows the demand is structural, not a typical inventory-driven cycle.

EffectStory 編輯部 ·
FEATURE

Sovereign AI: How Gulf States, Nordic Funds, and Private Capital Are Racing to Own the World's Compute

Saudi Arabia's HUMAIN, the UAE's G42-led Stargate UAE, and Qatar's QIA have each committed multi-billion-dollar deals to build national AI compute, led by HUMAIN's $10 billion AMD partnership and a plan for up to 600,000 NVIDIA GPUs. But US export approval currently caps HUMAIN's chip access at roughly 35,000 GB300-class units, while Norway's NBIM and SoftBank show that not all "sovereign AI" exposure means building a data center.

EffectStory 編輯部 ·
BRIEF

SK Hynix Eyes ₩10 Trillion Japan Plant as Semiconductor ETFs Rally

SK Hynix (SK海力士) is planning a ₩10 trillion plant investment in Japan, a move that coincided with gains in two Japan-focused semiconductor ETFs. The plan follows a first half of 2025 in which the company posted ₩39.87 trillion in revenue, ₩15.10 trillion in net profit, and a debt-to-equity ratio that fell to 48.13% from 62.15% at the end of 2024.

EffectStory 編輯部 ·