Big Tech Will Spend $725 Billion on AI This Year. The Profit Depends on a Number Nobody Checks.

The scale nobody is arguing about

Start with the number everyone agrees on. Google, Amazon, Microsoft, and Meta are on pace to spend $725 billion combined on AI infrastructure in 2026 — up 77% from $410 billion the year before. It is the largest peacetime capital buildout in corporate history, compressed into twelve months.

Source: Yahoo Finance, capex spending analysis / company Q1 2026 earnings guidance.

Nobody serious disputes that this money is being spent. The argument is over what it means for the income statement — and that argument runs through a single line item that most earnings-call listeners skip past: depreciation.

The lever: useful life

Useful life is the number of years a company assumes a piece of hardware keeps earning before it gets replaced. Stretch that assumption, and the same capital expenditure gets spread across more years — which shrinks the depreciation expense hitting the income statement every quarter. Same hardware, smaller number on the P&L. Every major hyperscaler has been quietly doing exactly this.

Source: Cerno Capital / Level-Headed Investing, “Are AI Chip ‘Useful Lives’ Creating Useless Earnings?”

This works on paper. It only breaks if the chips do not actually last that long — and Nvidia's own release calendar is the strongest evidence that they do not.

Nvidia's calendar says otherwise

Nvidia has moved to an annual product cadence: Hopper shipped in 2022, Blackwell in 2024, Rubin lands in 2026, and Rubin Ultra is already slated for 2027 — a new architecture roughly every 1.3 years, each one an order-of-magnitude jump in performance per watt, not an incremental bump. Investor Michael Burry has put a specific number on the gap: he argues the real usable life of this hardware is about 3 years, not the 5 to 6 years hyperscalers are booking.

What the gap costs, in dollars

Run the math on a 3-year assumption instead of the 5-to-6-year one hyperscalers currently use, and the gap is about $176 billion of understated depreciation between 2026 and 2028 alone. At firms like Meta and Oracle, analysts estimate that overstates reported profit by more than 20%. One-fifth of the profit, give or take, is a scheduling choice rather than a cash result.

# depreciation_gap.py — one year's AI capex under two useful-life assumptions
CAPEX_2026 = 725_000_000_000  # $725B in AI infrastructure capex, 2026 alone

for life_years in (6, 3):
    annual = CAPEX_2026 / life_years
    print(f"{life_years}-year schedule: ${annual/1e9:,.1f}B of depreciation per year")

gap = CAPEX_2026 / 3 - CAPEX_2026 / 6
print(f"\nAnnual depreciation gap for this one cohort: ${gap/1e9:,.1f}B/year")
print("Multiply across every year hyperscalers keep raising capex, and the")
print("gap compounds into the ~$176B (2026-2028) and ~$1T (2026-2031) figures")
print("reported by Cerno Capital and Goldman Sachs respectively.")

$ python depreciation_gap.py
6-year schedule: $120.8B of depreciation per year
3-year schedule: $241.7B of depreciation per year

Annual depreciation gap for this one cohort: $120.8B/year
Multiply across every year hyperscalers keep raising capex, and the
gap compounds into the ~$176B (2026-2028) and ~$1T (2026-2031) figures
reported by Cerno Capital and Goldman Sachs respectively.
✓ OK: single-cohort math checks out ($725B / 3 − $725B / 6 = $120.8B)
⚠ FIX: real schedules stagger multiple years of overlapping capex —
  that's why the reported multi-year gap ($176B, 2026-2028) is larger
  than any single year's cohort gap shown here

Goldman Sachs ran the sensitivity

Goldman Sachs published the full sensitivity through 2031 in a May 2026 note. Assume a 5-year GPU life, and cumulative industry depreciation comes to roughly $3 trillion. Assume 3 years — the number Nvidia's own release calendar points to — and it is closer to $4 trillion. One trillion dollars, hinging on a single accounting assumption.

A trillion-dollar swing on the expense side has to be matched by an even bigger number on the profit side for the math to still work. Goldman's analysts estimate hyperscalers would need to generate roughly $1 trillion a year in profit to hold the return on capital they have historically delivered at this spending level. Current analyst consensus sits near $450 billion. That is not a rounding error — it is a 2.2x gap between what the spending requires and what the market currently expects.

The market already flinched

Gaps that size do not stay quiet indefinitely. Alphabet dropped 10% in a single trading day on June 22, 2026, after guiding $180 to $190 billion in 2026 capex against a Street estimate of $119.5 billion — its Q1 free cash flow had already fallen 47% year over year to $10.1 billion. Amazon's free cash flow went from $26 billion to $1.2 billion, year over year, in the same quarter it committed to a $200 billion 2026 capex plan.

Source: Techtimes, “Alphabet Drops 10% on AI Capex Fears” / 24/7 Wall St, Amazon FCF collapse

The money hasn't stopped moving

None of this has slowed capital deployment. Together AI raised $800 million at an $8.3 billion valuation in the same week — up from $3.3 billion just 16 months earlier. Nobody is betting that AI stops mattering. They are betting on which set of numbers they believe: Nvidia's release calendar, or the balance sheet's useful-life assumption.

If you want to run this math yourself

The depreciation_gap.pysnippet above is intentionally simple — a single-cohort, straight-line model. If you are building a dashboard that tracks hyperscaler capex and depreciation disclosures against Nvidia's release calendar in real time, the pattern is the same one an always-on research agent uses for any recurring data pull: fetch the latest 10-Q depreciation footnote, re-run the sensitivity, flag when the gap moves.

That is the kind of task a BrainClawagent is built for — it runs on its own isolated VM with shell access, scheduling, and persistent memory, so it can pull new earnings disclosures on a cron job and re-run this exact comparison every quarter without you opening a terminal. And if you are wiring that agent up to multiple models for research, extraction, and summarization, MegaBrain routes each call to the right model at zero markup, so a recurring research job like this one does not quietly turn into a frontier-model bill for a task a mid-tier model handles just as well.

Sign up at getmegabrain.com to route your own agent's calls at cost, or spin up a BrainClaw agent to keep watch on the numbers in this piece long after they change.