Fiat Elpis · AI capital cycle note 06
AI capex is accelerating because the revenue backlog is accelerating too
Free cash flow is under pressure, but the financing question has changed: hyperscalers are spending into capacity constraints while cloud revenue and contracted demand accelerate.
From 300 recent @FiatElpis posts · expanded with primary-source research
The bearish AI-capex narrative looks only at the cash-flow statement. The bullish narrative looks only at revenue growth. The right comparison is the growth and duration of contracted demand against the cost and useful life of the assets being built.
01 · The information delta
The latest results were demand confirmations, not capex retreats
Microsoft reported that Azure grew 43% in its June quarter and passed $100 billion in annual revenue. Amazon said AWS grew 37% to $42.2 billion, its fastest growth in 18 quarters. Alphabet disclosed that more than half of its cloud backlog should convert within 24 months. Meta continued to raise infrastructure commitments while reporting 28% revenue growth.
At the same time, the cash demands are extraordinary. Amazon’s first-half cash capex reached $96.3 billion. Microsoft said roughly two-thirds of quarterly capex was short-lived assets, mainly GPUs and CPUs. Meta disclosed $68 billion of new long-term data-center lease obligations signed in July.
That combination is not evidence that returns are guaranteed. It is evidence that the cycle has not rolled over. Demand is still arriving fast enough that the largest platforms prefer near-term free-cash-flow pressure to the strategic cost of lacking capacity.
“Backlog and recurring-revenue indicators are growing faster than capex.”
The original Fiat Elpis market note on X
02 · Signals to track
How to test whether the spend is productive
Cloud revenue is reaccelerating
AWS and Azure grew faster on already enormous bases. Capacity is not being added into a visible revenue contraction.
Committed demand provides a conversion clock
Backlog and remaining performance obligations show how much contracted revenue can arrive as power, chips and data centers come online.
Operating leverage has not vanished
Cloud growth and efficiency gains can support operating income even while consolidated free cash flow is consumed by the build.
Balance sheets and debt markets remain open
The hyperscalers can issue investment-grade debt, sign leases and use operating cash. Funding is expensive, but it is not currently scarce.
The correct comparison
Capex is not justified by enthusiasm. It is justified when incremental contracted cash flow outruns the full cost of capacity. The evidence is improving, but the test unfolds over 2027–2028 as today’s construction begins generating revenue and depreciation.03 · What may be mispriced
The market may be early to call the peak
A capex peak requires either saturation, falling customer demand, unavailable financing or a strategic decision to slow. The July results showed the opposite: accelerating cloud consumption, larger backlogs and continued capacity constraints.
The real risk is duration mismatch. GPUs and CPUs turn over faster than concrete, power and network infrastructure. If model efficiency compresses compute pricing or customers over-reserve, the short-lived assets can earn poor returns even while the long-lived campuses remain valuable.
- Compare cloud and AI revenue growth with capex growth on an 18–24 month lag.
- Track backlog conversion, not backlog growth alone.
- Separate land, power and buildings from short-lived accelerators.
- Measure free cash flow after finance leases and capitalized infrastructure commitments.
04 · What would change my mind
What would signal a genuine AI-capex rollover
The thesis changes when demand or financing weakens before the assets monetize:
- Azure, AWS and Google Cloud decelerate together despite new capacity.
- Backlog growth stalls or customers push delivery dates outward.
- Cloud margins compress as depreciation rises faster than revenue.
- Hyperscalers cut capex guidance because utilization, not supply, becomes the constraint.
- Investment-grade spreads or lease terms tighten enough to change project economics.
Bottom line
The burden of proof has shifted to execution
The AI build is no longer supported only by forecasts. It is supported by accelerating cloud revenue, large contracted backlogs and customers paying for scarce capacity.
That does not settle the return-on-capital debate. It means the next bearish catalyst must be evidence of failed conversion—slower utilization, weaker pricing or deteriorating margins—not the size of the capex number by itself.
Sources & method
Primary sources, thesis separated from fact
- Microsoft — fiscal 2026 fourth-quarter earnings call
- Amazon — second-quarter 2026 results
- Alphabet — second-quarter 2026 Form 10-Q
- Meta — second-quarter 2026 results
- Amazon — second-quarter 2026 Form 10-Q
This note expands themes from the author’s recent X posts. Reported facts are linked to their sources; market interpretation is explicitly the author’s view. Market levels may change after publication.