随笔

"Cloud Earnings Confirm It: Compute Is Still Scarce, and TSMC Sits at the Choke Point"

2026-07-31

I. What Microsoft, Amazon, and Alphabet Just Said

Between July 22 and July 30, all three major hyperscalers reported Q2/fiscal-Q4 2026 earnings, and the message from each was the same: cloud demand is growing faster than they can build capacity for it.

  • Microsoft (Q4 FY2026, reported July 29): Revenue of $90B (+18% YoY); Azure grew 43% and crossed $100B in annual revenue for the first time; Microsoft Cloud overall surpassed $214B for the year. CFO Amy Hood told analysts plainly: "Demand continues to exceed available supply." Microsoft added 31 new datacenters in the quarter (88 for the year) and another gigawatt of capacity — doubling its total capacity in just two years. Quarterly capex hit $41B, and management guided fiscal 2027 capex to roughly $175B, a sharp step up from fiscal 2026.
  • Amazon (Q2 2026, reported July 30): Revenue of $200.6B (+20% YoY); AWS sales jumped 37%, beating the 31% analysts expected and marking AWS's fastest growth since 2021. AWS's AI and custom-silicon (Trainium) businesses each now exceed a $25B annual revenue run rate. Amazon raised its 2026 cash capex guidance to roughly $220B (up from ~$200B, partly on higher memory costs) — and said that even at that spending level, it still won't have enough capacity to meet demand in 2026, and the same will be true in 2027. Capacity for 2027 is already largely reserved by customers; some 2028 capacity is already spoken for.
  • Alphabet (Q2 2026, reported July 22): Revenue of $119.8B (+24% YoY); Google Cloud revenue surged 82% to $24.8B, with contracted backlog ballooning to $514B. CFO Anat Ashkenazi told analysts, "We're still in a supply-constrained environment," and said Alphabet will lean more on third-party (rented) capacity in Q3 as a "bridging strategy" while it builds out its own infrastructure. Alphabet raised full-year 2026 capex guidance to $195–205B, up from $180–190B just one quarter earlier.

II. The Pattern: Three Companies, One Bottleneck

Three different management teams, three different businesses, and all three used almost identical language: demand is outrunning supply, and it isn't close to resolving. None of them offered a date when the constraint might ease — Hood explicitly declined to. Each also raised its capex guidance during the quarter it reported, which is the more telling signal: they are not spending ahead of demand as a bet, they are spending in reaction to demand they can already see and can't fully serve. Alphabet even had to reach for a stopgap (renting third-party capacity) specifically because its own buildout can't keep pace.

This matters for one reason: all of this capacity is built on chips, and a large share of the highest-value chips — and virtually all of the advanced packaging needed to turn them into working AI accelerators — comes from one company.

III. Where the Bottleneck Actually Lives: TSMC

TSMC's own Q2 2026 report (July 16) reads like the supply side of the same story:

  • Revenue hit a record $40.3B (NT$1.27 trillion), up 36% YoY, beating expectations. TSMC raised its full-year 2026 revenue growth guidance to above 40%.
  • High-performance computing — which includes the AI datacenter accelerators that Microsoft, Amazon, and Google are all racing to buy — made up roughly two-thirds of Q2 revenue and grew about 20% sequentially.
  • CoWoS, the advanced packaging step required to bundle compute dies with HBM memory into a finished AI accelerator, is "extremely tight and sold out through 2026," per CEO C.C. Wei. Lead times at TSMC's advanced-packaging backend run 52–78 weeks, with bookings already extending into 2027.
  • TSMC has been expanding CoWoS capacity aggressively — roughly 80% per year, from ~35,000 wafers/month at the end of 2024 to ~75,000 by the end of 2025, targeting 125,000–130,000 by the end of 2026 — and even so, the gap between demand and available capacity is only expected to narrow from ~20% today to ~10% by year-end.
  • On the durability of the trend, Wei said: "Our conviction in the multi-year AI megatrend remains very high," and when pressed for an update to the company's prior 5-year AI-revenue CAGR guidance (mid-to-high 50s%, given in January), he wouldn't give a new number but said it's "stronger and stronger."

Put together: the three biggest buyers of AI compute all say they can't get enough of it, and the company that manufactures and packages the chips underneath that compute says its own expansion — running at roughly 80% capacity growth per year — still can't fully close the gap until at least late 2026. That is about as direct a confirmation as this kind of thesis ever gets from two independent sides of the same supply chain, in the same two-week window.

IV. Why This Belongs in a Long-Term Position, Not a Trade

None of this means TSM's stock won't fall on a given earnings day — the July 23rd entry in this journal already noted that TSMC's shares dropped after this same report, purely on capex-anxiety sentiment, even though the underlying quarter was excellent. That is exactly the kind of noise a long-term holder should expect and ignore.

What actually matters for a DCA position is the structural picture, and it hasn't changed — it has strengthened. Three independent, deep-pocketed customers are telling their own investors that demand for compute outstrips what they can build, for years out, not quarters. That demand routes overwhelmingly through TSMC's advanced nodes and packaging, which remain sold out well into 2027. A company sitting at a durable supply chokepoint for a good its biggest customers say they structurally cannot get enough of is precisely the kind of moat this journal keeps returning to — not a promise of quarterly stock gains, but a compounding advantage that only needs time to keep working.

References