Zhihu Review: For Every $100 AI Companies Earn, ~$40 Flows to Cloud Giants

Background

Zhihu @旷野 answered the question “How should we view the phenomenon that for every $100 AI companies earn, nearly $40 flows to cloud giants Amazon, Microsoft, and Google?” His core thesis: the 40% headline number is the least important sentence in the whole report. What actually matters is: ①of the $35-40, about $35 is genuine inference compute cost; the rest is legacy equity/channel arrangements (the OpenAI-Microsoft 20% revenue share, which goes to zero after 2028); ②the clouds’ 35-45% operating margin is on the income statement, not the cash flow statement — depreciation lags capex by 5-6 years and lands in a concentrated wave around 2028; ③Nvidia is the only link in the chain that gets paid before delivering; ④gross vs. net revenue recognition on indirect API will distort any OpenAI/Anthropic revenue comparison; ⑤the core risk is the denominator — AI lab revenue from $7B in 2024 to $690B in 2028, a 5x in two years. If that curve fails, all margin discussion becomes impairment discussion.

Verdict

Bottom line first: the author is one of the few people who actually read the Barclays report carefully — the 40% headline is indeed the least important sentence, and his digging into depreciation, revenue recognition, and the denominator is directionally right. Most of the data checks out, but the “40% is good news” framing is itself a bit of a sleight of hand, and he missed several harder counter-signals from the same week. Score 7.0/10.

📊 Data Verification (as of 2026-09-01)

Verdict: virtually all key data in the article checks out — a rare quality in Zhihu’s finance section.

🔍 Logic Deconstruction

Three points that hold up:

1. Depreciation lagging capex is the most valuable angle in the whole piece — Big-four capex $433.9B vs depreciation $149B leaves roughly $285B/year of cash already spent but not yet on the income statement. The 35-45% cloud margin is front-loaded and diluted. Amazon already cut servers from 6 to 5 years in 2025, demonstrating on its own statements what this looks like — just early and small.
2. Nvidia is the only link in the chain that gets paid before delivering — 75% data-center gross margin, cash collected upfront, clean receivables. A sharp observation.
3. Gross vs. net on indirect API — before comparing OpenAI/Anthropic revenue, you must ask about accounting treatment. The Uber/Lyft analogy is apt.

Three flaws:

1. ⚠️ “153%→90% is good news” is a sleight of hand — The 2024 ratio of 153% was never a steady-state economic ratio: that was OpenAI burning Azure commitments while its own recognized revenue was only $7B — a denominator too small to mean anything. The falling ratio is partly mechanical denominator catch-up, not evidence that unit-economics improvement is all real. The author’s own closing line — “what matters is the denominator, not the split” — quietly undermines his own “good news” thesis.
2. ⚠️ Inference margins jumping from teens to 55-65% in a year, direct API >80%, is Barclays’ model output, not disclosed data — An industry that was subsidizing inference a year ago doesn’t become Datadog-grade in one year; discount that number. There’s also an internal contradiction: if inference margin were truly 80%, after paying 35% to the cloud you’d still have 45%+, and the “40% to cloud” wouldn’t be a story at all — the real burden would have to be training, which is exactly the 48% he identifies. That judgment is right, but don’t treat 80% as realized fact.
3. ⚠️ The China section’s 134% figure is fragile — the model assumes 100% GPU utilization and 90% cache hit rate, which even the authors admit are theoretical upper bounds (industry actual 70%). Drop utilization to 70% and the result goes from -65K/month to -1020K/month. Direction unchanged, magnitude entirely assumption-driven. Reference it, don’t cite it as fact. The structural point — big Chinese tech treats API as a cloud-ecosystem funnel, independent vendors have no cloud to fall back on and can only raise prices — is genuine.

What the author missed — things happening in 2026 right now:

1. Alphabet did a $84.75B equity raise in June 2026 — the largest ever by a listed company. Bond issuance can be called “pre-funding”; dilution on this scale says management itself isn’t confident in the FCF payback timeline. The article talks about bonds throughout and never mentions the equity.
2. Oracle was downgraded to BBB- by S&P in July — reason: surging capex, negative FCF, customer concentration. The credit story is already biting.
3. Barclays’ own models give Meta negative FCF in 2027/2028.
4. Meta was reported to be preparing a cloud business (Zuckerberg: “definitely on the table”) — a signal that internal AI demand can’t absorb capacity, i.e. another player about to dump capacity into the market. Meta +17% in the first half of July as the market rewards a second monetization path; I’d rather read it as an early oversupply signal.

A counterpoint on depreciation:

The author frames 5-6 year depreciation as an accounting trick, but FactSet offers a counter: H100/A100 secondhand prices and contracted GPU lease rates remain firm, meaning the chips still have economic life in the market — 5-6 years isn’t pure fantasy. The real risk isn’t “fraudulent depreciation,” it’s “depreciation betting that scarcity lasts to 2028” — and scarcity is a cyclical phenomenon. The “prepaid receipt” metaphor is right, but the collateral on that receipt is today’s scarcity, and scarcity dissipates.

📌 My Independent Assessment

What “the $700B turning into concrete” actually means

The clouds’ income statement today is a prepaid receipt — every dollar of profit recorded today is borrowed from 2028. Big-four capex $433.9B vs depreciation $149B leaves ~$285B/year of “cash spent, cost unrecognized.” Depreciation over 5-6 years bets the GPUs last 5-6 years; their real economic life is ~3 years (Nvidia ships a new architecture every year; two generations back, there’s no secondhand market). If technology or demand disappoints in 2027-2028, equipment retires early and the excess of book value over recoverable value must be impaired — not slowly, but in one quarter, tens of billions at once. “Concentrated” because all four clouds bought in the same 2024-2026 window, and 2028 is simultaneously the depreciation peak, the retirement point, and — per Barclays’ own forecast — when self-built capacity comes online and clouds lose share. Weak demand plus expiring assets, squeezed from both ends.

Impact on the AI industry

1. Accounting lag buys overinvestment a 5-6 year grace period — at the cost of a more concentrated pop. This directly explains why clouds can borrow $700B and Alphabet can do the largest equity raise in history without stopping: management has done the math — grab territory first, trouble is five years out. The difference from the dot-com bubble: in 2000, the burn showed up on income statements the same year and popped fast; this cycle’s flaw is deferred 5-6 years by accrual accounting, but the misallocation accumulates larger. The bubble isn’t prevented, it’s postponed — with interest.
2. The three players in the capital chain get hurt completely differently — Clouds go from cash printers to debt carriers (~$285B/year unrecognized cost + debt + dilution + impairments); Nvidia goes from “the only one paid upfront” to “the next one worrying about orders” (demand entirely riding on cloud capex); AI labs go from “the ones being drained” to “beneficiaries of a buyer’s market” (self-built compute comes online, compute supply releases, prices fall, training-cost share declines faster — not because AI companies got stronger, but because compute shifted from a seller’s market to a buyer’s market).
3. The denominator decides everything — AI lab revenue $137B (2026) → $690B (2028), 5x in two years. If it delivers, the depreciation wall is just “an affordable expansion cost”; if it delivers half, all margin discussion becomes impairment discussion and the first to bleed are the companies that already signed $700B of debt. What drives that curve is not GPU shipments and not cloud margins — it’s end-user paid demand.

Can the revenue curve deliver? Reality vs. Barclays

• End-2026 ARR $200B: likely achievable (60-70%). The two labs alone are at $105B; add xAI, Google/Meta external API and others, and $200B by year-end is realistic. Critically, growth is accelerating — Anthropic Q2 revenue $11.5B, +143% QoQ, first quarter of positive adjusted operating income. Enterprises pay for output, not tokens; incremental unit economics are genuinely repairing.
• 2028 revenue $690B / ARR $782B: roughly 30% probability. Three reasons: ①the math demands 7.4x in two years, ~170% annualized off a $100B+ base — no precedent; ②the hardest contradiction: Anthropic’s own 2028 guidance is $19-20B (Reuters; its IPO valuation rests on it) — just 3% of Barclays’ $690B pie — meaning the bulk of that denominator isn’t OpenAI/Anthropic at all, either Big Tech’s own reported “AI revenue” or a much wider definition that can’t be independently verified; ③cross-check fails: Nvidia data-center revenue alone is $89B/quarter, ~$360B annualized, while AI lab revenue is $137B — the Barclays model implicitly requires compute-efficiency leap + large price cuts + demand explosion all at once.
• Incremental economics are repairing; the stock is far from it — OpenAI gross margin is only 33%; 2026 inference costs est. $14.1B; cash burn ~$27B in 2026, doubling to ~$63B in 2027. Don’t be fooled by the “80% inference margin”: that’s the best-case incremental scenario; on the stock, training plus total inference cost still pins overall gross margin to the floor.

Which names not to touch — risk tiers

The standard for “don’t touch” is not whether it’s expensive; it’s whose statement blows up first when the denominator fails.

🔴 Tier 1 (first to blow when the denominator fails — don’t touch):

• ORCL — negative FCF, capex at 174% of operating cash flow, already downgraded to BBB-, 15-19 year lease commitments. Highest leverage and weakest cash flow in the chain, building capacity for OpenAI while carrying its own depreciation wall. The 12.9 forward P/E is a trap — it’s earnings projected on the assumption the depreciation wall doesn’t exist.
• PLTR — 70x P/S, 159x trailing P/E. The purest pricing of the AI narrative. No capex, no depreciation — but 70x P/S means perfection is fully priced in; the moment the curve slows, the de-rating is the most violent.
• SNOW — loss-making, 22x P/S, EV/EBITDA -98. Doesn’t build its own compute, pays clouds for compute — the “pure cost side” of the chain, hit from both ends.
• AMD / MRVL — 30/31x forward P/E, 77/65 EV/EBITDA. Second sources and component makers. When clouds cut capex in 2027, second sources get cut first. NVDA eats meat while they drink soup; when NVDA gets hit, they die first.

🟡 Tier 2 (not untouchable, but bad odds — wait for signals):

• META — 13.6 EV/EBITDA, 16.5 forward P/E, looks cheap; but Barclays models negative FCF for 2027-28, and the reported plan to sell cloud capacity means internal demand can’t absorb it. Cheap for a reason.
• AMZN — 20.9 P/E, but forward P/E of 24.5 is higher than trailing — an anomalous signal that the market expects earnings to fall. Capex at 102% of operating cash flow; FCF already negative.
• GOOGL — 17 P/E looks like the cheapest of the clouds, but just did the largest equity raise in history ($84.75B) — management voted with real money that cash isn’t enough. Its profits get eaten by the depreciation wall.
• NVDA — best cash quality, 75% gross margin, 14x forward P/E, the only “paid-before-delivery” player — but that certainty is already priced in. The risk isn’t the statement; it’s the valuation on the day growth drops from 117%.

🟢 Most resilient (not saying buy — relatively): MSFT — 28 P/E / 21 forward, moves capex off the balance sheet via finance leases, FCF down only 28%, Azure backstopped by the OpenAI revenue share. The depreciation wall hits it least among the four — but it’s also the most expensive.

Score

7.0/10. Solid data, the two key holes (depreciation and revenue recognition) dug out precisely, independent judgment (self-build is not free). Deductions: the “good news” framing is a sleight of hand, the 80% inference margin is model output treated as fact, the China 134% figure is fragile, and it missed the same-week counter-signals of equity dilution and credit downgrades.

Sources: Barclays (2026-08-27/28), SEC 10-K/10-Q, Silicon Analysts, FactSet, Bloomberg, Reuters, CNBC, yfinance (2026-09-01)