Goldman Sachs published an equity strategy note this week laying out the exact performance target required to keep the artificial intelligence trade intact. Strategist Ben Snider stated that hyperscalers need to deliver another quarter of accelerating cloud computing adoption to prove the buildout has not peaked. Specifically, Goldman equity analysts expect combined cloud revenue growth across Amazon, Google, Microsoft, and Oracle to accelerate from 48% year-over-year in the second quarter to 55% in the third quarter.
To support the thesis that this acceleration is already locked in, market commentary pointed to Micron's latest earnings print. The memory chipmaker added $43 billion in sales compared to the prior-year quarter, citing voracious demand for high-bandwidth memory. Sell-side desks promptly treated Micron's multi-year capacity backlog as verification that end-market cloud monetization is expanding through 2028.
As an options trader looking at balance sheet flows, that deduction makes me uncomfortable.
The primary error in this line of reasoning is treating upstream component absorption as empirical evidence of downstream economic return. When Micron beats consensus and sells out its wafer allocation, that confirms that cloud hyperscalers placed capital expenditure orders six to nine months ago. It shows that server racks are being populated with silicon. It says nothing about whether corporate enterprise budgets are generating cash return on the software running across those racks. Selling picks and shovels measures the purchasing volume of miners. It does not measure the gold content of the ore.
By setting a 55% year-over-year growth hurdle, Wall Street has turned hyperscaler earnings into a digital call option.
In quantitative pricing, a digital option pays out a fixed premium if an underlying asset clears a specific strike price, and drops to zero if it falls short by a single tick. Hyperscaler valuations have been bid to multiples that price perpetual top-line acceleration. If aggregate cloud revenue prints at 56%, the multiple expands and the capital cycle is declared healthy. If aggregate revenue prints at 47%, growth remains massive in absolute dollar terms, but the operating leverage cuts in reverse.
That reverse operating leverage stems from depreciation accounting.
When a hyperscaler buys accelerators and power infrastructure, the cash leaves the balance sheet immediately under capital expenditure, while the income statement amortizes the asset over three to five years. If a company spends $80 billion on compute infrastructure, the annual depreciation expense arrives like a fixed annuity payment regardless of server utilization. When cloud revenue growth accelerates at 55%, the top-line expansion easily absorbs the incoming depreciation wave. If revenue growth decelerates while the fixed amortization charge continues to climb, operating margins compress sharply.
There is also the matter of circular revenue recognition.
A noticeable share of current hyperscaler cloud acceleration is driven by contracted backlog from venture-backed foundation model developers. Hyperscalers injected tens of billions of dollars of balance-sheet cash and equity into frontier research labs. Those labs then committed to spend equivalent sums on the sponsoring hyperscaler's cloud infrastructure. That mechanism converts equity financing into top-line software revenue. It is a legitimate accounting treatment, but it reflects capital market liquidity rather than organic corporate productivity gains. Once those equity reserves burn down, cloud growth must rely entirely on external enterprise software budgets.
The Street's consensus model treats capacity tightness through 2028 as proof that pricing power will remain permanent. That assumption models a single deterministic path with zero parameter uncertainty.
I want the distribution, not the point estimate.
The right tail of the distribution contains the scenario where autonomous agent workflows achieve production reliability, enterprise software buyers replace legacy headcount costs with compute budgets, and hyperscalers easily sustain 55% cloud growth.
The fat left tail contains the reality of IT budget cycles. Enterprise software buyers do not possess infinite price elasticity. If pilot programs fail to generate measurable operational savings, chief information officers freeze additional model routing spend. At that point, the high chip prices celebrated on supplier earnings calls stop being a bullish signal of market health. They become a heavy fixed carrying cost stranded on hyperscaler balance sheets.
Originally published at deanlee.info.
Top comments (2)
The digital-option framing is useful, and I'd add that the barrier is blurrier than "55%" suggests, because the combined number is not one clean series:
So a 55% vs 54% call can flip on how the desk weights four different calendars, which makes the payoff less digital than it looks, and the "who decides it cleared" part carries real model risk.
On the depreciation side, the lever to watch is the useful-life estimate in each 10-K's property and equipment note: moving servers from 5 to 6 years lowers annual depreciation by about a sixth on the same capex, so margins can look better without any change in demand. That is checkable in the filings' XBRL facts.
If you want to pull those facts without a key: canli-mcp is completely free (MIT, runs locally),
npx -y canli-mcp, repos at github.com/arhancanli/canlicapital and github.com/arhancanli/alphac. I built it.The settlement friction on synthetic hurdles is where structured bets quietly fall apart. When the underlying benchmark requires reconciling four staggered fiscal calendars and estimating a dollar base for Azure from percentage prints, the contract stops being a digital payoff on cloud demand. It turns into basis risk on the calculation agent. At seven points away from the strike, discrepancies in calendar weighting or constant-currency adjustments can easily flip the payout before anyone can verify the underlying volume.
The depreciation extension is the other structural blind spot. Stretching server useful life from four to six years creates an immediate accounting tailwind that flatters operating margin, but it conceals accelerated economic obsolescence. If frontier compute cycles render legacy clusters uncompetitive in year three, the remaining book value becomes a deferred write-down waiting for a replacement cycle.