Where Does the AI Build-Out Top First: Peaking Is a Wave

The word suggests a single event: one quarter where the number stops climbing. But the build-out spans many companies, and it tops in stages. Nvidia, Micron and the utility that sells the power to run them both are on different clocks. What looks like one peak is really a wave, starting at the GPU and rolling outward through the chain.

The season also opens on a contradiction. Estimates are still rising almost everywhere, and the analyst cuts that would confirm a peak have barely started. Yet the tape has already repriced most of the complex. Oracle trades 57% below its high, Microsoft 29%, while their consensus numbers climb. The earnings and the share prices tell different stories. What follows is an attempt to read both.

The spenders: a peak already in the price

The capex piece asked whether the spending could fund itself. On a trailing-twelve-month basis, which smooths the seasonality, the answer has not improved. Amazon runs capex at 102% of operating cash flow. Oracle runs 174%, on a capex bill of $4.20 for every $1 of revenue growth, with trailing free cash flow deeply negative. Those two are at the cash limit. Microsoft (57%), Alphabet (63%) and Meta (61%) sit below it, and their strain is subtler.

That subtler strain is depreciation deferral. Capex does not hit the income statement when it is spent; it is capitalised and then charged as depreciation over the asset's useful life. A company spending far faster than its existing kit depreciates still books depreciation today. It just books far less than the capex implies, and the rest lands in future quarters. How quickly each dollar is charged depends on one number the company sets itself: the useful life.

That assumption has been moving. Between 2021 and 2023 the hyperscalers extended the assumed life of their servers and network gear toward five and six years. Alphabet went to six, Meta in steps to about five, Amazon to six. That lowered annual depreciation and lifted first-year operating income by more than $10bn across the group. The cost of the kit itself did not change. It was pushed later and spread thinner. But Nvidia's cadence may make an AI accelerator economically spent in two or three years. If so, a five- or six-year book life under-depreciates it and overstates today's profit. Michael Burry put a figure on that in late 2025: roughly $176bn of understated depreciation across the group over 2026–2028. And the practice has split. In early 2025 Amazon broke ranks, shortening a subset of servers back to five years and taking a ~$677m hit, citing the pace of AI. Meta and Alphabet held their longer schedules. The group no longer agrees on what its own assets are worth, which is itself a yellow flag on earnings quality.

Our screen for it is the depreciation-deferral gap: capex as a share of revenue, minus depreciation as a share of revenue. It is a rough directional read, for three reasons: capex covers property and equipment only, reported depreciation also carries intangible amortisation, and leased capacity sits outside both. It marks where the wedge between spending and booked cost is widest: 26% of revenue at Meta, 21% at Alphabet, 17% at Microsoft. Today's depreciation is already in the margin. This wedge is the part still to come, and it arrives as the assets age, sooner if the sceptics on useful life are right.

The last two columns of that table disagree. Every spender's next-year EPS estimate is still being raised or held; the depreciation wave is nowhere in consensus. Yet the tape has already marked the shares down, Oracle and Microsoft hard. The peak is showing up in the price before the earnings. The beneficiaries are the more interesting half of the story, and the rest of this piece is about them.

The peak is a wave

The supply chain tops in sequence, ordered by proximity to the GPU, which is to say the capex dollar. We placed each name on that sequence and scored it on three axes: how fast its growth rate is decelerating, which way its margin is trending, and how much runway its backlog buys.

PEG = forward P/E over expected growth; below ~1 is cheap for the growth, above ~1.5 dear. Micron is marked n/m: its earnings are inflecting off a cyclical trough, so its growth rate makes an unstable denominator, and a low memory multiple flags a peak rather than value.

The two ends of the chain are moving in opposite directions right now. At the tip, Nvidia posted its slowest growth since early FY24 and slid on a data-centre line that merely met expectations at $81.6bn. Its forward growth fades from roughly 80% next quarter to 42% a year out. The level still rises; the rate has rolled.

At the other end, memory is inflecting up. DRAM contract prices are set to rise 62% in 2026 and NAND 75%. SK Hynix ran a 72% operating margin last quarter, and the majors are sold out through the year. Micron's gross margin is up 47 points year on year. When Meta raised its capex guidance it cited memory prices as a reason. The spender's cost line is Micron's revenue line.

But memory rides the same train, a few cars back. Its prices are spiking because supply is physically locked short. The fabs that would relieve it, SK Hynix's M15X, Micron's Idaho plant and Samsung's P5, add no real capacity until mid-2027. High-bandwidth memory (HBM), the stacked DRAM that sits on the accelerators, tightens it further. It takes two to three times the wafer area per bit and starves commodity DRAM alongside it. Bulls note that HBM content per chip climbs each GPU generation, so demand grows even without more units. That growth still needs the buildout to keep buying, though. And the constraint that is the whole bull case has an expiry date. New supply lands around mid-2027, just as the capex growth driving memory demand may itself be cresting. Fresh supply meeting demand that has stopped accelerating is the textbook memory top, and this one is roughly scheduled. A sold-out order book buys time, and the time runs out when the fabs arrive. The wave reaches memory too, on a timetable.

Between the two ends sit the names that gain from the very force pressuring the tip. As hyperscalers design their own silicon to escape Nvidia's margin, Broadcom and Marvell book the design wins. Custom accelerators move share around inside the chain while the chain keeps growing.

The physical layer peaks last: power equipment and generation. Construction and the grid gate it, and they move slower than any chip cycle. Vertiv's backlog is $15bn, up 80% year on year. GE Vernova's is $163bn. Eaton counts 228 gigawatts of data-centre backlog, roughly 12 years of demand at current build rates. Beyond the equipment sits the electricity itself, contracted through Constellation, Vistra and Talen on power-purchase agreements that run for decades. This is the longest-dated claim on the build-out.

What the revisions say

Markets reprice on the change in the growth rate, and that change has a confirming signal: the analyst revision. When a peak is real, next-year estimates stop rising and start to fall. So we checked the 90-day trend in each name's next-year EPS estimate, and the breadth of revisions behind it.

Almost none has fallen. Nvidia's number is up 15% over the quarter, 98% of revisions positive: the growth rate is peaking while the estimate is still being marked higher. Micron is up 53%, most of the spenders higher too. By the one measure that would confirm a peak, a falling estimate, almost nothing has peaked.

A few estimates have gone flat: Oracle, Eaton, Talen. Exactly one has been cut, Constellation, the nuclear utility. Nothing in the AI-power story would put the first crack there, which is why the Constellation story is worth watching.

What the tape has priced

The divergence sits in a single table. The estimates say the peak has not arrived; the price says it is most of the way in.

The tape is repricing a peak the estimates have not yet acknowledged. Price is falling while consensus rises. That gap is what positioning looks like at a top: the marginal holder sells before the analyst downgrades.

Read against the fundamentals, that gap complicates the obvious trade. The instinctive short into a peak is the peaking name. But Nvidia is already down 10% and trades at a PEG of 0.65, so the market has done much of that work. Oracle, the sharpest structural crack in the model, has already broken. Shorting that peak now is a post-mortem.

The mismatches are more interesting than the confirmations. Arista is decelerating and carries the group's richest multiple, a PEG of 2.38, yet it sits at its 52-week high. The roll is in the numbers but not the price. It is the one peak in the complex the tape has not touched.

Broadcom is the mirror. It is accelerating, it gains from the same in-housing that squeezes Nvidia's margin, it trades at a PEG of 0.45, and it has pulled back 17%. The tape has marked its improving fundamentals down.

Constellation is the name where both signals agree: estimates cut, stock broken. It is 38% off its high and still falling while the rest of the power complex rises. The model flags the break but does not explain it. Whether it is power-price normalisation, the cost of restarting Three Mile Island, or contract timing is the next thing worth knowing.

Turning tides

The peak, then, is real, and it is narrow. It has reached the growth rate of the GPU tip and the earnings quality of the heaviest spenders. It has not reached the earnings level of anything, the memory cycle, or the physical build-out. None of it is confirmed by an estimate cut, and yet most of it is already in the price.

That is the pattern across the whole table. Where the peak is visible on the tape, the move is largely behind us. Where it shows only in the fundamentals, it is still ahead.

The earnings are still rising, as they usually are at a top. The turn shows up first in the growth rate and in the share price, and only later in the reported number. In the AI complex, this season, the first two have already turned.

Method notes

• Universe: 5 spenders (MSFT, AMZN, GOOGL, META, ORCL) plus 13 beneficiaries across six cycle positions (GPU → custom silicon → foundry → memory → power equipment → power generation).
• Fundamentals from yfinance quarterly financials. Cash-flow ratios (capex/OCF, capex/revenue, D&A/revenue) are trailing-twelve-month to remove seasonality; single-quarter figures are too noisy (a low-OCF quarter puts Amazon at 170%, TTM at 102%).
• Depreciation-deferral gap = capex/revenue − D&A/revenue (TTM). A directional proxy, not a GAAP measure: capex is PP&E only while the D&A line also includes intangible amortisation, and leased capacity sits outside capex, so the true PP&E-only wedge is generally wider. The gap is positive for any firm investing ahead of its installed base; size and trend are what matter. Useful-life history, the >$10bn first-year benefit, Burry's ~$176bn estimate, and Amazon's ~$677m shortening charge are from company filings and reporting (2023–2025), not computed here.
• Estimate revisions: yfinance eps_trend (current vs 90-days-ago FY+1 EPS) and eps_revisions (up/down breadth, last 30 days). A cut is a 90-day move below −2%.
• Price reaction: yfinance 1-year daily history — drawdown from the 52-week high, trailing return, distance from the 50/200-day average, 3-month relative strength versus the SMH semiconductor ETF.
• Valuation: forward P/E and PEG from yfinance. Memory and foundry carry a cyclical caveat: a low P/E is normal at a cyclical earnings peak (the Molodovsky effect), and it does not signal value. Micron is shown n/m rather than cycle-averaged — with an AI-driven supply shock, past memory cycles are too small and unrepresentative a sample to normalise against.
• Limits: this is a triage framework, not a backtested signal. The revision and price data are a single snapshot (July 2026) in one regime. Forward growth estimates are consensus and can be stale. Merchant-power earnings are volatile, so the power-generation names' readings are carried mostly by the qualitative overlay. Company names illustrate the framework; nothing here is a recommendation.

Not investment advice.