Nvidia's guidance is already inside its customers' filings
The buyers have already told you what they will spend; the seller's Wednesday call is a formality with a stock price attached.
The contradiction sits in plain sight. On August 26, Nvidia will report fiscal second-quarter results and the market will treat its own forecast, roughly $91 billion for the quarter, as the news of the week (UBS preview, Aug 15). But the news already happened. In the last three weeks of July, four companies that together buy most of Nvidia's chips filed earnings documents stating exactly what they intend to spend in 2026, and those numbers are larger than anything Nvidia's guidance could add. The call is theater; the filings are the play.
Start with what the customers wrote down. Microsoft closed its fiscal year on June 30 with Azure growing 43 percent and a contracted backlog of $678 billion, and spent a record $41 billion on capital expenditure in that single quarter, which puts its spending at roughly $164 billion on an annualized run-rate (Microsoft fiscal Q4 earnings release via Blockspace Media, Jul 29). Amazon used its July 30 earnings call to raise 2026 capital spending to about $220 billion, up from roughly $200 billion earlier in the year, and chief executive Andy Jassy said AWS capacity stays constrained through 2027 (Cloud Computing News, Aug 2026). Meta lifted the floor of its 2026 spending range to between $130 billion and $145 billion after burning $31.1 billion in one quarter, and its free cash flow collapsed to $784 million, the weakest in at least five years (Meta Q2 2026 earnings release, Jul 29). Alphabet raised its full-year range again to $195 billion to $205 billion, against a Google Cloud backlog around $462 billion (Alphabet Q2 2026 earnings release, Jul 27).
Put the four together and you get something north of $700 billion of intended 2026 capital spending: Amazon near $220 billion, Alphabet up to $205 billion, Meta up to $145 billion, and Microsoft running near $164 billion annualized (each company's Q2 2026 earnings release, Jul 27-30). That picture was fixed before Jensen Huang opens his mouth on Wednesday. The only genuinely open questions are whether Nvidia beats its own $91 billion guide, which UBS expects because Rubin GPUs are ramping faster than planned, and how much of the upside was already priced when the customers' filings moved their shares (UBS preview, Aug 15).
Backlogs are the marketing department of capital expenditure.
Separate the trigger from the pressure. The trigger this week is an earnings print and the options market's habitual overreaction to it. The pressure underneath is older and stranger: a small set of buyers is spending money faster than it arrives. Meta took in $61 billion of revenue in the quarter and spent more than half of it on data centers (Meta Q2 2026 earnings release, Jul 29). Amazon tied part of its increase not to new demand but to higher memory and component costs, meaning the bill for the same ambition keeps rising (Yahoo Finance UK, Aug 2026). When your customers must spend more each quarter just to stand still, your growth stops being a forecast and becomes an invoice.
The actors want different things, and the filings expose the seam. Nvidia wants the backlog converted before anyone asks hard questions about returns; its incentive is to guide high enough to hold the multiple but low enough to leave beats in the bag. The hyperscalers want investors to read their spending as evidence of demand they cannot serve fast enough, which is why every one of them volunteers a backlog number: Microsoft's $678 billion, Amazon's $496 billion in contracted AWS business (Amazon Q2 2026 earnings release, Jul 30), Alphabet's $462 billion. Backlogs are the marketing department of capital expenditure. What none of them can show, Meta above all, is the revenue the machines will eventually produce; Meta discloses no AI revenue line anywhere in its results while spending like an AI company (Meta Q2 2026 earnings release, Jul 29).
History offers one bounded comparison. Between 1998 and 2000, Cisco's guidance was treated as the single cleanest read on the internet economy, and its largest customers, the telecom carriers, kept signing purchase commitments right up until the debt that funded them stopped rolling. The filings looked like confirmation all the way down. The lesson was never that demand was fake; it was that customer commitments financed by borrowed confidence reverse faster than any factory can slow. The counterexample argues the other way: Microsoft's cloud buildout of the 2010s ran enormous capital budgets for years against skeptical analysts, and the spending was absorbed by real enterprise demand plus fat software margins. This time the difference is who holds the bag. Microsoft then printed free cash flow to fund its data centers; Meta now does not, and Alphabet posted its first negative free cash flow quarter in 22 years (Alphabet Q2 2026 earnings release, Jul 27).
Walk the chain forward as plain sequence. Next comes Nvidia's August 26 print, landing inside a range the market has already computed from the filings, so the surprise trades smaller than usual and the after-hours move whips both ways. Then the binding constraint migrates from chip supply to everything the chips sit inside: Amazon is raising spending partly because memory costs rose, Jassy says power capacity doubles by end-2027, and Amazon is already shifting workloads between regions to chase electricity (Cloud Computing News, Aug 2026). The scarce goods are HBM memory, transformers, grid interconnects and construction crews, which is why Micron and the utilities trade the story as hard as the chipmaker does. After that comes the squeeze inside the buying club itself: the hyperscalers bid against each other for the same constrained inputs, pushing each other's costs up, which shows up next quarter as another round of guidance increases justified by component inflation rather than demand. That loop feeds on itself until someone's financing breaks or someone's revenue catches up.
Who pays and who profits meanwhile? Nvidia profits now, with a record $75.2 billion in data center revenue last quarter and an $80 billion buyback authorization behind the stock (Nvidia Q1 FY27 investor materials, May 2026). The customers pay now and collect later, if at all. The people in the middle absorb the squeeze twice: tech workers, since Meta and others are cutting staff even as the machine budgets swell (Meta Q2 2026 earnings call, Jul 29), and anyone holding the debt issued to fund data center construction, whose repayment depends on AI revenue lines that mostly do not exist on any income statement yet.
The observable sequence, if this read is right: Nvidia guides at or above $91 billion on Wednesday, the hyperscaler backlogs keep growing faster than the revenue they are supposed to become, and at least one more buyer raises 2026 spending citing component costs before October. The break: any of the four trims a range, or a backlog figure grows slower than the prior quarter's. One trim, not a miss by Nvidia, is the tell, because the filings move first and the supplier's number merely confirms what its customers already confessed on paper.
The judgment the numbers earn: Nvidia no longer tells the market anything the market's biggest spenders have not already filed. Its guidance is a receipt for decisions made in Redmond, Seattle, Menlo Park and Mountain View, and receipts do not lead cycles. When this buildout eventually turns, watch the customers' January filings, not Nvidia's August call, for the first honest word.