Nvidia's "Sky-High" P/E Already Collapsed to About 27. Now It's an Execution Bet

Nvidia reported the largest quarter in its history on Wednesday — $96.2 billion of revenue, up 106% from a year earlier and more than $4 billion above the average analyst estimate — and the shares, after jumping almost 9% in the follow-through, gave back half the gain by Friday, ending the week up just over 1%. This is the pattern now: five consecutive quarters of beating estimates, with the stock having fallen after four of the last five reports. The market keeps refusing to pay up — and that refusal, not the multiple, is the story.

To anyone who knows Nvidia only from the headlines, that looks backwards. For two years this stock has been sold as "sky-high P/E," the price you pay for perfection. The market data says otherwise. Around $218, Nvidia trades at roughly 27 times trailing earnings and about 22–25 times forward earnings — a far cry from the 40-plus multiples it wore through the 2023–25 boom. For scale, the S&P 500 sits near 25 times trailing earnings and about 20 times on a forward basis. The most valuable AI franchise on earth now costs about the same multiple as the average large company.

The mechanism, which most "is it cheap or expensive" takes miss, is arithmetic: a P/E falls whenever earnings grow faster than the price. Nvidia's non-GAAP EPS rose about 120% over the past year while the stock rose about 28%. Price went up, and the multiple still fell by roughly half. The market has been quietly de-rating the dominant supplier of AI compute while that supplier kept winning.

That is not proof the stock is cheap. It is proof the old question — is the P/E too high? — is finished. What replaced it is harder: can the growth the market now treats as ordinary survive what Nvidia itself put into its own guidance? The answer runs through a margin number and a country.

The margin number first. Gross margin for the quarter was 75.0%, but Nvidia guided current-quarter gross margin down to about 74%, and management said it expects margins to bottom near 71–72% the quarter after, settling at 72–73% through fiscal 2028. On a business running past $400 billion a year, three points of gross margin is over $12 billion of gross profit. The earnings growth that paid down the P/E is expected to cool precisely where the market set its new, more ordinary expectations.

The reason is the product cycle itself. Nvidia's next platform, Vera Rubin, is expected to deliver around $20 billion of revenue in its first real quarter — about a fifth of total data center sales — a pace the company's CFO called the fastest product ramp in Nvidia's history. Rubin racks carry more expensive memory and a full networking stack, and every new platform launches with costs. So near term, revenue accelerates, margin compresses, and the earnings line gets squeezed between the two.

The country is China. The $108 billion current-quarter guide assumes zero data center compute revenue from China. Chinese data center shipments were already under 1% of data center sales last quarter, so no collapse is being hidden — but a market that could have been an accelerator is being deliberately set to zero.

Then there is the deeper reason a record quarter draws a shrug: who is funding the growth. Data center is about 92% of Nvidia's revenue, sold largely to a handful of hyperscalers planning close to $700 billion of AI capital expenditure in 2026 — a figure the megacaps kept raising through the year, past $750 billion by some July tallies. And Nvidia is now seeding that spending itself: it announced partnerships with BlackRock, Apollo, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion of third-party capital for AI infrastructure, and its supply commitments have swollen to $279 billion, mostly memory. Nvidia returned $26 billion to shareholders in the quarter — then committed hundreds of billions more to the machine that buys its chips.

This is the circular-financing pattern that spooked the sector in late July, when chip stocks shed about $1.3 trillion in market value within days on worries that Nvidia was financing the demand that finances Nvidia. The same facts that make demand look durable — AWS alone agreed to add 2 million Nvidia GPUs across 2027 and 2028, on top of a million already announced — also make the revenue cycle lean on debt-funded construction that has never been through a downturn.

The genuinely encouraging counter-signal is in the small print. Within data center, Nvidia's re-segmented non-hyperscale bucket, ACIE — AI cloud, industrial and enterprise — grew 25% sequentially and 138% year over year, roughly twice the hyperscale growth rate, with sovereign AI more than tripling. The customer base is broadening beyond a few frontier labs. And the recurring software layer that could someday justify a premium multiple again is still small: management has pointed to $10 billion a year of software revenue as a milestone, against a hardware business on a run rate beyond $400 billion. Hardware has set the ceiling; the multiple will not move until the software layer is large enough to matter.

Here is how I read it. Nvidia is no longer priced to reward a bull case; it is priced to reward execution. At a market-average multiple on a business carrying a preliminary view of roughly 70% revenue growth for the next fiscal year, the return story is almost entirely about whether earnings deliver, not about whether the multiple expands. That is a better starting price than most investors have been told to expect — but it is an execution bet, and the conditions are now named: margins bottom near guidance and recover, China returns or ACIE keeps covering the gap, the financing engine holds, and the software layer scales.

A selloff after a record quarter has historically been a place to add rather than a verdict — but only when you know what the thesis depends on. The trailing multiple already prices the doubt; that is the whole point of the 27. What the guidance tells you is what must now go right: the margin line, the China line, and the financing line. Watch those, not the P/E. That is where this thesis lands.