print-icon
print-icon
Add ZeroHedge as a preferred source on Google

AMD Soars After Blowout $17 Billion Guide... As Free Cash Flow Quietly Shrinks To 4% Of Revenue

Tyler Durden's Photo
by Tyler Durden
Authored...

Three months ago, AMD beat on every line and the stock crashed anyway. Tonight, the market got the number it wanted.

After a brutal month in which AMD shares lost more than a quarter of their value, pummeled first by Intel's PC warning and then by the broader AI capex hangover, Lisa Su delivered the one thing that matters in this tape: a guide that clears the whisper, not just the consensus.

That is $700 million above the $16.3 billion sellside consensus, and right on top of the $17 billion Jefferies' Blayne Curtis had been pounding the table on in the days before the print. The stock, which closed up 3.7% at $483.90 into the event, promptly jumped another 10% to roughly $530 after hours, a move almost exactly in line with what the options market was implying.

And yet, as so often happens with AI darlings in late 2026, the more interesting numbers are buried further down the release. Like the fact that a record $13.4 billion quarter produced just $554 million in free cash flow, a 4% margin, down from 17% a year ago. Or that AMD took a $1.1 billion hit on the warrants it hands out to its own customers.

Beat, Raise... And Then Some

First, the headline numbers, all of which came in ahead of expectations (consensus per Bloomberg/FirstSquawk):

  • Revenue: $13.38 billion vs $13.08 billion expected, up 45% Y/Y and above the high end of the $12.7-$13.3 billion guided range
  • Adjusted EPS: $2.09 vs $1.97 expected, up 74% Y/Y
  • Adjusted operating income: $3.92 billion vs $3.68 billion expected; operating margin 29% vs 28% expected
  • Adjusted gross margin: 56%, in line with guidance

Then the segments, where the AI story did all the heavy lifting:

  • Data Center: a record $8.55 billion, up 97% Y/Y and 27% Q/Q, comfortably above the roughly $8.3 billion the Street was looking for (Polymarket traders, for what it's worth, had 79% odds on a print north of $8 billion). Segment operating income jumped 168% to $2.88 billion, a 34% margin vs 25% a year ago.
  • Client & Gaming: $3.72 billion, down 8% Y/Y. Client rose 7% to $2.95 billion, but gaming revenue cratered 40% to $773 million.
  • Embedded: $1.11 billion, up 29% Y/Y, helped by aerospace & defense, auto and industrial.

Put differently, data center is now 64% of AMD's revenue, up from 47% a year ago, while the legacy PC and gaming businesses are shrinking in absolute terms.

Which, considering what Intel just told the market about PC demand, is probably a feature, not a bug... as it's doing all the heavy lifting...

Lisa Su was predictably triumphant:

"Demand for our Instinct and EPYC products continued to strengthen, and we expect growth to accelerate further in the fourth quarter as we scale Helios and 'Venice' production. With our leadership product portfolio and deep customer and supply chain relationships, we are confident in our ability to sustain very strong revenue growth and expanding profitability over the coming years."

Helios Hits The Gas: The Napkin Math On Q4

The $17 billion guide implies 27% sequential growth in a single quarter, and roughly 65% Y/Y growth from last December's ~$10.3 billion. For a company AMD's size, that is extraordinary.

Where does it come from? AMD doesn't guide by segment, but the math is not hard. Assume Client & Gaming stays flat-to-down at around $3.6 billion (holiday season or not, Intel's warning and "unprecedented" memory pricing per Apple's Ternus are not exactly tailwinds) and Embedded grows modestly to ~$1.15 billion. That leaves roughly $12.2 billion for data center, a sequential jump of about 43% (approximate, our math, not AMD's).

That would be the first full quarter of the Helios rack-scale ramp built on the MI450, which AMD started shipping in September. On that front, the slide deck name-checks Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure as already running models and workloads on the first Helios racks, which is about as complete a list of the AI money-spenders as one can assemble.

It also puts Su's target of more than doubling data center revenue in 2027 in context. Through nine months, data center revenue is $21.0 billion; add our ~$12 billion Q4 estimate and 2026 lands around $33 billion, which means "more than double" implies something like $66-70 billion next year. Susquehanna's Christopher Rolland is already at $74 billion for 2027 ($33 billion of server CPUs and $41 billion of GPUs).

The one blemish in the guide: adjusted gross margin of about 55%, below the 55.5% consensus and down from 56% in Q3.

That is probably the cost of selling full racks (with plenty of third-party content inside) rather than just chips: more revenue, less margin per dollar. Expect Su to be asked about it on the call; expect the answer to be some version of "mix."

The Part Nobody Will Mention On The Call: Where's The Cash?

Regular readers will recall that last time around, in "AMD Crashes Despite Big Beat, CapEx Almost Triple Expectations" (Aug 4), the stock puked after hours because Q2 capex came in at $808 million, almost triple the $298 million expected, and free cash flow missed by 20%.

This time capex actually eased to $648 million. Cash flow got worse anyway: operating cash flow halved to $1.2 billion from $2.37 billion in Q2, and free cash flow collapsed to $554 million from $1.56 billion. The FCF margin? 4%, down from 14% last quarter, 17% a year ago and 25% in Q1.

The culprit is working capital, and it's all on the balance sheet:

  • Inventories jumped 35% in one quarter, to $11.46 billion from $8.47 billion
  • Accounts receivable rose 26%, to $9.21 billion from $7.28 billion
  • Total debt more than doubled to $7.08 billion, thanks to the $4.75 billion bond deal AMD sold in August

In other words, AMD built about $5 billion of inventory and receivables in a quarter in which it generated $1.2 billion of operating cash. The bulls will say that's what a ramp looks like: you build racks before you ship them, and hyperscalers pay on their own schedule. They may well be right, and the Q4 guide suggests the inventory is headed somewhere. But it's worth remembering who the customers are: per Goldman, the hyperscalers just posted a combined negative free cash flow of $31 billion in Q3, down from roughly $60 billion a quarter as recently as Q3 2025.

When your customers are burning cash and your receivables are growing faster than your revenue (26% vs 16% Q/Q), the question of who funds the gap becomes rather important. Something we discussed at length last week in "Debt Window Slams Shut: Meta Taps Shareholders For Up To $95 Billion, With Its Own Chip Suppliers Buying In".

A $1.1 Billion Warrant "Footnote"

Which brings us to the other line tucked into the GAAP-to-non-GAAP reconciliation: a $1.097 billion charge for the "change in fair value of warrant liability relating to the quarterly remeasurement of warrants issued to certain customers." That's $0.66 per share, or the entire gap between GAAP EPS of $1.46 and the adjusted $2.09 everyone is cheering.

The best-known of those customers, of course, is OpenAI, which back in October 2025 got a warrant for up to 160 million AMD shares as part of the chip deal that sent the stock up 24% in a day. We had a few thoughts at the time:

A year later, the circle has only gotten tighter. In October alone, data center developer Fermi jumped 20% after striking a credit support pact with AMD, and AMD joined Nvidia and Broadcom in indicating interest in buying up to $5 billion of Meta's stock offering:

So, to recap: AMD gives customers equity (warrants), gives developers credit support, buys its customers' stock, and then books the resulting sales. The irony is that the warrant charge grows as AMD's stock rises, so the better the AI trade does, the bigger the GAAP hit. Diversification at its finest. (AMD says customer warrants are not expected to affect its Q4 share count, for what it's worth.)

None of this is new to regular readers: we have been warning since last October, in "AI Is Now A Debt Bubble Too, Quietly Surpassing All Banks To Become The Largest Sector In The Market", that the AI buildout increasingly runs on borrowed (and recycled) money. AMD's August bond deal and today's balance sheet are just the latest data points.

Goldman: In This Tape, Beats Don't Get Paid

Of course, none of that matters tonight, because the guide was big enough to clear what Goldman's desk had flagged as a very high bar. Earlier today, Goldman TMT specialist Peter Callahan wrote in his "quick hits" note (available to pro subs) that AMD was the name everyone wanted to talk about:

"all eyes on earnings tonight (esp AMD + ANET)... bogeys (getting tons on AMD and DDOG)... most asked on: AMD 4Q guide expects, ANET CY27 framing and BKNG 4Q guidance"

Callahan also flagged the problem that has dogged every AI infra name this earnings season: visibility beyond 2027.

"quite clear at this point that the market is applying low(er) multiples to 2028 estimates in Semis / AI infra given the lack of perceived visibility – against that backdrop, it does feel like we are getting plenty of '2028' visibility datapoints that 'could' help the AI trade whenever risk appetite improves"

Nelson Armbrust's morning "What Matters Today" note (also available to pro subs) framed the AMD print as one of "two important reads after Intel and the optical sell off," adding that "in this tape, the reaction to the guidance may matter as much as the guidance itself." And the reaction to TMT guidance has been ugly: while the median S&P stock beating EPS estimates outperformed the index by 152bps the next day, TMT beats underperformed by 76bps (Communication Services by a stunning 419bps) per Goldman's portfolio strategy team.

Positioning probably helps tonight's move. Per Goldman Prime, hedge funds rotated out of TMT and into defensives in October, with Info Tech seeing the largest dollar net selling since June. Combine a 26% drawdown, light positioning and a guide $700 million above the Street, and you get a 7% pop. The harder question is whether it sticks: Goldman also sees AI capex growth "peaking," with hyperscaler spend slowing from +96% this year (~$800 billion) to +44% in 2027 ($1.2 trillion), and EPS growth for AI infrastructure stocks peaking at 80% this quarter before slowing to about 40% by late 2027.

Rally Now, Receivables Later

As for the bigger picture, AMD's quarter is a near-perfect microcosm of the AI trade in late 2026: revenue is real, growth is accelerating, and the cash is going somewhere else. The Q4 guide says the Helios ramp is happening; the balance sheet says AMD is fronting a growing share of the bill, in inventory, in receivables and in warrants.

Bulls will argue all of that unwinds in Q4 as the racks ship and the customers pay. Maybe. But with the hyperscalers now free cash flow negative, Meta raising equity from its own suppliers and OpenAI still years from profitability, we suspect the working capital build is the new normal rather than a one-quarter blip. That's fine for as long as the stock keeps going up. Then again, as AMD just showed, when the stock goes up, so does the warrant bill.

Goldman Sachs analyst James Schneider expects the stock to trade higher following stronger results and guidance. 

We believe investor expectations had been lowered significantly following Intel's weak guidance last week. Key highlights from the results included continued strength in Datacenter, gross margin outperformance relative to expectations, and in-line OpEx outlook. 

On the conference call, we expect investors to focus on: (1) the shape of MI450 ramp at key customer accounts in 4Q/2027; (2) visibility into sustainability of Server CPU growth in 2027; (3) gross margin trends; (4) PC demand environment.

Goldman is Buy-rated on AMD with a 12-month target price of $700 based on a 32X P/E multiple applied to our normalized EPS estimate of $22.00. 

Key downside risks include: (1) slower-than-expected adoption of agentic AI; (2) slower-than-anticipated deployment of AMD GPUs related to the Meta deal; (3) share erosion of x86 architecture in enterprise AI and (4) lack of operating leverage.

The call starts at 1730ET.

0