AMD beat every number that makes a headline. Then the market read the cash flow statement and took back the entire day's rally in one evening.
Record revenue of $11.54B, up 50% from a year ago and ahead of the $11.31B consensus. Non-GAAP EPS of $1.66 against $1.61 expected. Data Center more than doubled to $6.7B, up 107%, and is now 58% of the whole company. Q3 guided to roughly $13B versus $12.51B consensus. Nothing in the headline missed.
Capital spending came in at $808M against a $298.6M consensus — 2.7 times the number, and double last quarter. Free cash flow fell to $1.56B from $2.57B in Q1 even though revenue grew 13% over the same stretch. Free cash flow margin halved, from 25% to 14%.
AMD rallied 7% into the print and closed at $518.58. It trades at $471.00 after hours, down 9.2% from that close, erasing roughly $78B of market value. Going in at about 54 times forward earnings left zero room for a cash-flow surprise, and this was one.
Good news for 2028 is a bill that arrives in 2026. The market pays for cash, not ambition.
Picture a food truck that clears $100 a day. The owner announces she is buying a second truck for $60. Nothing about that is bad news — next year she serves two neighbourhoods instead of one. But this year, the cash she hands you drops from $100 to $40. If you bought into the business on the assumption of $100 a day, you paid too much, and the price has to fall to reflect what you are actually receiving right now. AMD just did the corporate version of that.
Capex is money spent on things a company keeps — test equipment, validation labs, buildings, internal compute. It does not hit reported profit all at once. It gets spread over years as depreciation. That is why earnings per share can beat while cash quietly walks out the door, and it is exactly the gap that opened here: the income statement looked excellent, the cash flow statement did not.
Eight lines. Six beat, two blew out. The two that blew out were both spending.
| Line item | Reported | Consensus | Read |
|---|---|---|---|
| Revenue | $11.54B | $11.31B | Beat · +50% yr/yr, record |
| Non-GAAP EPS | $1.66 | $1.61 | Beat |
| Non-GAAP gross margin | 56% | ~56% | In line |
| Non-GAAP operating income | $3.09B | — | +245% yr/yr, 27% margin |
| Data Center revenue | $6.7B | — | +107% yr/yr, 58% of total |
| R&D expense | $2.53B | $2.45B | Above plan |
| Capital expenditure | $808M | $298.6M | 2.7x above plan |
| Free cash flow margin | 14% | — | Was 25% last quarter |
| Q3 revenue guide | ~$13.0B | $12.51B | Beat · +41% yr/yr |
Not a rounding error. A step change in how capital-hungry this business is.
The street modelled $298.6M. AMD spent $808M. That is $509M of unmodelled cash out the door in a single quarter — about 18% of the non-GAAP net income the company just reported. And it is not a one-quarter blip in the trend: first-half capex is $1.20B against $494M in the same half last year.
Revenue grew 50%, so of course the dollar figure grew. The number that strips that out is capex as a share of revenue, and it went from a steady 3.7–3.8% to 7.0% — it roughly doubled in intensity, not just in size. That is the difference between a company scaling and a company changing shape.
AMD did not itemise the $808M in the release, so the composition is inference from what changed in the business — and one thing changed a lot. AMD is fabless, so this is not a fab. It is test and assembly capacity, validation and lab infrastructure, and internal compute for chip design. Critically, Helios is a rack, not a chip. Selling a full rack-scale system means integrating, burning in and validating physical hardware at scale — work a chip designer never had to own. Companies that ship systems carry structurally more capital on the balance sheet than companies that ship silicon.
If that read is right, then the $808M is not an overrun to be apologised for next quarter. It is the price of admission for competing with Nvidia at the rack level instead of the die level, and it should stay elevated. That is a bigger deal than a single-quarter miss, and it cuts both ways: it explains the spend, and it means the spend does not go away.
Up 7% at the close. Down 9% by dinner. The print was not the problem — the details were.
That last box is the one worth sitting with. Against Monday's close, AMD is down 2.8%, not 9%. The 9% headline is measured from a price the stock only held for a few hours, set by traders positioning into the event. Most of what got "lost" after hours was never really banked. That does not make the capex question go away — but it does mean the market's verdict on the quarter is a shrug, not a rejection.
What management said to justify the money.
Neither executive addressed the capex line head-on in the release. What they did instead was point at demand and at 2027 — the argument being that the spend is sized to an order book, not to a hope.
Hu also flagged that the Helios ramp creates near-term margin pressure. The Q3 guide holds non-GAAP gross margin at roughly 56%, so it has not landed yet — but it has been telegraphed. Between that and a doubled capex intensity, management has now signalled that the next few quarters cost money before they make money.
AMD nearly tripled its own forecast for the AI chip market. That is the real justification.
The spending only makes sense next to how big AMD now thinks this gets. In July the company put the total semiconductor market at roughly $2 trillion a year by 2028, with $1.4 trillion of that in AI accelerators alone — against a prior company estimate of about $500 billion. That is not a tweak. It is a near-tripling of the addressable pot, and it reframes $808M a quarter as a small entry fee.
AMD also raised its view of server CPU market growth from 18% to 35% a year, to more than $120B by 2030. The logic: agentic AI does not just need accelerators. Something has to orchestrate the agents and move the data around, and that work runs on CPUs. If AMD is right, its EPYC business — the boring half — gets a second growth curve that nobody is modelling.
| Commitment | Detail |
|---|---|
| Anthropic | Up to 2 gigawatts of MI450 deployed in Helios racks, plus ROCm software work |
| Core Scientific | 15-year agreement valued at roughly $14B, announced earlier this year |
| Microsoft | Helios at scale on Azure, alongside Pensando DPUs |
| Helios launch customers | Meta, OpenAI, Oracle, HUMAIN, Cirrascale, Tensorwave, Vultr |
| Cerebras | Collaboration on low-latency inference |
Four things worth knowing before calling this an overreaction.
What the bears have
- The guide beat consensus but missed the whisper. Roughly $13B cleared the $12.51B published number, but some buyside had penciled in as much as $14B. When a stock runs 9% into a print, the published consensus stops being the bar.
- The year-over-year comparisons are flattered. Q2 2025 carried an $800M MI308 export-control inventory charge. That is why GAAP operating income shows up as +1,585%. Strip it out and the improvement is still large, just not cartoonish.
- Gaming fell 31% to $779M, and not only on the console cycle — supply constraints on Xbox and PlayStation hardware contributed.
- The valuation has no give. 54 times forward earnings going in, about 49 times at $471, and 162 times trailing. At that price a cash-flow wobble is not forgiven, whatever revenue did.
What the bulls have
- Every operating line beat. Revenue, EPS, operating income, and the forward guide. Margins held at 56% while the business grew 50%.
- Data Center more than doubled and is now 58% of revenue. The mix shift toward the highest-value segment is happening faster than the model assumed.
- First-half free cash flow is still up 116% year-over-year, $4.12B against $1.91B. The sequential dip is real, but the trend is not broken.
- The capex has names attached to it. Anthropic, Microsoft, Meta, OpenAI, Oracle, Core Scientific. This is spending against contracted demand, not speculative capacity.
Both readings are defensible. They just have different time horizons.
The bear case is a 2026 case: cash generation halved sequentially, margin pressure has been pre-announced, and the stock is priced for neither. The bull case is a 2027 case: data center revenue more than doubles, Helios ships, and a $1.4 trillion accelerator market makes this quarter's $808M look like nothing. Neither side is wrong about the facts. They are arguing about which quarter matters.
Three things, in order of difficulty. Helios has to ship on schedule and at a gross margin that does not break the 56% line. Data Center has to more than double in 2027, as Su said on the call. And the $1.4 trillion accelerator forecast has to be a market, not a slide. Watch the Q3 print for the first of those — capex staying at or above $800M with gross margin holding is the bullish version; capex staying high while margin slips is the version that hurts.
My levels on the chart.
Johnny's read.
Want the entries, stops, and sizing?
Join the Discord to find out! →AMD Q2 2026 earnings press release and financial statements (Aug 4, 2026, AMD investor relations and GlobeNewswire) · AMD Q2 2026 earnings call · AMD Q1 2026 earnings release for prior-quarter capex and free cash flow · Q2 2026 earnings coverage from CNBC, Reuters, WSJ, Investopedia, TipRanks and BBN Times · capex and R&D consensus figures as reported in Aug 4 earnings coverage · prices, market cap, forward P/E and 52-week range via stockanalysis.com (S&P Global Market Intelligence and CBOE), Aug 4, 2026, 7:25pm ET. Compiled Aug 4, 2026.