AMD (AMD) Q3 2026 earnings review
AMD's AI rack sales jump, but inventory drains its cash
AMD's move into full artificial intelligence server racks is driving a big jump in sales while tying up billions in cash. Third-quarter revenue rose 45% from a year ago to $13.4 billion as Data Center sales nearly doubled, and chief executive Lisa Su set a $17 billion fourth-quarter sales plan that points to 66% growth. Profit per share, excluding one-offs, beat what the company's component plans pointed to by our math, because engineering and overhead costs came in $90 million below plan.
| Revenue | $13.38B +45% from a year ago |
|---|---|
| Data Center revenue | $8.55B +97% from a year ago |
| Profit per share, excluding one-offs | $2.09 +74% from a year ago |
| Q4 revenue outlook | Raised to $17B +66% from a year ago at the middle of the range |
⚖️ Verdict: 🟢 Bullish
The story got better because the first Helios AI racks are now running at five major cloud customers and powering a much larger fourth-quarter sales step. Industrial chip sales also grew 29% at a 41% operating margin, giving the chipmaker a second profit engine. The bad news: the rack ramp trims the fourth-quarter gross margin plan to 55%, a $3 billion inventory build cut free cash flow by 64%, and customer warrants triggered a $1.1 billion reported accounting charge.
The question now is how much profit and cash reach shareholders once the new AI racks ship in volume. One view is that multi-billion-dollar rack deliveries will quickly turn the inventory pile into cash and outrun warrant dilution; the other is that lower-margin racks and 320 million penny warrants will cap per-share gains. Fourth-quarter gross margin and cash flow will provide the first real test.
🐂 Bull Case
AI Racks Lift Q4 Sales Plan to $17B
AMD sells central processors, which are EPYC server chips, and graphics accelerators, which are Instinct artificial intelligence chips. It now bundles them into complete server racks called Helios. “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,” chief executive Lisa Su said.
Data Center sales nearly doubled from a year ago to $8.55 billion, making up 64% of the chipmaker's business. Anthropic, Meta, Microsoft, OpenAI and Oracle are now “running models and workloads on the first AMD Helios” systems, and Hewlett Packard Enterprise will deploy the racks for cloud provider Vultr.
That customer rollout backs a fourth-quarter revenue plan of $17 billion, up 66% from a year ago. A year ago, the fourth quarter included $390 million of China AI chip sales that U.S. rules later halted. Excluding those China sales, the plan points to 72% underlying growth by our math.
What to watch: fourth-quarter Data Center revenue and the first-quarter outlook. Another sequential jump will show whether full-rack deliveries are converting customer gigawatt pledges into multi-billion-dollar quarterly sales.
Industrial Chip Recovery Lifts Profit Margins
The Embedded division makes specialized, programmable chips for aerospace, defense, automotive and industrial equipment. After shrinking last year, the unit is growing fast again and carries the highest profit margin in the company. The business began sampling a new space-grade chip designed for 15-year missions, and autonomous-truck maker Kodiak AI chose AMD processors for its next vehicle platform.
- Embedded revenue: $1.11 billion, up 29% from a year ago and 14% from last quarter
- Embedded operating profit: $453 million, up 60% from a year ago
- Embedded operating margin: 41%, up from 33% a year ago
This recovery matters because AI server racks carry lower gross margins than the company average. High-margin industrial chip growth helps pay for the AI ramp and keeps overall profitability steady.
What to watch: the full-year Embedded design-win total in the fourth-quarter report. Clearing the $18 billion mark management targeted last quarter would support another year of high-margin growth.
Revenue Growth Outpaces R&D and Overhead Costs
For the past two quarters, critics worried that rising engineering and sales budgets were eating up the profit from AMD's revenue boom. Finance chief Jean Hu said this quarter's 74% jump in profit per share, excluding one-offs, was “demonstrating the significant operating leverage in our business model as our Data Center business drives scale.”
Operating expenses excluding one-offs came in at $3.56 billion, $90 million below the $3.65 billion the Santa Clara company had planned. On a reported basis, research and development fell by 2.2 percentage points of revenue from a year ago, while marketing and overhead fell by 1.9 percentage points. That cost discipline lifted the operating margin, excluding one-offs, by 5 percentage points to 29%.
What to watch: fourth-quarter operating expenses against the $3.75 billion plan. Hitting that target would lift the operating margin, excluding one-offs, to about 33% by our math and settle the spending concern.
🐻 Bear Case
AI Racks Trim Q4 Gross Margin and Drain Cash
As AMD shifts from selling individual chips to shipping complete Helios server racks, it has to buy expensive memory and components upfront. Management argues that scaling the Data Center business drives expanding profitability, but the new fourth-quarter plan trims gross margin, excluding one-offs, by 1 percentage point to 55%.
- Inventories: $11.46 billion, up 35% from last quarter and 57% from a year ago
- Quarterly inventory cash drain: $3.0 billion, nearly five times last year's $636 million outflow
- Free cash flow: $554 million (4% of revenue), down 64% from $1.53 billion a year ago
The company is tying up billions of dollars in working capital to build systems that carry a lower gross margin than its current average. If AI rack production yields are slow to improve, gross margin will face further pressure.
What to watch: fourth-quarter gross margin and inventory levels. A drop below 55% or another multi-billion-dollar cash drain would show that full-rack economics are tougher than chip economics.
Customer Warrants Cut GAAP Profit by $1.1B
To lock in multi-year AI chip orders from OpenAI and Meta, AMD gave them warrants — options to buy up to 320 million shares at one cent each as purchase milestones are met. The chipmaker excludes warrant costs from its headline non-GAAP numbers, and its slides state that “Q4’26 share count is not expected to be impacted by customer warrants.”
Even before any warrants vest, they hit the reported income statement this quarter. The semiconductor group booked a $1.10 billion charge ($0.66 per share) to revalue the warrant liability, “driven by changes in AMD’s stock price and other valuation assumptions.” Meanwhile, the fourth-quarter share count will rise by 7 million to 1.668 billion, including “potential shares to be issued for the acquisitions of Taalas Inc. and World Labs Technologies, Inc.”
What to watch: the warrant footnote in the quarterly filing. It will show whether the first Helios deliveries triggered any warrant vesting that will dilute future per-share profit.
Risks this quarter didn't answer
Several standing risks around customer concentration, off-balance-sheet commitments and consumer demand remain open or await the full quarterly filing.
- Gaming slump: gaming revenue fell 40% from a year ago to $773 million, pulling Client and Gaming operating margin down 6 percentage points to 15%; watch for console-cycle stabilization in 2027.
- Contracted backlog vs intentions: the release gives no update on firm contracted backlog to compare against the 14 gigawatts of customer deployment intentions; the quarterly filing's revenue footnote will update the figure.
- Data-center lease obligations: last quarter's filing showed over $14 billion of uncommenced leases and $4.1 billion of partner lease guarantees; check the new filing's commitments footnote for additions.
- Export controls: international billing exposure remains; any new trade rules would appear in regulatory filings.
👓 Other Themes
AMD Buys Two AI Startups to Expand Stack
AMD agreed to buy two AI startups before year-end to broaden its technology beyond standard server chips. It will acquire World Labs, “bringing leading model research expertise to AMD” to help design future infrastructure. The chipmaker will also buy Taalas, “adding specialized inference silicon technology” — custom chips built specifically to run trained AI models — as cloud customers increasingly weigh custom chips against general-purpose accelerators.
💲 Other KPIs
Total debt more than doubled from $3.23 billion last quarter after AMD issued $4.75 billion of senior notes in August and repaid $875 million of maturing debt. The borrowing lifted cash and short-term investments to $15.78 billion, giving the company extra liquidity while inventory consumes operating cash.
Uncollected customer bills rose 26.5% from last quarter, outpacing the 16.0% sequential rise in revenue. That build pulled $1.93 billion out of operating cash flow during the quarter, suggesting that a large share of Data Center shipments went out late in the period.
Spending on property and equipment rose 151% from $258 million a year ago, though it dipped from $808 million last quarter. At 4.8% of revenue, up from 2.8% a year ago, the company is spending more on internal lab and testing gear to support rack-scale systems.
🔮 Guidance
Raised. The plan calls for revenue to jump 27% from this quarter and 66% from a year ago at the middle of the range, confirming and sharpening management's earlier call for faster second-half growth. Excluding $390 million of year-ago China AI chip sales, underlying growth reaches 72% by our math. AMD has beaten its quarterly revenue midpoint for four straight quarters by an average of 4.4%; if the pattern holds, Q4 revenue would reach $17.75 billion by our math.
New. The plan has gross margin, excluding one-offs, slipping 1 percentage point from this quarter's 56% and 2 percentage points from the 57% recorded in the fourth quarter of last year. Last year's fourth quarter benefited from a $360 million release of previously reserved China chip inventory, while the upcoming quarter absorbs the early production ramp of Helios AI racks.
New. The company plans to increase operating costs by about $190 million, or 5.3%, from this quarter's $3.56 billion. Because revenue is planned to grow 27% sequentially, operating expenses would drop from 26.6% of revenue in Q3 to 22.1% in Q4, lifting the operating margin, excluding one-offs, to about 32.9% by our math.
New. The estimate of 1,668 million shares is up 7 million from this quarter's 1,661 million. The increase comes from a higher stock price Assumption for employee equity grants and shares to be issued for the World Labs and Taalas acquisitions, with zero customer warrant vesting assumed. Together with the 13% tax rate and $4 million net other income plan, the outlook implies profit per share, excluding one-offs, of $2.92 by our math.
❓ Key Questions
What was standalone Instinct AI revenue in Q3 and Q4?
Data Center revenue reached $8.55 billion, up 97% from a year ago, but management combines server CPUs and Instinct AI GPUs into one number. Without the split, investors cannot tell how much of the $3.6 billion Q4 revenue step comes from Helios racks versus Venice CPUs.
When does the first tranche of customer warrants vest?
AMD took a $1.10 billion charge for customer warrant liabilities in Q3, yet its Q4 share count estimate assumes zero warrant dilution. With OpenAI and Meta now running workloads on the first Helios racks, at what delivery milestone do the 320 million penny warrants begin entering the share count?
How much of the $11.5B inventory is finished racks?
Inventories rose by $3.0 billion in a single quarter, cutting free cash flow by 64% from a year ago. Investors need to know how much of that stock is high-cost memory and components waiting for assembly versus finished Helios racks waiting for customer data centers to open.
Does buying Taalas mean custom chips threaten Instinct GPUs?
AMD agreed to buy Taalas to add specialized inference silicon just as cloud giants build their own custom AI chips. We want to hear on the call whether customers asked for fixed-function inference chips because general-purpose GPUs are too costly to run.