In this record quarter, AMD story has clearly shifted: now it is a data center company that also sells PCs!
Revenue of $11.5B, +50% y/y, +13% q/q, with data center at 58% of total revenue versus 42% a year ago.
EPS grew ~82% y/y on a comparable basis — meaningfully faster than revenue, which is the operating-leverage proof point management wanted to land.
One important framing note from IR: all y/y comparisons exclude the ~$800M of MI308 China export-control inventory charges taken in Q2 2025.
The numbers
| Metric | Q2 2026 | Change |
| Revenue | $11.5B | +50% y/y, +13% q/q |
| Gross margin GAAP | 54% | +360bps y/y, +90bps q/q |
| Adjusting Operating income | $2.3B | 20% margin |
| Diluted EPS | — | ~+82% y/y (comparable basis) |
| Operating cash flow | $2.4B | — |
| Free cash flow | $1.6B | — |
- Gross Profit GAAP of $6.2B, a 54% Gross Margin (note the chart does not exclude MI308 impairment in 2Q25).
- Operating Margin
By segment
- Data center — $6.7B, +107% y/y, +16% q/q. Op income $2.1B (31% margin). EPYC up >70% y/y with cloud and enterprise each above 70%; fifth consecutive record server CPU quarter. Both units and ASPs grew double digits, with units the bigger driver. Instinct more than doubled on the MI350 ramp.
- Client & Gaming — $3.8B, +6% y/y. Client $3.1B (+23% y/y, record mobile; Ryzen Pro +50% y/y). Gaming $779M (−31% y/y on the console cycle, though +8% q/q). Op income fell to $582M / 15% from $767M / 21% — the one clear margin blemish, attributed to go-to-market investment and roadmap expansion.
- Embedded — $977M, +19% y/y, +12% q/q. Best growth in over three years. Op income $386M / 40% margin, up from 33%. Tracking to >$18B of new design wins for the year.
TAM revisions (vs. six months ago)
The forward numbers that actually moved the stock narrative. Management raised essentially every long-term marker:
- Data center AI accelerators: >45% CAGR to ~$1.4T by 2030
- Server CPU: >50% CAGR to ~$220B by 2030
- Total high-performance/AI compute: ~40% annually, approaching $2T by 2030
Q3 2026 guidance
- Revenue ~$13B ±$300M → +41% y/y, ~+13% q/q
- Gross margin ~56% (flat)
- OpEx ~$3.65B
- Other income ~+$55M; tax rate 13%; share count ~1.66B
- Mix: strong double-digit growth in data center and embedded; client roughly flat-to-slightly-up, offset by a strong double-digit gaming decline
AMD-specific targets:
- Server revenue: >80% y/y in 2H26, >70% for full-year 2027 off a much higher base
- Data center segment: more than double in 2027 — and under questioning, Lisa Su repeatedly upgraded this to “well over 100%“
- Company revenue growth: substantially above the prior “>35%” target
- EPS: expect to significantly exceed the $20 target from November’s Analyst Day, within the strategic timeframe
The phrase “tracking materially ahead of the long-term financial model” was the deliberate message of the call.
Product and strategy substance
- Venice (6th-gen EPYC, Zen 6, 2nm) — in production now. Claimed >2x perf/watt vs. leading x86 and up to 3.3x vs. leading ARM CPUs; 30+ SKUs. Every major OEM launching, leading clouds deploying from later this year. Demand described as stronger than any prior EPYC generation.
- Helios (rack-scale: Venice CPUs + MI450 GPUs + Pensando networking + ROCm) — in production, initial shipments beginning this quarter, ramping through Q4 into 2027. Claimed up to 15% more throughput at equal rack power and up to 30% more tokens per dollar than competition. Customer pull “tracking ahead of our initial forecast.”
- Customer commitments: OpenAI and Meta (multi-generation, gigawatt-scale) plus a new Anthropic partnership — up to 2GW of MI450 in Helios, first gigawatt starting 1H 2027, including joint engineering using Claude to optimize workloads and accelerate ROCm. Microsoft will deploy Helios at scale on Azure for frontier model inferencing.
- 2027 roadmap: MI500 + Verano CPUs + Pensando, with expanded scale-up domains in copper and optical. Positioned as the largest generational leap in Instinct history — a claimed >2,000x inferencing performance improvement over four years. Annual rack-scale cadence going forward.
- ROCm — described as at an inflection point. 3M+ models run out of the box, day-0 support for leading open models, 10x increase in open-source contributions. New ROCm.ai development platform lets developers use Claude, Codex and Cursor to port/optimize code for Instinct; claimed >2x training and 3x inferencing performance vs. ROCm 7.
- Client/embedded color: Gorgon Halo (192GB unified memory, runs up to 300B-parameter models), Hugging Face Pro bundled with Ryzen AI Halo systems; Ryzen AI embedded x100 and the Kria robotics platform on the embedded side.
Highlights from Q&A
- CPU vs. GPU crossover and 2027 mix: Su declined to name a crossover date but said both grow substantially; the AI TAM is larger, so as Helios ramps with strategic customers, AI growth in 2027 is the bigger dollar contributor — but “both businesses are very significant drivers.”
- 2027 gross margin puts-and-takes: Jean Hu was direct about the tension: server CPU growth is accretive to gross margin; data center AI carries gross margin slightly below corporate average but adds large absolute gross profit; embedded recovery is an additional tailwind. Net: mix-dependent, but she “feels really good” about navigating it. This is the most important margin disclosure on the call.
- What’s in the Q3 data center guide? Su: both server and AI grow nicely in Q3, and both grow more in Q4 — more server supply comes online and the Helios ramp only starts at the end of Q3, becoming much more substantial in Q4. She rejected the implication that server CPU flattens in Q4.
- Can AMD outgrow the ~40% market CAGR to 2030? Yes — explicitly, company revenue growth above 40%, with above-market growth expected in every segment.
- Gigawatts and monetization. Su didn’t confirm the analyst’s ~3GW / ~$15B-per-GW math directly. She reiterated: server >70% in 2027, total data center >100%; the three anchor customers (OpenAI, Meta, Anthropic) are all ramping; “we have supply to more than meet the guidance… and the upside potential is there.” On monetization: “double-digit billions” per gigawatt, still in that range. Ramp cadence: Q3 is the beginning, Q4 a step-up, Q1 a further step-up.
- Is 50% more HBM a memory-cost exposure? A candid answer. AMD has “very good visibility into HBM allocation” for 2027 from multi-year supplier work. More memory is a real TCO advantage for large models — but the benefit is workload-dependent, and for medium-sized models AMD is willing to modify the memory footprint if the TCO case isn’t there. That’s a notable flexibility admission in a tight memory market.
- Server units vs. ASPs. In Q2, double-digit growth in both, with units the larger contributor. ASP growth comes from higher core counts. AMD has spent several quarters expanding wafers, back-end capacity and substrates — which is why 2H guidance could be raised, with more capacity landing in 2027.
- Was “Helios ahead of forecast” about volume or yields? Volume — demand, specifically for 2027, particularly for inference. On yields: Q3 is the starting quarter, and yields should improve over the following quarters as expected for a highly complex product.
- Sizing the three server buckets. Of the $220B 2030 TAM, the agentic AI “sandbox” segment is the largest and fastest-growing piece by 2030 — but the smallest today. Near-term growth exists there; the real payoff is 3–5 years out. Su emphasized that different workloads need differently optimized CPUs, not one SKU, and claimed Venice leads across general purpose, agentic, and AI head-node categories. Key insight: workload breadth expands from Turin to Venice, which underpins the above-market share claim.
- On client 2026/2027. 2026 client growth y/y is intact. The PC market has held up better than expected, though a 2H decline is planned for. For 2027, optimism rests on the product portfolio, AI PCs, and rising enterprise content — with the honest caveat that the market itself depends on component costs.
- What does “accelerating” mean when Q3 y/y growth is below Q2’s 107%? Jean Hu clarified cleanly: acceleration means 2H versus 1H y/y growth, not q/q comparisons against Q2’s 107%.
- How much work is “more” doing in “more than 70%” and “more than double”? Su, amused, upgraded the language on the spot: server >70% is a strong statement; total data center will be “well over 100%,” driven by data center AI being “well over 100%.”
- Customer diversity in the MI400 ramp. OpenAI, Anthropic and Meta will consume through multiple CSPs, plus additional customers interested in Helios at “more regular scale than gigawatt scale.” Good diversity expected, particularly by Q4/Q1. Timing is being matched to when customers’ data center build-outs are ready.
- Any land/power/shell constraints that could threaten targets? Nothing that gives pause about hitting targets — but Su acknowledged a range, and the range depends on bringing capacity online in time. Silicon and Helios component supply she feels “very, very good” about. The gating factor is data center operators, and she framed the ecosystem as actively trying to accelerate builds.
- On Instinct gross margin entry vs. exit 2027. Hu declined to isolate Instinct margins, repeating that mix between server CPU and AI determines the quarterly path, that the two “have a good offset in general,” and that embedded and client are additional levers. Formal 2027 guidance comes later.
- Is server CPU supply constrained? Yes, and it has been all first half, because much of the demand was unforecasted. 2027 demand is better forecasted, so supply should be better than 2026. AMD feels good about supporting >70% growth, and “there may be opportunities for that growth to go higher.”
- Follow-up on 2nm ramp difficulty amid 3nm tightness. New nodes are always hard, but chiplet architecture means AMD needs fewer wafers on the leading-edge node — a structural advantage. Attention is on the full chain: wafers, back-end, packaging, substrates.
- The Cerebras partnership and fast inference. Helios plus Cerebras’ wafer-scale engine targets the fast-inference niche, available in the Cerebras cloud in Q4 and extending into 2027. Framed as workload-specific optimization within an open ecosystem rather than a major revenue line.
- Follow-up on OpEx growth vs. the 40% market CAGR. Hu: investment continues, but OpEx will grow slower than revenue, by design, to drive operating leverage — consistent with the “significantly above $20” EPS message.
Reading between the lines
- What’s genuinely strong. The server CPU story is the underappreciated one. Cloud and enterprise both above 70%, five consecutive record quarters, unit and ASP growth, a fourth consecutive record enterprise sell-through quarter, and supply — not demand — as the constraint. Embedded returning to 19% growth with a 40% operating margin adds a quiet margin cushion. And management upgraded its own long-term model publicly, which companies rarely do casually.
- What management was careful about. Gross margin. Hu never gave a 2027 number, and the reason is visible in her own framing: data center AI runs below corporate average, and it is the fastest-growing piece. The offsets — accretive server mix, embedded, client — are real but not guaranteed to fully cover. Watch whether the 56% level holds as Helios scales. Similarly, no one isolated Instinct margins despite being asked directly twice.
- The soft spots. Client & Gaming operating margin fell 600bps y/y. Gaming is down 31% and guided to decline strongly again. The PC market outlook for 2H is explicitly softer on memory and component costs — and those same costs are pressuring graphics card prices. Inventory rose to $8.5B, which management frames as demand preparation, though it’s worth tracking.
- The real risk isn’t demand. Across the Q&A, three separate analysts probed supply and execution, and every answer pointed the same direction: demand exceeds what AMD can currently ship. The gating items are HBM allocation, leading-edge wafer and substrate capacity, packaging, and — most outside AMD’s control — whether customers’ data centers have land, power and shells ready on schedule. Su’s “there is a range” answer was the most honest moment on that.
Net. A quarter that beat and a guide that raised, but the more consequential content was the forward framing: TAMs revised sharply upward, 2027 data center guidance verbally escalated from “more than double” to “well over 100%” during the call itself, and an analyst’s ~$30B Instinct estimate told outright it was too low. The bull case rests on execution and supply; the thing to actually monitor is gross margin as the mix tilts toward AI.
Thank you!
CoreValue Research Team
