15 min read

Advanced Micro Devices (AMD) — Q3 FY2026 Analyst Q&A Breakdown

Fiscal period: Q3 FY2026 · Report date: November 3, 2026

Question-by-Question

  • Tom O’Malley, Barclays

    • The Question: Opened on the raised 2027 outlook — company revenue growing more than 80% with data center becoming more than 80% of the business (his reading of the new markers) — asking first whether that growth is backed by supply (wafers, packaging, substrate, memory), and second whether demand would outstrip even that supply if AMD could source more.
    • Management’s Response: Lisa Su confirmed multi-year joint planning with customers and supply-chain partners — “We’ve been planning for it for the last several years” — then stated the scarcity plainly: “It is true that demand is higher than supply.” She committed to keep working supply-chain partners to increase supply and reaffirmed the markers: total company growth greater than 80%, server approximately doubling “from a fairly high point in 2026,” and continued strong customer visibility in data center AI.
    • The “Analyst’s Take”: Direct, and the scarcity admission is the news — management effectively said the 2027 markers are supply-capped, so any supply upside is upside to the markers. One quoted fragment carries the whole answer; no new numbers were needed.

  • Tom O’Malley, Barclays (follow-up)

    • The Question: With client guided down more than 20% in Q4 — in line, he notes, with what others in the market have signaled (analyst-provided market color) — how long does the PC correction run given normal Q1 seasonality, why has AMD outperformed so far, and how does it exit the trough into next year?
    • Management’s Response: Lisa Su framed the downturn as component-price driven: higher memory and component costs raise system prices, and consumers are the sensitive buyers. AMD had signaled a sub-seasonal second half on prior calls and is aligning sell-in with sell-through to keep inventory healthy. New planning detail: the 2027 assumption of a double-digit PC unit decline, with focus segments named — enterprise (where AMD is underrepresented), premium consumer, desktop channel. The genuinely new structural point: “our supply chain is quite fungible across our PC business and our data center business,” letting AMD shift supply as market signals move.
    • The “Analyst’s Take”: Direct and candid about a down year — no hedging on the decline. The fungibility point is the strategic tell: PC weakness reallocates toward data center rather than idling, which is why management can hold a bullish company-level marker against a bearish PC assumption.

  • Vivek Arya, Bank of America Securities

    • The Question: Two-part. (a) What confidence does AMD have that anchor customers have the financing and physical infrastructure — land, power, shells — to support “roughly 3 gigawatts or so” of planned deployments into 2027 (an analyst-provided figure management did not correct)? (b) As competitors invest in the ecosystem — residual-value guarantees, product fungibility — does that change the 2028 competitive landscape?
    • Management’s Response: On readiness, Su offered deployment-level visibility: “we are very much working hand in hand with our customers. So we know where the GPUs are going in terms of which data centers and which locations.” On 2028, she set a hierarchy — first principle is investing in AMD’s own growth (R&D, supply chain, capacity), then “selectively investing with our lead partners” to lock in data center infrastructure multiple years in advance. She did not engage the residual-value-guarantee comparison beyond that framing.
    • The “Analyst’s Take”: Confident and direct on 2027 readiness — knowing GPU-by-GPU data center locations is the strongest readiness claim available short of disclosed contracts — and deliberately measured on 2028: “selective and strategic” quietly distances AMD from open-ended checkbook commitments without criticizing rivals.

  • Vivek Arya, Bank of America Securities (follow-up)

    • The Question: How do frontier-lab customers conceptually choose between custom ASICs and AMD products — cost per token, power efficiency, software flexibility, time to deployment — and are there workloads where an AMD GPU works better than the alternatives?
    • Management’s Response: Anchored on TCO: “That is fundamentally the most important factor” — tokens per dollar at deployment scale. New milestone framing: Q3 was the first quarter MI455 and Helios racks shipped for revenue, with “a big step-up in Q4” and a continued ramp through 2027. New roadmap detail: co-design beyond MI450 — MI500 and MI600 are being tuned to what frontier labs need. She also conceded a positive: “from a general purpose GPU standpoint… we have more fungibility” across workloads.
    • The “Analyst’s Take”: Direct, with the co-design disclosure (MI500/MI600 co-engineered against frontier-lab workloads) as the substantive new counter to the ASIC thesis. She offered no workload-specific win claims, keeping the answer at the economics level.

  • Joshua Buchalter, TD Cowen

    • The Question: How should HBM costs flow through the Instinct model — he observes AMD’s revenue-per-gigawatt outlook has held consistent in the “$15 billion to $20 billion per gigawatt range” (analyst-provided figures; management did not confirm them) — is that long-agreed program pricing, and can AMD pass higher input costs through even to its “big three” customers?
    • Management’s Response: Su confirmed long-term agreements exist with suppliers and customers, then widened the frame beyond memory: “not just the memory market, by the way, other component costs are also seeing some inflationary pressures.” The new commitment: AMD is “working with our customers to… ensure that those costs are shared as we go forward” — cost pass-through is being arranged, not silently absorbed. She closed on TCO: leading tokens-per-dollar for AI workloads even with the inflationary pressures.
    • The “Analyst’s Take”: The cost-sharing admission is the news — pricing is actively managed to protect TCO rather than defended as fixed. But the actual question, why revenue per gigawatt has held so consistent, went unaddressed; she answered the pass-through half only.

  • Joshua Buchalter, TD Cowen (follow-up, to Jean Hu)

    • The Question: OpEx came in light in Q3 and the Q4 guide — but with 2027 revenue expected at more than 2x the long-term target model (analyst’s framing), how should OpEx growth be modeled, and can AMD hold the target gross-margin range, “sort of that 55% floor” (the analyst’s characterization of the target), as Helios ramps?
    • Management’s Response: Jean Hu gave a new quantified OpEx commitment: for 2027, OpEx “to grow less than the top-line revenue growth rate, less than half of the top-line revenue growth rate,” driving very significant operating-margin expansion. On gross margin: Q3’s data center AI came in better than expected entering the quarter; Q4 guides data center AI to surpass server — “a little bit dilutive to corporate average, but the gross margin dollars are increasing very significantly.” For 2027 she declined a floor: the mix between data center AI and server “actually determine what the company’s gross margin will be. Right now, I actually don’t have that visibility,” offering Q4 as the benchmark and noting embedded strength helps the corporate average.
    • The “Analyst’s Take”: Direct on OpEx — the less-than-half commitment is directly modelable — and evasive-by-design on gross margin: she reframed a floor question into a mix narrative and admitted no visibility, a real concession dressed as confidence.

  • Timothy Arcuri, UBS

    • The Question: How does 2027 data center split? He cites the server CPU raise from up 70% to up more than 100% and asks whether the GPU business will still be bigger in dollars than server CPU — “but not much” — and whether the GPU business, though lumpy, is ratable across the year.
    • Management’s Response: Su reframed the numbers as “markers”: total company greater than 80%, data center CPU “about double,” and — the new quantifier — data center AI will “significantly more than double.” She confirmed Q4 is the first quarter data center AI crosses over the CPU business in size and expects “both businesses to grow nicely throughout the year,” addressing shape-of-year without a quarterly path.
    • The “Analyst’s Take”: Direct, with “significantly more than double” the sharpened disclosure the prepared remarks lacked (they said only “grow even faster than we projected 90 days ago”). The instruction to “just consider them markers” is the hedge to carry into models.

  • Timothy Arcuri, UBS (follow-up, to Jean Hu)

    • The Question: How do the customer warrants layer through the year — any guardrails for rolling them into a model through 2027?
    • Management’s Response: The fullest warrant framework of the call, in three parts: (1) reported revenue is always net of the warrant revenue reduction, fixed at grant-date fair value — “that one you should not worry about it, it’s always incorporated”; (2) shares enter the diluted count only once all vesting and exercise conditions are met — shipments plus commercial conditions — with explicit share-count guidance promised in 2027; (3) the quarterly fair-value remeasurement is non-cash and does not affect non-GAAP results or free cash flow.
    • The “Analyst’s Take”: Direct and genuinely clarifying — the three-bucket split (revenue / share count / P&L remeasurement) is exactly the modelable guardrail Arcuri asked for, with the timing uncertainty honestly parked in 2027 rather than papered over.

  • C.J. Muse, Cantor Fitzgerald

    • The Question: Decompose the raised ~100% 2027 server growth: how much is units versus mix (higher core counts) versus higher Venice ASPs?
    • Management’s Response: Su gave a clean decomposition: “the growth in revenue is mostly in units and core count, and less so in ASPs, although there’s an ASP uplift as we increase core count.”
    • The “Analyst’s Take”: Direct and unusually concrete for a guidance-composition question — a volume-and-core-count story with ASP as a secondary tailwind, which matters for modeling server gross margin.

  • C.J. Muse, Cantor Fitzgerald (follow-up, to Jean Hu)

    • The Question: Q1 2027 seasonality — is the right read a normal seasonal decline in client and gaming with data center ratable above December, and what is the embedded view?
    • Management’s Response: Hu confirmed embedded typically sees “some seasonality” in Q1, but for full-year 2027 embedded grows significantly and strengthens through the year. The client/gaming and data center premises of the question were not disputed.
    • The “Analyst’s Take”: Brief but direct — the Q1 embedded dip against a strengthening 2027 is the one new, modelable nugget.

  • Stacy Rasgon, Bernstein Research

    • The Question: Size the Q4 GPU-versus-CPU crossover within the ~50% data center guide — barely at the line, or significantly higher — to get the entrance rate into next year.
    • Management’s Response: Jean Hu described the MI450 ramp as “a very significant step up sequentially from Q3 to Q4”; server is growing strongly, but data center AI “will be much higher than server business in Q4” — confirmed “in terms of dollars” when Rasgon pressed.
    • The “Analyst’s Take”: Direct on direction, still unquantified — “much higher” is a magnitude claim, not a number. The entrance-rate question stays open until the Q4 segment print.

  • Stacy Rasgon, Bernstein Research (follow-up, to Jean Hu)

    • The Question: Gross margin in 2027 relative to the Q4 rate — any color on a good entrance rate and a good exit rate given the product ramps?
    • Management’s Response: Hu declined to give a quarterly path: with data center above 80% of company revenue, the mix between data center AI (ramping, supply-constrained) and server (doubling) “actually determine what the company’s gross margin will be. Right now, I actually don’t have that visibility.” Q4 is “a good benchmark” because data center AI is much higher than CPU there; quarterly variation expected in 2027.
    • The “Analyst’s Take”: A clean non-answer — forthcoming about why she can’t say, but zero new numbers. The Q4-as-benchmark framing repeats her answer to Buchalter’s floor question, which is itself the signal: management will not anchor a 2027 gross-margin level.

  • Joseph Moore, Morgan Stanley

    • The Question: Given strong feedback on Helion racks — what is different about the CPU-only version for agentic workloads versus a traditional rack, and does AMD anticipate a GPU-blade attach rate across many of these deployments?
    • Management’s Response: Su described configurability as the point: hyperscalers run all-CPU configurations (“quite dense… Venice is — from a performance per socket standpoint… the best in the industry”), while enterprises mix CPUs and GPUs to optimize for their workloads. The design is fully open — contributed to OCP — which she said OEMs, ODMs, and customers view as a positive. The attach-rate question received no figure or expectation.
    • The “Analyst’s Take”: Direct on design philosophy, silent on attach economics — no attach-rate guidance offered, an easy and worthwhile follow-up next quarter.

  • Joseph Moore, Morgan Stanley (follow-up)

    • The Question: The ~100% 2027 server growth — is AMD meeting demand at that level, or is there a shortage condition through all of next year?
    • Management’s Response: “For us to grow that amount, we’re meeting a lot of demand.” Su re-grounded the TAM: roughly 50% CAGR over the next three to four years across three areas — traditional cloud and enterprise, agentic CPUs and workloads, and head nodes — with share gains claimed in each. But she was explicit that unmet demand remains: “I do believe there’s more demand, and we are actively working with our supply chain partners to satisfy additional demand,” flagging sockets, DDR5, and other components as the gating items.
    • The “Analyst’s Take”: Direct, holding two truths at once — the marker assumes AMD meets a lot of demand, and demand above it still exists. The component checklist (sockets, DDR5) is the concrete scarcity detail.

  • Harsh Kumar, BHO Capital Markets (firm name as transcribed)

    • The Question: Helios — AMD is one of two companies shipping complete racks; what are the share aspirations, given the server-CPU track record against a difficult competitor? And does AMD aspire to address the enterprise market with Helios and the associated setup?
    • Management’s Response: Aspirations stated directionally: “to be a very significant compute supplier to our largest customers, both frontier labs as well as hyperscalers,” with “very active dialogues with our largest customers on MI500 and MI600 and beyond” and the aim of “a larger and larger piece of their compute needs.” On enterprise: adoption is developing slower than frontier labs and hyperscalers, but it is “quite important” — enterprise customers are asking AMD for deployment advice — and AMD is expanding the portfolio with “enterprise-specific GPUs.”
    • The “Analyst’s Take”: Directionally confident but unquantified — no share target offered, consistent with the marker discipline elsewhere on the call. The enterprise-specific GPU line is the one new product signal.

  • Harsh Kumar, BHO Capital Markets (follow-up)

    • The Question: On client — historically well-managed through downturns: what is the share-keep strategy if the chief competitor cuts ASPs — match them, or gauge the situation as it goes?
    • Response: Su framed PCs as strategically important to “end-to-end AI deployment… from cloud to edge to client,” but emphasized discipline: focus on “the highest value and most profitable segments of the PC market — enterprise, premium consumer, desktop channel” — and noted “there’s a good amount of volatility in the low end of the market, and that’s the place where we have the least exposure.”
    • The “Analyst’s Take”: Direct and consistent with the O’Malley exchange — AMD is signaling it will concede low-end price moves rather than chase them: a de-facto mix-over-share stance for a down market.

Synthesis

Clusters

Concern Analysts (Q#s) Uniform management answer Signal
2027 growth composition and supply O’Malley (Q1), Arcuri (Q7), Rasgon (Q11), Moore (Q14) Markers, not forecasts: company >80%, server ~double, data center AI significantly more than double; demand exceeds supply and supply work continues The markers are supply-capped — every upside scenario runs through supply-chain execution, not demand discovery
Client/PC correction and pricing discipline O’Malley (Q2), Muse (Q10), Kumar (Q16) Down-market planning assumption, premium/commercial focus, fungible supply across PC and data center, no ASP matching AMD will trade low-end share for mix and reroute supply to data center
Margin and OpEx trajectory Buchalter (Q6), Rasgon (Q12) OpEx growth under half of revenue growth; gross margin mix-dependent with no 2027 floor, Q4 the benchmark Management models operating leverage confidently but refuses to anchor a gross-margin level through the data center AI mix shift
Customer-warrant mechanics Arcuri (Q8) Three-part split: net revenue (already reduced), share count (deferred to 2027 guidance), non-cash remeasurement Dilution timing is a 2027 disclosure event, not a Q4 modeling item
Competitive substitution (ASICs, guarantees, infrastructure) Arya (Q3, Q4) TCO/tokens-per-dollar first; co-designed MI500/MI600; selective, strategic infrastructure co-investment AMD answers competition with economics and co-design, not guarantees

Deflections

Topic (Q#) What was asked Deflection technique
GPU-blade attach in Helion racks (Q13) Expected GPU attach rate in CPU-only agentic racks Described configurability and OCP openness; no rate given
2027 gross-margin path (Q12) Entrance and exit rates for 2027 Mix explanation plus “I actually don’t have that visibility”; Q4 offered as benchmark
Warrant share-count timing (Q8) When warrant shares enter the diluted count Deferred to explicit 2027 guidance
Helios share aspiration (Q15) Quantified rack-scale share target Directional “very significant compute supplier” framing; no number
Revenue per gigawatt consistency (Q5) Why $15–20B/GW holds under HBM inflation (analyst-provided figure) Answered cost-sharing mechanics; never engaged the per-GW figure

Contradictions

Contradiction Prepared-remarks claim (source) Q&A walk-back
“Secured” supply vs unmet demand Prepared remarks: wafers, substrates, advanced packaging, memory and other components “secured… to support much higher growth next year” Q1/Q14: “It is true that demand is higher than supply” — additional supply still being worked; securing components is not the same as satisfying demand
Markers presented as expectations Prepared remarks: server to double and data center AI to grow “even faster than we projected 90 days ago” Q7: “you should just consider them markers” — directional language, not guidance, even as data center AI sharpened to “significantly more than double”
Margin confidence vs no visibility Prepared remarks: Q4 non-GAAP gross margin ~55% with “gross margin dollars… increasing very significantly” Q6/Q12: no 2027 gross-margin floor or path — “Right now, I actually don’t have that visibility”; the leverage story is secure at the OpEx line, unanchored at gross margin

Unanswered

  • GPU-blade attach rate for Helion deployments (Q13, Moore) — no rate or expectation offered.
  • 2027 gross-margin entrance and exit rates (Q12, Rasgon) — explicitly no visibility; Q4 benchmark only.
  • Warrant share-count timing into the diluted count (Q8, Arcuri) — deferred to 2027 guidance.
  • Quantified rack-scale share aspiration (Q15, Kumar) — direction only.
  • Consistency of the $15–20B revenue-per-gigawatt range under HBM inflation (Q5, Buchalter) — figure unconfirmed; mechanics answered instead.