AMD (Trans): DC to double; 2x+ again by 2027

Below is Dolphin Research's transcript of AMD FY26 Q2 earnings call

I. Key takeaways

1. Q3 outlook: revenue around $13.0bn (±$0.3bn), implying +41% YoY and ~+13% QoQ at the midpoint. By segment, Data Center and Embedded are guided to deliver strong double‑digit growth YoY and QoQ. Client and Gaming are guided down YoY (client growth offset by a material double‑digit decline in gaming) and slightly down QoQ (a small uptick in client offset by a strong double‑digit decline in gaming).

2. Other Q3 financials: non‑GAAP GPM ~56%; non‑GAAP Opex ~$3.65bn; non‑GAAP other income (net) ~+$55mn; non‑GAAP effective tax rate 13%; diluted shares ~1.66bn.

3. 2H and 2027 segment outlook raised: server revenue in 2H26 to grow >80% YoY; in 2027, off a higher base, grow >70% YoY. Data Center revenue in 2027 to grow 'well over 100%' YoY, with Data Center AI also 'well over 100%'. Mgmt says supply is sufficient to meet this outlook, with upside potential.

4. Mid‑/long‑term model and TAM up: Data center AI accelerators TAM to grow >45% CAGR to ~ $1.4tn by 2030; server CPU TAM to grow >50% CAGR to ~$220bn. High‑performance and AI compute market to grow ~40% CAGR in coming years, approaching ~$2tn by 2030. AMD says it is tracking well ahead of the long‑term model shared at last Nov Analyst Day, expects revenue growth to be well above the prior '>35%' target, and to exceed $20 EPS meaningfully over the strategy cycle.

5. This quarter's key metrics (YoY comps exclude ~ $800mn inventory and related charges last year due to U.S. restrictions on MI308 exports to China): Total: record revenue of $11.5bn (+50% YoY, +13% QoQ), marking the sixth straight quarter of >30% YoY growth. GPM 56%, up >200bps YoY and +80bps QoQ; Opex $3.4bn (+40% YoY); OP $3.1bn, OPM 27%; comparable diluted EPS up ~82% YoY.

Data Center: record revenue of $6.7bn (+107% YoY, +16% QoQ), 58% of total (vs. 42% a year ago). OP $2.1bn, 31% margin.

Client and Gaming: revenue $3.8bn (+6% YoY, +7% QoQ). Client $3.1bn (+23% YoY, +6% QoQ) with record mobile processor revenue; Gaming $779mn (‑31% YoY, +8% QoQ). Segment OP $582mn, OPM 15% (vs. $767mn, 21% a year ago).

Embedded: revenue $977mn (+19% YoY, +12% QoQ), the strongest growth in over three years. OP $386mn, OPM 40% (vs. $275mn, 33% a year ago).

6. Cash flow and balance sheet: operating cash flow from continuing ops $2.4bn; FCF $1.6bn. Inventory rose QoQ to ~ $8.5bn to support Data Center demand. Cash, cash equivalents and ST investments at quarter‑end were $13.1bn.

II. Details from the call

2.1 Management remarks

1. Server CPU: demand and share

Server CPU revenue hit a record for the fifth straight quarter, with both cloud and enterprise each up >70% YoY, above prior outlook. x86 server revenue share rose YoY. Customers are expanding deployments of 5th‑gen EPYC Turin and 4th‑gen EPYC Genoa; Turin now powers nearly one‑third of the 1,600+ EPYC public cloud instance types globally.

In the prior quarter, healthcare, financial services, media and tech companies added tens of millions of new cloud instances. Enterprise revenue hit a record with the fourth consecutive quarter of record channel shipments, and on‑prem deployments accelerated; AMD secured multiple large wins across FSI, manufacturing, telco, retail and tech.

More than 230 platforms based on 5th‑gen EPYC from HPE, Dell, Lenovo and Supermicro mark AMD's broadest enterprise lineup to date.

2. 6th‑gen EPYC Venice

Venice, built on the new Zen 6 cores and 2nm process, delivers over 2x perf/W vs. leading x86 CPUs and up to 3.3x vs. leading ARM CPUs, making it one of the biggest gen‑over‑gen EPYC performance jumps. The family includes 30+ processors balancing per‑core and single‑socket performance with diverse memory and IO options.

Venice is in volume production, with major OEM platforms rolling out on plan and leading cloud providers set to deploy starting later this year. Customer demand exceeds any prior EPYC generation. Mgmt sees Agentic AI opening a new growth vector for server CPUs across high‑frequency AI head nodes, high‑density Agentic servers and general cloud/enterprise workloads.

3. Data Center AI and Instinct

Data Center AI revenue more than doubled YoY, with MI355X adoption expanding as leading AI companies scale deployments for both inference and training, and cloud providers broaden MI350 availability. The customer base now spans the largest AI labs, hyperscale clouds, leading AI startups, national labs and sovereign AI builds.

4. Helios platform and strategic commitments

Helios is a rack‑scale AI platform integrating EPYC Venice CPUs, MI450‑series GPUs, Pensando networking and ROCm software. On a wide range of inference workloads, at equivalent rack power, throughput is up to 15% higher than peers, and tokens per dollar up to 30% higher. Helios is in production, with initial shipments slated for later this quarter, ramping in Q4 and continuing through 2027; customer pull‑through is very strong and above AMD's initial forecast.

New strategic partnership with Anthropic: up to 2 GW of MI450‑series GPUs to be deployed in Helios, with the first 1 GW starting in H1 2027. The partnership includes multi‑year co‑engineering to optimize Claude workloads on Instinct GPUs and accelerate ROCm development. AMD is also expanding its long‑term collaboration with Microsoft, which will deploy Helios at scale on Azure for Microsoft, its AI customers, and Azure AI Services inference of frontier models. Prior multi‑gen GW‑scale deployments with OpenAI and Meta continue.

5. Next‑gen platform cadence

AMD plans to launch one new rack‑scale AI platform each year, with meaningful gains in performance, efficiency and TCO per generation. The 2027 platform integrates MI500‑series GPUs, Verano CPUs and Pensando networking, expands the scale‑up domain, and supports both copper and optical interconnects.

Customer engagement on MI500 is very high, with multiple customers co‑planning next‑gen AI infrastructure. AMD expects MI500 to be the largest gen leap in Instinct history, driving >2,000x inference performance over four years.

6. ROCm and software ecosystem

ROCm has reached a key inflection, meeting customer requirements for performance, capability and developer experience for large‑scale AI production. Over 3mn models now run out‑of‑the‑box on AMD, and leading open‑source models offer Day‑Zero support for Instinct; open‑source contributions to ROCm rose over 10x in the past year. Last month AMD launched ROCm.ai, an AI‑assisted dev platform for AMD GPUs that supports building, porting and optimizing code for Instinct using popular coding agents such as Claude, Codex and Cursor.

Across a broad set of models, ROCm.ai delivers >2x training performance and >3x inference performance vs. ROCm 7. AMD is co‑optimizing ROCm with OpenAI, Anthropic and Meta, with improvements flowing back to the broader AMD ecosystem.

7. Client

Client revenue rose 23% YoY, driven by record mobile processor revenue and continued share gains. Ryzen Pro sales grew >50% YoY, with new wins across large healthcare, tech, auto and FSI customers. Dell, HP, Lenovo and Asus launched commercial PC portfolios featuring the latest Ryzen AI Pro 400‑series processors; the Ryzen AI Halo developer systems that went on sale this quarter also saw strong demand.

In Jul, AMD introduced the next‑gen Ryzen AI Halo platform with the new Gorgon Halo processor and an industry‑leading 192GB unified memory, capable of running models up to 300bn parameters. Starting later this year, each Ryzen AI Halo system will include a one‑year Hugging Face Pro subscription to lower the barrier to building and testing local LLMs. AMD is planning the back half against a softer PC market given rising memory and component costs pressuring demand, but expects to outperform the market on Ryzen strength and higher commercial penetration.

8. Gaming

Gaming revenue fell 31% YoY, mainly on lower semi‑custom sales at this phase of the console cycle. Gaming GPU revenue also declined YoY as higher industry component costs lifted GPU prices and weighed on overall demand.

9. Embedded

Embedded revenue rose 19% YoY, the strongest growth in 3+ years, with broad‑based demand across networking, aerospace & defense, test & measurement/simulation and communications. Embedded x86 grew meaningfully as hyperscalers and networking customers increasingly use AMD CPUs for critical DC networking and control‑plane functions, making it a key growth driver.

AMD introduced Ryzen AI Embedded X100 processors for demanding real‑time edge‑AI workloads and the Kria AI robotics platform for physical AI. Design‑win momentum remains strong, putting AMD on track for another record year, with new design wins to exceed $18bn, led by networking, data center, communications, test and A&D customers.

2.2 Q&A

Q: The accelerator TAM is ~$1.4tn and CPU TAM is ~$220bn, but shares are very different. When do DC GPU and CPU curves intersect in dollars—this year? Into 2027, which contributes more in dollars?

A: Compute demand is very strong at both ends—DC accelerators and server CPUs. The server CPU line is a more recent change, so we have been updating frameworks. For Q3 and Q4, both server and DC AI show strong growth—servers up >80% YoY in 2H, DC AI also strong, flowing into 2027. Both businesses will grow substantially; DC AI TAM is larger, so as Helios ramps with large strategic customers in 2027, DC AI will see meaningful growth. But both are key growth engines into and beyond 2027.

Q: With both lines growing next year, how do you trade off GPM—Helios scale is a headwind on GPUs, while server CPU margins are above corporate Avg.? Do pricing and mix shifts into CPUs make 2027 more tailwind than headwind?

A: GPM is primarily mix‑driven. We are pleased with H1‑26 GPM progression and the Q3 guide. In 2027, moving pieces include: (1) servers expanding meaningfully, a positive for margins; (2) the DC AI ramp timing—once DC AI scales, it brings very large revenue and margin dollars, albeit at a slightly lower margin than corporate Avg. The relative cadence will shape 2027, but we feel good about navigating this transition while improving and balancing margins. Also, Embedded is recovering strongly and should add incremental margin tailwind in 2027.

Q: What's the Q3 Data Center mix—between CPU and GPU, which adds more dollars from Q2 to Q3? Modeling '>80% 2H' suggests server CPU growth may be limited in Q4.

A: Server CPUs will grow significantly in both Q3 and Q4. Demand from hyperscalers and enterprises is very strong, and we have been adding supply all year. We guide the segment to double‑digit QoQ growth in Q3, with both servers and DC AI up nicely; in Q4, as more server and DC AI supply comes online, both grow further. Helios starts shipping late Q3 and is much larger in Q4. So Q3 is strong with solid double‑digit Data Center growth, and Q4 is higher than Q3.

Q: Relative to the ~40% TAM CAGR to 2030, can AMD outgrow that? Are you planning revenue growth >40% over the period?

A: The compute market is at a very exciting stage. Every piece—server CPUs, DC AI, Embedded and PCs—benefits from AI tailwinds, and in each sub‑segment we see share‑gain opportunities. So on top of a TAM growing >40%, we expect company revenue growth to exceed that pace.

Q: Across announced deals, how many GW of capacity does AMD see into 2027? What is revenue per GW? The market infers ~3 GW and at least $15bn per GW from warrants—do you agree, and where are upside/downside risks?

A: At a high level for 2027, we guided server CPU to grow >70% YoY off a higher base, and the overall Data Center to grow well over 100% driven by the DC AI ramp. We are pleased with progress at our three anchor customers—OpenAI, Meta and Anthropic—as they scale; demand is very high, and our ramp aligns with their DC builds and cloud partnerships. We have sufficient supply to exceed our guidance, with upside tied to close alignment with customer build plans. Helios ramps meaningfully over the next few quarters: Q3 start, step‑ups in Q4 and Q1, then continued scaling through 2027. Revenue per GW remains in the double‑digit billions and appears to be roughly at that level. The key is partnering closely through customer ramps, as demand for compute far exceeds that, and we aim to meet it.

Q: Helios uses ~50% more HBM than peers; AMD traditionally has more HBM. How much benefit comes from the extra HBM, does this increase exposure to memory inflation, and how do you protect margins and secure HBM allocation?

A: We work closely with memory partners on HBM for GPUs and system memory across DC, with long‑standing collaborations to secure ramp. We have good visibility into HBM allocations needed for 2027 deliveries. Memory bandwidth and capacity are core advantages of the AMD solution—larger models benefit from larger memory footprints, which factors into TCO. That said, in today's memory environment, every customer is optimizing configurations. Where larger memory is less impactful for TCO on certain workloads—e.g., mid‑sized models—we can adjust memory configs accordingly. The benefit is workload‑dependent.

Q: For 2027 server CPUs, supply looks secured with headroom. What are the implied unit and ASP assumptions? Should investors model by core‑count growth or something else?

A: Growth so far has been in both units and ASPs. In Q2, cloud and enterprise each grew >70%, with double‑digit growth in both units and ASPs, but units grew faster, consistent with the nature of our business and strong market demand. ASP uplift is driven by mix to higher core counts. Looking forward, both units and ASPs will grow. Over recent quarters we unlocked the supply chain—wafers, back‑end, substrates and components—to raise total server capacity, which is enabling the 2H raise; more capacity comes online into 2027 to support the growth. Model both units and ASPs up.

Q: You said Helios is 'above initial expectations'—does this refer to shipments or yields? If yields, should we expect GPM to improve as volumes ramp after the first couple of quarters?

A: 'Above initial expectations' refers to overall shipments—i.e., Helios demand into 2027. Customers who have looked deeply at Helios and the broader AMD ecosystem are confident it will be an important addition to AI portfolios, especially for inference—that is the context. On yields and the ramp, overall yields should improve over the next few quarters. Q3 is the start, followed by continued scaling; for a product of this complexity, yield improvement in the early quarters is expected.

Q: You split the $220bn server TAM into general‑purpose, Agentic 'sandbox' servers and AI front‑end nodes. Can you size these, especially the Agentic sandbox?

A: By 2030, Agentic AI/Agentic sandbox is the largest and fastest‑growing piece, though today it is the smallest. It is growing now, but over the next three to five years it becomes the biggest growth driver in servers. Our portfolio can grow in each sub‑segment—different workloads need different optimizations, requiring a full CPU family, not a single part. Venice leads across categories, from per‑core to socket performance, positioning us strongly in general‑purpose, Agentic AI and AI accelerator head‑nodes. Turin already sees very broad hyperscaler adoption; Venice expands workload coverage, with more workloads moving to next‑gen EPYC, underpinning our confidence to outgrow the market materially.

Q: You said client would grow YoY this year even with PC pressure—still true? How do you view client CPUs into 2027?

A: Client CPUs remain important. On‑device AI will play a growing role in how users experience AI, making AI PCs more important. H1‑26 was strong, and while we expect the market to soften in 2H on higher memory/component costs, the market has been more resilient than many expected, so we still see YoY growth in 2026. Into 2027, we have a strong slate not only for traditional notebooks/desktops but also for more AI‑centric PC experiences like Ryzen AI Halo. We are constructive on PCs as a broad user‑reach vehicle and on our ability to outgrow the market. We are monitoring component cost trends, but product strength and rising commercial Ryzen value per unit support a favorable client mix.

Q: The release says Data Center 'accelerates' in 2H, but Q2 was +107% YoY and Q3 guide is ~80% YoY—that is not acceleration. How should we read this?

A: 'Acceleration' refers to 2H vs. 1H. Comparing YoY growth in 1H Data Center to 2H, we see acceleration—that is the intent.

Q: For next year—'>70%' servers and 'well over 100%' overall DC—how much headroom does 'over' imply? Your trajectory seems well above those figures.

A: We are providing a framework for 2027. '>70%' for server CPUs is already a strong statement given where we are. For the overall Data Center, 'well over 100%' reflects that DC AI itself is expected to be well over 100%, driven by strategic customers, Helios ramp and the factors discussed.

Q: For MI400‑series early ramp over the next few quarters, how diverse is the customer set? Of the four to five large customers mentioned, how many contribute in Q4 vs. H1‑27?

A: We have discussed large frontier‑model players like OpenAI, Anthropic and Meta, who will consume via multiple cloud providers, including hyperscalers and others. Beyond them, many customers are interested in Helios at standard (non‑GW) scale. Diversity should improve over the next few quarters. We are aligning to each customer's DC readiness and build plans, ensuring we meet them; from Q4 into Q1, diversity improves further.

Q: From a DC readiness standpoint—land, power, shells—how capable are customers to take deliveries? Any constraints in 1H or 2H 2027 that could jeopardize targets?

A: We do not see factors that would make us hesitate on our goals; we feel good. There is a range driven by how quickly more capacity can come online. We are building the full supply chain—CPUs, GPUs, networking and all Helios components—and feel good about supply for both servers and DC AI. In parallel, we work closely with DC operators to maintain visibility, underpinning confidence in our DC AI outlook. We are also exploring ways to accelerate parts of the build, with the ecosystem aligned on bringing AI compute online faster.

Q: Based on guidance, Instinct could be roughly ~$30bn in 2027. How is revenue split between H1 and H2?

A: If you are referring to 2027 Instinct, your DC AI number may be low. Without giving specifics, 'over 100%' should be read as 'well over 100%'. This is a stair‑step over the next few quarters, with Helios expected to grow sequentially each quarter.

Q: For Instinct alone, how should we think about GPM from early 2027 to year‑end?

A: 2027 GPM depends on the relative ramps of server CPUs and DC AI. By quarter, the mix differs and direction can vary. Overall, we are optimistic about managing the MI450 ramp while server margins improve with strong growth in 2027, providing balance. Remember we have other levers—Embedded and Client continue to improve margins, and Ops execution is strong. We will provide more detail when guiding 2027.

Q: Are server CPUs supply‑constrained now? The market looks tight—does this persist, and can the supply chain support next year's CPU growth?

A: Server CPU supply has been tight, particularly in H1, as a large portion of demand was not previously forecast. Into 2027, demand visibility improves, so we expect server supply to be better than in 2026. We feel good about meeting the >70% YoY growth target, with potential upside depending on how things evolve.

Q: You are ramping heavily on 2nm while the market worries about tight 3nm—does this help AMD, or is ramping a new node still tough?

A: Ramping a new node is always challenging. Our chiplet approach means fewer wafers on the new node, giving us more flexibility. We are securing supply to meet strong demand, looking across the full chain—wafers, back‑end, packaging, substrates—and feel good about supporting strong ramps on both servers and DC AI.

Q: You announced a disaggregated compute collaboration with Cerebras at Advancing AI. Qualitatively, how does fast‑inference revenue grow from this year into next?

A: Inference grows materially from this year into next, and fast inference is becoming increasingly relevant. Cerebras has strong technology, and Helios combined with their Wafer‑Scale Engine offers a compelling solution. We expect availability on Cerebras Cloud starting in Q4 and extending into 2027. We will keep tailoring and optimizing for diverse workloads under an open ecosystem approach.

Q: With overall market growth at ~40% in coming years, how should we think about Opex growth?

A: With a ~40% TAM CAGR that we intend to outgrow, we will continue to invest given the opportunity, but Opex growth should trail revenue growth—our model is built for operating leverage to drive EPS. This aligns with our earlier view that EPS will be materially above the $20 target from Analyst Day.

<End of text>

Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure