By now, you should have had the opportunity to review a copy of our earnings press release and the accompanying slides. We delivered very strong second quarter results, with revenue exceeding the midpoint of guidance as higher EPYC and Ryzen processor sales more than offset headwinds from export controls that impacted Instinct sales. We set records for both EPYC and Ryzen CPU sales, reflecting the broad-based demand for our differentiated high-performance data center, PC, and embedded processors. Second quarter revenue increased 32% year over year to a record $7.7 billion, and we delivered over $1 billion in free cash flow.

Excluding the $800 million inventory write-down related to data center AI export controls, gross margin was 54%, marking our sixth consecutive quarter of year-over-year margin expansion led by a richer product mix. Turning to the segments, data center segment revenue increased 14% year over year to $3.2 billion. We saw robust demand across our EPYC portfolio to power cloud and enterprise workloads, and increasingly for emerging AI use cases. In particular, adoption of agentic AI is creating additional demand for general-purpose compute infrastructure, as customers quickly realize that each token generated by a GPU triggers multiple CPU-intensive tasks.

Against this backdrop, fifth-gen EPYC turn shipments ramped significantly, and we had sustained demand for our prior generation EPYC processors. In cloud, adoption expanded with the largest hyperscalers as they deployed EPYC to power more of their mission-critical infrastructure, services, and public cloud products. EPYC adoption also grew with telecom customers as providers modernize their infrastructure for next-generation networks. EPYC enterprise deployments grew significantly from the prior quarter, supported by new wins with large technology, automotive, manufacturing, financial services, and public sector customers.

What went well
  • Record second-quarter revenue of $7.7 billion, up 32% year over year and 3% sequentially, exceeding the midpoint of guidance as record EPYC and Ryzen sales more than offset the Instinct export-control headwind.
  • Client and gaming segment revenue rose 69% year over year to $3.6 billion, including record client revenue of $2.5 billion (+67%, led by record desktop CPU sales) and gaming revenue of $1.1 billion (+73%).
  • Data center segment revenue grew 14% year over year to $3.2 billion on strong EPYC demand and share gains, marking the 33rd consecutive quarter of year-over-year server share gains.
  • Record free cash flow of $1.2 billion and $1.5 billion of cash from operating activities of continuing operations; returned $478 million to shareholders in the quarter ($1.2 billion in the first half).
  • Excluding the $800 million charge, non-GAAP gross margin was approximately 54%, the sixth consecutive quarter of year-over-year margin expansion driven by richer product mix.
  • Began MI-350 series volume production ahead of schedule in June with faster-than-expected adoption; seven of the top 10 model builders now use Instinct, and Oracle is building a 27,000-plus node cluster on MI-350, fifth-gen EPYC, and Pollara NICs.
What went wrong
  • An $800 million inventory and related write-down tied to U.S. export controls effectively eliminated MI-308 sales to China and reduced diluted EPS by approximately $0.43.
  • Data center segment swung to a $155 million operating loss (versus $743 million operating income a year ago) and revenue fell 12% sequentially, driven by the export-control impact on MI-308.
  • GAAP gross margin was 43%, down 10 points from 53% a year ago, due to the $800 million charge.
  • Diluted EPS fell to $0.48 from $0.69, and operating income declined to $897 million (12% margin) from $1.3 billion (22%) a year ago, primarily on the inventory charge.
  • Embedded segment revenue declined 4% year over year to $824 million with end-market demand still mixed, and operating income fell to $275 million (33% of revenue) from $345 million (40%) on product mix.

Guidance Changes

MetricPeriodCurrent guidance
Total revenueQ3 2025Approximately $8.7 billion, plus or minus $300 million (~28% YoY growth, ~13% sequential growth), with no MI-308 China revenue included
Non-GAAP gross marginQ3 2025Approximately 54%, despite the strong MI-350 ramp
Non-GAAP operating expensesQ3 2025Approximately $2.55 billion
Net interest and otherQ3 2025A gain of approximately $10 million
Non-GAAP effective tax rateQ3 202513%
Diluted share countQ3 2025Approximately 1.63 billion shares
Instinct revenueQ3 2025Expected to grow year over year, driven by the MI-350 ramp at multiple customers (no MI-308 China revenue assumed)

Performance Breakdown

MetricYoYNote
Total revenue +32% ($7.7B) Record EPYC and Ryzen processor sales and higher semi-custom shipments more than offset the export-control impact on MI-308 China sales
Data Center +14% ($3.2B) Strong EPYC CPU revenue and cloud/enterprise share gains, partially offset by a year-over-year decline in AI GPU revenue from MI-308 export restrictions
Client +67% ($2.5B) Record desktop CPU sales, strong Ryzen 9000 / X3D demand, and richer product mix
Gaming +73% ($1.1B) Strong demand for newly launched Radeon GPUs and higher semi-custom console revenue as inventory normalized ahead of the holiday season
Client and gaming (combined) +69% ($3.6B) Record client CPU sales and strong PC and console gaming product demand
Embedded -4% ($824M) Embedded end-market demand remained mixed; flat sequentially

Earnings Call Themes & Trends

TopicPrevious mentionCurrent periodTrend
MI-350/MI-355 Instinct rampMI-300/MI-325 ramp with smaller initial deploymentsVolume production began ahead of schedule in June with a steep second-half ramp; MI-350 matches or exceeds B200 in key training/inference workloads and delivers up to 40% more tokens per dollar, with adoption faster than expected
MI-400 series and Helios rackDevelopment progressing toward a 2026 launch; up to 40 petaflops FP4 and 50% more memory/bandwidth/scale-out than competition, with Helios racks connecting up to 72 GPUs for up to a 10x generational performance increase and significant customer interest
China / MI-308 export controlsMI-308 sales being restricted to ChinaCommerce Department is reviewing license applications; company expects to resume MI-308 shipments as licenses are approved but includes no MI-308 revenue in Q3 guidance; most inventory was work-in-process, so ramp will take a couple of quarters
ZT SystemsAnnounced intent to acquire ZT SystemsAcquisition closed early in Q2 and the team integrated seamlessly; agreed to sell ZT's U.S. manufacturing business to Sanmina for $3 billion in cash and stock (including contingent payment), expected to close near the end of 2025
Sovereign AIAnnounced a multi-billion-dollar collaboration (Saudi Arabia / HUMAIN starting with MI-355) and has more than 40 active global engagements; viewed as additive to the hyperscaler/AI-company opportunity, with contribution expected later than initial hyperscaler ramps
EPYC server CPU momentum33rd consecutive quarter of year-over-year share gains; more than 100 new cloud instances launched (nearly 1,200 total globally), 28 new enterprise platforms, and AMD now powers more than one-third of the world's fastest supercomputers
ROCm softwareLaunched ROCm 7 with more than 3x higher inference and training performance, nightly builds, a first developer cloud, day-zero support for Llama 4, Gemma 3 and DeepSeek R1, and ROCm Enterprise AI

Q&A Summary

Did Q2 client results include pull-forwards, and will the second half stay roughly flat with the first half? (Barclays)
Lisa Su said client was up 68% year over year in the first half with broad-based strength (desktop channel, X3D, Ryzen AI, enterprise sell-through) and does not believe much was pull-forward; end-user consumption is strong. Q3 client is expected to grow modestly but a little below seasonal given uncertainties, with much of the revenue uplift coming from higher ASPs as AMD sells up the stack and gains enterprise share (HP, Lenovo, now Dell ramping).
On China MI-308: once licenses are approved, what will contribution look like versus the ~$700M in Q2 and ~$800M second-half figures cited earlier? (Barclays)
Su said the company is pleased with progress with the administration and views China as an important market; several licenses are under Commerce Department review. Most inventory is work-in-process rather than finished goods, so it will take a couple of quarters to run through, and timing of revenue depends on when licenses are granted — a better position than 90 days ago.
How should investors size the 2026 sovereign opportunity, and is it incremental to current MI growth? (Bank of America)
Su said sovereign is additive to the hyperscaler and leading-AI-company opportunity, with some regulatory items still being worked through. The HUMAIN / Kingdom of Saudi Arabia collaboration starting with MI-355 is a good example, and AMD's open ecosystem resonates with sovereign customers; the opportunity spans both MI-355 and the MI-400 series.
What sales level does the MI product need to reach to stop being dilutive to margins, and will Q4 gross margins hold at Q3 levels? (Bank of America)
Jean Hu said MI gross margin is a little below corporate average; the priority is capturing the large, fast-growing revenue opportunity and giving customers better TCO. Margin improvement is a longer-term trend driven by operational efficiency rather than tied to any quarterly revenue level, and the focus is maximizing gross-margin dollars in one of the fastest-growing markets.
Can data center GPU reach around $7 billion for the year, and what is assumed for Q3? (UBS)
Su declined to give a full-year number but said MI-355 started production in June with a strong ramp through Q3 and into Q4, growing year over year. Unlike the MI-300 ramp which began with smaller deployments, MI-355 is very competitive with B200/GB200, strong for inferencing and increasingly training, with customers wanting to deploy at scale.
Will the new developer cloud (leasing back capacity) be a material portion of MI-355 revenue in the second half? (UBS)
Su said the developer cloud is meant to make it easy for developers to access Instinct GPUs with ready-to-deploy containers and no long-term commitments; it will not add meaningfully to second-half revenue but builds customer experience. The larger revenue driver is large-customer deployments ramping, and MI-355's similarity to MI-300 infrastructure should let it ramp quickly.
How has MI-350 family adoption changed versus a quarter ago given the early launch? (Deutsche Bank)
Su said adoption is a bit faster than expected, with significant new broad-based customer interest over the past 90 days, plus strong excitement around MI-400 and the Helios rack, prompting customers to engage earlier in the lifecycle.
How should we think about second-half seasonality for client and gaming? (Deutsche Bank)
Su guided Q3 client to single-digit sequential growth and gaming to flattish versus a strong Q2. In Q4 the console business is expected down strong double digits as holiday builds complete, so the overall client and gaming segment would likely be down in Q4.
How much visibility and lead time is needed from customers for MI-350 and especially MI-400 large-scale deployments? (TD Cowen)
Su said lead times run roughly eight to nine months with a strong supply chain prepared for both ramps. MI-350 can drop into existing data centers, while MI-400 requires extensive rack-scale co-development with the largest customers to ensure Helios compatibility with their 2026 build-outs; the ZT team has been very helpful on internal platform build-out and customer engagement.
Adding back the Q2 charge, guidance implies roughly flat sequential gross margin despite data center GPU up meaningfully — what are the offsetting drivers? (TD Cowen)
Hu confirmed Q3 gross margin guided around 54% versus ~54% ex-charge in Q2. Tailwinds — expanding server business (nice margin), growing commercial/enterprise PC mix, and ongoing operational efficiency — offset the unfavorable mix from strong gaming and the margin-dilutive MI-350 ramp.
On the 'tens of billions' MI-400 opportunity — what is the timeframe, and could it be $20B-plus by 2027? (Morgan Stanley)
Su would not commit to timing but said confidence is increasing on positive AI customer adoption signals across MI-350 and MI-400; reaching tens of billions implies significant gigawatt-scale deployments, and AMD is engaged with the right customers to enable that ramp.
Is rack-scale becoming more important to inference than expected as reasoning models add complexity? (Morgan Stanley)
Su agreed, saying model proliferation keeps GPUs the compute of choice and that distributed inference is raising the importance of scale-up/scale-out architecture, which AMD is investing in; the goal is to be a full-scale solution provider validated across the competitive multi-generation roadmap.
What server-side growth is embedded in the data center guide and where is enterprise server share today? (Wells Fargo)
Hu said both server and MI are growing sequentially with MI the most significant; server share is believed up again versus Q1 (no third-party report yet). Su added that cloud CapEx includes significant CPU spend since AI content requires traditional CPUs, forecasts are strengthening, and Turin/Genoa adoption and enterprise growth make the team bullish on servers into 2026.
On the $800M write-down and MI-308 — is there finished inventory, and could the charge reverse when a license is approved? (Wells Fargo)
Hu said the majority is work-in-process with no on-the-shelf finished goods to ship immediately, so it will take time to ramp once a license is granted.
How will the Instinct ramp progress into 2026 — the MI-350-to-400 handoff, Helios contribution, and customer mix? (Cantor Fitzgerald)
Su said the second half of 2025 is all MI-350, ramping into the first half of 2026, with MI-400 and Helios development on track and significant Helios revenue expected in 2026. Hyperscalers and Neo clouds likely lead the initial ramp, with sovereign coming a bit later given build-out timing.
With ZT Systems being sold for $3B and only $3B of debt outstanding, how do you think about use of proceeds? (Cantor Fitzgerald)
Hu said the model generates strong free cash flow ($1.2B in Q2) and proceeds add more; capital-allocation priority is investing in the large AI opportunity first, then continuing to return cash via share repurchase ($1.2B repurchased in the first half).
Is the ~$1B sequential revenue increase mostly GPUs, implying roughly a $2B GPU run rate with nothing from China? (Melius Research)
Hu said the sequential increase spans data center (both GPU and CPU, with GPU the largest incremental driver), client, and a return to growth in embedded, but the majority is driven by the strong MI-350 series ramp.
When the MI-308 license comes, can you reach the prior ~$800M run rate immediately or will it take several quarters? (Melius Research)
Su said it will take some time to ramp; sitting in early August, she would not expect much MI-308 in Q3, and shipments would be scheduled after license approval with a gradual ramp.

More on Advanced Micro Devices Inc

Reported 2025-08-05 · figures from the Advanced Micro Devices Inc Q2 2025 earnings call.

See how VectorShift works for your firm

Request Demo