Earnings Call Highlights: AMD (AMD) | Q2 FY2026 - Data Center Tops $6.7B in a Quarter and Hits 58% of Revenue; Helios Opens Up the 2027 Upside

2 days ago

7 min read

This wasn't a "pretty good" quarter for AMD - it was a gear change. Quarterly revenue hit $11.5 billion, up 50% year over year, and the Data Center segment alone took 58% of revenue, meaning the company has officially shifted its center of gravity from "PC + server" to the "AI data center". The real signal isn't this quarter's numbers but management's language on 2027: Data Center segment revenue to "well more than double" next year, server CPU up "more than 70%", data center AI "well over 100%" - with Anthropic, Microsoft, OpenAI and Meta all betting on the same Helios rack. For the optical communications supply chain, this is a report in which "the demand side pushes visibility straight into 2027".

This article is for technology and industry trend analysis only and does not constitute investment advice.

1. Three Signals to Remember First

First, AI is now AMD's core business, not a theme. Data Center revenue rose 107% year over year to $6.7 billion in the quarter, 58% of total revenue (versus only 42% a year ago). When one segment's share jumps from roughly 40% to nearly 60% in a year, the company's valuation logic has already been rewritten.

Second, server CPUs are not a supporting act - they were the source of this quarter's beat. EPYC set a record for quarterly server CPU revenue for the fifth straight quarter, with cloud and enterprise each up "more than 70%", and management said outright that second-half server revenue will grow "more than 80%" year over year. The market kept staring at GPUs, and the CPU line blew up first. The same week, Broadcom also extended AI order visibility to 2028 (see our "Earnings Call Highlights: Broadcom (AVGO) | Q2 FY26"); the entire AI compute supply side said the same thing this quarter: demand is bigger than forecast six months ago.

Third, the 2027 guidance is more worth writing down than this quarter's numbers. Data Center segment revenue to "more than double" next year, server CPU growth of "more than 70%" - this isn't the usual "reaffirming the outlook", it's "raising and nailing down the numbers". Management even said it has "significantly exceeded" the long-term model from last year's Financial Analyst Day, and that EPS will "substantially exceed" the original $20 target.

2. Revenue and Financials

Hard numbers first (all figures below use AMD's press-release comparable / non-GAAP basis; growth rates exclude the roughly $800 million MI308 export-control inventory charge in the year-ago period; revenue is GAAP = non-GAAP):

• Total revenue: $11.54 billion ($11,536M), up 50% YoY and 13% QoQ, a quarterly record.

• Non-GAAP gross margin: 56%, up about 200 bps YoY on a comparable basis and 80 bps QoQ. (GAAP gross margin 54%)

• Non-GAAP operating income: $3.09 billion ($3,094M), operating margin 27%.

• Non-GAAP net income: $2.76 billion ($2,760M).

• Non-GAAP diluted EPS: $1.66, up about 82% YoY on a comparable basis - the key point is that EPS is growing far faster than revenue, which is operating leverage at scale.

The segment breakdown makes the quarter's "fire and ice" very clear:

• Data Center: $6.72 billion, up 107% YoY, with segment operating income of $2.1 billion and a 31% margin. EPYC up more than 70% and Instinct revenue doubling are the twin engines.

• Client: $3.06 billion, up 23% YoY, with record mobile processor revenue and Ryzen PRO commercial up more than 50%.

• Gaming: $779 million, down 31% YoY, mainly due to the late console cycle and declining semi-custom sales.

• Embedded: $977 million, up 19% YoY, the strongest growth in more than three years, with a segment margin as high as 40%.

The financials in one sentence: growth is carried by Data Center, profit is pulled up by the high margins of Data Center + Embedded, and Gaming is the only drag - but it's now too small to matter.

3. Technology and Business Highlights: Helios Is the Star of the Quarter

If you could pick only one thing, it's the Helios rack-scale AI platform entering volume production. Helios integrates EPYC Venice CPUs, MI450 series GPUs, Pensando networking and ROCm software into a full rack; management said that across a range of inference workloads, it delivers 15% more throughput at the same rack power and 30% more tokens per dollar. First shipments begin this quarter (Q3), ramp in Q4, and continue into 2027.

For STT readers, the real story in Helios is how to feed scale-up bandwidth inside the rack. AMD's head of silicon photonics, Streshinsky, put it bluntly: scale-up bandwidth in AI racks will ultimately have to be handed to optics - we break down that thread in "AMD's Streshinsky on Silicon Photonics: AI Rack Scale-Up Bandwidth Must Ultimately Go Optical". Helios uses copper in this generation, but for the 2027 next-generation platform (MI500 series GPUs + Verona CPUs) AMD has explicitly said it will feature "both copper and optical interconnect" - the timeline for optics entering the rack is now pinned to AMD's roadmap.

On the server side, 6th-gen EPYC Venice is the headline: Zen 6 cores, TSMC 2 nm, and performance per watt claimed to be more than 2x x86 rivals; it is already in production, with every major OEM scheduling it. AMD splits server CPU growth into three pieces - general-purpose cloud, agentic AI sandboxes and AI head nodes - with agentic sandboxes called out as the "fastest-growing but smallest today" piece of a $220 billion server TAM in 2030. This points in the same direction as Jensen Huang's re-carving of compute and connectivity in the agent era (see "Jensen Huang Breaks Compute Apart: In the Agent Era, Optical Connectivity Is the Critical Path to Revenue"): agentic workloads are expanding demand for both CPUs and interconnect at once.

On software, ROCm moves from ROCm 7 to ROCm.AI, claiming 2x training and 3x inference performance, with support for Claude, Codex and Cursor to generate and optimize code directly for Instinct - AMD's step toward closing its biggest gap against the CUDA ecosystem.

4. Management Outlook: 2027 Is the Real Script of This Call

Q3 guidance first: revenue of about $13.0 billion ± $300 million, up about 41% YoY and 13% QoQ at the midpoint, non-GAAP gross margin about 56%, and non-GAAP operating expenses about $3.65 billion. On Q3 alone, that's already a strong quarter.

But management concentrated its firepower on 2027, and the wording is worth parsing sentence by sentence for certainty:

• Data Center "segment" 2027 revenue to grow "well over 100%" - note that after repeated questioning in the Q&A, Lisa Su strengthened "more than double" to "well over 100%", a phrasing that raises certainty.

• Server CPU to grow "more than 70%" in 2027 and "more than 80%" YoY in 2H26 - and "off a higher base".

• Data center AI "well over 100%" next year, with Helios stepping up each quarter (starting in Q3, a step up in Q4, another in Q1).

• Monetization per GW in the "double-digit billions" of dollars; Anthropic alone has signed for up to 2 GW of MI450, with the first GW deploying starting in 1H27.

The TAM raise matters too: AMD now estimates the data center AI accelerator market at about $1.4 trillion in 2030 (>45% annual growth) and the server CPU market at about $220 billion (>50% annual growth), with the overall high-performance + AI compute market approaching $2 trillion by 2030 at about 40% a year - and AMD says it will "grow faster than the market". This lines up exactly with capex signals on the supply side: after the four major CSPs pushed 2026 capex to astronomical levels, the bottleneck is moving upstream, as we discussed in "After the $725B Capex Decision, the Battleground for Optics Has Shifted to the Supply Side": the demand side isn't short of orders; what's short is the speed of adding capacity and interconnect.

5. Supply Chain and Customer Clues

This call hid several threads for supply-chain readers:

HBM and memory: Helios uses about 50% more HBM than its rival, and AMD said its 2027 HBM allocation has "good visibility", but it also unusually admitted that in the current memory price environment it will "adjust memory content" by workload - a direct response to memory cost inflation, which also hints that the PC side will be held back in the second half by higher memory and component costs (AMD has already guided the 2H PC market "softer").

Scale-up interconnect: inside Helios it's Pensando networking + rack interconnect, and AMD, Meta and Broadcom have in fact long been working together to define line-side specs for scale-up. We broke down this whole scale-up optical interconnect bet that "AMD has a stake in" - single-fiber bidirectional, microring DWDM, external lasers - in detail in "Meta, Broadcom and AMD Jointly Define the OCI 200G Line-Side Spec": wherever AMD's rack roadmap goes, this line-side spec gets pulled along.

Foundry and packaging: Venice uses TSMC 2 nm and Instinct uses chiplets; AMD stressed that chiplets let it ramp volume on new nodes "with fewer wafers", one of the supply-side reasons it feels confident calling big 2027 growth. AMD says both front-end (wafer) and back-end (packaging, substrate) capacity are being ramped in parallel.

Customer concentration: the main 2027 data center AI customers are three frontier model companies - OpenAI, Anthropic and Meta - consumed via multiple CSPs (including Microsoft Azure). The upside is high visibility; the risk is concentration - the data center build-out pace of these few mega-customers will directly decide which end of the guidance range AMD lands on in 2027.

6. Conclusion

This quarter AMD made a turning point clear: it is no longer "the second source chasing NVIDIA" but an AI infrastructure company with 58% of revenue from data center, twin engines in server CPU and AI GPU both surging, and the confidence to nail down its 2027 numbers. For the optical communications supply chain, the key takeaway is that AMD has laid out the timelines for both the rack-scale platform (Helios) and the next generation's copper-out, optics-in shift (MI500 generation) - demand-side visibility now reaches 2027.

Three metrics to watch next quarter (and beyond): first, the actual shipments and ramp slope of Helios / MI450 (Q4 is the first checkpoint); second, whether server CPU supply loosens materially in 2027 as management says (supply is the only ceiling on this line); third, how HBM allocation and memory costs affect gross margin and PC demand. The numbers look great, but what really decides 2027 is how fast capacity and interconnect catch up.