Business

Know the Business

AMD is a fabless design house whose product portfolio sits on top of two of the most economically rivalrous franchises in technology: x86 server CPUs and AI accelerators. Everything else — Ryzen client chips, semi-custom console SoCs, Xilinx FPGAs, Pensando networking — scaffolds those two engines. The company prints $34.6B of revenue, $6.7B of free cash flow, and at the June 17 close of $520.89 commands ~$849B of market value — priced as if the next leg of execution is largely in the bag.

This page maps the parts behind that price tag: the economic engine, the segment-by-segment moat, the customer concentration that comes with hyperscaler dependence, the OpenAI warrant that transfers ~9% of the company to a single customer, the structural margin climb from 27% a decade ago to 55% non-GAAP today, and the lens for underwriting the stock.

The economic engine — three segments doing three different things

AMD reorganised in Q1 FY2025: Client and Gaming merged into one reportable segment, alongside Data Center and Embedded. Each segment is a different competitive game with different cyclicality, customers, and margin profile.

Three things in this table do most of the work for valuation:

• Data Center is now the largest segment (48% of FY2025 revenue, on track to roughly 56% in Q1 FY2026). Its operating margin of 21.7% looks lower than you'd expect — that is because the 2025 number absorbed roughly $440 million of net inventory charges from the MI308 China export ban. Strip that out and segment operating margin runs north of 24%; Data Center segment margin was 28% in Q1 FY2026, which is closer to the structural number.
• Embedded is the highest-margin segment at 36% segment operating margin (39% in Q1 FY2026). This is the Xilinx franchise. FPGAs are sold into aerospace/defence, industrial, and comms infrastructure where designs lock in for a decade and pricing is sticky. The segment fell 3% in 2025 on cyclical industrial/comms softness — it is, in effect, AMD's countercyclical anchor and its highest-quality earnings stream.
• Client and Gaming roughly tripled its operating margin from 12.3% in FY2024 to 19.6% in FY2025 on a 51% revenue jump. This is the Ryzen share-take story plus a Sony/Microsoft console cycle restart and the new Valve Steam Machine win — but it is also the most cyclical of the three segments, with management already flagging H2 PC demand softness from memory-cost inflation.

The shape of the FY2024 → FY2025 step-up makes it clear how unevenly the AI cycle is being distributed across the company.

In FY2025, Client and Gaming added $4.9B of revenue and $1.7B of segment operating income — more incremental segment OI than Data Center, which absorbed the MI308 inventory hit. That split is largely transient; Q1 FY2026 has Data Center back as the dominant earnings engine. The mix moves quarter-to-quarter; a consolidated multiple that ignores the segment shape will misprice this business.

Why fabless? The model that built the margin curve

AMD is a fabless designer. It owns design, IP, software, and customer relationships. It does not own fabs — it buys leading-edge wafers from TSMC, packages through Taiwanese ATMP partners (SPIL, ASE, KYEC, Tongfu JVs), and licenses some legacy 12/14nm capacity from GlobalFoundries via the long-running wafer supply agreement. This single architectural choice — to not run fabs — explains roughly everything about the company's economics.

Two facts to hold. AMD's capex is roughly 3% of revenue, not 30%; Intel's is north of 30%. AMD converted 19% of revenue to FCF in 2025; Intel ran negative free cash flow on higher revenue. Second, fabless does not equal high gross margin — NVIDIA's 75% GM and AMD's 50% GM are both fabless. The 25-point gap is mix, software lock-in, and pricing power, and it is the distance between today's valuation and the bull case.

The cost of fabless is a single-source supplier risk concentrated in TSMC at the leading-edge node (3nm, 2nm), TSMC CoWoS-S advanced packaging (the binding constraint on Instinct production), and HBM3E memory from SK Hynix/Samsung/Micron. Every dollar of incremental AMD revenue from data-center products has to win allocation against NVIDIA, Apple, Broadcom, and others at the same set of bottlenecks. Read the industry primer for the full chokepoint map — for this page, hold the single fact that AMD does not control its own supply.

The fabless choice has produced the cleanest fact in AMD's last decade: a structural climb in gross margin from the high-20s to ~50%, with management guiding non-GAAP gross margin to a 55–58% range over the medium term.

The flat read between FY2024 and FY2025 (49.4% → 49.5%) understates the underlying trajectory — the FY2025 figure is the post-MI308 charge number. Excluding the $440 million of net inventory write-downs, the underlying figure is roughly 50.7%, and the Q1 FY2026 non-GAAP gross margin of 55% sits above the historical range. Management's stated long-term operating margin target is north of 35% on a non-GAAP basis, against ~25% delivered in Q1 FY2026 — i.e. there is another ~10 points of margin embedded in the plan that has yet to be earned.

Moats — segment-by-segment, with the mechanism behind the score

A moat in semiconductors is usually a combination of (i) IP/architecture leadership at a leading-edge node, (ii) software lock-in, (iii) customer switching cost, and (iv) supplier preference at scarce manufacturing capacity. AMD's franchise is a stack of partial moats with very different durability profiles.

Two interpretive points:

• The semi-custom monopoly is the moat nobody talks about. AMD has 100% of current-generation Sony and Microsoft console silicon. Console contracts run 7–10 years from design start to console end-of-life and require integrated CPU+GPU+software co-engineering that no other public semis firm can offer at AMD's price point. Gaming margins are well below corporate average — Wall Street therefore underweights it — but it provides predictable cash for a decade and effectively pre-funds R&D for the broader graphics roadmap. The Valve Steam Machine win added in 2025 is meaningful in the same way.
• ROCm is the bear case in one acronym. NVIDIA's CUDA software platform is the actual moat in AI accelerators — not silicon. AMD has been investing heavily in ROCm (its open counterpart), and the 2025 release notes show real progress: day-0 model support, agent-based coding integration, MLPerf wins at MI355X. But this is a 15-year head-start that AMD is trying to close. Bulls cite the multi-gigawatt OpenAI and Meta commits as evidence that ROCm is "good enough" for inference at scale; bears cite the inference-skewed deployment mix as confirmation that training still belongs to CUDA. The truth is probably in between, with AMD's wedge in inference widening and training share remaining heavily NVIDIA-leaning.

Customer concentration — hyperscalers, two consoles, and a warrant for OpenAI

AMD discloses that "a small number of customers will continue to account for a substantial part of AMD's revenue and receivables in the future." The 10-K does not put names on the concentration, but the disclosed deals, the hyperscaler-led EPYC and Instinct ramps, and the semi-custom business make the customer set legible:

The OpenAI warrant — the most important economic fact most readers miss

In October 2025, AMD signed a 6-gigawatt product purchase agreement with OpenAI for Instinct MI450 deployment. Concurrently, AMD issued OpenAI a warrant to purchase up to 160 million AMD shares at an exercise price of $0.01 per share, vesting in tranches tied to GPU purchase milestones and stock-price targets, exercisable through October 5, 2030. None of the warrant shares had vested at FY2025 close, so there is no GAAP impact yet — but the economics are highly significant, and a serious investor must price this in explicitly.

AMD committed roughly $34B of equity value to OpenAI in exchange for an undisclosed multi-year GPU revenue stream. Whether the math is accretive depends on Instinct gross margin and revenue scale — neither has been disclosed. The Meta 6 GW commit announced in Q1 FY2026 carries no comparable warrant, suggesting the OpenAI structure was the cost of anchoring a first hyperscaler commit, not a recurring tax.

Geography — half the business ships internationally, but invoiced in dollars

International sales were 67% of FY2025 revenue (66% FY2024). Substantially all sales are denominated in US dollars, so there is no FX-translation drag. The economic geography is dominated by Asia, where ATMP/OSAT partners assemble and ship to OEM/ODM customers and forward to global end markets.

The more material geographic exposure is in manufacturing: TSMC's leading-edge wafer capacity is overwhelmingly in Taiwan. Cross-strait risk is industry-wide, not AMD-specific. The MI308 China export-ban (~$440M net inventory charges in 2025, partially reversed in Q4) is the more proximate regulatory drag — ~1.3% of revenue, but indicative of how a single BIS action can flip a quarter.

The cash machine — and where the cash is going

AMD's cash generation has stepped up sharply. Operating cash flow more than doubled in FY2025 to $7.7B (from $3.0B), and free cash flow tripled to $6.7B. Q1 FY2026 alone generated a record $2.6B of FCF — 25% of revenue.

Where does the cash go? The pattern is a tell about how management is allocating capital — heavy R&D investment, growing capex to support the data-center ramp, persistent stock-based compensation that runs roughly equal to buybacks, and the major Xilinx and ZT Systems M&A spend.

Two patterns to call out:

• Stock-based comp now runs at roughly $1.6B/year, against buybacks of ~$1.9B/year. Net buyback effort is therefore only ~$300M after offsetting SBC dilution — a fraction of the headline. For a roughly $800-850B-cap firm, that is not a meaningful capital-return policy. The OpenAI warrant tightens this further: future buybacks will de facto be funding warrant dilution rather than per-share earnings growth.
• R&D ran at 23% of revenue in FY2025 ($8.1B). That is unusually high for a 50%-gross-margin business and reflects the fact AMD is funding three simultaneous roadmaps (server CPU, AI GPU, AI PC) against an entrenched leader on the GPU side. Management has flagged that R&D will continue to outgrow SG&A as the AI investment cycle continues, which means operating-margin expansion has to come through revenue leverage, not cost discipline.

Return on capital — book ROIC is misleading, tangible ROIC tells the truth

The reported FY2025 return on equity of 7.1% and ROIC of 6.6% would suggest a mediocre business. They are not the right numbers. AMD's balance sheet carries $25.1 billion of goodwill and $16.7 billion of acquired intangibles — almost entirely from the $48.8 billion Xilinx acquisition closed in February 2022. Strip those out, and the picture is very different.

Peer economics — where AMD sits on the spectrum

The cleanest test of whether AMD is a high-quality business is to set its margin and growth profile against direct peers. The table below uses the latest reported annual figures for each company; market caps and enterprise values are as of early June 2026.

Margin ladder in one chart — the order of merit is unambiguous:

Three things to take away:

• AMD sits in the middle of the fabless peer set on gross margin and at the bottom on operating margin. It is not a NVIDIA, but it is not an Intel either. The 50% gross margin is well below NVIDIA's 75% and Broadcom's 63%; this is the closeable gap if AI accelerator mix scales.
• Operating margin tells a sharper story: AMD's 10.7% GAAP op margin is hit hard by R&D intensity (23% of sales) and acquisition-related intangible amortisation. On a non-GAAP basis AMD ran 25% op margin in Q1 FY2026. The cleanest framing is that the gap between GAAP and non-GAAP op margin is the Xilinx amortisation tail — about 6 points of margin — that runs off mechanically over the next several years.
• EV/sales of ~22x at AMD is in the same neighbourhood as NVIDIA's 23x. This is the part of the page that pays attention: the market is already paying NVIDIA-like multiples on AMD's revenue, on the assumption that the next leg of AI accelerator share gain and margin expansion will arrive. If you do not believe that, the stock is not cheap.

The scatter below is a quick way to see relative positioning — bubble size is market cap.

The shape is the bull case for AMD: it is in the upper-right quadrant alongside NVDA and AVGO (the businesses earning premium multiples) and well above INTC and the slower-growth fabless peers. The bull case is that AMD migrates up — toward NVDA-like gross margin as Instinct mix scales — while staying in the high-growth band. The bear case is that gross margin compresses as MI450 ramps at below-corporate-average margin while R&D stays elevated, leaving AMD stuck below 50% gross margin for longer than the multiple implies.

Cyclicality and the AI override

Semiconductors are deeply cyclical historically (industry primer covers this in detail). AMD's segment mix means cyclicality enters the consolidated P&L unevenly:

The composite cyclicality of the consolidated company is now moderate, not high — because the AI-driven Data Center revenue (and Q1 FY2026 mix at 56% of company revenue) is large enough to absorb cyclical drawdowns in Client/Gaming/Embedded. But this is conditional on the AI cycle. If hyperscaler capex normalises, the Data Center segment becomes cyclical at the same time as everything else, and the consolidated business looks more like the 2022–23 dip (revenue down 4% in 2023). A serious investor sizes that risk before paying 22x sales.

How to value this business

There is no single multiple that captures AMD. The right way to frame this is a sum-of-the-parts where each segment is valued on its own economics, then haircut for warrant dilution. The lens varies because the underlying businesses are unalike.

The single chart to sit with before forming a view:

This is the central tension. AMD trades at NVDA-like EV/sales but with materially lower current gross margin and operating margin. The implicit assumption in the multiple is that AMD closes a large fraction of the gap toward NVDA's margin profile over the next 3–5 years. That assumption rests on (a) Instinct gross margin migrating toward corporate average and (b) ROCm becoming "good enough" to widen inference share. If either fails, the multiple compresses.

Synthesis — what an intelligent investor should walk away with