Business

Know the Business

AMD is a fabless semiconductor company that designs CPUs, GPUs, and adaptive computing chips for data centers, PCs, gaming consoles, and embedded systems. What matters most right now: the company is in the middle of a platform shift from "Intel's scrappy challenger" to "one of three companies that can build AI infrastructure at scale." The market is betting AMD can close the gap with NVIDIA in AI accelerators while continuing to take server CPU share from Intel — and the $498B valuation requires both bets to pay off simultaneously.

How This Business Actually Works

AMD designs chips but owns no fabs. TSMC manufactures nearly everything, which means AMD's gross margin is determined by chip complexity, ASPs, and wafer pricing — not by factory utilization. This is the right model for the current era: it freed AMD from the capital trap that nearly destroyed Intel, and lets the company ride TSMC's process leadership without bearing the $20B+ annual capex burden.

The revenue engine has three gears:

Data Center (~48% of FY2025 revenue): The highest-value segment, combining EPYC server CPUs (where AMD now holds record market share against Intel) and Instinct AI accelerators (where AMD is a distant second to NVIDIA). Data center grew 32% in FY2025 and management targets 60%+ CAGR over the next 3-5 years. Segment operating margins run 25-33%.

Client and Gaming (~42%): Ryzen CPUs for PCs (record share, ~$3.1B in Q4 alone) and semi-custom console SoCs for PlayStation/Xbox plus Radeon discrete GPUs. Client is growing steadily on share gains; gaming is entering a transition year as the current console cycle matures. Segment margins run 17-21%.

Embedded (~10%): FPGAs, adaptive SoCs, and embedded processors inherited from the 2022 Xilinx acquisition ($49B deal). Highest operating margins (~33-39%) but recovering slowly from an inventory digestion cycle. Design wins of $17B in 2025 point to long-term growth.

The cost structure is dominated by two items: wafer purchases from TSMC (~45-50% of revenue) and R&D (~23% of revenue, or $8.1B in FY2025). R&D is the real operating leverage story — AMD spent $1.5B on R&D in 2019 and now spends $8.1B, but revenue has grown 5x. As data center scales and amortizes that R&D base across higher-ASP products, operating margins should expand toward 30%+ over the strategic timeframe.

One critical point: $25B of goodwill sits on the balance sheet from Xilinx. This represents nearly a third of total assets. If the embedded segment doesn't eventually deliver the growth AMD paid for, that goodwill becomes a write-down risk — but at $17B in annual design wins, the pipeline suggests the acquisition is working.

The Playing Field

The peer table reveals three things worth internalizing:

AMD's profitability is thin for a company valued at $498B. A 12.5% net margin and 7% ROE are anemic compared to NVIDIA (56% margin, 101% ROE) and Broadcom (37% margin, 33% ROE). The valuation is paying for where margins are going, not where they are. Management's target of more than $20 EPS (vs ~$2.60 today) implies roughly 8x earnings growth — that's the promise embedded in the stock price.

NVIDIA is not just ahead — it's playing a different game. NVIDIA's revenue is 6x AMD's with margins 4x higher. The CUDA software ecosystem creates switching costs that AMD's open-source ROCm stack hasn't matched. AMD's advantage is price-performance and openness — hyperscalers want a credible alternative to prevent NVIDIA from extracting monopoly rents.

Intel is in crisis, and that's AMD's tailwind. Intel's negative net margin and operational stumbles continue to feed AMD's server CPU share gains. Every EPYC socket won is a high-margin annuity. AMD now has record server share with more than 3,000 enterprise solutions available — a position unimaginable five years ago.

Is This Business Cyclical?

Deeply cyclical — and the current AI supercycle risks making investors forget that.

The cycle hits AMD in three distinct ways:

PC demand cycles compress client revenue. The 2022-2023 post-COVID hangover showed how fast the client business can stall — revenue barely grew as channel inventory bloated.

Console generational transitions create predictable gaming revenue cliffs. Semi-custom SoC revenue will decline "significant double digits" in 2026 as the current 7-year PlayStation/Xbox cycle matures. The next Xbox launches in 2027, which will restart the cycle.

AI infrastructure capex cycles are the new risk. Hyperscalers are collectively spending hundreds of billions on AI infrastructure in 2026, but if ROI disappoints, this spending could decelerate sharply. AMD's data center segment — now nearly half of revenue — would be directly exposed.

The historical record is instructive: AMD's operating margin has swung from -19.5% (2012) to +22.2% (2021), back to +1.8% (2023), and now sits at 10.7%. This is not a company with stable economics. The bull case requires believing that data center AI revenue provides structural ballast that previous product cycles didn't — and that's possible, but unproven through a downturn.

The Metrics That Actually Matter

The conventional ratios (P/E, ROE) are nearly useless for AMD right now because the business is mid-transformation. Trailing P/E of 117x reflects $2.60 in EPS that management expects to grow to more than $20 within 3-5 years. ROE of 7% is depressed by $63B of equity (inflated by Xilinx goodwill). The metrics that actually predict whether the stock works are:

Data center revenue trajectory. This is the single most important number. The 60%+ CAGR target implies data center revenue growing from ~$17B to $60-70B by 2028-2029. The MI450/Helios ramp in H2 2026, the OpenAI 6-gigawatt deployment, and the Oracle partnership are the near-term proof points.

Gross margin direction. Q4 2025 normalized gross margin was ~55%. If AMD can sustain 54-55% as data center scales, operating leverage alone drives the earnings story. If gross margin stalls at 50%, the path to $20 EPS gets much harder.

Free cash flow. FCF tripled from $2.4B in FY2024 to $6.7B in FY2025 despite heavy capex investment. This is the clearest signal that the business model is inflecting — AMD is finally converting revenue growth into cash at a high rate.

What I'd Tell a Young Analyst

The entire AMD investment case comes down to one question: can the MI400/MI500 series GPUs and Helios rack-scale platform capture enough AI infrastructure spending to justify the current valuation? Everything else — EPYC share gains, Ryzen momentum, embedded recovery — is supporting evidence, not the main act.

Watch three things closely:

The MI450 ramp in H2 2026. This is AMD's most important product launch in a decade. OpenAI is the anchor customer, with the first gigawatt of deployment starting in the second half. If the ramp is smooth and AMD secures additional multi-gigawatt commitments, the stock works. If there are delays or performance shortfalls versus NVIDIA's Blackwell/Rubin, the 117x P/E compresses fast.

NVIDIA's CUDA moat versus AMD's ROCm openness. The market likely underestimates how much software ecosystem lock-in protects NVIDIA's pricing. ROCm 7 is real progress — Hugging Face, vLLM, and SG Lang now contribute directly — but "good enough" software with lower hardware cost is AMD's pitch, and that only works if hyperscalers remain committed to multi-sourcing.

The gap between GAAP and non-GAAP. AMD reports $4.17 non-GAAP EPS versus $2.60 GAAP. The difference is ~$3B in annual intangible amortization from Xilinx. This is a real economic cost (AMD paid $49B for those assets), but it doesn't consume cash. Understand this gap before comparing AMD's earnings to peers.

The biggest risk isn't competition from NVIDIA — it's an AI capex slowdown. If hyperscaler spending decelerates before AMD's data center business reaches critical mass, the company will be caught with $8B in annual R&D spend and declining revenue leverage. That's the scenario the market isn't pricing in.