Post Highlights
-
AMD Earnings Analysis: Q4 2025 and the Numbers Behind the Record
-
TSMC Macro Exposure:
-
TSMC Insights: Where the Market May Be Getting the Story Wrong
-
NVIDIA Business Model: How CUDA Became the Real Moat
-
TSMC Business Model: Why Every Chip Company Depends on One Foundry
AMD reported $16.6 billion in Data Center segment revenue for FY2025. Up 32% year-over-year. A record. Management called it evidence of "accelerating adoption" of AMD's AI franchise.
None of that tells you what is actually happening inside the segment.
AMD combines EPYC server CPU revenue and Instinct AI GPU revenue into one number. It has never separated them. So when Data Center grows 32%, you do not know whether that growth came from AMD taking server CPU market share from a weakened Intel (something it has done consistently for five years) or from Instinct GPUs scaling into a genuine AI accelerator franchise. Both outcomes produce the same line in the filing. They do not produce the same long-term business.
This is the central analytical problem in AMD stock. The margins, the valuation, the competitive positioning, all of it flows through this one unresolved question. And AMD has structured its reporting in a way that means you cannot answer it from public filings.
AMD's business in plain terms
AMD designs chips and outsources manufacturing entirely to TSMC in Taiwan. It does not own a single wafer fab. This keeps capital expenditures below $1 billion annually and lets AMD access TSMC's most advanced process nodes without the $20-30 billion annual capex that owning a foundry requires.
The chiplet architecture AMD pioneered, which assembles processors from multiple smaller dies rather than one large monolithic piece of silicon, let it improve yields and iterate faster than Intel could on a single monolithic die. EPYC server CPUs went from near-zero market share in 2017 to clear leadership in price-performance benchmarks by 2024. That is what makes AMD a serious company.
Whether AMD is also a serious AI GPU company is what the current multiple is pricing. That question is what drives the current valuation. And the filing does not answer it.
For context on AMD's foundry dependency and what TSMC concentration actually means for semiconductor investors, see the Finovian TSMC Business Model analysis.
Four segments, and what the opacity costs you in each one
Data Center: $16.6 billion, 48% of total revenue
This is where the blind spot lives.
AMD's Data Center segment covers EPYC server CPUs, Instinct AI GPUs, data processing units, networking silicon, and FPGAs for data center use. All reported as one number. The Q4 2025 press release described growth as being "driven by strong demand for AMD EPYC processors and the continued ramp of AMD Instinct GPU shipments." Both are growing. How much each one contributed is not stated.
Consider what you are actually buying when you own AMD at 22x forward earnings. You are betting that Instinct GPU revenue will scale into a business that competes with NVIDIA's data center franchise at scale. That bet may be right. But AMD's filing structure means you cannot track it quarter by quarter. If EPYC CPUs are doing the heavy lifting, driven by Intel's continued stumbles on Granite Rapids transitions and ongoing cloud provider qualification cycles, Data Center could keep growing at 20-30% annually with Instinct GPU adoption barely moving. The headline number would look identical.
AMD's reporting structure benefits AMD. A combined Data Center line lets management show strong segment growth regardless of which product is responsible. The AI narrative stays intact. The CPU narrative stays intact. Investors have no clean way to challenge either one because the data does not separate them.
That is not a conspiracy. It is standard segment reporting. But the analytical consequence is that every Data Center growth number AMD reports is partially unverifiable, and investors who treat it as confirmation of the AI GPU story are making an assumption the filing does not support.
Client and Gaming: $14.6 billion, 42% of total revenue
Client (Ryzen CPUs for desktops and laptops) and Gaming (Radeon GPUs and semi-custom SoCs for Sony and Microsoft consoles) are reported together.
Client revenue hit a record $10.6 billion in FY2025, up 51% year-over-year. Ryzen AI processors with integrated neural processing units drove higher average selling prices and continued market share gains from Intel. Gaming came in at $3.9 billion, also up 51%, with semi-custom console chips providing the bulk of the uplift.
The semi-custom business is predictable within a console generation and will naturally taper as the current PlayStation and Xbox hardware ages. It is not a growth driver. It is a steady income stream with a known decline slope.
This segment does not have an opacity problem in the same way Data Center does. You know what Ryzen and Radeon are doing because there is no second, harder-to-track business sitting inside the same number. Client and Gaming is what it is.
Embedded: $3.5 billion, 10% of total revenue
Embedded is the Xilinx FPGA business, plus adaptive SoCs for industrial automation, automotive, and defense applications.
FY2025 came in at $3.5 billion, down 3% from $3.6 billion in 2024. Recovery from a multi-year inventory digestion cycle is underway, with the back half of 2025 showing sequential improvement. Long design cycles mean new design wins take years to appear in revenue, and existing wins generate sticky revenue once qualified.
The Xilinx acquisition paid roughly $35 billion for this business and the broader adaptive computing thesis. Three years later, Embedded segment revenue is flat versus where Xilinx was before the deal. That is not a write-off. The thesis may still pay out over a longer horizon. But calling it validated at this point would require more evidence than the current numbers provide.
This matters in one specific way for the AMD investment case: it is the primary reason the GAAP-to-non-GAAP gap is as large as it is, which the next section covers directly.
The GAAP problem: $4.1 billion is not noise
AMD's FY2025 GAAP operating income was $3.7 billion. Non-GAAP operating income was $7.8 billion. The $4.1 billion gap is larger than what many semiconductor companies earn in total.
Two items drive most of it.
Stock-based compensation: $1.638 billion. AMD, like most technology companies, argues this is non-cash and not indicative of core performance. Reasonable people disagree on whether that framing is accurate. But the argument is at least disclosed and consistently applied.
Amortization of acquisition-related intangibles: $2.254 billion. This deserves more scrutiny than it typically gets.
AMD paid roughly $35 billion for Xilinx above and beyond tangible book value in 2022. That premium was allocated to acquired intangibles (customer relationships, developed technology, trade names) and is now being written down at $2.25 billion per year. Calling that amortization "not indicative of core operating performance" is a presentation choice. The capital was spent. The opportunity cost was real. The question of whether AMD got $35 billion worth of value out of Xilinx, given that Embedded segment revenue is currently flat, is not answered by moving the amortization below the non-GAAP line.
The opacity problem shows up here too. If Instinct GPU margins are structurally below NVIDIA's data center margins (which is likely, given the competitive pricing AMD has used to win hyperscaler accounts), then gross margin expansion over the past year may be driven more by EPYC's favorable mix shift than by Instinct GPU scaling. You cannot confirm that from the filing because the segment does not break it out.
The full GAAP reconciliation and earnings quality analysis is in the AMD Earnings Analysis.
NVIDIA's moat is not a gap AMD can close by shipping better chips
Most AMD coverage frames the competition with NVIDIA as a hardware race AMD is gradually winning. That framing is wrong in a way that matters.
CUDA is not a feature. It is an ecosystem that has been compounding for 15 years.
Every AI model trained on NVIDIA hardware, every library written to CUDA's API, all of it accumulates into a software base AMD's ROCm has to match before it can substitute.. The gap does not shrink as AMD ships better GPUs. If anything, it widens. Every quarter NVIDIA holds its hardware position, the CUDA codebase grows larger and the switching cost for enterprises goes up.
The hyperscalers have managed this. Microsoft, Google, and Meta have engineering teams large enough to They have engineering teams large enough to work around ROCm's limitations and still extract cost savings running AMD hardware. For those three customers, AMD is a real alternative and a genuine pricing lever against NVIDIA.
The broader market (enterprises, research labs, startups, any company without a thousand ML engineers) largely does not have that option. They hire CUDA-trained engineers. They use CUDA-native libraries. Switching to ROCm is a project, not a purchase decision.
AMD shipping better GPUs doesn't close this. NVIDIA's moat deepens every quarter it holds the hardware position, because the software ecosystem keeps accumulating. AMD can compete on price and performance for a defined set of customers. Becoming a platform company in AI requires winning the software layer, and there is no roadmap from AMD that addresses that directly.
NVIDIA's CUDA moat and what it means for the competitive structure of AI infrastructure is covered in the Finovian NVIDIA Business Model analysis.
What actually breaks the AMD thesis
Not the growth rate. Not a missed quarter.
The AMD thesis breaks if Instinct GPU adoption does not expand beyond hyperscale.
The hyperscalers are buying Instinct GPUs. That is real. But they are buying them for two reasons that are not the same as "AMD is becoming a platform in AI." First, they have engineering resources to work around ROCm limitations. Second, AMD gives them pricing leverage against NVIDIA. If AMD stopped shipping competitive hardware tomorrow, the hyperscalers would pay NVIDIA prices without much hesitation. AMD is useful to them - not irreplaceable
For the bull case to be right, AMD needs Instinct GPU adoption to spread into enterprise accounts, research institutions, and cloud customers that are not Microsoft, Google, or Meta. That requires ROCm to reach a level of compatibility and tooling depth that makes "deploy on AMD" a rational default decision rather than a specialized engineering project. There is no evidence that threshold has been crossed. AMD does not report metrics that would let you track whether it is being crossed.
A second break condition: the EPYC ceiling.
Intel is not staying still. Granite Rapids successors are in development. AMD's server CPU market share gains over the past five years happened against an Intel that was operationally compromised on process technology. If Intel stabilizes its foundry position and ships competitive next-generation Xeon silicon, the rate of EPYC share gains slows. Add an Instinct GPU adoption ceiling to that scenario and Data Center is growing 10-15% annually. Priced at 22x forward earnings, that is a different stock.
The third break condition: the disclosure trap.
If AMD eventually separates EPYC and Instinct GPU revenue, whether under investor pressure or SEC scrutiny, and the GPU contribution is smaller than the market assumed, the narrative does not just adjust. It breaks. Every quarter of Data Center growth that investors attributed to the AI story gets reread. The valuation framework applied to the stock was built on an assumption that turns out to have been unverifiable. That kind of reset is not gradual.
These are not just business risks. They are the conditions under which AMD stops being valued as an AI company.
Margins, capital structure, and what the numbers actually show
FY2025 non-GAAP gross margin was 52% for the full year, reaching 57% in Q4 2025. That Q4 figure included a $360 million one-time reversal of reserved AMD Instinct MI308 inventory, previously written down due to U.S. export restrictions on China sales. Strip out that reversal plus the $390 million in MI308 China revenue and AMD's own adjusted non-GAAP gross margin lands at approximately 55%.
Q1 2026 guidance is $9.8 billion in revenue at approximately 55% non-GAAP gross margin, which already embeds a 5% sequential revenue decline from Q4 2025's record $10.3 billion. The 57% headline from Q4 is not the number to model forward.
On cash: AMD generated $5.5 billion in free cash flow in FY2025, up from $2.4 billion in 2024. Cash and short-term investments ended the year at $10.6 billion against $3.2 billion in total debt. Cash and short-term investments of $10.6 billion against $3.2 billion in debt. Net cash positive by $7.4 billion.
R&D was $8.1 billion for the year, up 25%. AMD is running two major competitive battles at once. $8.1 billion in R&D is what that costs. The consequence of that spend shows up in non-GAAP operating margins of 22%, not the 28-30% some bull cases project as a near-term target.
One more number worth noting: the "All Other" category in AMD's segment operating income absorbed $4.007 billion in FY2025, covering stock compensation, acquisition amortization, and related charges. This is why Data Center's reported segment operating income of $3.603 billion and Client and Gaming's $2.855 billion look cleaner than the consolidated $3.694 billion GAAP operating income. The wedge is large. Knowing it is there helps you read the segment-level figures accurately.
The market prices AMD as an AI GPU company. At roughly 22x forward earnings, that is the story embedded in the valuation. The problem is that the filing does not let you confirm or challenge that story quarter by quarter.
My working thesis is that AMD's Data Center segment YoY revenue growth rate drops below 20% by the Q2 2026 report, with Q1 2026 guidance already pointing toward that trajectory. But the growth rate itself is not the most important part of this thesis.
What actually matters is what happens to the narrative when that deceleration arrives.
Right now, investors are applying a multiple to AMD that reflects a belief it is building a durable AI GPU platform to compete with NVIDIA. If Data Center growth decelerates and AMD still has not disclosed the EPYC/Instinct split, the market will have to confront a question it has been deferring: how much of AMD's Data Center growth was actually the AI GPU business, and how much was EPYC continuing to take share from Intel in a market where Intel was operationally disadvantaged?
If the answer turns out to be mostly EPYC, AMD re-rates. Not into a bad business. EPYC is genuinely good. But "Intel's best rival in server CPUs" is not priced at 22x forward earnings. It is priced closer to 15-17x. That spread, applied to AMD's current earnings base, is not noise. At current EPS it prices in several dollars per share.
I could be wrong on the timing. Hyperscaler AI capex is running at a pace that could sustain Data Center growth above 20% for longer than the Q1 2026 guide suggests. And if AMD eventually discloses the GPU/CPU split and the Instinct number is large, this thesis is wrong.
But the core analytical problem - that AMD's current reporting structure makes the AI GPU story impossible to confirm or challenge from public filings - does not go away regardless of the short-term revenue trajectory. Investors paying 22x forward earnings deserve to know which business they are actually paying for. Right now they do not.
The long-term question for AMD has two parts that are usually asked separately and should be asked together.
First: does the Data Center growth deceleration arrive on the timeline the Q1 2026 guide implies? My threshold is YoY growth below 20% in the Q2 2026 report as confirmation that the trajectory has shifted. Above 20% and the bull case has more runway than I currently see.
Second: does AMD ever disclose the EPYC/Instinct revenue split? Voluntary disclosure before Q2 2026 earnings, whether in a press release, investor day, or SEC filing, would force a direct reread of every prior Data Center quarter. That disclosure event, not the growth rate itself, is where the valuation reset happens.
If growth holds above 20% and the split stays undisclosed, the opacity thesis remains unresolved rather than wrong. If growth decelerates and the split remains hidden, the narrative question becomes unavoidable. Both answers land within the next two earnings cycles. That is a short enough horizon that this thesis resolves rather than drifting indefinitely.
This Take will be verified against AMD's Q2 2026 earnings report, expected late July 2026. Data Center segment YoY growth at or above 20% = this Take is wrong. Below 20%, with the GPU/CPU split still undisclosed = confirmed.
Frequently asked questions
How does AMD make money?
AMD earns revenue through three segments. Data Center ($16.6B in FY2025, 48% of total) covers server CPUs and AI GPUs combined into one reported number. Client and Gaming ($14.6B, 42%) covers Ryzen processors, Radeon graphics cards, and semi-custom console chips. Embedded ($3.5B, 10%) covers FPGAs and adaptive SoCs for industrial and automotive applications.
Does AMD disclose EPYC CPU and Instinct GPU revenue separately?
No. Both are reported inside the Data Center segment as a single number. AMD has not disclosed the split in any public filing or earnings call. This means there is no direct way to verify from AMD's financial statements how much of Data Center growth is driven by its AI GPU business versus its CPU business.
What is AMD's gross margin?
FY2025 GAAP gross margin was 50%. Non-GAAP gross margin was 52% for the full year and 57% in Q4 2025. That Q4 figure included a one-time $360 million inventory reserve reversal tied to export restrictions on Instinct MI308 sales to China. Adjusted for that item, run-rate gross margin was approximately 55%, which is also what AMD guided for Q1 2026.
Why can't AMD beat NVIDIA in AI GPUs?
The barrier is software, not hardware. NVIDIA's CUDA platform has 15 years of developer tooling, library support, and application compatibility. AMD's ROCm alternative is improving but cannot be substituted by better chip specifications alone. The moat is self-reinforcing: every new AI library written in CUDA and every new model trained on NVIDIA hardware increases the switching cost for anyone considering AMD. The hyperscalers manage around this. Most of the market does not.
What would break the AMD investment thesis?
Three conditions. First, Instinct GPU adoption fails to expand beyond hyperscale accounts into enterprise and mid-market. Second, EPYC server CPU market share gains stall as Intel ships competitive next-generation Xeon processors. Third, AMD eventually discloses the EPYC/Instinct revenue split and the GPU contribution turns out to be smaller than the market assumed. At that point every prior quarter gets reread and the valuation framework resets.
More on AMD and Other
- AMD Earnings Analysis: Q4 2025 and the Numbers Behind the Record
- TSMC Macro Exposure: TSMC Earns 74% of Revenue From the US, One Tariff Decision Rewrites the Investment Case
- TSMC Insights: Where the Market May Be Getting the Story Wrong