I’ve been wrong about AMD three times in my investing life.
First time was 2019. I thought the EPYC server story was overhyped and that enterprise IT procurement was too conservative to switch away from Intel en masse. I passed. The stock went up 150% over the next two years.
Second time was late 2021. I bought in around $155, watched it drop to $56 during the semiconductor correction, and sold near the bottom convinced that the cyclical downturn would last longer than it did. It didn’t.
Third time was early 2025. Looked at the valuation, thought it was priced for perfection, didn’t add when I should have. The stock has since roughly tripled.
I’m telling you this upfront because it matters for how you read everything that follows. My track record on AMD specifically is humbling. The company has consistently outperformed what I thought was possible, and the reason it keeps doing that is worth understanding in detail — because it’s not luck, and it’s not just riding the AI wave the way everyone assumes.
The Q1 2026 Number That Stopped Me Cold
Total revenue for the quarter ending March 28 came in at $10.253 billion. Up 38% year over year. The data center segment alone printed $5.8 billion — 57% growth year over year — which now represents 55.9% of AMD’s total revenue. That’s a company that has genuinely transformed its business mix in roughly four years.
But the number I keep coming back to isn’t the revenue figure. It’s the market share data.
Mercury Research published their Q1 2026 CPU statistics and AMD’s EPYC line captured 46.2% of server CPU revenue. Not units — revenue. That matters because revenue share is harder to achieve than unit share; it means you’re winning the high-value end of the market, not just the commodity deals. For context, AMD’s server CPU revenue share was below 10% in 2019. Below 10% to above 46% in seven years, against a competitor — Intel — that controlled this market for two decades.
Intel’s GPU market share, by the way, dropped to 3.7% in April 2026 from 8.22% in March. One month. That’s not a trend, that’s a collapse.
What Lisa Su Actually Built Here
The standard narrative about AMD is that it’s the scrappy underdog that got lucky when Intel stumbled on their manufacturing roadmap. That’s wrong, or at least it’s incomplete in a way that leads investors to misunderstand the durability of the current position.
Intel didn’t just stumble. They spent years prioritizing near-term margin over foundry investment, got caught flat-footed when TSMC’s process technology pulled decisively ahead, and are now attempting a manufacturing turnaround that will take years to complete even if everything goes right. AMD saw that window coming and designed the EPYC product line specifically to exploit it — not just in performance, but in the total ecosystem: software compatibility, platform consistency, power efficiency, and the kind of predictable roadmap that enterprise buyers need before they’ll commit to a platform.
The Turin generation of EPYC, built on TSMC’s 3nm/4nm nodes, delivers performance-per-watt numbers that have accelerated cloud provider adoption in a way the previous generation couldn’t. When your chip does more work per watt, data center operators save money on electricity and cooling — two costs that have become significant budget line items as power constraints tighten around AI cluster deployments. That’s not a spec sheet argument. That’s a real economic argument that procurement teams at hyperscalers run through actual financial models.
Lisa Su guided that the server CPU market will reach $120 billion by 2030. AMD expects to capture a meaningful portion of that. At 46.2% revenue share today, “meaningful portion” looks conservative.
The AI GPU Story: Real, But Different From What People Think
AMD’s Instinct GPU line — the MI300X, MI350, and now the MI455X — has become one of the more misunderstood parts of the AMD thesis.
The bull case that circulated in 2024 was essentially “AMD is NVIDIA but cheaper, so hyperscalers will switch.” That was always too simple. NVIDIA’s CUDA ecosystem has fifteen years of developer momentum behind it. Models are written in CUDA. Frameworks are optimized for CUDA. The switching cost for training large language models is genuinely high, and nobody running a $100 million training run wants to be the person who chose the cheaper option that turned out to have a 15% performance gap at scale.
Where AMD’s Instinct chips have actually gained traction is inference — running already-trained models to generate outputs for users. Inference workloads are less dependent on CUDA’s training-specific optimizations. The performance gap is narrower. The price-performance math favors AMD in certain configurations. And inference is where AI compute demand is heading as the industry matures past the initial training-heavy buildout phase.
OpenAI committed 6 gigawatts of infrastructure to AMD. Meta committed up to 6 gigawatts. These are not pilot programs. These are multi-year, multi-billion-dollar infrastructure commitments. They won’t displace NVIDIA at the top of the training hierarchy. They don’t need to. They just need to capture a meaningful slice of the inference market that is growing faster than training because deployed AI applications are scaling to hundreds of millions of users.
The MI455X — part of the MI400 series, shipping in Q3 2026 via the Helios rack system — delivers 40 petaflops at FP4 precision with 432GB of HBM4 memory. These are numbers that competing with NVIDIA’s best on inference workloads in ways that the prior generation couldn’t match.
The Part of This Story That Makes Me Uncomfortable
AMD is currently trading at roughly 58 times forward earnings. NVIDIA is at about 25 times. Broadcom is around 38 times.
Think about that for a second. AMD is trading at more than twice NVIDIA’s forward multiple despite NVIDIA having higher revenue, higher margins, deeper software moat, and more entrenched customer relationships. The market is pricing AMD as if it will grow significantly faster than NVIDIA over the next two to three years.
That’s possible. AMD’s revenue growth rate in Q1 2026 was 38% versus NVIDIA’s 85% — so actually, on recent evidence, NVIDIA is still the faster grower by a wide margin. The AMD multiple requires either an acceleration of AMD’s growth rate or a deceleration of NVIDIA’s that closes the gap. Neither is guaranteed.
| Metric | AMD (Q1 2026) | NVIDIA (Q1 FY2027) |
|---|---|---|
| Total Revenue | $10.3B | $81.6B |
| Revenue Growth YoY | +38% | +85% |
| Data Center Revenue | $5.8B | $75.2B |
| Gross Margin (GAAP) | 53% | 75% |
| Forward P/E | ~58x | ~25x |
| Market Cap | ~$685B | ~$5.7T |
The gross margin gap is the one I come back to most. AMD at 53% versus NVIDIA at 75%. That 22-point difference isn’t noise — it reflects the structural advantage NVIDIA’s software-driven pricing power gives them. When you have a monopoly on the software stack that developers build everything on, you can charge monopoly prices. AMD doesn’t have that. Their chips have to compete on hardware merit, which means lower margins.
I’m not saying the AMD investment case is broken. I’m saying the current multiple requires a lot of things to go right simultaneously, and the margin for error is thin.
What the Gaming and Embedded Segments Tell You
Client and Gaming revenue was $3.6 billion in Q1, up 23% year over year. Gaming specifically came in at $720 million, up 11% but partially offset by lower semi-custom revenue — the chips AMD makes for PlayStation and Xbox consoles, which are in the later phase of the current console cycle.
The gaming GPU business is a different competitive situation from the data center. NVIDIA dominates the discrete GPU market for PC gaming with somewhere around 80% of the discrete market. AMD’s Radeon line is real competition but hasn’t been able to close the gap in a way that threatens NVIDIA’s position at the high end where most of the revenue sits.
I don’t think this matters much to the AMD thesis at this point. Gaming is roughly 7% of total revenue. It’s not driving the stock. What it does tell you is that AMD has multiple business lines generating real revenue even if you ignore the AI story entirely — which is a healthier position than companies whose entire valuation rests on a single product cycle.
The embedded segment came in at $873 million, up 6% year over year. This is AMD’s industrial, automotive, and IoT business — slower growth but sticky revenue from customers with long design cycles who are slow to switch once they’ve built AMD chips into their products.
The Intel Situation Is More Significant Than It Looks
Intel reported Q1 2026 results that showed data center and AI segment revenue of $4.1 billion — still substantial, but down from prior levels and significantly below AMD’s $5.8 billion in the same period. For the first time in a very long time, AMD has surpassed Intel in data center revenue.
That sentence is worth reading again slowly. AMD has more data center revenue than Intel.
This is not the outcome Intel’s management team was predicting two years ago. It’s not the outcome most analysts were predicting two years ago. Intel has a manufacturing turnaround underway and their Granite Rapids server processors are a genuine competitive response — not a surrender. But the server CPU market has developed a dynamic where AMD’s multi-year head start in TSMC’s leading-edge nodes is now translating directly into procurement decisions at major cloud providers, and those decisions take two to three years to reverse even if Intel’s products become competitive.
The procurement cycle in enterprise IT is slow by design. Once a cloud provider qualifies a chip architecture for their infrastructure, tests it exhaustively, builds software stacks around it, and starts deploying at scale, they don’t rip it out the following quarter. AMD’s current 46.2% server revenue share is partly locked in through existing deployments that will continue to generate replacement and expansion purchases.
Where This Leaves the Investment Case
The question I ask before buying anything is whether I can construct the bull and bear case clearly enough to know which scenario I’m actually betting on.
The AMD bull case is that server CPU share continues its climb toward 50%+ as Intel executes imperfectly on their turnaround, that inference GPU deployments at scale create a durable second revenue pillar alongside EPYC, and that the MI400 series specifically closes enough of the performance gap with NVIDIA that AMD captures 15-20% of the AI GPU market by 2027. If those three things happen, AMD’s earnings grow fast enough to grow into the current 58x multiple, and potentially beyond.
The AMD bear case is that 58x forward earnings is simply too much to pay for a company that still generates 22 points less gross margin than its primary competitor, that Intel’s recovery is faster than AMD’s current market position suggests, and that NVIDIA’s software moat proves deeper than inference-market share gains can penetrate. In that scenario, multiple compression is brutal because there’s very little cushion in the current price for any earnings disappointment.
My own position is that AMD is a better business than it’s been at any point in its history, that Lisa Su has built something genuinely durable in the server CPU market, and that the inference GPU angle is real enough to deserve a meaningful valuation premium over traditional semiconductor companies.
But 58 times forward earnings is asking me to have very high conviction in a specific outcome at a time when NVIDIA is growing revenue more than twice as fast and generates margins that AMD hasn’t demonstrated it can match. I can hold what I have. I can’t add aggressively at this price. The risk-reward at current levels requires more optimism about AMD’s competitive trajectory than I’m willing to commit to in writing.
The stock has made me look wrong before. It might again. That’s genuinely my honest position.
| Scenario | AMD Revenue (2027E) | EPS | Justified P/E | Implied Stock Price |
|---|---|---|---|---|
| Bull: GPU share grab + EPYC holds | ~$50B | ~$10+ | 35x | $350+ |
| Base: Steady execution, modest GPU share | ~$40B | ~$7.50 | 28x | ~$210 |
| Bear: Intel recovery + GPU disappointment | ~$32B | ~$5 | 22x | ~$110 |
| Current price (~$421) implies | — | — | 58x fwd | Requires bull scenario |
The table is uncomfortable if you bought recently. The bull scenario — which requires GPU market share gains and sustained EPYC dominance simultaneously — is what the current price assumes. The base case, which is a good outcome by most semiconductor company standards, implies the stock is materially overvalued today.
That’s where I stop. Not because AMD is a bad company — it’s one of the better-run businesses in semiconductors and the EPYC story specifically has more runway than the current market share numbers suggest. But because being a great company and being a great stock at any given price are two separate evaluations, and the current price requires a precise set of outcomes that I can’t hold with enough conviction to ignore the downside scenarios.
I’ll keep watching the MI400 ramp. If the Helios rack shipments in Q3 show genuine hyperscaler adoption at scale, the GPU thesis gets meaningfully stronger. That would be the signal to revisit.