AMD closed Friday at four hundred and fifty-five dollars and change.
An all-time high.
A number that, twelve years ago, would have sounded like a typo.
The stock has roughly tripled in a year. Year-to-date it’s up somewhere between two-thirds and double, depending on which morning you happen to check the chart. Tuesday’s Q1 print — $10.25B in revenue, $5.8B from data center, the kind of release where the after-hours screen flashes green so fast you assume the data is stale — sent the stock 16% higher the next session.
Lisa Su walked into the call and guided next quarter to $11.2B.
Forty-six percent year-over-year.
She then casually told analysts to expect 35% revenue growth, compounded, for the next three to five years.
Open any feed on any platform and the explanation is the same three words.
Artificial intelligence demand.
It’s true.
It’s just not the whole truth.
And the parts that aren’t being repeated are the parts that decide whether you make money from four-fifty-five, or hand it back.
The story
There is a perfectly reasonable bull case here, and it deserves to be laid out cleanly before anyone starts poking at it.
Hyperscalers — Microsoft, Google, Amazon, Meta — have committed to roughly $725 billion of capex in 2026. Up 77% from last year’s record. About three-quarters of that funds AI infrastructure. Half a trillion dollars chasing a finite supply of accelerators is a tide that lifts boats whether or not the boats can swim.
Then there are the contracts.
OpenAI signed up for six gigawatts of AMD GPUs in October. Meta followed in February, with a custom MI450 derivative co-engineered for its own workloads. Independent analysts price the combined commitment north of one hundred billion dollars over five years — roughly one and a half times AMD’s entire 2025 revenue.
Both deployments begin in the second half of this year.
And then there’s the inference argument, which is genuinely strong. The world is shifting from training models to running them, and inference rewards the things AMD happens to be good at: more high-bandwidth memory, better cost-per-throughput, denser CPU-to-GPU ratios in the rack. ROCm 7 finally ships as a clean PyTorch install. The MI350 is in production. The MI450 has anchor customers signed before the chip has even taped out.
Last quarter, AMD did something it has never done in fifty years of corporate existence.
It overtook Intel in data center revenue.
If you read only that paragraph, you buy AMD.
That is also where almost every piece of coverage you’ve read this week stops.
The part printed in light gray
When AMD signed OpenAI in October, the press release led with six gigawatts, the MI450, and a deployment timeline. Headlines wrote themselves.
What the headlines didn’t lead with — what sat further down the page, in the kind of language lawyers write and journalists skip — was that OpenAI received a warrant for up to 160 million AMD shares at an exercise price of one cent. Vesting tied to deployment milestones. Stock-price thresholds peaking around six hundred dollars. Exercisable through 2031.
Four months later, AMD signed Meta.
Same structure.
Same one-cent warrant.
Another 160 million shares.
Add them up. Three hundred and twenty million shares. Roughly twenty percent of the company, conditionally pledged to the two customers that also happen to be the loudest validation points in the entire AMD rally.
This is the sentence I keep waiting to read in mainstream coverage, and never quite do.
The demand signal everyone is celebrating is partly co-financed by AMD itself.
It’s not a scandal. Strategically, it’s elegant. You align your largest customers with your roadmap. You discount in equity rather than in cash. You keep the headline ASP intact. Nvidia couldn’t do this deal. Nvidia doesn’t need to.
AMD did it because it had to.
But there is a meaningful difference between a hyperscaler validating you by writing a check at list price, and a hyperscaler accepting your equity as part of the consideration.
The market is pricing the first.
The contracts are the second.
Both can be true. Only one is in the consensus.
The price target
The average sell-side price target on AMD sits at $382.
The stock is at $455.
We are seventy-three dollars above the average forecast of the people whose job it is to be bullish.
This is not unusual after a print like Tuesday’s. The walk-up always lags the rally — targets get nudged in five-percent increments while the screen has already moved thirty. Catch-up revisions are coming. They always do.
The question is which side of those revisions you want to be standing on.
Forward P/E lands somewhere between fifty-two and sixty-three times, depending on whose model you trust. The semiconductor peer median sits around twenty-four. AMD is trading at roughly 2.2 times its sector on next year’s earnings.
That premium is defensible in a world where AMD compounds revenue at thirty-five percent for half a decade, which is exactly what Lisa Su told the call.
It is also one of the more ambitious things any chip CEO has said into a microphone in this cycle. Semiconductors are cyclical. Three to five years of unbroken 35% growth, with no digestion year in the middle, is something this industry has roughly never delivered.
You can believe it. You can be skeptical of it.
What you cannot do is pretend the price is undemanding.
The ten percent
ROCm 7 is good now. Genuinely. PyTorch installs cleanly. Local LLM inference on supported AMD silicon runs at roughly 85 to 90 percent of comparable Nvidia throughput.
Eighty-five to ninety.
Not parity.
In a market where buyers compare throughput-per-dollar to the second decimal, ten to fifteen percent is not a rounding error. It is the difference between a hyperscaler standardizing on AMD and a hyperscaler diversifying with AMD.
Those two procurement decisions look identical on a press release.
They are not the same trade for the stock.
The four phone numbers
Strip out OpenAI, Meta, Microsoft, and a small cluster of additional cloud customers, and the AI revenue story gets thin fast.
Four or five decision-makers anchor a meaningful piece of the data-center growth.
Any one of them blinking on capex — or, more realistically, shifting workloads onto their own in-house silicon (Trainium, TPU, MTIA, the parade of merchant-replacement projects that never quite make the headlines) — re-rates the entire stock.
Cerebras, by the way, just signed a twenty-billion-dollar inference deal with OpenAI.
The same OpenAI.
SRAM-first architectures, custom ASICs, merchant alternatives. All aiming at the same inference budget AMD is trying to win. The MI400 and MI500 roadmaps were designed for a world where merchant GPUs dominate inference.
Some of that world is being quietly built somewhere else.
What is actually working
The doubts are not the whole picture. The bull case is not nothing.
Data center grew 57% last quarter, faster than Nvidia’s data center segment in percentage terms — a sentence I never thought I’d type. Server CPU TAM was just raised to $120B by 2030. The rack architecture is moving from one CPU per eight GPUs toward one-to-one, and EPYC is the only product in the world positioned for that shift at scale. The MI450 has paying anchor customers before tape-out, which, in this industry, is abnormal in the best possible way.
If you believe inference is the defining workload of the next decade, AMD gives you exposure that Nvidia gives you, at a quarter of the market cap.
That is the cleanest version of the bull case.
It is intact.
What is actually risky
A forward multiple at 2.2x the sector. A price seventy-three dollars above the average sell-side target. A 320-million-share dilution clock running through 2031. Four to five customers carrying most of the AI revenue. Custom-silicon competition coming from those exact customers. A software gap that is closing, but not closed. A CEO publicly committing to a growth rate the history of this industry has rarely sustained without a stumble.
One of those flashing yellow is digestible.
Two at once is uncomfortable.
Three at once is the kind of moment retrospectives are written about.
So — buy, hold, or short?
The honest answer doesn’t fit on a bumper sticker.
It depends on where you enter. AMD at $300 and AMD at $455 are not the same trade, even though they are the same company.
It depends on how long you hold. By the time the dilution fully vests, the contracts have already paid for themselves.
It depends on what the technicals are saying that the fundamentals aren’t, which is the conversation I want to have with you next.
In the premium section, I lay out the valuation framework I’m using, the framework I think the Street is using wrong, the specific levels where I become a buyer, a seller, or a shorter, the two catalysts that will decide which side of that line we land on by year-end, and — at the very end — the position I’m taking…