AMD Is Winning the Turnaround. The Price Already Assumes It Finishes.

AMD trades at about 186 times trailing GAAP earnings right now. Nvidia, the company that actually dominates the market AMD is trying to break into, trades at about 32 times. Broadcom trades at 66.6 times. Something about that ordering looks wrong, and it's worth figuring out what.

Nvidia holds an estimated 80% of the AI accelerator market. AMD holds an estimated 5% to 7%. The company with the sliver of share is priced richer than the company that owns the room.

That's not the usual story of a beaten-down stock getting a speculative premium on hope alone. AMD's turnaround already shows up in the numbers: data center revenue hit a record $5.8 billion last quarter, up 57% year over year, and server CPU revenue share hit a record 46.2%.

Under CEO Lisa Su, this is a genuinely improving business. The problem: the price has already banked the improvement, and then some.

The turnaround is not a story anymore. It's a set of numbers.

Since Su took over in October 2014, when AMD was a $3 billion company four years into losses with cash reserves sliding toward $785 million, the stock has gone from $2 to $3 a share to a market cap of roughly $900 billion: a run built on a rebuilt product roadmap (Zen, RDNA) and a coherent reinvestment strategy, not one lucky product cycle.

The table below is the short version of what that turnaround looks like today. One line worth flagging on its own: the $4.17 non-GAAP EPS figure adds back non-cash Xilinx acquisition costs. It's AMD's own preferred number, and it's higher than the GAAP earnings figure the market multiples later in this piece are built on. Two honest numbers, answering two different questions; keep the distinction in mind.

Metric	FY2025	Q1 2026 (YoY)
Total revenue	$34.639B (+34.3%)	$10.3B (+38%)
Data center revenue	$16.635B (+32%)	$5.8B (+57%, record)
Non-GAAP diluted EPS	$4.17 (+26%, record)	not yet reported
Free cash flow	$6.697B (+178%)	not yet reported

The server CPU number deserves its own line, separate from the table, because it mixes two different measurements that are easy to conflate. AMD's share of x86 server chip revenue hit a record 46.2% in Q1 2026, up from an estimated 40% a year earlier. The simpler unit measure (how many chips shipped, not what they sold for) moved from 27.2% to 33.2% over the same period. Revenue share runs ahead of unit share because AMD's gains skew toward the highest-core-count, highest-price server chips.

That gain came against real resistance. Hyperscalers don't swap server CPU platforms casually, and the multi-quarter qualification cycles involved are themselves a barrier that favors the incumbent, Intel. AMD's 96-core Turin chip matching Intel's 128-core Granite Rapids on real workload performance is a demonstrated advantage, not a promised one, which is why the switching happened anyway.

The balance sheet backs up the operating picture too. AMD carries roughly 5% debt to equity, a net cash position, and a current ratio of 2.85, meaning it could absorb a real cyclical shock without any solvency risk.

One real structural moat, and its limits

AMD's most durable competitive protection has nothing to do with AI. It's a 2009 antitrust settlement: AMD and Intel hold a mutual cross license letting each build x86-compatible chips using patents the other one owns.

No third company can legally build an x86 chip without licensing it from one of these two. The agreement automatically terminates for both parties if either undergoes a change of control, which is exactly the kind of poison pill that discourages an acquirer from trying to buy AMD in the first place.

That's a genuine moat. It's also narrower than it looks: it protects AMD from a hypothetical new x86 competitor, a threat that barely exists.

It does nothing to stop Arm-based chips, like Amazon's Graviton or Nvidia's own Grace CPU, from taking server workloads without ever touching x86 licensing. And in AMD's biggest growth market, AI accelerators, instruction set licensing is irrelevant entirely: that fight runs on GPUs and custom ASICs, not CPU architecture.

The software gap nobody's being fully honest about

Nvidia's dominance in AI chips was never really about the silicon. It's CUDA, the software layer that lets developers actually program a GPU to do AI work, and it has an 18-plus-year head start in trained developers, optimized libraries, and enterprise tooling. AMD's answer is called ROCm, and the honest read on it in 2026 is more specific than the online narrative allows.

Where ROCm has actually caught up: standard inference workloads. PyTorch officially supports ROCm, and for a straightforward large language model inference pipeline with no custom kernels, ROCm is now production ready. AMD's MI355X is benchmarked at roughly 30% faster than Nvidia's B200 on one specific inference workload (Llama 3.1 405B), with about 40% better cost per token, according to vendor and analyst benchmarks that haven't been independently audited.

Where CUDA still wins outright: anything built on Nvidia-specific tooling like TensorRT-LLM or FlashAttention-3, and training workloads generally. Independent testing puts ROCm's training throughput at roughly 70% to 80% of equivalent CUDA hardware, and most broad machine-learning benchmarks still show CUDA ahead by 10% to 30%.

The "ROCm caught up" narrative circulating in 2026 runs ahead of what the benchmarks actually show.

The cleanest evidence that this gap still carries real economic weight: AMD's gross margin sits at about 50%, versus Nvidia's 75%. That 25-point gap is the difference between a company that has closed a technical gap and one that has closed the ecosystem gap that actually determines pricing power.

The bottleneck isn't the chip. It's the packaging slot.

Even a perfect chip has to get built, and the build step that matters most right now isn't wafer fabrication. It's advanced packaging, specifically a TSMC process called CoWoS (chip-on-wafer-on-substrate) that stacks memory and compute together into the finished AI chip. By multiple industry analysts' account, it's the single tightest chokepoint in the entire AI supply chain today, tighter than raw chip fabrication itself.

TSMC is scaling CoWoS capacity aggressively, from roughly 12,000 wafers a month in late 2024 toward a targeted 120,000 or more per month by the end of 2026. Nvidia has reportedly locked up more than 60% of that 2026 allocation, and AMD's own commitment, around 105,000 wafers, is described by industry analysts as enough to support only "a mid-teens market share" (a single-source estimate, not independently corroborated, but the only figure available).

AMD's realistic AI-chip ceiling for the next couple of years is being set by a supply contract, not by how good the MI400 turns out to be.

Your biggest customers are also building your replacement

Here's the part of AMD's AI story that gets the least attention and matters the most.

AMD's two largest new AI commitments are OpenAI (6 gigawatts of deployment, with the first 1 gigawatt of MI450 chips going live in the second half of 2026) and Meta (a separate 6 gigawatt deal).

The OpenAI deal comes with a warrant: a contract giving OpenAI the right to buy up to 160 million AMD shares at a strike price of one cent, as OpenAI hits deployment milestones. That's worth potentially close to 10% of the company in dilution, a real cost to existing shareholders contingent on the continued fundraising and execution of a company that is itself still burning cash.

Zoom out past the dilution question and there's a bigger structural issue. Meta, Microsoft, Google, and Amazon are simultaneously AMD's most important AI customers and the four companies building their own competing chips: Meta's MTIA, Microsoft's Maia, Google's TPU, and Amazon's Trainium, mostly designed with help from Broadcom and Marvell.

None of these chips are for sale to anyone else. They exist purely to reduce their builders' dependence on merchant silicon, industry shorthand for chips sold on the open market to any buyer, which is exactly what AMD and Nvidia sell and a captive in-house chip is not.

AMD's biggest customers today are also, by design, its longest-term competitive threat.

The growth rates show which way this is heading: custom AI chip shipments are projected to grow roughly 45% in 2026, against roughly 16% for merchant GPUs like AMD's and Nvidia's. One data point on scale: Broadcom's own AI chip revenue hit $20 billion in FY2025 and is guided toward $100 billion by FY2027, about six times the size of AMD's entire 2025 data center segment, with Broadcom and Marvell between them already designing most of the custom chips being built.

This isn't a fight AMD can win by shipping a better chip. It's a fight over whether hyperscalers keep buying merchant silicon at all once their in-house alternatives mature, and AMD has no real lever to pull against it.

What the price already assumes

None of the above means AMD is a bad business. It means the price already assumes AMD wins the good version of every one of these fights.

Peer valuations lay out just how far ahead of the pack AMD's multiple sits:

Company	Trailing P/E	Forward P/E	EV/EBITDA
AMD	186.0x	63.5x	121.3x
Broadcom	66.6x	25.4x	46.3x
Nvidia	32.3x	21.2x	30.6x
Qualcomm	20.7x	19.2x	15.8x

(EV/EBITDA compares a company's total value, debt included, against its cash operating profit: a yardstick less distorted by financing and depreciation choices than P/E, which is why it gets tracked alongside P/E rather than instead of it. Figures as of July 9-10, 2026.)

AMD is priced richer than every peer in this table, including the company that actually owns the AI accelerator market.

A three-scenario projection makes the same point in dollar terms. The mechanism worth understanding first: a stock's price is earnings multiplied by whatever multiple the market is willing to pay for them, so if the multiple shrinks enough, growing earnings can still mean a falling price. That's exactly what happens in the base case below.

Assume AMD sustains 30% annual earnings growth for three straight years, an aggressive assumption for a company already this size, and the market's multiple compresses from today's rich level to a more ordinary 30 times. The stock falls roughly 51% from today's price anyway. Only the most aggressive bull case, 50% sustained annual growth with the market still paying 45 times exit earnings, produces any upside at all, and a modest one at that.

Scenario	Assumed growth	Exit multiple	3-year price change
Bull	50%/year	45x	+13.5%
Base	30%/year	30x	-50.7%
Bear	5%/year	20x	-82.7%

(This model starts from AMD's FY2025 non-GAAP EPS of $4.17 and the $557.89 share price referenced throughout this piece, not the GAAP earnings base behind the 186x figure above. That's why the multiples implied here run lower than the peer table's GAAP multiple: two different, both-honest measures of the same stock.)

By the traditional definition of a margin of safety, buying at a meaningful discount to a conservative estimate of value, there isn't one here. The base case alone implies roughly half the stock's value disappearing, and that's the moderate scenario, not the pessimistic one.

The smaller cracks worth watching

Start with one thing that looks like a crack but isn't, once you check it. AMD's reported return on equity has averaged just 3.83% to 3.84% over the past three years, well below the roughly 15% investors usually want from a high quality business. The explanation is AMD's $49.8 billion, all-stock Xilinx acquisition, which added $22.8 billion of goodwill and $27.3 billion of intangibles to the balance sheet in 2022, inflating the equity side of that ratio. Strip those accounting artifacts out (a measure called return on tangible equity) and the same three years of profit produce a 13.56% average return, more than three times higher: a warning about a specific ratio, not evidence the underlying business is weak.

Two spots in the portfolio haven't resolved as cleanly. AMD's gaming GPU share fell from about 15% to about 5% in two years, despite a genuinely well reviewed product launch. A good product wasn't enough to dent Nvidia's installed base and software ecosystem in gaming, a useful caution against leaning too hard on "AMD ships a great chip" as the whole thesis for the much bigger AI accelerator fight.

The Embedded segment, built mostly from the 2022 Xilinx and Pensando acquisitions, is still net negative three years after closing. Revenue peaked in 2023 on what looks like post-acquisition channel restocking, then fell 33% in 2024 and another 3% in 2025, with margin compressing from about 39% to about 35% along the way. That's not proof the roughly $51 billion combined bet failed. The segment is small and the window is short. But it's not proof it worked either.

And AMD's access to the China AI chip market runs through a 15% revenue share arrangement with the US government, agreed to in exchange for export licenses, that legal scholars have separately argued may function as an unconstitutional export tax. AMD's own management doesn't trust this revenue enough to include it in guidance, which tells you what the company itself thinks the odds are.

What this actually means

Nothing here says AMD is broken. If anything, the operating case is the strongest it's been since Su took over: real share gains, record data center growth, a management team with more than a decade of proof it can execute, and a balance sheet that can survive a bad year.

What's changed is the question worth asking. It isn't "is AMD a good company." The evidence says yes. It's "does the price already contain all the good news, plus some good news that hasn't happened yet." On the numbers above, the answer is yes, and by a wide margin.

If you want exposure anyway, size the position for the bear case, roughly 83% downside, not the bull case, roughly 14% upside, because that's the risk you're actually underwriting at this price. Two levels worth watching for a better entry, both this piece's own estimate rather than a market consensus target: a pullback toward $300 to $350, or independent, non-vendor-sourced proof that ROCm has closed the gap with CUDA on training workloads specifically, not just inference.

Until one of those shows up, this is a genuinely great business at a price that has already spent all the good news, and is now borrowing against news that hasn't arrived yet.

Sources

This article is LLM-assisted, disclosed per site policy, and does not constitute financial advice or a recommendation to buy or sell any security.

Key facts