Arm's Data-Center Pivot Is Real—the Shares Are Priced for Silicon It Hasn't Shipped

In late July, Arm Holdings reported its best quarter ever and the market turned on it anyway. Record revenue of $1.29 billion, an earnings beat, data-center royalties more than doubling from a year earlier — and the shares fell more than 8% in a day. That split, great business news meeting a falling price, is the whole story in miniature.

The company everyone still files under "smartphone chip designer" has spent the past year becoming something else. The part doing the growing — royalties from chips that run AI data centers — is real and profitable. The part that's stalling — the phone royalty base it was built on — is quietly being squeezed by memory-cost inflation that is trimming how many handsets get made. The investor's job is to notice that the two halves are not worth the same thing at all.

The engine that actually pays

Arm's model is unusual, and it's worth stating plainly because it explains both the profit margins and the new risks. Arm does not make chips. It licenses blueprints — the instruction set and core designs — to every chipmaker that wants them, and takes a royalty on each chip that ships. For two decades that meant one thing in practice: a tax on nearly every smartphone, from Apple to Qualcomm to MediaTek. Selling design work with no factories produces absurd economics; Arm's gross margin sits near 97.5%.

That same royalty has now become a tax on the data center. Management says data-center royalty revenue has been growing more than 100% year on year and expects that in a few years data center will be the largest of its businesses. The server-grade Neoverse line has passed 1.5 billion cores, and the latest 500 million shipped in about nine months — against six years for the first billion. The buyers are the hyperscalers who want off x86: AWS's Graviton, Google's Axion sitting beside its TPUs, Microsoft's Cobalt, Nvidia's Grace in the Blackwell lineup.

Be honest about attribution before crediting this to genius. A large share of the surge is hyperscalers fleeing Intel and AMD and reaching for a neutral architecture they can mold into their own custom chip; the "Graviton pattern" applies here. But from Arm's seat the point is identical either way: it collects a cut on every one of those cores at roughly 97% margins, with no factories and no channel conflict. The growth driver really has moved off the phone, and the measured part of that shift is cheap, high-margin, and durable. This is not the part of the story that made the stock double this year.

The part the market is actually paying for

In March, Arm unveiled the AGI CPU — its first production silicon in 35 years of business. Its founding CEO's rule was famously "make chips over my dead body". Arm is now, dead body notwithstanding, selling a finished data-center CPU, fabricated by TSMC on its 3nm process, up to 136 Neoverse V3 cores, aimed squarely at "agentic AI." The launch materials promise up to 2x performance per rack versus x86 and "$10 billion in capex savings per GW" of AI data-center capacity. Meta is the lead partner; Lenovo, Quanta and Supermicro are lined up as system makers, with broader availability flagged for the second half of the year.

Those are keynote numbers, and they should be labeled as such. Two-times-per-rack and ten-billions-per-gigawatt are marketing assets until an independent teardown shows a shipping part delivering that total cost of ownership. Nothing in this product has shipped at scale yet.

The bigger risk isn't the spec sheet; it's the business model. Arm's near-97% margin depends on being the neutral party everyone licenses from. The AGI CPU turns Arm into a direct competitor against its own best customers — the very hyperscalers whose custom-chip royalties are doubling. And the economics of selling hardware are nothing like the economics of selling blueprints: management guides early gross margins on the silicon line in the high-30s to low-40s percent, a fraction of the core's 97.5%. high-30s to low-40s percent

Management's own demand math shows how early this is. Initial AGI CPU demand is said to exceed $2 billion against a secured $1 billion of capacity across fiscal 2027 and 2028 — the company is supply-constrained before the part is even broadly available, and has traded design delivery for dependence on TSMC's manufacturing execution. It projects up to $15 billion a year from this silicon by 2031, against roughly $40 billion of total revenue across its entire 35-year history. up to $15 billion a year from this silicon by 2031That is a bet on execution, not a run-rate. The useful distinction for an investor is that the royalty shift is an earnings statement you can track, while the silicon line is a prophecy you have to take on faith.

What the price already assumes

The valuation forces the choice into the open. Even after the July selloff and a sharp pullback from mid-year peaks, Arm trades near $261 — about 267x trailing earnings, roughly 258x EV/EBITDA, and around 54x sales, on a market cap near $279 billion, for a company growing revenue 22% and still doing barely over a billion dollars a quarter. No multiple like that gets paid for measured royalties alone. It gets paid for the silicon future.

Arm's own guidance shows which story is currently decelerating and which is being front-loaded. For the current quarter it sees royalty growth in the low teens — the phone softness plus a natural normalization — while license revenue, where the AGI CPU commitments land, is guided up about 30%. royalty growth in the low teens In other words, the easy predictable money is slowing and the premium-priced hope is accelerating.

So the title is true in two different ways. The compensation engine has genuinely moved off phones and onto data-center royalties — that part is real, high-margin, and durable. But the price has moved onto the unproven part: a silicon business that dilutes the margins that made Arm special, competes with its own customers, and lives on keynote claims until a rack of them converts to delivered, profitable deployment. The way to tell which is which: keep watching the royalty line for whether data-center growth keeps doubling, and the silicon line for whether demand turns into shipped parts at the promised cost. Everything beyond that is the furthest extension of a launch slide.