The order book says $2 billion. The factory says $1 billion.

On the afternoon of July 29, the most repeated word on Arm Holdings' earnings call was supposed to be "record." Revenue, royalties, licensing — every line hit a first-quarter high. But that is not what the analysts kept circling back to. They kept coming back to one disclosure: customer demand for the Arm AGI CPU, the company's own silicon that it unveiled in March, has passed $2 billion across fiscal 2027 and 2028. And Arm has locked down manufacturing capacity for exactly half of that.

Normally when a chip company says demand doubled in ninety days, that is a confetti moment. This time the sentence came bundled with four other words: wafers, substrates, test capacity, memory. That is CEO Rene Haas' own bottleneck list, delivered on the call. Arm has entered the state every hardware company eventually learns to hate — it can sell more than it can build. For a company that spent thirty-five years drawing blueprints for other people and collecting a cut, this is an entirely new category of problem.

Here is why that gap matters more than the record quarter. It is the cleanest single number showing how heavy Arm's business model transition actually is. As a pure IP licensor, Arm never needed a fab, never carried inventory, never negotiated a substrate contract. It drew one design and sold it into a hundred sockets, and the hundredth sale cost roughly nothing to fulfill. The moment you start shipping finished silicon, your revenue ceiling gets set by how many wafers TSMC allocates you and how much DRAM SK Hynix and Micron decide to hand over. Demand can be $2 billion or $20 billion. If you can build $1 billion, you book $1 billion.

Haas sounded confident anyway. He said Arm's conviction that it can exceed the $1 billion AGI CPU opportunity "has increased over the past 90 days," and walked through the supply chain piece by piece — wafers, substrates, test capacity, memory — saying the confidence level on securing what it needs has improved across all of them. Confidence improved. Capacity secured: still $1 billion. That tense difference is the actual content of this quarter.

The four parties tangled up in one sentence

Start with Arm itself. Founded in Cambridge in 1990, it spent the entire smartphone era as the company that sold blueprints and nothing else. Apple, Qualcomm, Samsung, MediaTek, Nvidia, Amazon, Google — essentially every company on earth that builds chips holds an Arm architecture license. Arm's own framing is that its compute platform touches 100% of the connected global population, with more than 22 million developers in its ecosystem. Neutrality was the asset. Because Arm took nobody's side, everybody could be a customer.

Arm broke that neutrality itself, on March 24. At an event in San Francisco, Haas unveiled a 136-core data center processor called the Arm AGI CPU — the first silicon in company history that Arm designs and sells directly. Meta is the lead partner and co-designed it. Per CNBC's spec breakdown, it is a chiplet design built on TSMC's 3nm N3P process, with Neoverse V3 cores running at 3.5GHz and carrying 2MB of L2 cache per core. Arm claims more than 2x rack-level performance versus the latest x86 systems, at a 300W TDP.

Second party: the customer list. Beyond Meta, the launch partners included OpenAI, Cloudflare, SAP, F5, Cerebras, Positron, Rebellions and SK Telecom. OpenAI's Sachin Katti said the Arm AGI CPU "will play an important role in our infrastructure." Meta's Santosh Janardhan said the chip meaningfully improves data center performance density and supports a multi-generation roadmap. System builders lined up as ASRock Rack, Lenovo, Quanta and Supermicro. Early units already shipped; volume supply was slated for the second half of 2026.

Third: SoftBank. It bought Arm outright for $31.4 billion in 2016 and relisted it on Nasdaq in 2023, but it still holds the overwhelming majority of the stock. As of June 2026, SoftBank Group's stake stands at 922.73 million shares, in the 86% range. SoftBank also acquired Ampere Computing for $6.5 billion in 2025 and sits inside the Stargate data center project alongside OpenAI and Oracle. So OpenAI appearing on Arm's first-party silicon customer list is not purely coincidental. That said, Arm has never publicly quantified how much AGI CPU silicon goes into Stargate, and going further than that is speculation, not reporting.

Fourth: the regulators. Bloomberg reported on May 15 that the US Federal Trade Commission opened an antitrust probe into Arm. The question is straightforward. Once Arm started selling its own chips, did it acquire an incentive to give its hundreds of licensees — Apple, Qualcomm, Nvidia among them — worse architecture license terms, or to refuse them outright? That is the obvious question when a referee suits up as a player, and how the probe resolves could matter more to Arm's future than the $15 billion silicon roadmap does.

$715 million in royalties, and a business that has booked zero revenue

Let us get the numbers straight. Arm's fiscal year ends in March, so what landed on July 29 was Q1 of FYE27, the quarter ended June 30, 2026. Total revenue: $1.289 billion, up 22% year over year, a first-quarter record. Royalty revenue: $715 million, up 22%. Licensing revenue: $574 million, up 23%. Both are first-quarter records too. Non-GAAP operating income came in at $531 million for a 41.2% margin, roughly 200 basis points better than a year ago. On a GAAP basis it was $91 million and 7.1%. That spread is mostly stock-based compensation, and Arm's P&L has looked like this since the IPO.

The data center line did the heavy lifting. Arm said data center royalty revenue more than doubled year over year, and that cumulative Neoverse core shipments crossed 1.5 billion — with 500 million of those going out in just the last nine months. AWS Graviton, Google Axion, Microsoft Cobalt, Nvidia Vera: every hyperscaler's homegrown server CPU sits on Arm architecture. Whichever camp wins, Arm collects. That structure is the engine currently pushing data center royalties up and to the right.

Metric Q1 FYE27 (ended June 2026) YoY Notes
Total revenue $1.289B +22% First-quarter record
Royalty revenue $715M +22% First-quarter record
Licensing revenue $574M +23% First-quarter record
Non-GAAP operating income $531M — 41.2% margin
GAAP operating income $91M — 7.1% margin
AGI CPU customer demand Over $2B (FYE27–28 combined) 2x prior quarter Revenue recognition starts FYE28
AGI CPU secured capacity ~$1B worth — Half of demand
Q2 revenue guidance $1.33B–$1.43B Record at midpoint Licensing ~+30%, royalty low teens

There is one thing that trips people up here, so be careful with it. The $2 billion is not revenue. It is customer demand — pipeline. Per Arm's investor materials, actual revenue recognition for the AGI CPU business does not begin until FYE28. Which means the $1.289 billion Arm just reported contains essentially none of it. What sold well this quarter was the old business: smartphone and data center royalties, plus Armv9 and Compute Subsystem licenses.

The guidance deserves a second look too. CFO Jason Child framed Q2 revenue at $1.38 billion plus or minus $50 million (the company subsequently restated the range as $1.33 billion to $1.43 billion). He sees licensing up around 30% year over year and royalties up in the low teens. But the full-year royalty growth outlook got cut — from roughly 20% down to something closer to the high teens. The reason is phones. However hard cloud AI pulls, mobile softness is offsetting part of it. So Arm printed a record quarter and trimmed its annual outlook in the same breath, which is why post-earnings coverage of the stock reaction came out inconsistent across outlets.

The margin structure changes too, and this is the part investors tend to skip. Child described first-generation AGI CPU product margins as "high 30s, maybe low 40s," with a path to 50% within a few years. Set that against an IP business that runs north of 65% operating margin and the dilution is obvious. It is the same logic behind Arm's stated FYE31 target of $25 billion in total revenue, split as $10 billion from IP and CSS (65%+ operating margin) and $15 billion from AGI CPU (30%+ operating margin). Silicon generates more revenue; IP generates better margins. Arm is deliberately trading one for the other.

What each player actually walks away with

For Arm, the prize is ripping the roof off its revenue ceiling. An IP royalty is a few percent of a chip's selling price. A customer ships a $1,000 chip and Arm gets a slice. Sell the finished product yourself and the whole $1,000 lands on your income statement. That is the arithmetic that let Haas stand up in March and say $15 billion from AGI CPU alone by 2031 and $25 billion company-wide. A business that did a bit over $4 billion in total revenue in 2024 is calling for a 6x, and the model shift is the entire basis for it.

Meta is optimizing for something else. Designing a server CPU from scratch takes a team of thousands and several years. Co-design it with Arm and become the anchor customer instead, and you get silicon tuned to your workloads much faster while splitting the development bill. What Meta buys is rack density. Per Arm's own figures, a 1OU blade carries 272 cores, a standard 36kW air-cooled rack holds 8,160 cores, and a 200kW liquid-cooled configuration exceeds 45,000 cores per rack. For inference and agentic workloads — where you need to spread CPU cores wide — density translates directly into real estate cost and power cost.

OpenAI, Cloudflare and SAP are buying leverage. The data center CPU market was an Intel-AMD x86 duopoly for a very long time, with hyperscaler custom silicon only recently muscling in. Add Arm as a merchant supplier and there is one more column in the negotiation. For mid-tier operators who cannot afford a custom silicon program of their own, it means ordering hyperscaler-grade CPU performance out of a catalog rather than building it.

SoftBank's take is the bluntest of the group. Holding roughly 86% of Arm means every dollar of Arm's valuation flows straight into SoftBank's books. Layer on Ampere Computing (the $6.5 billion acquisition) and the Stargate project, and SoftBank is assembling design, silicon, land and power for AI data centers inside one group. But that cuts both ways. Arm's own chip and Ampere's server CPUs are aimed at the same market, and there is no public answer yet on how the group intends to reconcile them.

And to be fair, you have to name the parties who get nothing. Qualcomm, MediaTek and Nvidia — the large licensees — gain zero from this transition. Reuters reported that Arm competed against Qualcomm for the Meta data center CPU deal and has been hiring executives out of its own licensees. Yes, Broadcom, Marvell, Nvidia, AWS, Azure and Google Cloud all sent congratulatory video messages to the March keynote. Congratulatory videos and competitive bids are different things. The FTC opened its file at precisely this seam.

When platform companies started selling products: the receipts

The most famous failure is Google in 2011. As the neutral platform steward of Android, Google bought Motorola Mobility for $12.5 billion and started making handsets. The result was distrust from Samsung and HTC — the partners the platform actually ran on — and by 2014 Google had sold Motorola to Lenovo for $2.9 billion and backed out. It is the canonical demonstration that when a platform owner starts competing with the businesses on its platform, partners notice before the market does. Google did not abandon hardware entirely, though; it came back far more carefully with Pixel and TPU.

Second exhibit: Microsoft, 2013. It had been licensing Windows Phone to partners, then bought Nokia's devices unit for $7.2 billion — and the OEMs it had left walked away. In 2015 Microsoft wrote off $7.6 billion and effectively shut the business down. The lesson there is about timing. Microsoft attempted vertical integration after its platform had already slipped to a distant third. Arm is moving from a position of peak platform dominance, which is a genuinely different setup and could produce a different outcome. It also explains why the regulatory exposure is so much larger.

Third: the server CPU market's own graveyard. Qualcomm launched the Centriq 2400 in 2017 to attack Arm-based data center CPUs and pulled out in under a year. The chip was not bad. What killed it was the software ecosystem and the length of the customer qualification cycle. Everyone will tell you it is different now — 1.5 billion cumulative Neoverse cores, 22 million developers, and above all anchor customers like Meta and OpenAI raising their hands first. But the record still says "good chip" has never been sufficient in the server market.

There are successes on the other side of the ledger. Nvidia launched DGX systems in 2016 and put itself in direct competition with the OEMs buying its GPUs, and it neither lost the OEMs nor stunted its own systems revenue. Amazon acquired Annapurna Labs in 2015, shipped Graviton in 2018, and became both customer and competitor to Intel and AMD while keeping both relationships intact. The common thread is simple: neither company degraded the quality or terms of what it sold to its existing partners. That is exactly what Arm will be tested on for the next several years, and exactly what the FTC is watching.

How Nvidia, AMD and Intel swing back

Nvidia already has its answer. The Vera CPU shown at GTC 2026 uses custom Olympus cores with SMT to run 88 active cores, and it carries 1.8TB/s of chip-to-chip bandwidth — double Grace. The subtle part: Vera is built on an Arm architecture license. So every Vera that Nvidia sells pays Arm a royalty. The two companies are now simultaneously competitors and revenue partners, and how long that stays stable is a genuine risk for both sides.

AMD is answering with brute core count. EPYC Venice is reported to run Zen 6c-based cores on TSMC's 2nm process, scaling to 256 cores, with AMD claiming a large generational performance jump. Against the AGI CPU's 136 cores per chip, AMD holds the better card in a pure core-count comparison. Core count is only one dimension of performance, of course, and Arm is deliberately trying to shift the frame to rack-level performance and power efficiency. Worth remembering that the "2x per rack" claim is Arm's own estimate, not a third-party benchmark.

Intel has the hardest seat at the table. Clearwater Forest packs 288 E-cores on the 18A process, and the P-core line, Diamond Rapids, is on the 2026 calendar too. But SemiAnalysis has flagged that Clearwater Forest lands only about 17% faster than the prior-generation Sierra Forest at equivalent core counts, and that hybrid bonding yield issues are weighing on cost. Intel's defense of the x86 line requires proving its process is competitive again, and that proof is still in progress.

There is one more competitor, and it is a quiet one: the hyperscalers' own silicon. AWS Graviton, Google Axion and Microsoft Cobalt are all Arm-architecture parts, and all of them are reasons for their owners not to buy Arm's finished chip. The harder Arm pushes the AGI CPU, the louder the question "why would we buy Arm's part when we build our own?" gets. And the better those in-house parts sell, the more royalty Arm collects. Arm now has to balance between the ecosystem it grew and the product it sells.

Finally, Qualcomm. It has been angling for a data center CPU re-entry and reportedly went head to head with Arm over the Meta contract. Arm and Qualcomm also have a litigation history over license terms. With an FTC probe live, licensees now hold a real lever — "if our terms get worse, we have somewhere to take that" — and that is a material constraint on Arm's negotiating position.

So what actually changes

If you are an infrastructure or backend engineer — the moment to revisit your x86-assumed deployment pipeline is getting close. If you have already run on AWS Graviton or Google Axion, this is not a shock. But from the second half of 2026, standardizing on arm64 servers in on-prem or colocation environments becomes a real option rather than a thought experiment. Multi-architecture container image builds, JIT runtime tuning, and vendor binary dependencies are the three things that usually trip teams up. Wiring an arm64 lane into CI now is cheap; doing it under deadline is not.

If you are an investor — the thing to take from this quarter is not the revenue record, it is three direction changes. First, the full-year royalty growth outlook came down from 20% to the high teens, meaning cloud AI is not fully covering smartphone weakness. Second, AGI CPU revenue does not book until FYE28, so for at least the next two or three quarters you are judging a pipeline number, not a P&L line. Third, silicon margins are structurally below IP margins, so the bigger that business gets, the more it dilutes total company profitability. Any assessment of the $25 billion FYE31 target has to hold all three at once.

If you are in enterprise IT or procurement — supply at half of demand shows up directly in your purchase terms. Early volume goes to anchor customers like Meta and OpenAI, and everything after that is a lead-time fight. Stack the 2026 memory shortage on top of it: producing a bit of HBM consumes roughly 3x the wafer capacity of DDR5, which keeps server DRAM allocation tight, and server OEMs have been taking reduced allotments. If you have new rack deployments planned, lock your memory procurement schedule alongside your CPU order, not after it.

If you are a founder — the direction of travel is good for you: hyperscaler-class CPU performance becomes something you order rather than something you build. But you will be near the back of the allocation queue for a while, so arm64 cloud instances remain the practical move for now. What matters is keeping your codebase architecture-neutral starting today. Being able to move to whichever camp is cheapest later is itself a cost-reduction option with real value.

If you are just a regular user of AI products — you will not feel this directly. The indirect path is real but slow. A meaningful share of AI service cost sits in data center CPUs and memory, so more competition and better power efficiency can eventually show up in pricing. Then again, when supply lags demand the way it does right now, input costs rise and service prices hold or climb instead. Which way it breaks takes a few more quarters to read.

🥄 Three Things You’re Probably Wondering

— How likely is it that the $2 billion in demand turns into actual revenue? What Arm disclosed is "customer demand," not a contracted backlog. The company explicitly said secured capacity covers about $1 billion, so converting the other half requires winning additional TSMC wafers, substrates, test capacity and memory allocation. Haas saying his confidence "increased" is not the same as saying it is locked, so the only real checkpoint is what the secured-capacity number reads next quarter.

— Could Qualcomm or Nvidia actually leave Arm? Not in the near term. The software assets stacked on top of the Arm architecture are enormous, and migrating to RISC-V is a multi-year project. What is genuinely variable is how license terms get settled depending on the FTC outcome, and large licensees have been quietly increasing RISC-V investment as a hedge for several years now. This more likely shows up as gradually reduced dependence rather than departure — but it is too early to call.

— Will SoftBank merge Arm and Ampere? SoftBank holds roughly 86% of Arm and owns Ampere Computing outright, and both sell data center CPUs, so there is real overlap. Some group-level tidying looks necessary eventually, but neither company has confirmed any integration plan publicly. On the Stargate link, what is verifiable is that SoftBank is an investment and site partner; how much Arm AGI CPU silicon actually goes in is not in any public document, so there is no honest answer to give yet.

Further Reading

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!