Arm Says Its Chips Power Over Half of Top Hyperscaler Servers: What Is Verified and What Is Claimed

A headline now says Arm-based chips power over half of the top hyperscaler servers. The claim is partly backed by Arm's own filing and partly stronger than it. This post separates what Arm wrote, what an executive said, and what nobody has measured independently.

Short answer: Arm told investors it holds "about 50%" of CPU compute among top hyperscalers. "More than 50%" comes from an executive interview. Both are company claims.

TL;DR: claim by claim

What Arm actually wrote

In the shareholder letter filed with the SEC, Arm wrote: "Arm's market share of CPU compute now represents about 50% share among top hyperscalers, demonstrating that Arm is at scale in the data center." The same letter ties the scale to Arm Neoverse compute subsystems (CSS) and reports a record $1.49 billion quarterly revenue, with data center royalty more than doubling year over year.

Three details in that letter are worth keeping apart from the headline.

• The wording is "about 50%", not "over 50%".
• The unit is "CPU compute", which is not defined. It could mean installed cores, shipped chips or capacity.
• The audience is investors. The letter is a promotional document, even though it is a regulatory filing.

What the Nikkei interview added

AI Weekly, summarizing the June 25, 2026 Nikkei Asia report, says a senior Arm executive put Arm's share of the hyperscale cloud computing market above 50 percent, in an interview in Taipei. The same summary notes two caveats: the figure comes from a company executive rather than an audit, and how share is measured is not specified. It also cites Tom's Hardware reporting Arm servers at over 45 percent of data center market revenue, a related but different metric.

We could not open the Nikkei article, so we rely on that relay for the executive's words. That is a secondary source.

The chips behind the number

Arm's letter names the products it counts as proof of scale.

• AWS Graviton. Arm says AWS reported its custom silicon business, including Graviton, Trainium and Nitro, is running at more than $20 billion annually and growing triple digits. It quotes up to 40 percent better price performance than x86 for Graviton, and says Anthropic uses Trainium alongside tens of millions of Graviton cores.
• Google Axion. Google's TPU 8t and 8i will replace x86 host processors with Axion CPUs, which Arm says are up to 2x better in performance per watt. We covered the TPU line in Oracle's plan to offer Google TPUs in 2027.
• Microsoft Cobalt. Deployed across a substantial portion of Azure regions, per the letter.
• NVIDIA Grace and Vera. NVIDIA's Vera CPU is described as up to 50 percent faster and 2x more efficient than x86 CPUs, and designed into a rack of 256 Vera CPUs. These are vendor figures.
• Arm AGI CPU. Arm's first production data center chip, with Meta as lead partner. Arm says it has more than $2 billion of customer demand across fiscal 2027 and 2028 and claims more than 2x performance per rack versus x86 platforms.

Note that most performance figures here are vendor claims about specific workloads, not neutral benchmarks.

How the claim evolved: a short timeline

The 50 percent figure did not appear overnight. Each step below is a company statement.

1. Mid-2025 guidance. In its fiscal Q1 2026 materials, Arm projected that Neoverse-based chips at top hyperscalers would reach nearly 50 percent this year. This was a forecast.
2. February 2026. On the fiscal Q3 call, management said Neoverse had passed 1 billion deployed cores and that share among top hyperscalers was expected to reach 50 percent.
3. May 6, 2026. The fiscal 2026 shareholder letter moved from "expected" to "now represents about 50%". This is the first time in the documents we read that the figure is stated as current.
4. June 25, 2026. The Nikkei interview, as relayed, raised it to "more than 50%".
5. November 4, 2026. Headlines repeat "over half". Arm's scheduled earnings release the same day may carry new wording, which we could not read.

The pattern is a forecast that became a present-tense claim, then rounded up in retelling. That is common in company metrics, and it is why we ask for the denominator.

The business behind the claim

Arm has reasons to publish a clean number. The same letter reports fiscal 2026 revenue of $4.92 billion, with royalty revenue of $2.61 billion and licensing revenue of $2.31 billion. It also says the company has begun selling its own silicon, the Arm AGI CPU, which changes its position: until now Arm licensed designs to customers such as AWS and Google. A high hyperscaler share makes the case that customers already trust Arm at data center scale, which supports demand for the new chip.

That context does not make the number false. It does mean a reader should weigh it as a sales and investor message. The figure of more than $2 billion of customer demand for the AGI CPU across fiscal 2027 and 2028 is likewise a demand statement, not shipped revenue. Arm's separate target of $15 billion in that business is a forecast.

What would make the claim checkable

A neutral check needs three things: a stated unit, a named set of "top hyperscalers", and a date. If Arm published its method, analysts could test it against known deployments. Cloud providers could also help by reporting the share of new server capacity that is Arm-based, as some have done for single product lines. Until then, the safest wording is the one in the filing: Arm says about half, by its own measure.

Why "share" is hard to pin down

Four different measurements all get called share.

1. Units shipped. How many server CPUs a customer bought in a period. Hyperscalers buy and deploy in batches, so this swings.
2. Installed cores. Arm cores are often smaller, so a core count can overstate or understate compute.
3. Capacity or power. Share of megawatts or throughput, which needs data the clouds rarely publish.
4. Revenue. Share of server CPU spend. Tom's Hardware's 45 percent, as relayed, is of this type.

A 50 percent claim can be true under one and false under another. Until Arm or an analyst publishes the method, it is a company signal rather than a statistic.

Why this matters for AI builders

Arm's argument is that agentic AI shifts demand to CPUs: "more than 4x current CPU capacity per gigawatt" in its words, and a market above $100 billion by 2030. That is a forecast, not a measurement. The practical effects are already visible, though.

• Build for arm64. If your cloud host nodes or serverless runtimes are Arm, container images and wheels need arm64 builds. Graviton, Axion and Cobalt instances are often cheaper per vCPU.
• Check price-performance on your own workload. The 40 percent and 2x figures are vendor numbers for chosen tests.
• Watch the host-CPU shift in AI racks. Arm hosts next to accelerators change firmware, drivers and tooling. For a view of local AI hardware, see our look at Qualcomm's Cloud AI 100 Ultra and the Gigabyte GB300 deskside workstation.

The supply side also matters. Memory and power limits shape how fast any CPU architecture can grow, as in Altman's visit to SK Hynix and Samsung and Anthropic's compute plans. Power deals are covered in every hyperscaler's nuclear deal.

How to check the claim yourself

1. Read the fiscal results letter on Arm's investor relations page and search for "market share of CPU compute". Note the wording and the footnotes.
2. Compare it with the earlier statement that Neoverse chips would reach nearly 50 percent at top hyperscalers in 2026. That was guidance, not a measurement.
3. Look for third-party server CPU trackers and check whether they count revenue, units or cores.
4. Ask the hyperscalers. AWS has said publicly that Graviton is a large share of new CPU capacity in some cases, but check each statement for its denominator.
5. Treat any "over half" headline as unverified unless it links to a document with the unit stated.

What we could not verify

• Today's earnings. Arm was scheduled to report fiscal Q2 2027 results on November 4, 2026. We could not retrieve that release, so we cannot confirm whether it updated the share wording to "over half".
• The Nikkei text. We relied on a relay of the interview.
• Any independent figure that matches 50 percent on the same definition.

We will update this post if the company's own documents show a new wording.

Figures and quotes are accurate as of November 4, 2026. Company-reported figures are marked as such.

Related reading

• Oracle will offer Google TPUs on OCI in 2027
• Qualcomm Cloud AI 100 Ultra specs
• Gigabyte GB300 deskside workstation vs DGX Spark
• Anthropic compute: NYT vs September claim
• Every hyperscaler's nuclear deal
• Altman's Korea memory talks