This piece separates what primary sources confirm from what vendors claim. It also gives you the checks to run before you move anything.
The headline claim: "more than 2x performance per rack"
Arm launched the Arm AGI CPU on March 24, 2026. The launch was the first time the company extended into production silicon products. Arm describes it as its first Arm-designed data center CPU, aimed at agentic AI infrastructure, and says it delivers more than 2x performance per rack versus x86 CPUs.
Arm's published specifications are:
- Up to 136 Arm Neoverse V3 cores per CPU
- 6GB/s of memory bandwidth per core at sub-100ns latency
- A 300-watt TDP
- Air-cooled deployments of up to 8,160 cores per rack, and liquid-cooled systems with 45,000+ cores per rack
Two caveats matter here. First, the press release footnotes its headline comparisons with "Based on estimates." Treat "more than 2x" as an Arm claim, not a lab result. Second, "per rack" is a density-and-power metric. It says nothing about per-core speed, per-socket throughput or laptop behavior.
The secondary coverage describes the comparison as modeled and workload-specific, with a validation partner involved. I couldn't confirm those methodology details from a primary document. I'm leaving them as unverified, and I'd suggest you do the same.
Timing also matters. Per the secondary report of Arm's November 4 results, first commercial deliveries are targeted for Arm's fiscal Q4 2027. That coverage also says Oracle plans AGI CPU instances in 2027. I couldn't fetch Arm's shareholder letter, so these are company targets reported secondhand. Neither means the chip is widely shipping today. No cloud pricing was verifiable.
What x86 has actually shown
AMD reported on November 3, 2026 that Data Center segment revenue was a record $8.5 billion, up 97% year over year. AMD attributes it to strong demand for EPYC processors and the launch of Instinct MI450 Series GPUs.
That is demand evidence, not a benchmark. It doesn't show x86 is faster or holds more share. It does show x86 server demand is climbing even as Arm pushes in. Revenue, specs, rack density and cost per unit of work are different kinds of evidence, so don't swap one for another.
Windows: where Arm is ready, and where it isn't
For Windows readers, the practical question is compatibility. Microsoft's documentation gives a clear picture:
- Prism is Microsoft's emulation technology that lets x86/x64 apps run on Windows PCs with Arm processors.
- Windows 11 Arm-based PCs running version 24H2 include the Prism emulator for improved performance on emulated apps.
- Microsoft says Prism gained support for translating more x86 extensions, including AVX and AVX2, plus BMI, FMA and F16C. That rolled out with Windows 11 versions 24H2 and 25H2.
The hard limit is drivers. Microsoft's support FAQ says drivers for hardware, games and apps only work if they're designed for a Windows 11 Arm-based PC. It adds that if a driver doesn't work, the app or hardware that relies on it doesn't work fully either. Microsoft's Surface FAQ says specialized hardware or older peripherals might require verification for Arm64 compatibility.
This is the part the broader Arm-vs-x86 debate tends to skip. Emulation covers applications, not low-level system code. I didn't find a Microsoft source naming specific security products. Even so, the driver rule means anything that installs a kernel driver needs an Arm64 build. Typical examples are VPN clients, EDR agents and anti-cheat. Verify each one with its vendor.
Apple's Rosetta timeline is a Mac issue, not a Windows one
Apple's developer notice dated September 1, 2026 sets out its own timeline:
- On macOS 26.4 or later, users may get a system notification when launching apps that rely on Rosetta.
- macOS 27 is the final release to support Rosetta, and Intel-only apps will no longer run on Apple silicon Macs after that update.
- Rosetta functionality for older, unmaintained gaming titles that rely on Intel-based frameworks will continue to be supported.
- Apple's advice to developers is to build and distribute apps as universal binaries.
This is a platform-specific migration deadline. It is not a Microsoft policy and it doesn't signal an end to x86 on Windows. It's useful mainly as a template for what a vendor-forced architecture shift looks like. Windows has no equivalent deadline.
Picking by workload
These are decision criteria drawn from the verified facts. They are not benchmark rankings.
| Workload | What the evidence supports | What to check first |
|---|---|---|
| Windows apps, drivers, peripherals | Arm is workable for mainstream apps via native builds and Prism. Drivers are the hard limit. | Arm64 drivers for every peripheral, plus your security and VPN agents |
| Legacy or specialist Windows software | x86 avoids the emulation and driver question entirely | Whether the vendor ships an Arm64 build |
| Linux services and containers | Arm is worth piloting where your images and dependencies build for arm64 | Build arm64 images and measure cost per unit of work |
| Agentic AI server orchestration | Arm positions the AGI CPU here, but delivery targets run to 2027 and neutral tests are missing | A proof of concept on your own code, with equivalent rack configurations |
| Existing x86 estates | Strong AMD data-center demand shows x86 isn't standing still | Whether a measured gain justifies migration cost |
A sensible migration approach
Treat Arm as a second target before treating it as a replacement.
- Inventory binaries and agents. List native libraries, base images, drivers and endpoint agents. Flag anything that is x86-64 only.
- Build for both architectures. For containers, a multi-platform build with a single tag keeps rollback easy. On Windows, confirm which apps are native Arm64 and which run emulated.
- Test on real Arm hardware. Emulated test environments can hide timing and driver problems.
- Run a pilot slice. Compare latency, error rates and cost per completed unit of work against your x86 baseline.
- Keep a rollback path. Hold x86 capacity until the numbers stay steady through a full business cycle.
If an emulated Windows app misbehaves, Microsoft documents a per-app fallback. Right-click the app's executable and choose Properties. On Windows on Arm, the Compatibility tab has a Windows on Arm section, where "Change emulation settings" opens the Arm emulation settings window. Moving from Default to Safe to Strict to Very strict trades performance for potentially better compatibility.
Bottom line
The defensible conclusion is narrower than the "Arm wins" or "x86 wins" headlines. Arm's more-than-2x rack claim is real as a claim, but it comes from Arm and rests on estimates. The AGI CPU is still a delivery-targeted product. AMD's 97% data-center growth shows x86 demand is strong. On Windows, Arm is viable for mainstream apps, and drivers and kernel-level software remain the deciding factor.
My view, from general industry experience rather than any single source, is simple. Pilot Arm where your software already builds for it, and keep x86 where compatibility is the cost driver. Wait for independent, same-workload benchmarks before committing a data center to any vendor's rack-level ratio.