Arm doubled its licensee compute subsystem and named two customers for its own CPU on the same day

Arm launched Neoverse CSS N4 on September 8th: 8 to 128 Neoverse N4 cores on a single die built for TSMC N3P, up to 3.8GHz, 256MB of L3 per die, 2MB of L2 per core, DDR5 or LPDDR6, and up to 128 lanes of PCIe Gen7 with CXL 4.0. Armv9.3 becomes the baseline, and the cores add FP8 and MMLA.

Against the N2 subsystem it replaces, almost everything doubles or better. N2 stopped at 64 cores, 1MB of L2 per core, 64MB of L3, and 64 lanes of PCIe 5.0.

Those figures come with a configuration attached — 128 cores at 3GHz with 2MB of L2 each — and Arm has not said what clocks a 128-core part actually holds. Nor has it named a single partner. Which is normal for CSS launches, and not much of a risk: a handful of large contracts is the whole business.

Read the I/O list and you can probably date the product yourself. PCIe Gen7 and CXL 4.0 are not going into anything shipping next year. ServeTheHome makes the same point about the memory support, which stretches to MRDIMMs at 8000 to 12000 MT/s, and the interconnect is a CMN S4 mesh scaling to 16×16 with coherent multi-chiplet links over CHI C2C and UCIe. So this is likely a subsystem for silicon that tapes out in 2027 and appears in 2028 (which is also roughly when LPDDR6 volume arrives).

And now the awkward part.

On the same day, Arm said Oracle and ByteDance will deploy its own AGI CPU — a dual-die part with up to 136 Neoverse V3 cores, 272MB of L3 and up to 6TB of memory — alongside Meta, Lenovo, SAP, OpenAI and Cloudflare. Arm shipped an IP product for the companies that build server CPUs, and a customer list for the server CPU it builds itself, in one press cycle.

We would not call that a betrayal. Arm has been open about AGI serving as a demonstration vehicle, and its V-series is still under everyone's biggest chips: AWS Graviton, Google Axion, Azure Cobalt 200. The tension is quieter than a land grab, and it seems to show up in what CSS N4 is optimised for. N-series parts go into DPUs, IPUs and cost-sensitive cloud cores. Intel's IPU Adapter E2100 runs Neoverse N1. XSight Labs built a 64-core 800G DPU on CSS N2, which is exactly the case for CSS: if you are a small team, Arm has already done the base design. The V-series seat, the one Arm now occupies with its own product, is where the margin is.

The better argument against worrying is Nvidia, which points the other way entirely. Grace used Neoverse V2. Vera, announced two weeks ago, uses Nvidia's own Olympus cores. Arm's largest and most visible licensee has already stopped buying Arm's cores while keeping Arm's architecture (the licence that matters, and the one nobody gives up). A company losing core sales to its own customers has an obvious reason to sell a finished chip.

So which way does this go? Our read is that the architecture licence outlives the core licence, that CSS is Arm's answer for everyone who cannot afford a custom core team, and that this is a smaller and more durable business than designing CPUs against Oracle's other suppliers. We would expect the first named CSS N4 partner to be a networking or DPU vendor rather than a general-purpose server CPU, and to be announced before the middle of 2027.

One number Arm has not published for AGI, more than a year after first talking about it, is a benchmark. "More than 2x the performance per rack compared to the latest x86 systems" is an internal estimate. If you are buying, ask for the SPEC run.