Arm shipped a kit for its customers' chips and a customer list for its own, on the same day
Neoverse CSS N4 doubles the licensee subsystem to 128 cores, while Oracle and ByteDance sign up for AGI, the server CPU Arm designs and sells itself.
Arm sold two things on the same day. One was a kit for companies that want to build their own server CPUs. The other was a server CPU that Arm builds itself.
On 8 September 2026 the company launched Neoverse CSS N4, a compute subsystem: a pre-built bundle of cores, cache and interconnect that a customer drops into its own silicon instead of designing from scratch. It scales to 128 cores on a single die made on TSMC's N3P process. The subsystem it replaces stopped at 64.
Against that predecessor almost everything doubles or better. L2 cache per core, L3 per die, PCIe lanes. Arm's own headline claim is twice the socket performance of the Neoverse N3 generation, measured in a configuration it specifies down to the cache size per core.
You can date the product from its connector list
Read the I/O support and you can probably guess when this silicon ships. PCIe Gen7 and CXL 4.0 are not going into anything that arrives next year, and neither is LPDDR6, which the subsystem also supports. So our guess is a design that tapes out in 2027 and turns up in systems in 2028, which is roughly when LPDDR6 memory arrives in volume anyway.
Arm has not named a single partner for CSS N4. It also declined to say what clock a fully populated part actually holds, which usually means the answer is lower than the number on the slide (it generally is).
Neither omission is a red flag.
A handful of large contracts is the entire business.
And on the same day, a customer list for Arm's own chip
Arm also said Oracle and ByteDance will deploy AGI, the dual-die server CPU it designs and sells under its own name, joining Meta, Lenovo, SAP, OpenAI and Cloudflare. AGI carries up to 136 Neoverse V3 cores. Those are the big ones, the class of core that goes into the chips hyperscalers actually care about.
So Arm shipped an IP product for the companies that build server CPUs, and a customer roster for the server CPU it builds instead of them, inside one press cycle. Is that a betrayal of the licensees?
Not quite, and the spec sheets say why. N-series cores mostly land in DPUs, IPUs and cost-sensitive cloud CPUs: Intel's E2100 IPU adapter runs Neoverse N1, and Microsoft's Azure Cobalt 100 ran N2. XSight Labs built a 64-core 800G DPU on CSS N2, which is the honest case for the whole product line. If you are a small team, Arm has already done the base design work and you get to market faster.
Cobalt 200 moved up to V3.
The V-series seat, which is where the margin lives, is the one Arm has now taken for itself.
Nvidia already answered the question everyone is asking
The better argument against worrying is Nvidia, and it points the other way. Grace, its first data-centre CPU, used Arm's Neoverse V2 cores. Vera, its successor, uses Nvidia's own Olympus cores instead.
Arm's largest and most visible licensee stopped buying Arm's cores while keeping Arm's architecture licence, which is the licence nobody gives up. A company watching that happen has a very ordinary reason to start selling a finished chip.
Our read is that the architecture licence outlives the core licence, and that CSS is the answer for everyone who cannot afford a custom core team. That is a smaller, duller and more durable business than designing CPUs to sell into Oracle 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 still has not published for AGI is a benchmark. "More than 2x the performance per rack compared to the latest x86 systems" is an internal estimate, not a measured result. If you are buying, ask for the SPEC run.
