Arm now sells a chip, and its licensees are buying cores from a competitor

Divide the air-cooled rack's cores by its 30 compute blades and you get 272 a blade, which is two of the 136-core parts in each slot ^[5][16]. Move to the liquid-cooled design and 5.6 times the power buys 5.6 times the cores ^[17], or roughly 4.4W a core either way ^[18]. The 200kW rack is a floor-space purchase, not an efficiency one, and nothing in the two validated configurations suggests otherwise.

The silicon is Neoverse V3 cores at up to 3.7GHz across two dies, on TSMC's 3nm process ^[3][4]: a general-purpose server CPU carrying a name the parts list does not earn. Eddie Ramirez, a vice president in Arm's cloud AI unit, describes it as what cloud providers have been building for a couple of years, made available to the general market ^[7]. He is describing his own licensees' work.

The neutrality question follows from that. Arm collected the "Switzerland of chip design" tag by holding no products of its own and licensing to all comers without preference ^[1]. It now holds one, built to a single hyperscaler's requirements ^[2], and that hyperscaler is separately building custom silicon with Broadcom ^[10]. Any cloud operator running a Neoverse design programme is licensing cores from a company whose sales team also carries a finished part into the same rack decisions.

Arm's justification is a cost floor. Ramirez puts the entry price for a chip like this north of $500m, says that is unrealistic for a large part of the market, and concedes Arm's cloud wins have come mostly from firms able to fund full silicon teams ^[8][9]. Meta sits above that line and is already spending above it, but came to Arm because it was the last hyperscaler without significant Arm adoption and wanted to move quickly ^[10][11]. To its credit, Meta insisted the result be a general-purpose CPU rather than a Meta custom chip ^[12], and Paul Saab, a Meta engineer on the project, framed the work at OCP's 2026 EMEA summit in Barcelona in April as advancing open platforms and standardisation across the stack ^[13].

The customer that actually matches the pitch is Verda, the AI cloud formerly called DataCrunch, which runs facilities in Finland and Iceland and has committed to deploying the part without saying when ^[14]. Founder Ruben Bryon says the aim is to pair it with Nvidia GB300 and upcoming VR200 fleets for an Arm-native stack "from orchestration to inference" ^[15]. The CPU's job in that account is hosting somebody else's accelerators.

And the performance claim points at the wrong opponent. Arm's comparison is with Intel x86 platforms, on its own internal benchmarks ^[19], an incumbent it has been eating into for years. The designs this part will actually be quoted against are the Neoverse implementations Arm licensed to its own customers, and for that comparison there is no number at all.