AMD ships its biggest server part
AMD said on Tuesday that its 6th Gen EPYC server processors — the SP7 line built on the Zen 6 architecture — are generally available, with systems already shipping from server makers. The top general-purpose part, the EPYC 9996, carries 256 cores per socket, or 512 in a two-socket machine, according to the footnotes of AMD’s announcement.
The distance between the body of that announcement and its footnotes is where the story sits.
Every number in it is AMD’s
AMD compared the 9996 with a system built on Nvidia’s Vera CPU and reported up to 3.7 times the performance. Against an Intel Xeon 6980P system it claimed up to 2.2 times the performance and twice the performance per dollar. For what AMD calls agent sandboxes — many small isolated environments in which AI agents run — it put a high-density EPYC part at roughly 3.3 times as many agents per CPU watt as Vera, and about 2.5 times as many as the Xeon.
None of those is an independent result. They are AMD’s own measurements on AMD’s own configurations, and the machines they were compared against are described in footnotes rather than in the text.
One headline figure is not a measurement at all. AMD said its high-frequency host-node part delivers up to 1.8 times the language-model output throughput of the previous EPYC generation. The footnote supporting it describes a projected ceiling of 1.8 to 1.9 times, derived from a doubling of PCIe bandwidth — a model of what the extra bandwidth ought to buy, rather than a benchmark that was run.
AMD also qualified its “world’s best” language in the same document. The claim reflects the company’s own assessment of the general-purpose enterprise server CPUs it had evaluated as of 23 July 2026, and AMD says it does not mean leadership on every workload.
Racks, not only sockets
Alongside the processors AMD introduced Helion, a set of rack-scale systems aimed at running large numbers of agents. Helion Liquid is meant to sit beside AMD’s Helios accelerator racks; Helion Air is for air-cooled halls; and Helion Air Inference puts CPU execution and local GPU inference in the same rack. AMD said some Helion configurations reach 73,728 threads per rack.
That framing — the CPU as the place agents actually run, rather than as the part that keeps a GPU fed — is the argument the launch rests on. Dan McNamara, AMD’s senior vice-president and general manager for compute and enterprise AI, said AI “is significantly expanding the role of the CPU across every layer of the data center”.
Who has it
Systems are shipping now from server makers and contract manufacturers. AMD said Meta is validating 6th Gen EPYC platforms ahead of a planned deployment at scale. In the cloud, Microsoft’s Azure HDv2 is in preview, with Azure HXv2 and the Helios-based Azure ND MI455X v7 listed without dates, and Oracle’s OCI Compute E7 instances will use the processors. AMD said cloud instances from both providers would be generally available “soon”, which is the only timing it offered.
The company spent the same week arguing the other side of the host-CPU question too. On Monday it said its previous-generation EPYC parts serve as the host for Google Cloud’s TPU 8tx, which AMD describes as the first cloud TPU system with hardware-enforced confidential computing.
What to watch
The useful numbers will come from somewhere other than AMD. Each of the three comparisons becomes checkable the moment an independent party runs the same workload on an EPYC 9996 and on the Nvidia or Intel system AMD named. The second thing to watch is the gap between “shipping from server makers” and a dated general-availability notice from Azure or OCI, because that is when the parts reach buyers who are not assembling racks themselves.
