AMD ships its sixth-generation EPYC as Qualcomm silicon joins the same code stack

What AMD shipped

6th Gen AMD EPYC SP7 CPU-based platforms are now shipping from leading server OEMs and ODMs, AMD says, powered by the new AMD Zen 6 architecture. The EPYC 9006 SP7-based platforms are shipping now, and they are the first of four planned processor families in the portfolio, with SP8, 9006X SP7 and 9006 LP Server CPUs to follow. The launch is framed explicitly around agentic AI: the host CPU, not the accelerator, is the pitch. Meta is working with AMD as a lead customer to validate 6th Gen EPYC platforms, and Microsoft Azure is bringing 6th Gen AMD EPYC to customers across the AI infrastructure stack, while Oracle Cloud Infrastructure Compute E7 instances will be powered by 6th Gen AMD EPYC server processors. Three of the largest buyers of AI infrastructure are named as customers in the same announcement, which is the part of a launch that actually matters: a chip that ships is a product, and a chip that three hyperscalers are validating is a platform. m:…6845 m:…7086

The numbers AMD puts forward are comparisons, and they should be read as the company's own claims. The AMD EPYC 9996 CPU delivers up to 3.7x more performance compared with an Nvidia Vera CPU-based system, AMD says, and high-density AMD EPYC processors provide up to 3.3x agents-per-CPU-watt advantage over Nvidia Vera. Against Intel, the EPYC 9996 delivers up to 2.2x the performance of an Intel Xeon 6980P system. The underlying hardware: EPYC 9006 SP7 Server CPUs feature up to 256 cores and 512 threads per socket, support JEDEC-standard MRDIMM-12800 memory, provide up to 1.6 TB/s of memory bandwidth per socket, and feature PCIe Gen 6 I/O, with frequencies up to 5.3 GHz with 96 cores. AMD Helion racks can reach up to 73,728 threads in one rack. The metric worth pausing on is agents-per-CPU-watt, because it is a measure of the host side of agentic workloads: how many agents a CPU can feed and coordinate per watt, rather than how fast one model runs. If agents are the workload, the host CPU is no longer a bystander to the GPU, and AMD is pricing its pitch in exactly those terms. m:…6845 m:…7086

Qualcomm enters through the same code

The second half of the week's challenge comes from a different direction: not a faster chip, but the same code running on a different chip. Qualcomm Technologies' data center AI accelerators are now integrated into the Modular Platform, Modular says, and developers can target Qualcomm Technologies' silicon using the same code stack as NVIDIA and AMD GPUs. The Qualcomm AI accelerator roadmap includes the Qualcomm Cloud AI 100, Qualcomm Dragonfly AI 200 and future generations. The bring-up of the Gemma 4 31B model on the Qualcomm Cloud AI 100 took under six months, which is the number that tells a buyer whether porting is a project or a product decision. Six months to bring up a 31-billion-parameter model on new silicon is fast enough that a team can evaluate the hardware without betting the company on it, and that is the threshold that matters. m:…0708

The technical detail in Modular's account is about what had to be built to make the same code work. Qualcomm Cloud AI 100 Ultra features 4 devices, each with 32 GB of RAM and 16 NSPs, and Gemma 4 31B requires tensor parallelism across four chips due to memory constraints. Modular implemented MXFP6 weights and an int8-based KV cache for the Gemma 4 optimization, and optimization work on GPT-2 reached 6.7x the baseline performance. GPT-2 124M was the first complete model supported on the Qualcomm Cloud AI 100, and speculative decoding was implemented using smaller Gemma 4 assistant weights. The Qualcomm Dragonfly AI 200 supports native FP8 and BF16 formats. None of this is a claim that Qualcomm silicon matches Nvidia today; it is a claim that the porting cost is falling. And porting cost is the moat. Nvidia's advantage has never been only the chip, it has been that everyone's code already runs on it. A week in which AMD ships a host CPU pitched in agents-per-watt and Qualcomm silicon joins the same code stack is a week in which that moat gets measured from two sides at once. m:…0708

What a buyer should do with this

The practical move is not to switch stacks, it is to price the switch. Meta is validating AMD's platform as a lead customer, Microsoft is bringing it to Azure, and Oracle is building instances on it, which means a buyer can now ask for a real quote on the alternative instead of a theoretical one. On the accelerator side, the Modular integration means a team can try Qualcomm silicon without rewriting its models, which turns a strategic question into a benchmark: run the workload, measure it, and compare. The numbers in both announcements are the vendors' own, and independent testing is still the missing piece. But the shape of the challenge is now clear. It is not one challenger beating Nvidia at its own game. It is the host CPU getting bigger and the accelerator field getting wider at the same time, and a buyer who only watches GPU prices will miss both. m:…6845 m:…0708