AMD puts CPUs back at the center of agent workloads

The 6th generation Epyc, codenamed Venice, arrives with a simple argument: agentic AI gives CPUs renewed importance in data centers. AMD bets on more cores, more bandwidth, and a range designed for multiple server uses.

On November 10, 2026, AMD launched its 6th generation Epyc server processors, codenamed Venice, with systems already being delivered to several manufacturers and cloud instances expected soon on Microsoft Azure and OCI. The message is clear: agentic AI, which orchestrates tasks, queries databases, and calls tools, restores the CPU's central role in the data center. And the processor no longer just plays a supporting role while GPUs take center stage.

When agents put the CPU back in the spotlight

For AMD, the shift comes from the very nature of agents. A chatbot responds to a request; an agent breaks down a task, chooses an action, consults an API, verifies a result, then repeats. This cycle requires CPU-side computation for orchestration, security, data access, and tool execution. Dan McNamara, senior vice president and general manager of the Compute and Enterprise AI division, sums up the change in one sentence: "AI is significantly expanding the role of the CPU across every layer of the data center." AMD thus estimates that the CPU-GPU ratio moves from about 1:4 for chatbots to a ratio tending towards 1:1 for agentic AI.

This is not just a product marketing story. In its own projections, AMD sees the addressable market for data center CPUs rising from about $26 billion in 2025 to around $220 billion in 2030. This estimate remains [to be verified] until confronted with independent measurements, but it clearly shows where the company's bet lies: on the resurgence of CPU demand, not on a head-to-head race with Nvidia GPUs.

Venice, a family designed for multiple server profiles

Venice is not a single chip but a family. The 10 announced processors are divided into two main profiles: models with a very high number of cores, and versions with faster cores. The high-end SP7 reaches 256 Zen 6c cores and 512 threads, with up to 16 memory channels and a power envelope up to 600 W. The other end focuses on 96 Zen 6 cores with a boost frequency up to 5.3 GHz, for latency-sensitive workloads or software licensed per core.

AMD emphasizes that these platforms are not just variants of the same block. SP7 and SP8 are two distinct sockets, each with its own I/O and memory logic. SP7 targets density, consolidation, and GPU hosts; SP8, expected in the first half of 2027, will target general-purpose servers, storage, and edge. This segmentation matters because it indicates that the manufacturer no longer wants to sell just cores but deployment formats. It’s less flashy than a talking robot demo, but it bills better.

Highlighted gains: memory, bandwidth, and security

AMD also highlights technical improvements on Zen 6. The cores use TSMC’s 2 nm process and deliver an IPC increase announced between 27% and 31% depending on presentations. Zen 6c also closes the cache per core gap with Zen 6, which changes the game for dense workloads. The manufacturer adds FP16 support, better INT8 handling, PCIe Gen 6, CXL 3.1, and on the security side, Caliptra 1.2 as well as PCIe link encryption and post-quantum cryptography.

Memory remains a sensitive point. On SP7, AMD announces up to 1.6 TB/s memory bandwidth with MRDIMM, compared to 614 GB/s on the Turin generation, as well as capacity reaching up to 8 TB per socket. The manufacturer also highlights a “unified power provisioning” logic, intended to recover part of the power usually reserved for memory. According to its calculations, this could free enough to add one to two servers in a 35 kW rack [to be verified depending on configuration].

Comparisons with Intel, Nvidia, and Arm: figures to be taken with caution

AMD presents several comparisons against Intel, Nvidia, and Arm. In some scenarios, the company claims up to 3.3 times more agents per watt than Nvidia Vera, 1.8 times more agents per dollar than a 6th generation Xeon, and up to 2.8 times more agents per watt than an Arm AGI CPU. Other figures point in the same direction: 1.9 times the throughput of the Turin generation, 2.4 to 3.5 times the results of the Xeon 6980P on agentic pipeline tests, and up to 1.8 times more tokens per second on certain GPU host workloads. These measurements are provided by AMD and often concern the high end of the range; they therefore require confirmation by third-party tests.

The most useful comparison for buyers is probably the field one. On existing fleets, AMD claims that replacing Xeon servers deployed around 2021 with Venice could reduce the number of machines by about 86% and the five-year cost by 37% at equivalent throughput. This projection obviously depends on the actual price of licenses, cooling, and application profile. But it explains why CIOs are paying close attention: infrastructure modernization can help free up space for AI, not just swap boxes in a white room.

What this means for businesses

For businesses, Venice’s interest lies mainly in the convergence between server renewal and the rise of agentic AI. Many fleets launched between 2019 and 2021 are reaching the end of their life cycle, while energy bills and memory costs remain high. AMD wants to offer a single answer to these constraints: consolidate classic workloads, host agent sandboxes, run databases, and serve as a host for GPUs when inference acceleration is needed. Pilot deployments on Microsoft Azure HDv2, in preview, go in this direction.

One simple condition remains: proving gains on real workloads. AMD itself recommends testing total wall power consumption, actual cost under existing licenses, performance on representative workloads, and agent behavior when multiple sessions run in parallel. This is the only way to avoid choosing a very fast server... for software that mostly charges for patience.

Ultimately, Venice marks less an isolated breakthrough than a repositioning. AMD is not leaving the GPU battle, but it is reinstalling the CPU at the center of data center architecture at a time when agents become more demanding, more numerous, and more intrusive in enterprise systems. For a market long convinced that everything went through the GPU, this is already a change of scenery.

Key points

  • Launch on November 10, 2026, with first systems shipped.
  • AMD targets up to 256 cores and 512 threads on SP7.
  • Azure and OCI expected to offer Venice instances soon.
  • AMD aims for 3.3x more agents per watt than Vera.
  • SP8 expected in the first half of 2027.
  • Dan McNamara speaks of an expanded CPU role driven by AI.

By the numbers

  • 10 processors — the Venice family announced by AMD, November 2026.
  • 256 cores — maximum announced on SP7 models, AMD, November 2026.
  • 1.6 TB/s — memory bandwidth announced on SP7 with MRDIMM.
  • $26B to $220B — addressable data center CPU market, AMD estimate 2025-2030.
  • 86% — announced server count reduction in a replacement scenario.

To read