2026 OCP APAC Summit | From 1:4 to 1:1: AMD's OCP 2026 Keynote Wasn't About 'Open' — It Was About the CPU's Comeback in the Agentic Era

2 days ago

7 min read

At OCP 2026 APAC Summit, Ravi Kuppuswamy of AMD's data center business wrapped his keynote in the message that "open is the only answer." But only one signal really deserves underlining: agentic AI is pulling the data center CPU-to-GPU ratio from the old 1:4 back to 1:2 — and even 1:1. This isn't another round of the accelerator arms race. It's latency-sensitive work — orchestration, retrieval, tool execution — pushing CPU compute back to center stage. For Taiwan's supply chain, that one sentence carries far more weight than the row of three-letter open-standard acronyms on the slides.

1. The real signal of this keynote isn't the word "open"

Set the stage talk aside. Ravi's most emphatic line was "Open is no longer a choice, open is a must," but open has never been news to an OCP audience — everyone in that room already believes it. The line that actually deserves a pause came midway through:

In the past, each CPU was paired with roughly four GPUs; in the agentic era, that ratio is converging toward 1:2, and even 1:1.

For the past two years the narrative was "the GPU is everything; the CPU is an accounting rounding error." Training is batch work that runs on raw compute density; the CPU just has to keep the data flowing. Agentic workloads have the opposite profile: latency-sensitive, memory-heavy and inherently distributed. They need CPUs for orchestration, retrieval systems to feed context, and a low-latency network to stitch GPU inference together. Once the work pattern shifts from "one question, one answer" to "a swarm of remote agents each running tasks and then converging," the CPU is the component that gets called back up.

The fact that AMD took this stage to talk about a "ratio reversal" is itself the most direct industry signal: a company that sells both CPUs and GPUs is telling you the value of the CPU line is coming back. It's two sides of the same inflection as Jensen Huang stating openly that "optical connectivity is the critical path to revenue in the agent era." We broke down the demand-side logic in Jensen Huang Disaggregated Compute: In the Agent Era, Optical Connectivity Is the Critical Path to Revenue; this AMD keynote spells out what the rack looks like on the supply side.

2. Why agentic workflows pull the CPU back to the center

Ravi broke an agentic task into stages: it enters through an agent gateway, moves into reasoning, then into tool execution, and only then returns a verified answer. The key was the line that followed — AI is not a monolithic workload. You can't design "the single strongest CPU" and have it swallow the whole pipeline, because each stage has different compute needs and each pulls in a differently configured platform.

The real industry implication: in this seven- or eight-stage pipeline, CPU racks handle roughly five critical steps — gateway ingress and scheduling, context retrieval, the glue layer for tool calls, result verification, and the coordination that ties the whole agent swarm together. None of these are GPU jobs, yet every one directly sets end-to-end latency. That's the underlying reason 1:4 can't hold and the ratio has to move toward 1:1. It's not that GPUs matter less; it's that "single-chip peak performance" is now necessary but not sufficient. You have to layer on rack density, energy efficiency, cost, reliability — and the most easily overlooked line of all: don't pay the infrastructure tax.

In other words, the agentic data center is no longer a contest of "whose accelerator is fastest" but "whose entire pipeline jams the least." And the part of the pipeline most likely to jam is usually the fabric that links all those GPUs and CPUs together.

3. Unpacking AMD's "four pillars of open" to see what it's really selling

Ravi split open into four words: composable, verifiable, interoperable and deployable. It sounds like canned slideware, but map each layer back to real hardware and each pillar lines up with a slice of the business Taiwan's supply chain can capture:

Composable means modular server architecture plus standardized accelerator integration — OCP cards and management modules like DCSCM make "one mechanical design, different compute" possible. This layer maps directly to ODMs' modular design capability.

Verifiable was one of the few parts of the keynote delivered with real feeling. Ravi repeatedly stressed the importance of hardware Roots of Trust like Caliptra — if you can't even trust your own data and the platform isn't standardized, safeguarding a customer's workloads is nearly impossible. This layer turns security from a "nice-to-have" into the foundation, and it maps to RoT chips and the attestation chain.

Interoperable is about the network. Ravi named UALink and UEC on the scale-up and scale-out paths respectively — using open standards to stitch individual pieces of silicon into a fabric, so that "a swarm of agents talking across data centers" physically works. This is exactly the scale-up battleground STT has been tracking all along; AMD jumped in long ago, joining Meta and Broadcom in Meta, Broadcom and AMD Jointly Define the OCI 200G Line-Side Spec to bet on single-fiber bidirectional, micro-ring DWDM and external lasers for scale-up. Why scale-up is the endgame and why scale-out moves first is fully worked through in The Great Optical Packaging Transition (Part 2): CPO's Three-Stage Evolution.

Deployable is the rack layer — standardized racks are what let CPU and accelerator platforms roll out quickly. Ravi specifically mentioned extending ORv3 to a double-wide format (ORW) to serve both training and inference. OCS — once a technology only Google could afford to run internally — entering the broader industry toolbox is an extension of the same "deployable" logic, which we covered in OCP 2026 White Paper: OCS Enters the Whole Data Center Industry's Toolbox.

4. Putting numbers on the timing: how fast this upgrade cycle is moving

A trend's speed only carries weight when expressed in numbers. Pieced together, the hard metrics from this keynote form a timetable for a data center being rewritten:

On ratio, CPU-to-GPU goes from 1:4 toward 1:1 — in the same data hall, that lifts CPU procurement volume by an order of magnitude. On density, AMD showed a rack design with 4 CPUs per OU, which together with the ORv3 double-wide extension pushes per-rack compute density higher still. On process, the next-generation Venice CPU runs on TSMC 2nm, which Ravi positioned as "the best, most efficient" CPU in the world. Add DCSCM for system management, MRC (Multipath Reliable Connection) for fabric resilience, OAC's modular building blocks and Caliptra's security attestation, and this isn't a single-point upgrade — it's a synchronized refresh from silicon and firmware to fabric and rack.

When four CPUs are packed into one OU and each one jumps to 2nm, the rack's power delivery, cooling, connectors and board layer counts all have to be redrawn. That's where Taiwan's supply chain should really be watching — the more a refresh becomes "replace the whole rack at once," the more it rewards system integration capability rather than selling a single component.

5. The counterargument: AMD's position shapes the angle of this slide

STT's rule is that after the good news, we draw the knife. Three parts of this keynote deserve a discount:

First, the speaker's position carries an obvious bias. AMD sells both CPUs and GPUs, and a "1:1 ratio" conclusion is most favorable to it — the more CPUs sold, the better. A pure-GPU player would never draw the same slide from this angle. Whether the ratio actually converges to 1:1 will ultimately depend on real shipment BOMs, not a line in a keynote. Today this is still a "claim," not an "established fact."

Second, open has never been a purely technical proposition — it's positioning. The companies pushing UALink, UEC and OCP standards mostly want to use an open ecosystem to dilute the moat of a certain closed solution. That's good for the supply chain (more choice, more pricing leverage), but don't read "open is a must" as pure technical idealism — it is also a commercial containment play.

Third, a ratio reversal does not mean GPU demand is cooling. Ravi himself stressed that agentic "only adds to, and takes nothing away from" the elements of the traditional data center. The CPU's comeback is about a change in the relative ratio, not a drop in absolute GPU volume. Misreading this signal as "time to underweight the GPU supply chain" would be a serious directional error.

Conclusion

The significance of this AMD keynote for Taiwan's supply chain can be condensed into one sentence: the money in the next data center cycle will concentrate more on "whole-rack systems" than on "individual chips" — and the CPU, a line left out in the cold for two years, is being called back up by agentic workflows.

Specifically for Taiwan, three lines benefit most directly. First, rack system integration — designs like 4 CPUs/OU, ORv3 double-wide and DCSCM that "replace the whole rack at once" draw on ODM and EMS system capability, so the value weighting of players like Wiwynn, Quanta and Foxconn shifts upward. Second, the scale-up fabric — as agentic workloads push latency sensitivity to the limit, the open UALink/UEC interconnect battle is the front line for optical modules, CPO, retimers/DSPs and the high-speed copper cable supply chain. Third, the "infrastructure tax" segments of security, power and cooling — RoT chips driven by Caliptra, and the power and thermal upgrades forced by high-density 2nm racks, are underappreciated hidden beneficiaries.

The verdict: agentic AI hasn't pushed the CPU out — it has moved it from the lead role to the role of "orchestrator." Across the supply chain, what decides the winner is no longer whose chip is fastest, but who can stitch a whole rack of heterogeneous compute together with the lowest latency and the lowest infrastructure tax. Open is just the ticket in; system integration capability is the real stake in this game.

This article is for technology and industry trend analysis only and does not constitute investment advice.