OCP Global Summit 2026 | AMD | Lisa Su | Agentic AI Makes the CPU the Next Bottleneck

21 hours ago

8 min read

1. Why This Talk Matters: Lisa Su's First OCP Keynote Was About the CPU

AMD is a long-time OCP member, but this was the first time Lisa Su delivered a keynote at the OCP Global Summit herself. She opened with a "15 Years of Collaboration" timeline: the Open 3.0 modular server in 2012, Naples with Microsoft's Project Olympus in 2017, the Ultra Ethernet Consortium in 2023, the UALink Consortium in 2024, Helios debuting at OCP and the ESUN collaboration in 2025, and MRC (Multipath Reliable Connection) plus the Open Adapter Card in 2026.

The message of that slide is clear: AMD wants "open" to read as part of its roadmap, not as a marketing slogan. But the real substance came in the second half. A separate fireside chat already covered Helios engineering in detail; in this keynote Lisa Su spent most of her time on how agents are re-accelerating CPU demand, and on Helion, the new rack AMD built for it.

▲ Training compute has grown 5x a year since 2020, but from 2026 inference becomes the largest AI workload, and agent traffic is up more than 14x this year (Image source: AMD | 2026 OCP Global Summit presentation (slide 6))

She gave three demand numbers in quick succession:

Token consumption is up 95x in two years, with monthly inference tokens now approaching 40 quadrillion.
Training compute has grown 5x per year since 2020, with no sign of slowing.
Inference is now the largest workload, about 60% of the market in 2026; the fastest-growing part is agentic traffic, up more than 14x this year, while human traffic rises only steadily.

2026 is the year of inference, and within inference the fastest-growing users are not humans but agents.

2. Breaking Down the TAM: CPU Grows 8.5x, Faster Than Accelerators

Lisa Su stressed that this TAM covers silicon and systems, not the entire AI infrastructure build-out. The two curves on the slide:

Market	2025	2030	CAGR	Growth multiple
Data center CPU	~$26B	~$220B	>50%	~8.5x
Data center AI accelerator	~$200B	~$1.4T	>45%	~7x

The growth multiples are STT's calculation from the slide endpoints and are consistent with the CAGRs AMD shows.

The $1.4T accelerator figure was expected; the CPU line is the news. In Lisa Su's words, a year ago the market saw the CPU as a roughly $20B business growing in the single digits, and AMD has raised this TAM three times in the past nine months. The slide splits CPU demand into general purpose, AI host and agentic, and nearly all of the new demand comes from the last two.

▲ AMD estimates data center CPU TAM grows about 8.5x from 2025 to 2030, versus about 7x for AI accelerator TAM (Image source: Simple Tech Trend original; data from AMD 2026 OCP Global Summit presentation)

The logic: between model calls, an agent runs code, queries data and calls tools, and that work lands on the CPU. Scale agents by 10x and the CPU sandboxes must scale with them; the GPU does not do that work. The AI bottleneck is widening from "not enough GPUs" to "not enough CPUs either", and that is exactly the story a company selling both CPUs and GPUs most wants to tell.

3. Helios: Shipping Since September, Ramping Into 1H 2027

We only summarize the Helios specs here. Rack-level numbers on the slide: 2.9 EFLOPS FP4, 31 TB of HBM4, 1.7 PB/s memory bandwidth, 260 TB/s scale-up and 43 TB/s scale-out; AMD claims up to 30% more tokens/$ on leading models and 17,000 tokens/s per GPU on DeepSeek R1 inference (MLPerf).

More important than the specs is the timing:

Production shipments began in September 2026, and the first systems are installed at OpenAI, Meta, Microsoft, Oracle and Anthropic.
The ramp continues through the fourth quarter and into the first half of 2027.
The adopter list has widened from five model companies to neoclouds, OEMs and ODMs, with nearly 30 names on the slide.

▲ Helios adopters, expanding from OpenAI, Meta, Anthropic, Microsoft and Oracle to neoclouds and OEMs/ODMs (Image source: AMD | 2026 OCP Global Summit presentation (slide 18))

In Earnings Highlights: AMD Q2 FY2026 (Helios) we covered what these orders mean for revenue. What OCP added is delivery progress: the main window for revenue recognition is Q4 2026 through 1H 2027, and shipments over the next two quarters are the real test.

On networking, Helios uses open standards at all three layers, front-end, scale-up and scale-out: UALoE (UALink over Ethernet) and ESUN for scale-up, and UEC, MRC and MetaRoCE for scale-out, with Broadcom, Cisco, Arista, HPE and DriveNets as partners. This matches the direction in OCP APAC Summit 2026: Ethernet Starts Eating Scale-Up: AMD does not build its own switches and leaves the network layer to the Ethernet ecosystem.

4. Helion: A CPU Rack Built for Agents

The only new product in this keynote is the Helion Agentic AI Rack, which Lisa Su called a "cousin" of Helios. The design premise: agents are not one thing, and they need different compute.

Personal agents: simple tasks that need the most agents per watt, served by dense servers (Venice 256-core, 400 W).
Enterprise agents: varied tasks that need balance (Venice 256-core, 600 W).
Technical compute agents: need the highest per-core performance (Venice 96-core high-frequency, 5.0 GHz).

Rack-level specs on the slide:

Item	Helion spec
CPU	Up to 32,768 Zen 6 cores
GPU	Up to 16 Instinct MI350P, 74 PFLOPS FP4
Networking	Up to 102 Tbps, fully programmable
Rack	ORv3 44OU, native OCP spec compliant
Availability	November 2026, contributed to OCP

32,768 cores is the equivalent of 128 Venice 256-core CPUs. Lisa Su said verbally that this is 40% more cores than competing racks and the highest compute density in the industry.

▲ The Helion rack holds up to 32,768 Zen 6 cores, 16 MI350P and 102 Tbps of networking in an ORv3 44OU rack (Image source: AMD | 2026 OCP Global Summit presentation (slide 24))

Helion comes in two configurations. Hyperscalers get a CPU-only version: GPU racks sit in one part of the data center, agentic CPU racks sit beside them, and CPU nodes can be mixed by agent type. Enterprises get an all-in-one CPU plus GPU rack: CPU sandboxes, MI350P inference nodes, and database and storage nodes in one rack. Helion also integrates a token router that sends each agent's model calls to frontier or open-source models.

The configuration reveals AMD's view: hyperscalers will split CPUs and GPUs into separate racks, so CPU demand becomes an entire new row of racks, not just the host CPU inside a GPU server.

5. Venice and Verano: One Core Generation, Six Product Variants

Helion can mix and match because the Venice family is finely segmented: Venice HF and Verano for GPU host nodes, Venice 256-core for agent sandboxes, Venice SP7 and SP8 for general-purpose servers, and Venice-X for technical compute.

The new chip this time is EPYC "Verano":

Built on Venice CPU cores, with six CCDs of 32 cores each, up to 192 cores
Paired with LPDDR and positioned for the lowest system power
Higher CPU-to-GPU bandwidth, also suited as a host node in GPU racks
Production in the first half of 2027

LPDDR in a server CPU is a signal worth noting: an agent sandbox cares about how many agents it can run per watt, not about per-CPU memory capacity.

6. Open Is the Moat, and Also a Delivery Promise

On software, Lisa Su said the top 15 open-source AI communities now support AMD natively. She gave one example: an internal AMD agentic AI tool (transcribed as "Hyperloom") has been continuously optimizing more than 14,000 Hugging Face models with no humans in the loop, improving performance by almost 40%. Letting agents tune ROCm performance is AMD's new way of catching up with the CUDA ecosystem.

On hardware, Helios uses OCP Open Rack Wide, OCP power shelves and OCP GPU management; Helion uses an ORv3 rack and a standard modular compute board, and its spec will also be contributed to OCP. Lisa Su's line was that "AI at scale is a team sport": data centers will run CPUs, GPUs and custom ASICs side by side, and the more that is standardized, the less a customer has to change when switching accelerators.

For a challenger with lower market share, this is a rational choice: openness lowers customers' switching costs and hands rack manufacturing and delivery to the entire ODM ecosystem.

7. Risks and Counterarguments

The TAM is AMD's internal projection: raising the CPU TAM three times in nine months shows the demand is real, but also that forecasts are volatile. If agent business models slow, this curve gets revised first.
Helion's GPU count is inconsistent: the slide says up to 16 MI350P, while Lisa Su said 32 on stage; check the official spec sheet before relying on it.
The 30% tokens/$ claim has no public baseline: the slide only says "leading AI models" and does not specify the competitor or configuration.
The Helios ramp has only just begun: shipments started in September and volume only scales in 1H 2027; delivery risk across liquid cooling, power and networking in different customer facilities is not yet retired.
No GPU roadmap update this time: the cadence after MI400 and the timing of MI500 were not mentioned at all.

8. What It Means for Taiwan's Supply Chain

Helion is direct business for Taiwanese ODMs. Of the 13 Helion hardware partners on the slide, Wiwynn, QCT (Quanta), Foxconn, MiTAC, Inventec, ASUS, ASRock Rack and Gigabyte, 8 in total, are Taiwanese; the rest are Dell, HPE, Lenovo, Supermicro and Cisco. If the CPU TAM follows AMD's projection, a new row of CPU racks to assemble arrives from 2027, so it is worth tracking when these companies start reporting Helion shipments.
Wiwynn is the only Taiwanese name on the Helios adopter list. ODM concentration is clearly higher for GPU racks than for CPU racks, so the breadth of benefit differs between the two product lines.
CPU racks also mean networking demand. Each Helion rack carries up to 102 Tbps of networking, roughly the capacity of one 102.4T switch chip. By inference, as agentic racks scale, front-end and scale-out switch and optical module demand grows too, not only from GPU clusters.
On the wafer side: AMD has previously said Venice uses TSMC's 2nm process; the more finely the Venice family is segmented and the more it ships, the more direct the demand for advanced process nodes and advanced packaging.

9. Conclusion

In her first OCP keynote, Lisa Su chose to give the stage to the CPU. The CPU TAM grows 8.5x, faster than accelerators, and the reason is agents: inference is already the largest AI workload, agentic traffic is up 14x this year, and every agent needs a CPU sandbox.

The timeline is clear: Helios ramps from Q4 2026 through 1H 2027, Helion is available in November 2026, and Verano enters production in 1H 2027.

▲ AMD's three production milestones over the next nine months: Helios ships in September and ramps into 1H 2027, Helion is available in November, and Verano enters production in 1H 2027 (Image source: Simple Tech Trend original; data from AMD 2026 OCP Global Summit presentation)

Three things to watch over the next 12 months: whether Helios shipments scale each quarter as planned, whether hyperscalers place large orders for CPU-only Helion racks, and whether AMD raises its CPU TAM yet again.

The GPU decides whether a model can be trained; the CPU is starting to decide how many agents can run.

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