OCP Global Summit 2026 | Broadcom | Asad Khamisy | Ethernet Scale-up, Copper and OCI

1. The Two Questions This Keynote Set Out to Answer

Asad Khamisy opened by narrowing the talk to the two questions customers ask him most: which technology to use for scale-up, and whether cabling over the next two to three years should be copper, optics or hybrid. He gave a clear answer to both: use Ethernet for scale-up; extend copper's life inside the rack, go hybrid across racks, and bet on OCI for the long term.

The position itself is not new. We covered Broadcom's open-Ethernet argument in OCP APAC Summit 2026: Ethernet Starts Eating Scale-up, and its CPO reliability data in Broadcom's Third-Generation CPO Report Card. What is new at this Global Summit keynote comes in three places: production-status numbers, the next-generation copper spec, and the OCI versus pluggable power comparison.

One more detail worth noting: the word "CPO" does not appear anywhere in the deck; "Integrated OCI" takes its place. When Broadcom talks about scale-up optics, it has shifted the narrative from packaging format to interface standard.

The question is no longer whether Ethernet can do scale-up, but how many more generations copper can last and at what power optics can enter.

2. Where the Three Network Tiers Meet: Why Scale-up Cannot Grow Without Limit

Broadcom splits an AI cluster into three tiers:

• Scale-up: within a rack or a few racks; tightly coupled compute with a single memory address space, using memory load/store or packed DMA transactions, requiring ultra-low latency and small packets.

• Scale-out: a Clos network across racks using message passing; Khamisy said customers are already building 128,000 XPUs in a single cluster.

• Scale-across: once a region runs out of power, clusters span data centers hundreds of kilometers apart; Broadcom argues for deep-buffer switches to reduce packet loss, because retransmitting from the source costs an order of magnitude more latency than queuing in a buffer.

Why can scale-up not reach a hundred thousand XPUs? Khamisy named two bottlenecks: latency grows with network size, so tensor-parallel token output is held back by HBM, compute and network latency together; and switches that support small packets at low latency cannot carry tables that large. Scale-up therefore stops at one to a few racks. The slide sets the target as going from 64–144 XPUs in a single rack today to 1K–2K XPUs across multiple racks tomorrow.

The matching chip lineup is all in production: Tomahawk Ultra (51.2 Tbps, HPC and scale-up), Tomahawk 6 (102.4 Tbps, scale-up and scale-out), Jericho 4 (51.2 Tbps, scale-out and scale-across), with Thor Ultra 800G as the NIC. Khamisy stressed that pushing cluster utilization above 90% (versus about 60% in the cloud era) requires a NIC that supports enhanced RoCE load balancing and congestion control.

3. Breaking Down the Scale-up Protocol: ESUN Is the Foundation, SUE-T Sits on Top

This time Broadcom laid out the scale-up protocol as two clear layers:

• ESUN (Ethernet for Scale-up Networking): the data link layer, providing reliable lossless Ethernet (link layer retry plus credit-based flow control), compressed headers for small messages, and multi-hop topologies. Announced at the 2025 OCP Global Summit, the v1.0 spec was published in February 2026, with 175+ members, and ESUN racks are already in production.

• SUE-T: the transport layer, handling push/pull memory access, the memory ordering model, transaction packing, reliability, load balancing and scheduling, with spec publication targeted for 2H2026. Khamisy said SUE-T delivers end-to-end reliability and congestion control at far lower cost than RDMA.

▲ Broadcom splits scale-up into two layers: ESUN for the lossless Ethernet link layer (v1.0 released) and SUE-T for the transport layer (targeting 2H2026) (Image source: Broadcom | 2026 OCP Global Summit presentation (slide 10))

The point of the layering: XPU vendors can adopt ESUN alone and run their own RDMA or proprietary protocol on top, or take SUE-T as well. Broadcom also licenses SUE-T plus ESUN IP to XPU vendors, which lowers the bar for Ethernet scale-up from "build a new protocol" to "license a block of IP."

To the objection that Ethernet headers are too long and inefficient, Broadcom answered with variable packet sizes:

Use case	Link efficiency
Unpacked 256B posted write	92%
2 packed 256B posted writes	96%
4 packed 256B posted writes	98%
Native 2kB block transfer	99%

The next ESUN spec revision will further optimize the header and reduce the inter-packet gap (IPG).

4. Production Status: The Timeline from Concept to Shipment

This is the one slide to remember from the talk:

• 5+ hyperscalers are running Ethernet scale-up in production

• 10+ partners are building Ethernet-based scale-up racks

• Multiple custom and merchant XPU Ethernet scale-up racks enter production within the next 6 months

▲ Broadcom's Ethernet scale-up production status: 5+ hyperscalers in production, 10+ partners building racks, and multiple XPU racks entering production within 6 months (Image source: Broadcom | 2026 OCP Global Summit presentation (slide 11))

Behind these racks is Tomahawk Ultra: 512 lanes of 100G PAM4, single-hop scale-up of 256, 250 ns switch latency, 77 BPPS at 64-byte line rate, plus in-network collectives. For XPU vendors, "within 6 months" means the first high-volume non-NVLink scale-up racks should ship in 1H2027.

5. Breaking Down Copper: One More Generation In-Rack, at Double the Power

Broadcom's answer on copper is pragmatic: extend copper's life inside the rack. Current versus "possible next generation":

Item	Current	Possible next gen
Total rack bandwidth	1.04 Pbps	2.08 Pbps
Per XPU	14.4 Tbps (72 XPUs)	28.8 Tbps (72 XPUs)
Switch links	512 x 200G PAM4, 9-12 switches	576 x 400G PAM6, 9 switches
Copper reach	~2 m	~2 m
SerDes power	7 pJ/bit	7 pJ/bit
Total rack power	~250 kW	~500 kW

▲ Today's Tomahawk 6 copper backplane delivers 1.04 Pbps per rack; the next generation doubles it to 2.08 Pbps with 400G PAM6 while keeping twinax density at 1,024 pairs per OU (Image source: Broadcom | 2026 OCP Global Summit presentation (slide 16))

Two details stand out. First, twinax density is 1,024 per OU at 26 AWG in both generations, so the bandwidth doubling comes entirely from moving from PAM4 to PAM6 modulation, not from adding cables. The slide states plainly that the current generation is limited by backplane density. Second, 576 lanes of 400G works out to about 230.4 Tbps per switch (STT calculation), more than double Tomahawk 6's 102.4 Tbps; Broadcom did not say which chip that would be.

Converting the slide numbers to SerDes power (STT calculation): 1.04 Pbps times 7 pJ/bit is about 7.3 kW, and about 14.6 kW for the next generation, roughly 3% of rack power either way. Copper SerDes are not the big power consumer; what forces optics is reach: 2 meters cannot hold 1K XPUs.

6. Breaking Down Optics: Hybrid Is the Transition, OCI Is the Destination

Beyond a single rack, Broadcom presented three architectures:

• All-optical disaggregated racks: XPU racks and switch racks separated, scale-up domain growing to e.g. 576, lower power and weight per rack; the cost is optics pricier than copper and added cross-rack latency (5 ns per meter of fiber).

• Hybrid copper plus optics: N copper links in the rack and M fiber links across racks, splitting scale-up into a "low latency domain" and a "higher latency domain," with XPU software aware of both and optimizing for them, which cuts optics cost. Khamisy said this is what he sees customers doing.

• OCI disaggregated racks: preserve a large scale-up domain at the same power as copper, architecturally defined so cost and reliability approach copper KPIs.

▲ Broadcom's three-stage scale-up cabling: 2 m of copper in the rack, a hybrid architecture splitting low- and high-latency domains, and OCI reaching 500 m at 7 pJ/bit; every 50 m of fiber roughly equals one switch hop of latency (Image source: Simple Tech Trend original; data from Broadcom 2026 OCP Global Summit presentation)

Latency deserves a closer look: at 5 ns per meter, 50 meters of fiber is 250 ns, exactly one Tomahawk Ultra switch latency (STT calculation). In other words, every extra 50 meters of multi-rack scale-up distance is equivalent to one more hop. That is why the hybrid architecture hands latency domains to software.

The OCI power comparison is the core number of the talk:

Link option (200G)	Total link power	Reach
DAC	7 pJ/bit	2 m
Integrated OCI	7 pJ/bit	500 m
DR LRO front-panel pluggable	23 pJ/bit	500 m

▲ At the same 500 m, pluggables run 23 pJ/bit vs. 7 pJ/bit for OCI; for a 14.4 Tbps XPU that is about 331 W vs. 101 W per XPU, a gap of about 230 W (Image source: Simple Tech Trend original; data from Broadcom 2026 OCP Global Summit presentation)

Converted (STT calculation): at 14.4 Tbps per XPU, a pluggable link draws about 331 W and OCI about 101 W, a gap of about 230 W per XPU, or about 16.6 kW across 72 XPUs. The OCI MSA was established in March 2026; founding members on the slide are Meta, Microsoft, OpenAI, AMD, Broadcom and NVIDIA. We covered the line-side spec in Meta, Broadcom and AMD Define the OCI 200G Line-Side Spec.

7. Risks and Counterarguments

• OCI has no timeline: DAC and hybrid designs have production or customer deployments; OCI has only spec targets and an MSA, and the talk did not say when it ships.

• Unclear whether 7 pJ/bit is a target or a measurement: the table is labeled "total link power" but does not say whether the laser and allocated cooling are included. If an external laser is excluded, the parity-with-DAC claim weakens.

• Next-gen copper is "Possible": the slide explicitly calls it a possible next generation; 400G PAM6 SerDes and a 230 Tbps-class switch have not been announced.

• A 500 kW rack is another bottleneck: bandwidth doubles and so does power; whether power delivery and liquid cooling can keep up is harder than the copper itself.

• Hybrid pushes complexity into software: low- and higher-latency domains require domain-aware XPU software; custom XPUs can do it, general software stacks may not.

• ASR uncertainty: the transcript mentions a new packaging technology that puts copper and optics on the same package, transcribed as "ICA"; there is no matching slide, so the official name cannot be confirmed.

8. What It Means for Taiwan's Supply Chain

• Switch and rack assembly volume is real: the slides state Ethernet scale-up racks are in production at 5+ hyperscalers, with more XPU racks in production within 6 months, built on the OCP ORv3 rack plus switch trays. Taiwan is a key manufacturing base for white-box switches and AI rack ODMs, and non-NVLink scale-up switch trays are a new product line for them.

• Copper demand extends at least one more generation: 1,024 twinax pairs per OU at 26 AWG and a 400G PAM6 backplane mean a spec upgrade, not an exit, for cable assembly and high-speed connectors. The opportunity for Taiwan's high-speed cable and connector makers lies in PAM6 backplane qualification.

• OCI sets the ceiling for optical modules: the 23 versus 7 pJ/bit gap means cross-rack scale-up will not rely on front-panel pluggables for long. For Taiwan optical module and component makers that only do pluggables, the scale-up window is the hybrid transition; OCI laser sources, fiber connectivity and assembly are the next positions to secure.

9. Conclusion

This keynote laid Broadcom's scale-up roadmap out on a clear timeline: Ethernet scale-up in production on copper backplanes today, more XPU racks in production within 6 months, the SUE-T spec in 2H2026, a next-generation 400G PAM6 copper backplane doubling in-rack bandwidth, hybrid designs across racks first, and OCI to finish the job.

The key judgment: Broadcom is neither abandoning copper nor rushing optics; it has set optics' entry condition as "power must equal copper." Pluggables at 23 pJ/bit fail that test; OCI at 7 pJ/bit passes. Three things to watch over the next 12 months: whether the SUE-T spec ships on schedule in 2H2026, when the 400G PAM6 copper backplane and matching switch are announced, and OCI's first production date and whether 7 pJ/bit includes the laser.

The bar Broadcom sets for optics is not speed but power equal to copper; until optics clears it, copper stays in the rack.

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