Nvidia vs. Broadcom at OCP APAC 2026: The Battle to Own the AI Factory Network

The networking bottleneck has become the defining infrastructure problem for AI at scale. At the 2026 OCP APAC Summit in Taipei (August 11–12), Nvidia and Broadcom presented two fundamentally different answers to the same question: who should own the network inside an AI factory?

Nvidia argued for a vertically integrated, CPO-first, full-stack approach built around its own Spectrum-6 ASIC and proprietary Spectrum-X software. Broadcom countered with an open-Ethernet ecosystem of merchant silicon, positioning its Tomahawk 6, Tomahawk Ultra, Jericho 4, and Thor Ultra as the building blocks for hyperscalers who want to design their own networks .

The summit made one thing clear: the industry is no longer debating whether networking is the bottleneck. It is debating who gets to own it.

Nvidia's Vision: Vertically Integrated, CPO-First, Full-Stack Ownership

Nvidia's networking strategy is anchored by the Spectrum-6 Ethernet ASIC, which delivers 102.4 Tbps per chip and powers the SN6000 series of switches . The company ships switches in both co-packaged optics (CPO) and pluggable-optics variants, using 200 Gbps SerDes and 800 Gb/s per port .

Technical approach. The CPO-based Spectrum-X Photonics switches directly integrate silicon photonics optical engines into the ASIC package, eliminating pluggable transceiver cages . The top-end SN6800-LD is a 5U switch with 512 MMC-12 connectors supporting up to 2,048 ports at 200 GbE and achieving 409.6 Tbps aggregate throughput . Nvidia also offers the Quantum-X800 InfiniBand platform for GPU-to-GPU scale-up domains, including the Q3450-LD switch with 144 ports of 800 Gb/s .

Gilad Shainer, Nvidia's SVP of Networking, argued at the summit that with GPU supply normalizing, the interconnect layer has become AI infrastructure's binding constraint . Nvidia couples its networking tightly with its GPU roadmaps: the Vera Rubin platform launched alongside Spectrum-X Ethernet Photonics in mid-2026 .

Performance claims. Nvidia's CPO approach claims roughly 3.5x better power efficiency and ~10x better reliability compared to traditional pluggable optics . The SN6800's 409.6 Tbps bandwidth comes from four Spectrum-6 ASICs in a single chassis, each driving 32 silicon photonics engines . Nvidia also claims Spectrum-X delivers up to 1.6x higher AI networking performance than standard Ethernet, sustaining up to 95% network efficiency across clusters exceeding 10,000 GPUs .

Cost implications. Nvidia's integrated stack commands a significant premium: hyperscalers must buy into the full ecosystem of GPUs, SuperNICs, switches, cabling, and software to realize the performance gains. However, Nvidia argues total cost of ownership is lower at extreme scale because CPO cuts per-port power consumption by roughly 65% and dramatically reduces optical failure rates .

Strategic bet. Nvidia is betting that hyperscalers will pay for maximum performance and a single-vendor, proven stack. The manufacturing supply chain for its CPO switches runs heavily through Taiwan, a point Shainer emphasized at the summit . Nvidia guards its networking software tightly and ties it to its GPU roadmaps .

Broadcom's Vision: Open Ethernet Ecosystem, Merchant Silicon, Hyperscaler Choice

Broadcom showcased what it calls "the industry's only complete Ethernet portfolio for AI scale-up, scale-out, and scale-across" . The company's strategy is to sell merchant silicon that hyperscalers integrate into their own switch designs.

Technical approach. Broadcom's 2026 portfolio includes four key products:

• Tomahawk 6 — 102.4 Tbps switch ASIC, shipping in production volume, with a CPO variant called Tomahawk 6–Davisson
• Tomahawk Ultra — ultra-low-latency switching delivering 250 ns cut-through latency for scale-up domains
• Jericho 4 — an Ethernet fabric router for inter-data-center AI connectivity, designed to interconnect over one million XPUs across multiple data centers
• Thor Ultra — the industry's first 800G AI Ethernet NIC, a clean-sheet design purpose-built for AI backend networks and compliant with the Ultra Ethernet Consortium (UEC) 1.0 specification

Broadcom positions these as merchant chips that hyperscalers can integrate into white-box switches. The company aligns with the OCP's ESUN (Ethernet Scale-Up Networking) initiative and the Scale-Up Ethernet (SUE) framework, serving as an open-standards counterweight to Nvidia's proprietary NVLink and InfiniBand .

Performance claims. Tomahawk 6 matches Nvidia's Spectrum-6 on raw switching capacity (102.4 Tbps). Tomahawk Ultra claims 250 ns port-to-port latency for scale-up traffic, which is class-leading for Ethernet . Jericho 4 targets inter-data-center latency and massive routing scale; a single system scales to 36,000 HyperPorts each operating at 3.2 Tb/s . Thor Ultra provides 800 Gb/s NIC throughput with a "clean-sheet design for AI traffic patterns" .

Cost implications. Broadcom's value proposition is lower per-port silicon cost through merchant chips, multi-vendor competition, and no vendor lock-in. Hyperscalers like Meta, Google, and Microsoft can design their own switches around Broadcom ASICs, negotiate pricing independently, and mix vendors for NICs and optics . Broadcom offers CPO as an open reference (Tomahawk 6–Davisson) rather than a captive system .

Strategic bet. Broadcom is betting that hyperscalers — especially Meta, Google, OpenAI, and Anthropic — will prefer an open, standards-based Ethernet ecosystem over Nvidia's proprietary stack. At multi-gigawatt cluster scale, the argument goes, hyperscalers want to control their own network design, disaggregate hardware procurement, and avoid single-vendor lock-in . Broadcom's exit from the UALink consortium in favor of OCP's ESUN initiative signals its commitment to open governance .

Key Differences at a Glance

Dimension	Nvidia	Broadcom
Architecture philosophy	Vertically integrated, full-stack (GPU + NIC + switch + software)	Merchant silicon, open Ethernet ecosystem, hyperscaler DIY
Switching silicon	Spectrum-6 (102.4 Tbps, proprietary)	Tomahawk 6 (102.4 Tbps), Tomahawk Ultra (250 ns), Jericho 4 — all merchant
Optics strategy	CPO-first, tight integration with own ASICs	CPO available (Tomahawk 6–Davisson) but also supports pluggables; open reference
NIC / endpoint	ConnectX-9 SuperNIC (in-house)	Thor Ultra (800G Ethernet NIC, merchant)
Key performance claim	~409.6 Tbps CPO aggregate; ~3.5x power savings vs. pluggables	250 ns latency (Tomahawk Ultra); million-XPU fabrics via Jericho 4
Scale target	Up to hundreds of thousands of GPUs in a single AI factory	Millions of XPUs across DCs (scale-up/out/across)
Cost model	Premium integrated TCO; lower per-port power at extreme scale	Lower silicon cost; hyperscaler procurement optionality
Standards posture	NVLink + InfiniBand + proprietary Spectrum-X software	Open Ethernet: ESUN, SUE, OCP-aligned
Hyperscaler appeal	Best performance for those who want one throat to choke	Best for hyperscalers that design their own networks

Which Hyperscalers Lean Which Way?

The evidence from the summit and recent industry moves suggests a rough alignment:

• Meta has been a vocal advocate for open Ethernet-based AI networking and historically builds its own switches around Broadcom silicon. Broadcom's SUE/ESUN alignment directly serves Meta's agenda .
• Google similarly builds custom networking and has long preferred merchant silicon with open standards; Google was an early adopter of 200G/400G Ethernet around Broadcom ASICs.
• OpenAI and Anthropic — as AI-model companies rather than infrastructure operators — outsource much of their networking to their cloud partners (Microsoft Azure for OpenAI, Google Cloud for Anthropic). Their networking preference is therefore mediated by their cloud provider's choice .
• Microsoft uses both Nvidia InfiniBand internally and is exploring Ethernet alternatives. The Azure stack runs on Nvidia GPUs today, but Broadcom's open approach offers a second path.

Bottom Line

Nvidia is selling a turnkey, maximum-performance AI factory network. Broadcom is selling hyperscalers the tools to build their own. The 2026 OCP APAC Summit made clear that the AI networking market has bifurcated into these two camps, and the winners will be determined not just by technical specs but by whether hyperscalers ultimately value integration or optionality more at multi-gigawatt scale.