Nvidia and Broadcom Publicly Diverge on AI Network Scaling Roadmap
Nashnova编辑部
Nvidia and Broadcom clashed openly at the OCP Asia-Pacific Summit in Taipei over how to wire AI data centers — one insists on proprietary interconnect at the core, the other bets Ethernet can do it all — and the outcome will reshape who captures value in AI infrastructure.
What exactly are they fighting over?
At the OCP Asia-Pacific Summit in Taipei on August 11, Nvidia and Broadcom took the stage to lay out competing visions for the three-layer AI networking stack.
The three layers: in-rack scale-up → cross-rack scale-out → cross-data-center scale-across. Both companies agree networking is the core bottleneck for AI system performance — but they disagree fundamentally on which technology should stitch those layers together.
This means → the dispute is not about products. It is about architectural control — whoever defines the connectivity model dictates how value is distributed across AI data centers.
What does Nvidia's approach look like?
Nvidia SVP Gilad Shainer outlined the company's "AI Factory" architecture: NVLink for in-rack scale-up, Spectrum-X Ethernet for cross-rack scale-out, and Spectrum-XGS Ethernet for cross-data-center scale-across.
Shainer was explicit: "NVLink is not a network — it is compute infrastructure." In plain terms = Nvidia is pulling the in-rack layer out of the "networking" category entirely, folding it into its own GPU compute stack and keeping Ethernet away from the most critical tier.
On openness, Nvidia is expanding interfaces around the NVLink core — Spectrum-X supports SAI and SONiC open standards, and NVLink Fusion lets customers plug custom XPUs into the NVLink ecosystem. This reflects a strategy of "proprietary core + open periphery."
Why does Broadcom say Ethernet can handle all three layers?
Broadcom's Mohan Kalkunte argued that Ethernet can unify all three tiers, using the ESUN 1.0 standard — Ethernet Scalable Unified Networking, released in March, which supports lossless transport and multi-hop operations — for end-to-end deployment.
In plain terms = Broadcom's pitch is "one standard for everything" — customers don't need one technology inside the rack and a different one between racks. Ethernet from end to end is enough.
A telling detail: Nvidia and Broadcom are both among ESUN's 12 founding members. This means → the split is not a simple open-vs-closed fight. Both participate in the same open standard — yet each insists on a different architectural center of gravity.
What does the deep-buffer switch debate reveal?
The sharpest clash came over the cross-data-center layer. Nvidia deliberately avoids deep-buffer switches — switches with large on-chip packet storage — in Spectrum-XGS. Shainer's line: "Deep buffers are essentially shock absorbers, and shock absorbers create jitter."
Nvidia relies instead on adaptive routing and congestion control, claiming the approach nearly doubles cross-site NCCL communication performance.
Broadcom's Kalkunte fired back on stage: "Someone earlier said you don't need deep buffers — the results will speak for themselves." He noted that hyperscalers already deploy Broadcom's deep-buffer solution at scale. Broadcom's logic: as distance grows, bandwidth-delay product widens, AI traffic bursts become more synchronized, and congestion feedback loops lengthen — the cushion deep buffers provide becomes more critical, not less.
This means → the two companies tested under different conditions, and their data is not directly comparable. This debate will not be settled by benchmarks alone — real-world deployment by customers will be the judge.
Co-packaged optics — the one area where they agree?
On co-packaged optics — CPO, integrating optical engines directly onto the networking chip so light, not electricity, carries data over longer distances — the two companies' positions converge for once.
Nvidia announced its CPO-based Spectrum-X photonic switching platform is in production, supporting 200 Gb/s SerDes, and is partnering with TSMC on COUPE packaging and optical engines. TSMC's COUPE-based CPO solution is expected to enter mass production in 2026. Broadcom showcased its CPO system built around Tomahawk 6.
This reflects an industry-wide consensus: network power consumption is rising relentlessly and has become a hard constraint. Integrating optics into the chip package is a path neither company can avoid.
Who gets to be the final judge?
As more customers move toward custom XPUs — in-house AI accelerator chips — and demand greater control over system architecture, the choice of interconnect technology will determine which components remain central in AI data centers.
In plain terms = Nvidia is betting that proprietary interconnect is irreplaceable; Broadcom is betting that Ethernet will ultimately unify everything. Both bets cannot win in full.
This means → the final arbiter is not a standards committee. It is the actual deployment decisions of hyperscalers and enterprise customers — whichever approach gets adopted at scale will define the foundation of next-generation AI infrastructure.
Content is for reference only, not financial advice.