On July 30, 2026 local time, Broadcom (AVGO) dropped a bombshell at its annual technology conference - officially launching its next-generation AI network switching chip Tomahawk 6, with a single chip switching capacity of up to 51.2Tbps, using 3nm advanced process, supporting 400G/800G ports, and designed for large-scale AI training clusters. This launch directly addresses the biggest pain point in current AI computing expansion: network bandwidth and latency bottlenecks.
AI Computing Explosion, Network Becomes the "New Throat"
Over the past two years, the large model competition has pushed GPU cluster sizes from thousands to hundreds of thousands of cards, but simply stacking chips can no longer solve the performance doubling problem. The interconnection bandwidth demand for NVIDIA's GB200 and other super chips is growing exponentially, and traditional data center network architectures frequently experience congestion in ten-thousand-card clusters, with GPU utilization often less than 50%. Industry consensus: the computing network has become the "new throat" constraining AI training efficiency.
Tomahawk 6: The "Interchange" on the Data Highway
Broadcom's Tomahawk 6 switching chip is designed to break through this bottleneck. Its 51.2Tbps switching capacity is twice that of the previous generation, equivalent to transmitting the daily data volume of an entire city like Los Angeles every second on a 40mm² silicon chip. More importantly, it natively supports the new Ultra Ethernet Alliance (UEC) standard, which can improve data throughput efficiency in AI clusters by over 60% in congestion scenarios. Broadcom's CEO demonstrated measured data from ten-thousand-card clusters during his speech: after adopting Tomahawk 6, large model training task time was reduced by about 47%.
Not Just a Chip, But a System-Level Breakthrough
Alongside Tomahawk 6, Broadcom also launched supporting SerDes 112G PHY and optical DSP product series. Broadcom has integrated the switching chip, physical layer components, and optical communication solutions into an "AI network platform" for the first time, significantly reducing the deployment complexity of ultra-large-scale data centers when combined with its proprietary Sonic switching operating system. Microsoft Azure, Google Cloud and others appeared on-site, revealing they have begun pre-research on next-generation AI networks based on this chip.
Industry Change: NVIDIA's "Wall" and Broadcom's "Wall-Breaking"
For a long time, NVIDIA has built a closed GPU cluster ecosystem through NVLink and InfiniBand. Broadcom's move this time is undoubtedly breaking through this wall. Tomahawk 6 supports native Ethernet interfaces, allowing customers to build open AI networks at low cost, no longer bound to NVIDIA's proprietary solutions. Analysts point out that Broadcom has secured over 70% of AI switching ASIC orders from North America's three major cloud vendors, and with this general chip launch, Broadcom is poised to become the "invisible hegemon" of AI computing infrastructure.
Chain Reaction in the US Stock Market
The launch of this chip has directly caused resonance in the US stock market: Broadcom's stock price rose nearly 7% in after-hours trading, while NVIDIA, which relies on hardware sales, fell slightly. Optical module leaders Coherent and Lumentum received capital chasing due to deeper cooperation with Broadcom on high-bandwidth optical solutions. Chinese optical module concept stocks such as Zhongji Xuchuang and Xinyisheng also opened higher the next day. From the industry chain perspective, AI network equipment is moving from a "supporting role" to a "leading role," with the value of switches, optical modules, and high-speed copper connectors continuously increasing.
Investment Perspective: Who is Truly Benefiting?
Deeply analyzing Tomahawk 6's supply chain, three key directions deserve attention:
- Cloud vendors increasing capital expenditure: Microsoft, Google, and Amazon all raised their 2026 computing capital expenditure expectations in this quarter's earnings reports, with network equipment share increasing from about 8% to 15%. Broadcom and Arista Networks (ANET) will be the main winners.
- Independent rally of Chinese optical modules: The 800G optical module demand explosion brought by Broadcom's upgrade, Zhongji Xuchuang and Xinyisheng have locked in 2027 capacity, enhancing earnings certainty.
- Indirect benefits for semiconductor equipment: TSMC's 3nm process capacity is fully loaded, and using Broadcom's solution has further tightened advanced packaging capacity. With the perspective of Fangzheng Securities, packaging and testing valuation is expected to be reshaped.
However, risks also exist. Broadcom's new product mass production is scheduled for early 2027. If there are yield or customer switching delays during this period, short-term valuation may face pressure. Additionally, NVIDIA also plans to launch its Spectrum-X Gen2 Ethernet solution in 2027, and the duopoly game will redefine the computing network market.
Conclusion: Computing Competition Enters the "Second Half"
When the industry thought AI computing competition was only about the number of GPUs, Broadcom reminded the market with a single chip: what determines the upper limit of large model training is not just chip computing power, but the "highway" connecting these chips. The launch of Tomahawk 6 marks the official entry of computing networks into the 51.2T era, also indicating that the investment focus of the US AI computing industry chain is shifting from "computing fast" to "connecting stably." For investors, grasping this mainline may be the key to capturing the real long-term dividends of the AI era.
