Broadcom has focused on one of the major bottlenecks in AI infrastructure: optical interconnection. The company announced the availability of Taurus BCM83640, a 3 nm PAM-4 optical DSP with 400G per lane, which, according to the company itself, is the first of its kind in the industry and is designed for next-generation 1.6T modules. The proposal not only aims to increase speed but also seeks to lay the technical groundwork for future 3.2T transceivers and networks with up to 204.8T switching capacity.
The significance of this announcement lies in its timing. AI clusters continue to grow in size, energy consumption, and internal traffic density, demanding ever more bandwidth per optical link. Broadcom suggests that moving from 200G/lane to 400G/lane is the logical next step to scale connectivity in AI and cloud data centers. In their view, 1.6T pluggable modules based on Taurus double the bandwidth per optical lane and enable 102.4T systems in a 1RU format, while also paving the way for next-generation 3.2T modules and 204.8T switches.
Broadcom describes the BCM83640 as a monolithic 1.6T (8:4) DSP with integrated laser driver, 400G serial optical interfaces, and compatibility with optical modules ranging from 1.6T to 3.2T. The company also claims the chip offers “best-in-class” performance in power consumption and BER, though this comparison primarily comes from the manufacturer rather than an independent validation published in the announcement. It is clearer that the device has been designed to integrate with Broadcom’s own optical components, including their 400G EML and photodiodes, and complies with IEEE and OIF standards to support LR links between chip and module.
The value of 400G per lane in AI networks
Behind this launch is a very practical question: how to continue increasing capacity without significantly raising the number of lanes, power consumption, or system complexity. In AI networks, density is as important as raw speed. If each optical lane doubles in capacity, operators can pack more data in the same physical space or achieve the same throughput with fewer intermediate elements. Broadcom emphasizes this point by presenting Taurus as a key component for more efficient, higher-density pluggable transceivers.
This move also aligns with the evolution of industry standards. By the end of 2025, OIF announced an initial framework for CEI-448G, aimed specifically at defining the next generation of 448 Gbps electrical interfaces geared toward AI, HPC, and communications. While not exactly the same as Broadcom’s newly introduced 400G per lane optics, it confirms that the industry has been preparing for a new layer of lane speed advancements for months. Taurus, therefore, arrives in a context where the transition to 400G/448G per lane was already taking shape in forums and technical roadmaps.
Broadcom has not presented Taurus as an isolated announcement. In another statement related to OFC 2026, the company places it within a broader portfolio for scaling AI infrastructure, alongside solutions such as 102.4T Ethernet switches, co-packaged optics, and other high-speed connectivity components. This suggests that the BCM83640 is not just a stand-alone product but part of a larger strategy to cover multiple levels of AI networking, from optical modules to ultra-high-capacity switching.
From 1.6T today to 3.2T tomorrow
One of the key messages of the announcement is that Taurus isn’t limited to the immediate future. Broadcom emphasizes that its 400G/lane technology will serve as the foundation for future 3.2T modules with similar electrical interfaces, which could further enable networks of 204.8T. In essence, the BCM83640 aims not only to address the next wave of optical transceivers for AI data centers but also to anticipate the subsequent leap.
This is especially important because data center optical industries are experiencing increasingly shorter transition cycles. A provider that does not connect its 1.6T roadmap with its 3.2T plan risks staying limited to incremental improvements. Broadcom aims to avoid this pitfall by positioning Taurus as a transitional platform toward a new, comprehensive generation of optical interconnection.
Broadcom has already begun sampling the chip with early-access clients and partners, indicating that the product is entering an initial commercial phase rather than remaining a conceptual preview. However, the company has not disclosed pricing and directs interested parties to contact its sales channels for samples and conditions.
A significant step in the optical race for AI
It’s important to note that many of the most ambitious claims in the announcement come from Broadcom and are driven by commercial considerations. Even with that caveat, the announcement marks a tangible advance in the effort to boost optical bandwidth in AI data centers. The combination of 3 nm process, the leap to 400G per lane, and the explicit focus on 1.6T and 3.2T modules positions Taurus as a clear signal of where the interconnection market is headed in the coming years.
If AI continues to push clustering toward larger, denser, and more bandwidth-hungry configurations, success will depend not just on GPUs or switches but also on the ability to transmit more data with less power and friction between nodes and racks. This is precisely where Broadcom aims to compete strongly with Taurus.
Frequently Asked Questions
What exactly has Broadcom announced?
Broadcom announced the availability of Taurus BCM83640, a 3 nm PAM-4 optical DSP with 400G per lane designed for 1.6T modules for AI and cloud networks.
Why is the move to 400G per lane important?
Because it allows doubling the bandwidth per lane compared to architectures at 200G/lane, helping to improve density and capacity in high-performance AI networks.
What network capacity would this technology enable?
Broadcom states that 1.6T modules based on Taurus enable 102.4T systems in a 1RU, and the same technological foundation opens the door to 3.2T modules and 204.8T networks in the next generation.
Is the Broadcom Taurus BCM83640 already available?
Yes, Broadcom has indicated that sampling has begun with early-access clients and partners, although prices have not been publicly disclosed.
via: Broadcom