Over the past two years, the main storyline of the US AI rally has always revolved around GPUs: whose chips are more powerful, whose shipments are larger, and whose customers have bigger budgets. But a growing body of evidence suggests the bottleneck determining AI cluster efficiency is shifting from chips to the network. As GPU counts per cluster move from tens of thousands toward hundreds of thousands, the bandwidth, latency, and stability of data exchange between chips are becoming key variables in the cost and speed of large model training. Starting from industry logic, this article maps the upgrade path of optical interconnects and data center networking, and analyzes investment opportunities across the related US-listed supply chain.
1. Why the Network Is Becoming the New Bottleneck for AI Clusters
To understand this shift, we first need to distinguish the two ways AI clusters scale: scale-up and scale-out. Scale-up means tightly coupling more GPUs within a single rack through high-bandwidth interconnects, with NVIDIA's NVLink as the typical example; scale-out means linking hundreds or thousands of servers across racks and data halls into a massive cluster, relying mainly on InfiniBand or Ethernet switching fabrics.
The problem is that large model training uses multiple parallel strategies such as data parallelism and tensor parallelism, requiring GPUs to perform collective communications like gradient synchronization at high frequency. Any shortfall in network bandwidth or jitter in latency leaves expensive GPUs sitting idle, lowers actual utilization, and passively drives up training costs. It is widely observed in the industry that the larger the cluster, the higher the share of communication overhead, and the more significant the impact of network quality on effective compute output. This is precisely why leading cloud providers, while procuring GPUs, are willing to invest comparable levels of effort and budget in network equipment.
Moreover, as inference takes a growing share of demand, traffic patterns in hyperscale data centers are changing too: training traffic prizes stability, while inference traffic behaves more like the bursty patterns of traditional cloud workloads. This places higher demands on network flexibility and operability, further elevating the network layer's position in AI infrastructure.
2. The Optical Interconnect Upgrade Wave: From 800G to 1.6T, Then CPO
Gains in network bandwidth ultimately come down to physical-layer optical interconnects. Over the past few years, data center optical modules have iterated rapidly along a 400G and 800G cadence, and the industry is now transitioning to the 1.6T generation. As speeds double, the number of optical modules per GPU is also rising: larger cluster scale, flatter network architectures, and higher GPU-to-optical-module ratios are stacking up, making optical components one of the fastest-growing segments in AI infrastructure buildout.
On the technology side, several branches deserve close attention. The first is silicon photonics, which combines optical components with mature silicon-based processes and is gradually showing advantages in cost, power, and integration. The second is low-power approaches such as linear direct drive, which trim the DSP stage to cut power and latency. The third, and the most radical, is co-packaged optics (CPO) — packaging the optical engine together with the switch chip, greatly shortening electrical signal travel and substantially lowering power consumption. As single-chip switch capacity evolves toward the hundred-terabit level, the power pressure on traditional pluggable solutions keeps mounting, and CPO is seen as a candidate for next-generation hyperscale switching; Broadcom and others have already added it to their product roadmaps.
Competitive Landscape: Three Types of Players, Each Holding Their Ground
The current landscape falls roughly into three categories. The first is system-level players: NVIDIA, through its acquisition of Mellanox, gained full capabilities in InfiniBand and Ethernet switching, and NVLink and Spectrum-X bundle with its GPUs — networking has become a key part of its ecosystem moat. Broadcom, with its Tomahawk line of switch chips and CPO initiatives, has become the preferred supplier for hyperscalers building their own networks. The second is optical component makers: Coherent and Lumentum command high-end optical engine and component capabilities, while Fabrinet handles precision optical module manufacturing. The third is connectivity chip vendors: Credo's active copper cable solutions and Astera Labs' retimer and CXL offerings have found clearly differentiated positions in intra-rack interconnects.
Notably, Chinese vendors have long held the bulk of global share in traditional pluggable optical module manufacturing — the industrial basis for the high correlation between the A-share optical module sector and the US AI rally. If CPO adoption accelerates, value will migrate from module assembly upstream to optical engines and components, and the industry's profit distribution could be restructured. This is one of the most important storylines to watch in this space over the next two years.
3. The US-Listed Investment View: Three Themes and One Variable
- Theme 1: Network chips and switching equipment. Broadcom's switch chips and custom networking ASICs, NVIDIA's networking business, and Arista Networks' Ethernet switches for AI clouds directly benefit from cloud providers' steadily rising share of network investment.
- Theme 2: Optical components and optical engines. Speed iteration layered on the CPO trend means vendors with silicon photonics and highly integrated optical engine capabilities offer greater earnings elasticity; watch Coherent, Lumentum, and Fabrinet on the manufacturing side.
- Theme 3: Intra-rack interconnect. As rack power density climbs, the boundary between copper and optical interconnects within the rack keeps being redrawn; connectivity chip companies like Credo and Astera Labs hold clear strategic positioning advantages.
The biggest variable, however, is the progress of cloud providers' in-house networking and supply chain diversification. To cut costs and reduce dependence on a single vendor, hyperscalers are accelerating an Ethernet-based open ecosystem — the Ultra Ethernet Consortium's standardization work is a prime example. This trend poses a potential challenge to NVIDIA's networking advantage, yet benefits Broadcom, Arista, and optical component vendors in the open ecosystem. The same trend carries opposite implications for different names, so investors need to track it dynamically.
4. Risk Disclosure: Three Tests Beneath the Excitement
First, technology transition risk. CPO's reliability, maintainability, and cost curve have yet to be fully proven; if adoption lags expectations, high-beta names face valuation correction pressure. Second, customer concentration risk. AI network orders are highly concentrated among a handful of cloud providers and the NVIDIA ecosystem, and any swing in a single customer's capex gets amplified down the supply chain. Third, competition and margin risk. As more players enter, price competition in optical modules and components may intensify, putting the durability of excess profits in doubt. Moreover, the sector's overall valuation is not cheap, leaving little margin for error on earnings delivery — the risk of chasing highs blindly should not be overlooked.
Conclusion: From Chip Compute to Network Compute
The AI compute race is evolving from a contest of single-point chip performance into a comprehensive contest of whole-system engineering. Networking and optical interconnects are the undervalued link in that system: they lack the star power of GPUs, yet they determine whether every GPU gets fully utilized. For investors, rather than agonizing over the grand question of whether this is a bubble, it is better to return to industry details and clarify how specific variables — bandwidth, power consumption, interconnect architecture — will evolve. True excess returns often hide in these unglamorous links. Going forward, we will keep tracking how each company's networking revenue mix and guidance change through earnings season, bringing readers more timely compute intelligence updates.
