For three years, investors have been trained to count GPUs. More accelerators meant more AI capacity, more capital spending and, usually, another reason to revisit Nvidia's valuation. That story was right. It is also becoming incomplete.
The next constraint is increasingly what sits between the processors. Tens of thousands of accelerators only become one useful computing system if data can move between them with low latency, predictable throughput and tolerable power consumption. A $50,000 accelerator that spends too much time waiting for another accelerator is a very expensive idle asset.
That is why I think the next phase of the AI infrastructure trade is moving outward from the chip and into the network. Ethernet is being rebuilt for AI-scale workloads. Optical links are replacing electrical ones at shorter distances. Co-packaged optics are moving closer to the switch silicon. The market has spent years asking who sells the compute. It is going to spend more time asking who keeps the compute busy.
Ethernet is no longer the boring option
For years, InfiniBand had the cleaner high-performance-computing story. Ethernet had ubiquity; InfiniBand had deterministic performance. The AI buildout is forcing the industry to narrow that gap quickly.
The Ultra Ethernet Consortium published the first public version of its specification in June 2025 and version 1.0.3 on 16 July 2026. Its mission is explicit: an open, interoperable, high-performance Ethernet communications stack for AI and HPC at scale.
That matters commercially because open standards widen the supplier base. Hyperscalers spending tens of billions of dollars on infrastructure have every incentive to avoid unnecessary single-vendor dependence. If Ethernet can deliver the congestion control, lossless transport and scale AI clusters require, the addressable market for switching silicon, systems and optics becomes more competitive and potentially much larger.
Nvidia's own strategy tells you networking is becoming strategic
Nvidia is not waiting for Ethernet to become somebody else's opportunity. Its networking portfolio now spans NVLink for scale-up, Quantum InfiniBand and Spectrum-X Ethernet for scale-out, multi-data-center networking and next-generation silicon photonics.
The company describes the network as the fabric that allows tens of thousands, and eventually millions, of GPUs to behave like one distributed engine. That wording is important. Networking is no longer treated as plumbing around the compute. It is part of the compute system itself.
Nvidia's Spectrum-X platform is also a reminder that the Ethernet-versus-InfiniBand argument is becoming less binary. Nvidia can sell into both outcomes. For investors looking for more incremental exposure to the transition, however, the more interesting names may be suppliers whose addressable markets change dramatically as AI networking intensity rises.
Watchlist #1: Arista Networks
Arista Networks is the cleanest listed-company expression of the Ethernet side of this thesis in my view.
Arista reported its first quarter above $3 billion of revenue in Q2 2026, with non-GAAP earnings per share up 40% year over year. More importantly for this argument, it has introduced 1.6Tbps AI fabric platforms, including liquid-cooled systems designed for scale-up, scale-out and scale-across architectures.
That is a meaningful change in role. Arista is moving from selling high-performance switches toward supplying the network architecture around rack-scale AI systems. If Ethernet becomes the default fabric for a larger share of AI clusters, Arista benefits from both the number of ports and the increasing value of each networking layer.
The risk is obvious: none of this is secret. Arista already trades as an AI infrastructure beneficiary, its hyperscaler exposure is concentrated and Nvidia is competing aggressively in Ethernet. I would therefore treat Arista as high-quality exposure to a standards shift, not as an undiscovered bargain.
Watchlist #2: Coherent
The second name is where the networking story becomes an optics story.
Coherent reported $2.05 billion of revenue in its fiscal fourth quarter, up 34% year over year, alongside substantial margin expansion. Management has been unusually direct about what is driving the opportunity: AI data-center architectures are moving from copper toward optical connectivity.
That transition is fundamental. As bandwidth rises and clusters become larger, electrical links run into distance, power and heat constraints. Every generation of higher-speed networking moves more connections toward light.
Nvidia made the strategic importance explicit in March when it announced a $2 billion investment in Coherent, a multibillion-dollar purchase commitment and future capacity rights for advanced laser and optical networking products. Companies do not secure capacity that aggressively for components they consider peripheral.
Coherent's attraction is manufacturing scale across photonics. Its risk is also manufacturing scale: capacity expansions can eventually create pricing pressure, and optical-component markets have a history of moving from shortage to oversupply. The secular direction can be right while the cycle still hurts.
Watchlist #3: Lumentum
Lumentum is the more aggressive version of the optics thesis.
The company reported $1.01 billion of fiscal fourth-quarter revenue and guided the following quarter to between $1.225 billion and $1.275 billion. Management says AI compute workloads are pushing data-center architects toward optical links as a primary means of connectivity, with 1.6T adoption beginning to layer into growth.
Nvidia also committed $2 billion to Lumentum in March, again alongside a multibillion-dollar purchase commitment and future capacity access rights. That gives the company unusually visible strategic validation, but it also tells you how tight this part of the supply chain has become.
Lumentum could have more operational torque than a diversified supplier if optical capacity remains scarce. It could also have more downside if customers dual-source, capacity catches up or pricing normalises. I would rank it behind Coherent on risk-adjusted quality, but ahead of most smaller optics names on strategic relevance.
Broadcom is the giant that sits across both sides of the trade
Broadcom is impossible to ignore because it sits across networking silicon and custom AI accelerators at the same time.
Its fiscal third-quarter revenue reached $29.6 billion, up 86% year over year, and management guided the fourth quarter to approximately $34.8 billion. Those numbers reflect a business much broader than Ethernet switching alone, which is both the attraction and the limitation of using Broadcom as the pure expression of this thesis.
If I want to know whether the entire hyperscale AI capital-spending machine remains healthy, Broadcom may be one of the best indicators in the market. If I specifically want exposure to Ethernet becoming more important relative to raw compute, Arista and the optics suppliers give a cleaner read-through.
That does not make Broadcom weaker. It makes it less pure.
The deeper standard shift is from copper toward light
I think Ethernet is only half the story. The more profound change may be the distance at which electrical signalling stops making economic sense.
Traditional pluggable optical transceivers are entrenched and serviceable, but higher bandwidth increases their power burden. The industry is therefore pushing optical engines closer to the switching silicon through silicon photonics and co-packaged optics.
Nvidia's Spectrum-X Ethernet Photonics platform is designed around that direction, integrating optics directly with the switch and targeting much higher bandwidth with lower power requirements than traditional architectures. Nvidia says the platform is intended to support AI factories scaling toward millions of GPUs.
That transition will not happen overnight. Pluggable optics have an enormous installed ecosystem and easier replacement economics. But AI is compressing the timeline because power is now one of the most valuable resources inside a data center. Saving watts on the network can translate into more watts available for compute.
The bottleneck keeps migrating
This is the pattern I keep coming back to in AI infrastructure. First there was a shortage of accelerators. Then high-bandwidth memory. Then advanced packaging. Then transformers, power and data-center land. Networking and optics are joining the list.
The opportunity tends to migrate toward the constraint that is hardest to solve next. Nvidia's separate $2 billion strategic investments in Coherent and Lumentum are unusually strong evidence that optical capacity has become important enough to secure financially, not merely negotiate commercially.
That does not mean every optics company is suddenly a good investment. Essential technologies can still produce mediocre returns if capacity expands too quickly, customers integrate vertically or prices collapse. Memory chips have taught investors that lesson repeatedly.
The question is not whether the component is necessary. The question is who can remain scarce, differentiated and profitable once the industry understands that it is necessary.
How I would rank the names
If I were building a watchlist around this specific standards shift, rather than making a call on the entire AI market, my current order would be Arista, Coherent, Lumentum and Broadcom.
Arista has the strongest combination of Ethernet exposure, hyperscale relevance and operating quality. Coherent gives me the preferred balance of optical exposure, manufacturing scale and strategic validation. Lumentum offers greater potential torque to tight optical capacity but also greater normalisation risk. Broadcom may be the strongest company of the four, but too much of its investment case now comes from custom AI silicon and software for me to call it a pure networking-standard trade.
That ranking is not an instruction to buy four tickers. It is a hierarchy of how directly I think each company benefits if AI infrastructure keeps moving from a compute-scarcity problem toward a connectivity-efficiency problem. Valuation still matters. Customer concentration still matters. Execution still matters.
My view
The market still tends to measure AI infrastructure by GPU count. I think that number will become less useful on its own.
The next question is how efficiently those accelerators can be turned into one machine. That requires faster Ethernet, better congestion control, more optical links, lower-power switching and eventually much more silicon photonics.
The AI boom is not moving away from semiconductors. It is moving down the bill of materials and exposing every weak link around them.
That is why I am spending more time on the companies that connect the GPUs than on the next incremental forecast for how many GPUs hyperscalers will buy.
The compute still matters. The network is becoming what decides how much of that compute you actually get to use.