The AI infrastructure race is moving beyond one question: who has the fastest GPU?

As clusters grow toward hundreds of thousands—and eventually potentially millions—of accelerators, getting data between those GPUs quickly, reliably and efficiently becomes increasingly important. NVIDIA Corporation(NVDA.US) is now pushing that transition with Spectrum-X Ethernet Photonics, its co-packaged optics, or CPO, networking system for the Vera Rubin generation.

The key numbers explain why the technology deserves investor attention: 102.4 Tb/s of bandwidth per Spectrum-6 switch chip, roughly double the prior generation, about 5x better network power efficiency and as much as 10x higher reliability versus conventional pluggable-transceiver architectures.

The investment question is therefore bigger than a new NVIDIA switch.

If AI clusters continue scaling, networking and optical interconnects could become an increasingly important part of the AI infrastructure spending cycle.

Why AI Is Becoming a Networking Problem

A powerful GPU only creates value when it can stay busy.

In large AI clusters, thousands of GPUs constantly exchange data during model training and inference. As the number of accelerators increases, congestion, latency and failed connections can leave expensive GPUs waiting instead of computing.

NVIDIA Corporation(NVDA.US) describes this as a growing constraint on AI-factory performance: peak GPU performance alone is no longer enough to determine how effectively a large cluster operates.

Spectrum-6 is designed for that problem.

The chip delivers 102.4 Tb/s per switch chip, up from 51.2 Tb/s for the previous Spectrum-4 generation, while moving from 100G to 200G signaling. NVIDIA also offers a larger CPO system capable of 409.6 Tb/s aggregate switching capacity.

That creates a useful framework for investors:

GPU performance determines how fast individual accelerators can calculate. Networking increasingly determines how efficiently thousands of those accelerators can work together.

CPO Changes Where the Optics Sit

Traditional data-center switches generally rely on pluggable optical transceivers connected to the switch through electrical pathways.

CPO brings silicon photonics much closer to the switch ASIC itself. By shortening the electrical path and eliminating components such as separate DSP retimers, NVIDIA Corporation(NVDA.US) says the architecture reduces complexity, power consumption and potential failure points.

The company reports that Spectrum-X Ethernet Photonics can reduce optical loss from roughly 22 dB to about 4 dB, translating into as much as 64x better signal integrity. NVIDIA Corporation(NVDA.US) also says its external laser architecture can reduce the number of lasers required by approximately 4x compared with legacy designs.

For hyperscale AI infrastructure, those improvements matter economically.

Networking consumes power that could otherwise be allocated to compute. Failures can interrupt expensive training jobs. And more components create additional installation and maintenance complexity.

That is why 5x better network power efficiency, 5x longer sustained AI application runtime and higher reliability may ultimately matter more than the headline bandwidth number itself.

NVIDIA Is Expanding the AI Trade Beyond GPUs

For NVIDIA Corporation(NVDA.US), the strategic significance is straightforward.

The company is increasingly selling an AI factory architecture, rather than simply selling accelerators. GPUs sit at the center, but NVLink, Spectrum-X Ethernet, ConnectX networking, BlueField DPUs and software increasingly determine how the overall system performs.

That makes networking another layer through which NVIDIA Corporation(NVDA.US) can deepen its position inside AI infrastructure.

The next confirmation investors should watch is deployment rather than specifications.

CoreWeave(CRWV.US), Lambda and Oracle Corporation(ORCL.US) Cloud Infrastructure are among the first adopters of Spectrum-X Ethernet Photonics, while CoreWeave(CRWV.US), Microsoft Corporation(MSFT.US), Oracle Corporation(ORCL.US) and others are also among early users of the broader Spectrum-6 platform.

If CPO moves from early deployments into a standard architecture across large Rubin-era AI clusters, that would provide stronger evidence that optical networking is becoming a structural part of the AI capex cycle.

Why Lumentum May Be One of the More Direct Optics Stocks to Watch

For investors looking beyond NVIDIA itself, Lumentum Holdings, Inc.(LITE.US) provides one of the clearer U.S.-listed links to the optical side of the story.

Lumentum Holdings, Inc.(LITE.US) is part of NVIDIA's silicon photonics ecosystem, alongside companies including Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR(TSM.US), Foxconn, SPIL, Coherent Corp.(COHR.US) and Fabrinet(FN.US).

The relationship has also moved beyond ordinary supplier cooperation. In March, NVIDIA announced a multiyear strategic agreement with Lumentum covering advanced optics, including a $2 billion NVIDIA investment, a multibillion-dollar purchase commitment and future capacity-access rights for advanced laser components.

That makes Lumentum Holdings, Inc.(LITE.US) relevant to a different part of the AI investment thesis.

For NVIDIA Corporation(NVDA.US), investors are watching whether networking expands the company's AI platform.

For Lumentum Holdings, Inc.(LITE.US), the question is whether rapidly increasing demand for lasers and photonics translates into higher capacity utilization, stronger revenue growth and sustainable profitability as new manufacturing capacity comes online.

This distinction matters. Rising industry demand does not automatically guarantee higher margins if suppliers expand capacity too aggressively.

TSMC Adds Another Layer to the Supply Chain

Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR(TSM.US) is also part of NVIDIA Corporation(NVDA.US)'s CPO ecosystem.

TSMC has developed its own silicon photonics platform, while NVIDIA identifies the foundry as one of the technology partners supporting its co-packaged optical architecture.

For TSMC, silicon photonics is unlikely to rival leading-edge GPU fabrication as an earnings driver in the near term.

But strategically, the technology reinforces a broader trend: advanced AI infrastructure increasingly requires the integration of compute, packaging and optical connectivity rather than improvements in transistor performance alone.

That potentially increases the importance of suppliers capable of handling multiple layers of advanced semiconductor manufacturing.

CoreWeave and Oracle Are the Demand-Side Test

CoreWeave(CRWV.US) and Oracle Corporation(ORCL.US) play a different role in this theme.

They are customers rather than optical-component suppliers, which means Spectrum-X does not represent a direct product-revenue opportunity for them.

Instead, their deployments could provide some of the first real-world evidence of whether CPO improves AI-cloud economics at scale.

CoreWeave is already among the first companies deploying Spectrum-6 as switching infrastructure for Vera Rubin systems, while Oracle Cloud Infrastructure is among the first Spectrum-X Ethernet Photonics adopters.

For these companies, investors may want to watch whether next-generation networking helps improve:

• GPU utilization and cluster efficiency;
• power consumed by networking infrastructure;
• reliability during large training and inference workloads; and
• the speed at which additional AI capacity can be deployed.

If those benefits become measurable in commercial deployments, the investment case for CPO becomes stronger than laboratory specifications alone.

What Happens Next? Three Tests for the Optical AI Trade

The first test is adoption.

The critical question is whether CPO remains concentrated among a few leading AI infrastructure operators or spreads broadly through Rubin-generation deployments.

The second is capacity utilization across the optics supply chain.

Companies are expanding optical and laser manufacturing capacity to meet expected AI demand. Investors should therefore watch whether actual orders grow fast enough to absorb that capacity rather than focusing only on announced expansion plans.

The third—and perhaps most important—is AI economics.

If CPO helps operators keep more GPUs working while using less networking power and experiencing fewer interruptions, optical interconnect spending becomes easier to justify even as AI infrastructure budgets face increasing scrutiny.

That would strengthen the idea that the next phase of AI infrastructure spending is becoming broader:

Compute remains the engine, but networking increasingly determines how efficiently that engine can run.

Stocks & ETFs to Watch

Risk Warning: Leveraged ETFs can amplify both gains and losses and may not track their stated multiple over longer holding periods. Investors should fully understand the risks before trading.

Disclaimer: This article is for informational purposes only and does not constitute investment advice or a recommendation to buy or sell any securities.