Subscribe to newsletter
Thank you!
You have successfully subscribed.
Data center networking is undergoing a structural shift driven by the biggest infrastructure buildout in the history of technology. The world's largest hyperscalers are collectively spending north of $600 billion a year on infrastructure. And I expect that number to cross $1 trillion by 2028 and $2 trillion by 2030, based on projections driven by the adoption and diffusion of existing model capabilities and RL training regimes. If you believe we're approaching superintelligence sooner than consensus, the acceleration could be steeper still.
Networking sits at the center of this buildout as a critical enabler. Without dramatically more bandwidth, lower latency, and better energy efficiency in the networks that stitch AI clusters together, the next generation of AI infrastructure simply cannot be built.
What makes this particularly interesting is that the opportunity is growing along two independent axes simultaneously.
The first is the raw buildout itself. More data centers, more racks, more GPUs all of which need to talk to each other. Every new GPU cluster is a new network deployment.
The second is the shift from copper to optical. The industry has long operated on a simple heuristic: copper where you can, optical where you must. Copper is cheaper, simpler and more reliable. But copper has physical limits: signals degrade at high speeds and it can't scale to the distances and bandwidths that modern AI infrastructure demands. The shift to optical is not a question of if, but when and how fast.
When we solve the challenges around cost, reliability, and maintenance, optical adoption will accelerate dramatically, unlocking new scale-up cluster designs and thus more intelligence. This is the second growth engine, layered on top of the raw buildout, setting up the industry for explosive growth.
Despite the wild hype over the past months in public markets, the copper-to-optical transition won't happen overnight. Instead think of it as four overlapping phases, each following its own adoption curve playing out over the next 5+ years:
Phase 1 - optical scale-out (already underway). Optical links already dominate the connections between racks and across data-center fabrics. Optical Circuit Switching (OCS) is also deployed in selected architectures, notably Google’s TPU infrastructure, where it allows the physical network topology to be reconfigured for availability, utilization and workload requirements.
Phase 2 - optical scale-up spans racks. The next major transition is the expansion of the tightly coupled scale-up domain beyond a single rack. As AI systems grow from tens to hundreds or thousands of accelerators, copper can no longer provide the necessary reach, bandwidth density and signal integrity. Optical links will connect accelerator trays and racks into larger scale-up systems, potentially complemented by reconfigurable optical fabrics in some architectures.
Phase 3 - optics replace intra-rack copper. In parallel, optics will move progressively deeper inside the rack. Rising lane rates and aggregate bandwidth will make longer electrical traces and backplanes increasingly difficult to engineer. Optical links will first replace the longest and highest-bandwidth intra-rack copper connections, with the optical-electrical conversion point moving closer to switches and accelerators over time. Copper will likely remain for the shortest links where its cost and simplicity remain advantageous.
Phase 4 - optics reaches the package. Pluggable optical modules will increasingly be supplemented by Near-Package Optics (NPO) and Co-Packaged Optics (CPO). NPO places optical engines close to the compute or switch ASIC, while CPO integrates them into the same package assembly. Switch-side CPO is already beginning to commercialize; accelerator-side CPO for large scale-up domains is more technically demanding and will likely mature later.
These phases are not sequential. Phases 2 and 3 will overlap; Phase 4 will begin in pockets while Phases 1 and 2 are still maturing. Different hyperscalers will move at different speeds. But the end-goal is clear: building photonics superclusters through optically linking thousands of chips to unlock new generations of models.
The demand side of this market is extraordinarily concentrated.
The networking TAM we're discussing, roughly 15% of total data center capex and growing, is controlled by a handful of buyers. On the one side is NVIDIA that sells their vertically integrated solutions (under the product names of NVLink and Infiniband) mostly to neoclouds. On the other are the hyperscalers, including Google, Meta, Microsoft, Amazon, OpenAI, and now Anthropic and xAI, who have the engineering talent, capital and purchasing scale to avoid the NVIDIA system tax.
The hyperscalers design and build their own networking infrastructure. They source components (switch ASICs from Broadcom, transceivers from Lumentum and Coherent, assembly from ODMs like Foxconn) and deliberately avoid dependence on a single vendor. A startup going after these customers is selling to some of the world's most sophisticated engineering teams and needs to have a strong story of how they fit into their technical roadmaps.
The neoclouds don't have that capability. They buy the NVIDIA system and let NVIDIA design the stack, take the risks and manage the supply chain. A startup selling into this channel is selling to NVIDIA, not the end customer. NVIDIA's recent $4B investment split between Lumentum and Coherent to lock up photonics supply capacity is a signal of how tightly they intend to control that ecosystem.
What does that mean for startups? Win one of these, and you can scale to hundreds of million in revenue over night. Fail to understand how they buy, and you'll spend years in sales cycles that never close. This is high-stakes sales with extreme technical complexity.
Getting a hyperscaler to buy your product is just as much a qualification gauntlet as it is a sales process.
Qualification for a single device takes 3–6 months and costs millions before a single purchase order is placed. That process covers accelerated lifetime tests (running devices under extreme temperatures for 1,000+ hours), vibration and shock testing, electrostatic discharge, and interoperability testing across multiple switch vendors. The products get evaluated across a myriad of dimensions: Does the signal work reliably? What do your eye diagrams look like? Does it pass Bit Error Rate thresholds?
And just as importantly, do you have a credible supply chain, fab capacity and a yield story that can scale to millions of units? InP bottlenecks are just one of many challenges in today’s supply-constrained markets.
Landing your first design win with a hyperscaler is not the finish line. It leads to a bigger question: what kind of company do you want to build?
That first win matters enormously: it proves the product, generates the first real revenue, and creates switching costs that are valuable as they are hard to unwind. A hyperscaler that has spent 3–6 months and millions of dollars qualifying your device, co-developed it against their specific infrastructure roadmap, and integrated it into their supply chain is not going to replace you overnight. This design-in is a strong moat. But a single design win with a single customer is a single point of failure, not a durable company.
The path from here to building something enduring requires deepening within the customer you've won, and expanding to the next one. Deepening means staying relevant on the roadmap and when the industry turns over a speed generation in two years, you need to be ready with the next product, or they'll qualify someone else. Expanding means taking the credibility of your first win and using it to get into a second buyer's qualification process, then a third. It means building a portfolio of products across different parts of the optical stack, different performance tiers, potentially different applications.
History shows a consistent pattern: companies in this space can get acquired early, often at valuations well ahead of their financial maturity. Just recently, Credo bought DustPhotonics for roughly $1B and Marvell bought Celestial for $3B+. Neither company was shipping large volumes at that time. The acquisitive nature of well-resourced incumbents makes it genuinely hard for a new company to break through and fully capture the benefit of these overlapping growth curves.
While an early exit is tempting, it's also hard to imagine the optical networking landscape of 2035 looking exactly like today's. The technology transitions are too large, the capital inflows too significant, and the pace of change is accelerating. Copper will have been pushed out of the rack. Optics will be enabling the next generation of photonic supercomputers that span entire data halls. Supply chain bottlenecks will have eased. And the hyperscalers that today rely on Lumentum and Coherent as the dominant transceiver suppliers may well be buying from companies that are just getting started today.
The companies that resist the early exit, win multiple customers and diversify into adjacent products are well positioned to be the next Lumentum in 2035. That’s when a one-product winner becomes a generational company shaping the future of AI.