AI compute gets the attention, but the patent depth behind optical interconnect sits elsewhere. The result is a useful lesson in where ownership and leverage actually live.
The market has decided the next great AI moat is the optical interconnect, the layer that moves data between chips with light instead of copper. It is pricing that decision in real money. Ayar Labs raised 500 million dollars in March 2026 at a 3.75 billion dollar valuation. Lightmatter sits at 4.4 billion. Marvell paid up to 5.5 billion to absorb Celestial AI. So here is the uncomfortable question almost nobody is asking: if this layer is the moat, who actually owns it? Pull the patent data and the answer is not the company you would guess. It is not NVIDIA, and it is not the startups soaking up the funding.
Start with why this layer suddenly matters. For two decades, raw compute raced ahead of everything around it. Peak server compute scaled about 3x every two years. The memory bandwidth feeding it scaled about 1.6x, and the interconnect bandwidth between chips only about 1.4x (Gholami et al., AI and Memory Wall, arXiv:2403.14123, 2024). Model sizes, meanwhile, grew roughly 410x every two years against accelerator memory that barely doubled. The processor stopped being the bottleneck. The path between processors became it.
That is not an abstract academic point. It is why hyperscaler capital is now flowing into physical plumbing rather than model cleverness. Microsoft guided to about 190 billion dollars of capital spending in 2026, up roughly 61 percent year over year, and attributed about 25 billion of the increase to higher memory and component prices (CNBC, April 29, 2026). When the most valuable software companies on earth start spending like utilities, the constraint has moved from code to copper, and from copper to light.
The patent record shows the response was a long time coming. Optical interconnect for compute now spans 29,568 patent families worldwide (Minesoft Origin, extended-family count, run June 2026). New families roughly tripled across two decades and climbed from about 1,020 in 2015 to 2,668 in 2025. This is not a sudden gold rush. It is a 20-year build that the market only started pricing once compute hit the wall.
Now the part that should change how you read the headlines. When you rank who actually holds that patent depth, the AI names are missing.
The leaders are the optical and telecom old guard. Huawei holds 1,286 families. Nokia 666. Cisco 461. Marvell 354. Coherent 331, and that is before counting its II-VI and Finisar lineage, which adds well over a hundred more. Ciena 284. Intel 573. These are the companies that spent thirty years learning how to push light through glass for the telecom network, and they quietly carried that expertise into the data center.
Now find the AI darlings. NVIDIA, the most valuable company on earth, holds 46 families in this landscape, and its highest-ranked entries are not even organic. They are Mellanox patents, picked up in the 2019 acquisition. The photonics startups commanding billion-dollar valuations are thinner still: Ayar Labs has 12 families in this space, Lightmatter 14, Celestial AI a handful of recent filings. The market is paying startup-of-the-decade prices for patent footprints you could read in an afternoon.
This is the divergence that matters. Attention has piled onto NVIDIA and a short list of venture-backed photonics names. Ownership of the underlying layer sits somewhere else entirely.
Hold the valuations and the patent counts next to each other and the gap is almost comic.
Ayar Labs: 3.75 billion dollar valuation, 12 patent families in this landscape. Lightmatter: 4.4 billion, 14 families. That is hundreds of millions of dollars of valuation per family. Compare that to Coherent, which holds 331 families and makes the lasers that the whole co-packaged-optics supply chain depends on, or to Marvell, which holds 354 and just paid up to 5.5 billion to buy even more depth through Celestial AI.
Two readings are possible, and both are useful. One: the startups own something narrow but foundational, and a dozen well-aimed families can matter more than a thousand defensive ones. Two: the market is funding a moat that the patent record does not yet show, and the incumbents who do hold the depth can build or buy their way across it. The fact that Marvell chose to buy Celestial rather than out-file it, and that NVIDIA and AMD both fund Ayar rather than letting either own it outright, tells you the incumbents are taking the second reading seriously. They are consolidating the layer, on their terms.
You do not have to take the counts on faith. Read the claims, and the renting-versus-owning split jumps out.
Ayar Labs holds a patent application titled Remote Memory Architectures Enabled by Monolithic In-Package Optical I/O (WO-2021163653-A1, 50 claims). Its first independent claim describes a remote memory system built from a multi-chip package: an integrated-circuit chip with a high-bandwidth memory interface, paired with an electro-optical chip whose optical macros use microring resonators to convert electrical signals into light, push them across an optical link, and convert them back, with later claims reaching wavelength-division multiplexing and the CXL memory protocol. In plain terms, that is a claim aimed at the exact wound described at the top of this piece: it disaggregates memory from compute and reconnects them with light, attacking the memory wall at the architectural level. A 12-family startup is sitting on an architecture-level description of the problem everyone is now spending billions to solve.
Then look at NVIDIA’s strongest entry in the same landscape. It is a Mellanox patent, Distributed Optical Circuit Allocation in Optical Data-Center Networks (US-12323750-B2, granted 2025). Its first claim describes a method for allocating data paths across optical switches: maintaining distributed allocation tables, sending queries along candidate optical circuits, pruning links where channels are unavailable, and establishing the path that clears. Valuable engineering, genuinely. But notice what it is. It is network-management software for routing traffic across an optical fabric, inherited from an acquisition. It is not a claim on the physical optical engine that makes the light in the first place.
That is the whole thesis in two documents. The startup describes the optical-to-memory architecture. The compute king holds an inherited algorithm for managing the network. One is the engine. The other is the traffic controller. (This is a description of published claim language for competitive context, not a legal opinion on scope, validity, or infringement.)
The fair pushback: NVIDIA is not trying to win on optical patents, and it does not need to. Its moat is the integrated system, the GPU plus NVLink plus CUDA plus the rack, and it can assemble best-in-class optics from suppliers without owning a single waveguide patent. That is true, and it is exactly the point. NVIDIA’s co-packaged-optics switches, the Spectrum-X and Quantum-X Photonics line announced at GTC in March 2025, are built on TSMC’s 3D packaging, on lasers from suppliers like Coherent and Lumentum, and on optical I/O from partners like Ayar that NVIDIA funds rather than owns. That is a tenant’s position. A very powerful tenant, with the best address in the building and the leverage to dictate terms to the landlord, but a tenant. The value of the interconnect layer accrues, at least in part, to whoever holds the deed.
For most companies, the lesson is not "be NVIDIA." It is the opposite. You will not have the leverage to rent the critical layer and still capture its value. If the layer your product depends on is one you do not own, your margins live at the mercy of whoever does.
So here is the takeaway worth carrying out of all this, and it generalizes far past photonics.
In every boom, attention and ownership separate. Press releases track attention. Patent portfolios track ownership. The optical-interconnect story is a clean case: the attention sits on NVIDIA and a few funded startups, while the ownership sits with a telecom old guard that has been filing through this layer for twenty years. When those two signals diverge, the gap is not noise. It is where the leverage lives, for licensing, for acquisition, for pricing power. Marvell already acted on it. Others will.
For founders: build where you can own, not where the press is loudest, and aim your filings at the layer a competitor cannot simply buy. For investors: separate the layer that is getting funded from the layer that is owned, because they are not always the same, and the gap is a risk you are underwriting. For acquirers: the most interesting targets are the quiet ones whose portfolios sit under the layer everyone else is bidding up.
Read the portfolio, not the press release. The logo on the keynote stage is rarely the name on the deed.
This analysis is competitive and strategic intelligence based on published patent and market data. It is not legal advice. For opinions on patent scope, validity, or infringement, consult qualified patent counsel. Patent figures are extended-family, landscape-level counts from Minesoft Origin (run June 2026) and are research findings, not legal conclusions. Market and funding figures are attributed to the sources named in the text.
Work with ipCapital Group
From invention to monetization, our team has guided 2,000+ engagements across the full IP lifecycle. Start with a free 30-minute discovery call.
Written by
Seth Cronin