ETFs · Optics · Explainer

Roundhill launches LYTE photonics ETF — and why optics now runs AI infrastructure

6 August 2026 · 9 min read

Roundhill Investments began trading the Roundhill Photonics & Optics ETF (LYTE) on 6 August 2026. It is an actively managed fund that holds at least 80% of its assets in companies whose core technology involves generating, manipulating, detecting or transmitting light — across AI data centres, defence, medical imaging, industrial manufacturing and quantum computing, with a stated preference for businesses spending heavily on research and development.

The fund is the news. The reason it exists is the more interesting story, and it starts with a problem that has nothing to do with markets.

Copper ran out of road

Every AI cluster is thousands of accelerators that have to behave like one machine. That only works if they can talk to each other fast enough, and for most of computing history that meant copper.

Copper has a hard physical ceiling. Push the data rate high enough and the signal degrades over distance — the faster you go, the shorter the usable cable. At the speeds modern accelerators need, passive copper survives a couple of metres at best. Inside a single rack that is fine. The moment a workload spans racks, rows or buildings, copper simply cannot carry it.

Light does not have that problem. An optical link carries far more data, far further, with far less loss. So as clusters grew past the size of one rack, the industry did not choose optics because it was elegant. It ran out of alternatives.

That is the whole thesis in one sentence: optics is not an AI accessory, it is what makes a cluster bigger than a rack possible at all.

Why the bottleneck moved from compute to the fabric

The intuitive model of an AI data centre is that performance comes from the accelerators. That was true when models fit on a handful of chips. It stopped being true some time ago.

Training a large model means constantly synchronising state across every chip involved. Each step, huge volumes of data move between accelerators. If that movement is slow, the accelerators sit idle waiting — and an idle accelerator is the most expensive thing in the building.

This is why utilisation, not raw compute, is the number operators actually chase, and why interconnect spending has grown faster than almost anyone modelled. You cannot fix a network bottleneck by buying more GPUs. You make it worse.

Where the money actually goes

"Optics" in this context is not one product. It is a stack, and different listed companies sit at different layers.

Transceivers

The modules that convert electrical signals to light and back, plugged into switches and servers at both ends of every optical link. The industry has moved through 400G to 800G, with 1.6T now arriving. Each generation roughly doubles throughput per port.

The important dynamic is the attach rate — how many optical modules are needed per accelerator. It is not one. A single accelerator may need several links, and cluster architectures that add network layers multiply that further. Module demand can grow considerably faster than accelerator unit sales, which is why this part of the supply chain has repeatedly surprised on the upside.

Lasers and optical components

Someone has to make the light. Lasers, modulators, detectors and the exotic materials behind them — indium phosphide in particular — are a genuine constraint, with long lead times and a small number of capable suppliers.

Connectivity silicon

The chips that keep high-speed links reliable. As rates climb, even short in-rack copper needs active silicon to stay error-free, which created a component market that barely existed a few years ago.

Switching and transport

The switches tying the fabric together, and the long-haul optical transport connecting separate datacentre campuses. As training runs span multiple sites, the bandwidth needed between buildings has become its own line item.

The power problem, and what comes next

Every pluggable transceiver burns watts. Multiply by tens of thousands of ports and interconnect becomes a meaningful share of a facility's power budget — in a business where power is already the binding constraint on how much capacity can be built.

That pressure is driving the sector's next architectural shift: co-packaged optics, which moves the optical engine directly alongside the switch silicon instead of into a pluggable module at the faceplate. Shorter electrical path, less power per bit, more bandwidth per rack.

It is also a genuine risk to incumbents. A transition from pluggable modules to co-packaged designs redistributes value across the supply chain, and the companies best positioned for 800G pluggables are not automatically the winners in a co-packaged world. This is the single most important technical debate in optics, and it is unresolved.

The listed companies behind the theme

LYTE had not published holdings at launch, but the US-listed optical universe is well defined. ChipSentiment tracks it live:

Marvell (MRVL) — optical interconnect silicon and custom accelerator programmes
Credo (CRDO) — active electrical cables and rack-scale connectivity silicon
Coherent (COHR) — lasers, optical components and transceivers
Lumentum (LITE) — transceivers and laser components, pivoted from telecom to datacentre
Ciena (CIEN) — optical transport between datacentre campuses
Broadcom (AVGO) — networking silicon, switching and co-packaged optics work
Arista (ANET) — high-speed Ethernet switching for accelerator clusters

Smaller pure-plays including Applied Optoelectronics and POET Technologies rose when the LYTE prospectus was filed in May — a reminder that order books in this corner are thin enough that fund flows alone can move prices.

See live prices and sentiment across the optical group →

What the fund adds, and what it does not

LYTE is actively managed, so a manager selects and weights the holdings rather than tracking an index. For a theme this technical, where the pluggable-versus-co-packaged question could reshuffle winners, there is a reasonable argument for active selection.

The honest counterpoint: most of these companies already trade on US exchanges and can be bought individually at zero commission. Roundhill's memory fund, DRAM, solved a real access problem — SK Hynix and Samsung dominate global memory and are listed in South Korea, awkward for US retail investors to reach. Photonics has no equivalent barrier, so LYTE is selling selection and convenience rather than access.

It is also not first. Tema's LAZR launched on 30 June with a similar mandate and has gathered roughly $18 million, with shares down since inception. Whether LYTE follows that path or Roundhill's more spectacular one will be visible in its first fortnight of asset flows.

The risks worth naming

Roundhill's own prospectus is direct about the central dependency: it warns that commercialisation of photonic and optical technologies may hinge on continued adoption of AI, data centre infrastructure and advanced telecoms, and that a slowdown in those areas could materially affect the companies the fund holds.

It is a single bet on the buildout. Optics has no defensive characteristics here. If hyperscaler capital expenditure slows, this is among the first places it shows.
Technology transition risk. The shift toward co-packaged optics could favour different companies than today's leaders.
Concentration. Narrow thematic funds move hard in both directions. That is the trade, not a defect.
Valuation. Much of this group has already re-rated substantially on exactly this thesis.

The one-line version

Compute gets the headlines, but a cluster is only as fast as the links between its chips — and past a certain size, those links have to be light. Whether you buy the fund or the underlying names, that constraint is the thing you are actually investing in.

None of the above is a recommendation, and it does not account for your circumstances. Confirm fees, holdings and terms in the prospectus before acting on anything here.

Frequently asked questions

What is the Roundhill LYTE ETF?

The Roundhill Photonics & Optics ETF (LYTE) is an actively managed exchange-traded fund that began trading on 6 August 2026. It holds at least 80% of assets in companies whose core technology involves generating, manipulating, detecting or transmitting light, spanning AI data centres, defence, medical imaging, industrial manufacturing and quantum computing.

When did the LYTE ETF launch?

6 August 2026. Roundhill filed the prospectus with the SEC in May 2026.

Why is photonics important for AI data centres?

Copper connections degrade at high data rates over distance, so beyond a couple of metres they cannot carry the traffic modern accelerator clusters generate. Optical links carry far more data much further. Once a cluster spans more than a single rack, optical interconnect is what makes it possible at all.

What is co-packaged optics?

An architecture that places the optical engine directly beside the switch silicon rather than in a pluggable module at the faceplate. It shortens the electrical path, cuts power per bit and increases bandwidth per rack. It is also a transition risk, because the companies leading in pluggable modules are not automatically the winners in a co-packaged design.

Which photonics and optics stocks are listed in the US?

Marvell, Credo, Coherent, Lumentum, Ciena, Broadcom and Arista are the main US-listed names spanning optical silicon, lasers, transceivers, switching and transport, alongside smaller pure-plays such as Applied Optoelectronics and POET Technologies.

Is LYTE the first photonics ETF?

No. Tema's LAZR Photonics & Optics ETF launched on 30 June 2026 with a similar mandate and has gathered roughly $18 million, with shares down since inception.