Singular Link Intelligence · Issue 04 · September 2026
From Copper to Light: Rewiring the AI Factory — Photonics, Connectivity and the Deep Technologies Behind Them
Issues 01 to 03 followed the money into chips, compute and power. This issue follows the wires. AI chips are now so fast that the copper cables between them have become the limit. The answer is to send data as light — a shift that began between buildings, moved between racks, and is now moving onto the chip itself. We close on Ayar Labs, one of the companies building that last step.
01 · Executive summary — the network is now the bottleneck
Core thesis: an AI data centre is thousands of chips working as one machine, and how well they work together depends on how fast they can talk to each other. That conversation still runs largely on copper, and copper is reaching its limit. Whoever controls the lasers, the optical chips and the packaging will hold one of the next bottlenecks of the AI build-out.
Five headline findings
The faster a signal travels through copper, the shorter the distance it survives and the more power it wastes as heat. One NVIDIA rack already holds about 5,000 copper cables.
Optical links already connect racks. In 2026 they moved inside the switch. The next step puts them beside the processor, with volume expected from 2028.
Since March, NVIDIA has committed at least $6.5B to Coherent, Lumentum, Marvell and Corning, with purchase commitments and rights to laser capacity attached.
New materials, new lasers, micro-LEDs, optical switching and hollow-core fibre are moving from the laboratory to the data centre. About 70 venture-backed startups are chasing them.
Issue 03 showed that electricity is the binding constraint. Optics that halve the power spent on networking raise the output of every megawatt the region has planned — and Gulf capital is already on the cap table of the leading companies.
What an allocator should take from this issue. Connectivity is becoming a bottleneck of the same kind as chips and power. Proven revenue today is in lasers, modules and fibre. The venture upside is in the step beside the processor, which is about two years from volume and which the large incumbents intend to own. Copper has outlasted every forecast of its end, so timing matters more than direction.
02 · From copper to light — moving the light closer to the chip
For forty years, making chips faster was the hard part and connecting them was routine. That has reversed: the chips can now produce data faster than copper wire can carry it between them.
Distance: push a signal faster and it fades sooner. Heat: every watt spent moving data is a watt not spent computing. Space: thousands of thick cables per rack leave no room for more. AI makes all three worse, because the chips in a cluster exchange data constantly.
Scale-out connects racks across the hall and is already optical. Scale-up connects chips inside a rack so they act as one computer. It is still copper, it carries the most traffic, and replacing it is the prize.
Four steps, each shortening the copper
Plug-in modules
An optical unit plugs into the front of a switch. Proven and easy to replace. Today’s standard.
Linear plug-ins
The same module with its most power-hungry chip removed. Up to half the power; harder to tune.
Co-packaged optics
Optics move inside the switch. NVIDIA ships from H2 2026; Broadcom is already shipping.
Optical I/O
Optics sit beside the processor and replace copper inside the rack. Volume from 2028.
A caution from the industry itself. Pat Gelsinger, former Intel chief executive and now on Ayar Labs’ board, has said he declared the death of copper two decades too early. Copper will stay wherever it still works; the case for optics is strongest where copper has physically run out.
03 · The money and the market — NVIDIA is paying its suppliers to build
The clearest signal in any supply chain is a customer paying in advance. In three months this year, the company that sets the design of most AI data centres did exactly that across lasers, optical chips and glass.
What the money buys. These are not passive stakes. The Lumentum and Coherent deals carry multi-billion-dollar purchase commitments and rights to future laser capacity; Corning is building three US plants. The largest customer in the market is paying suppliers to build before it needs the capacity. That is what a real bottleneck looks like.
Broadcom is the open alternative to NVIDIA and already ships co-packaged switches. Marvell bought Celestial AI for about $3.25B upfront. AMD bought Enosemi and Ciena bought Nubis. TSMC supplies the packaging most designs depend on.
Revenue is in the proven parts: plug-in modules, lasers and fibre. Co-packaged optics and optical I/O are still a rounding error beside them. Their investment case rests on 2028 and after, not on current sales.
A warning on co-packaged optics forecasts. Published estimates for 2026 range from about $165M to over $300M, and ten-year forecasts from about $3B to more than $20B (Mordor Intelligence, Fortune Business Insights, IDTechEx). We treat the direction as better established than the size.
04 · Emerging deep technologies — what comes after today’s optics
Today’s optical chips are made of silicon and lit by conventional lasers. Both are approaching limits of their own. Behind them is a second wave of technologies at very different stages of readiness.
When each technology is expected to matter in AI data centres
Co-packaged switches. Optical circuit switching. Hollow-core fibre.
Comb and quantum-dot lasers. Optical chip-to-chip links. Thin-film lithium niobate. Micro-LED links.
Computing with light. Quantum photonics.
Start dates are Singular Link’s judgement from analyst and company sources and show first meaningful deployment, not peak volume.
| Technology | What it does, in plain language | Who is building it | When |
|---|---|---|---|
| Optical circuit switching | Routes the light itself, with no conversion back to electricity at each switch | Google (in-house), Lumentum, Coherent; nEye, iPronics, Oriole | Now |
| Hollow-core fibre | Light travels through air inside the fibre, about half as fast again as through glass, so sites can sit further apart | Microsoft, with Corning and Heraeus | Now |
| Comb and quantum-dot lasers | One source produces many colours of light, so one fibre carries many times more data | Xscape, Scintil, Pilot Photonics; Quintessent, Ranovus | Next |
| Optical chip-to-chip links | Replaces the copper between processors inside a rack | Ayar Labs, Lightmatter, Marvell (Celestial) | Next |
| Thin-film lithium niobate | A material that switches light faster than silicon, at lower voltage | Hyperlight, Lightium, Rapid Photonics | Next |
| Micro-LED links | Arrays of tiny LEDs send many slow signals side by side, using far less power than lasers | Avicena | Next |
| Computing with light | Uses light to do the calculation, not only to move the data | Lightelligence, Neurophos, Akhetonics, Lumai | Later |
| Quantum photonics | Quantum computers built from particles of light | PsiQuantum, Xanadu | Later |
How to read this list. The nearer a technology is to an existing standard and an existing factory, the sooner it earns revenue. Optical switching and hollow-core fibre are being bought today. New materials and light sources slot into the next product generation. Computing with light asks customers to change how they build computers, which is why it has always been ten years away.
05 · Startups and supply chain — seventy startups, five tight links
Money has followed the bottleneck. Most optical startups will not reach volume on their own, and the supply chain they depend on is younger and more concentrated than the one that makes processors.
Celestial AI went to Marvell, Nubis to Ciena and Enosemi to AMD — each to a large chip or networking company, and each before meaningful revenue. Buyers are paying for technology and engineers they cannot build in time.
A manufacturer will not risk an unproven part in the middle of a boom, and a part cannot be proven until one does. The companies that have broken through have their customers as shareholders. It is the first thing we look for.
The optical supply chain — five links
Materials
Indium phosphide for lasers. Raw supply concentrated in China.
Lasers
The scarcest part. Capacity is being bought in advance.
Optical chips
Made in ordinary chip factories — their main advantage.
Packaging
Joining fibre to chip. Slow, costly, centred on TSMC.
Modules, fibre
Assembly in Asia; fibre from a few makers such as Corning.
A co-packaged switch needs far more laser power than a plug-in module, which is why NVIDIA’s deals include rights to future laser capacity. Attaching glass fibre to a chip at volume is still closer to watchmaking than to chip manufacturing.
Laser raw material and the largest module maker are both in China. In response, Lumentum, Coherent and Corning are all adding US capacity, partly with NVIDIA’s money. Capacity outside the existing centres is being actively sought.
06 · Company focus — Ayar Labs: light beside the processor
Ayar Labs in one paragraph. Founded in 2015 and based in San Jose, California, Ayar Labs makes the parts needed to replace copper with light directly beside a processor. It is one of three leading approaches to the scale-up network, and the one with the broadest set of chip makers among its investors.
What it makes. The optical chip (TeraPHY) sits in the same package as a processor and turns its signals into light at the source. It is made in standard chip factories. The light source (SuperNova) sits in a cooler, reachable part of the rack and feeds the chips through fibre, which answers the reliability concern about lasers.
The road to volume
$650M of new capital; engineering bases in Hsinchu and Bengaluru.
Products ready for volume production by the end of the year.
Expected customer volume, in line with the wider industry timetable.
Strategic investors include NVIDIA, AMD, Intel, MediaTek, Alchip and Wiwynn — competitors who rarely share a round. Financial investors include Neuberger Berman, ARK Invest, Insight Partners, 1789 Capital and the Qatar Investment Authority.
Analysts describe Ayar as further along than most peers but still to be proven at scale. Watch for qualification on schedule and a named processor shipping with the chip. What could go wrong: a large chip maker builds its own, or the ramp slips past 2029.
The Singular Link view. Ayar Labs is the most direct expression of this issue’s argument: if copper inside the rack is the limit, the answer is light beside the processor. Its position rests on having the industry’s main chip makers as shareholders at the same time. Our stance is constructive — the right problem, credible backers and a manufacturable design — but the company is unproven at scale and the valuation has moved ahead of revenue.
Singular Link has exposure to Ayar Labs. Commercial terms of that position are not disclosed in this publication. Facts are from public announcements, Reuters and independent analysts. Company names shown, including current portfolio positions, are illustrative of Singular Link’s thematic focus and not a recommendation.
07 · The Gulf view, risks & watch items — more compute from every megawatt
This section is Singular Link’s own interpretation. Photonics bears directly on the three things the Gulf brings to AI: energy, new-build capacity and patient capital.
Power limits how much AI capacity can be built. Optics that halve networking power return that electricity to computing — a higher yield on every megawatt planned.
Most Gulf capacity is still to be built, so it can be specified for co-packaged optics from the start. Being late to build is, for once, an advantage.
Lasers, materials and packaging are short and being relocated. They are a more achievable industrial entry point than a processor factory.
Gulf capital is already on the cap table. The Qatar Investment Authority is an investor in Ayar Labs. The leading optical companies are financed by their own customers and a small group of institutions; regional capital is welcome for the same reasons as in compute — patience, scale and a home market that will buy the product.
What could change the picture
Watch items for Issue 05
First volume shipments of NVIDIA’s Spectrum-X Photonics switches.
Lumentum and Coherent results — is laser capacity still what limits revenue?
The first named processor to ship with optics in the package.
Sources. Company announcements and filings (NVIDIA, Lumentum, Coherent, Corning, Marvell, Ayar Labs) to 30 September 2026; Reuters; CNBC; The Next Web; optics.org; Forbes; Data Center Dynamics; Converge Digest; Cignal AI Optical Components Report and Startup Tracker (July 2026); LightCounting (March 2026); market-size ranges from Mordor Intelligence, Fortune Business Insights and IDTechEx. Technology timing, risk ranking and the Gulf lens are Singular Link’s own. Charts marked illustrative or indexed are not measurements. Disclosure: Singular Link holds positions in this sector, including exposure to Ayar Labs.
This material is provided by Singular Link for informational purposes only. It is not investment advice, nor an offer to sell or a solicitation of an offer to buy any security or interest in any fund. Company names shown, including current portfolio positions, are illustrative of Singular Link’s thematic focus and should not be taken as a recommendation, a complete list of holdings, or an indication of future portfolio composition. Terms of private portfolio positions are not disclosed in this publication. Past performance is not indicative of future results.