Everyone Wants Light. Almost Nobody Can Make Enough.

Palo Alto - October 5, 2026 - Every autumn, the optical industry gathers by the sea and announces that it has made light go faster. This year in Málaga, the better story was that light has become scarce. Speed is the scoreboard; supply is the game.

ECOC's exhibition ran September 21 to 23 at the FYCMA convention center, with Optica's Global Photonics Economic Forum following on the 24th and 25th. Speed headlines were easy to find. Marvell showcased 2nm technology driving 400G lanes, plus a 102.4T co-packaged optics platform. Coherent demonstrated a 3.2T pluggable built on eight 425G lanes. Impressive and expected. But notice what Coherent chose to headline: PhotonLink, a platform that sells components and finished optical assemblies, promising volume manufacturing. The pitch was not "faster." It was "available."

The real fight is about supply, not speed
The recurring debate was "slow and wide" versus "fast and narrow": many lanes at modest speeds from microLEDs or VCSEL arrays versus a few lanes at 200G or 400G PAM4. ams OSRAM CEO Aldo Kamper used his forum keynote to pitch wide and slow as the sweet spot between copper and fast-narrow optics: simpler, with no heavy serializing, and redundancy built in. FUNDA Research's ECOC notes counted Lumentum, Coherent, Credo, Marvell, Avicena, and even Nvidia's research team among those bringing hardware or concepts.

The stated argument is power. The unstated argument is materials. Fast, narrow lasers rely on indium phosphide, exactly what the industry can't get enough of. Lumentum and Coherent already run high-volume gallium-arsenide VCSEL capacity, outside that shortage, and Alfalume's Alexey Kovsh pitched quantum-dot lasers on gallium-arsenide wafers up to eight inches. When your suppliers are sold out, architecture becomes procurement strategy. Slow and wide is real engineering, but it is also a hedge, and the industry should say so out loud.

Skeptics have a point. Chris Cole, speaking in a personal capacity, argued that slow-and-wide volume packaging is four to six years away and called fast-and-narrow the only viable option for 400G. Others, like Peter Winzer of Ciena in his ECOC Plenary presentation, and Andy Bechtolsheim of Arista in multiple ECOC presentations, pointed out that 400G SerDes copper interconnect will be dominant within the rack, and therefore volume optics will follow suit. Many do not expect a named slow-and-wide deployment before the second half of 2027. A hedge isn't a replacement. But hedges are what you buy when the main position is sold out.

Hurlston and the sold-out sign
Lumentum CEO Michael Hurlston delivered the forum's afternoon keynote and was set to receive Optica's i4 Individual Achievement Prize from Málaga's mayor. It is a lovely irony: a celebration of a man whose chief problem is that he cannot make enough of what his customers want. His slides framed AI as pulling photonics inward through scale-across, scale-out, scale-up, and scale-in, then posed hard questions, including how to integrate indium phosphide with CMOS efficiently and how to scale VCSEL density without sacrificing yield, reliability, or power. Those are manufacturing questions, not physics questions.

His message on supply has been consistent all year and is not subtle. On the May earnings call, he put the EML imbalance at "somewhere greater than 30%," up from 25% to 30% a quarter earlier. At the RAISE Summit in Paris in July, he said telecom customers once ordered lasers by the hundreds, while Nvidia and the hyperscalers now want hundreds of millions, and that Lumentum is still shipping more than 30% below what customers want. He has said its newest indium phosphide fab in Greensboro, North Carolina, is about two years from production. In March, he told a television audience that Lumentum was sold out through the end of 2027.

Not everyone agrees. LightCounting's April forecast put transceiver demand about 30% above supply and expected the shortage to clear by the end of 2026. Hurlston is telling investors that his order book extends into 2027. Both can't be right. Scarcity hasn't hurt the business, to put it mildly: Lumentum reported $1.01 billion in fiscal fourth-quarter revenue and guided to $1.225 to $1.275 billion for the next quarter. Being unable to serve the whole market is the best problem in technology, right up until a customer finds a second source.

The pluggable fights back
While co-packaged optics grabbed the headlines, the surprise of the week was the plain old pluggable. Arista's Andy Bechtolsheim used his forum keynote to promote XPO, a liquid-cooled 12.8T pluggable module. His slides reported 150 member companies in the XPO MSA, with the 1.0 specification published on July 31 and production modules expected in the first quarter of 2027. The incumbent OSFP is already the highest-volume pluggable ever, with more than 120 million units shipping this year and 200 million projected for 2028. For a 400 MW AI datacenter, he claimed 75% fewer switch racks and 44% less floor space. That is a vendor's math, but it has 150 members behind it.

Lumentum's ECOC demo is telling: an eight-wavelength external laser module for CPO and NPO, delivering around 250 milliwatts per wavelength, with availability expected in the first half of 2027. When the laser is the scarce part, whoever sells the better light source holds the keys. My bet: co-packaged optics gets the keynote, pluggables keep the volume, and the winners are those who can ship either.

Four more bets worth watching
Salience Labs wants optical circuit switches to replace electronic packet switches in AI scale-up networks. It ran a live 64-port silicon-photonic switch at ECOC, and its architecture briefing argued that point-to-point optical paths avoid the two or three switch hops and the tail latency of electronic fabrics. The throughput and energy gains were simulated, so treat them as a pitch.

Resolight, an Israeli startup, goes further with an all-optical switch on thin-film lithium niobate. It says it switches in nanoseconds and can cut energy to one picojoule per bit. It has taped out a first chip but expects customer-lab products only by the end of 2027, with insertion loss and non-standard signaling still unresolved.

Seagate supplied the humbling data point. CTO John Morris showed lasers integrated into recording heads at the wafer level, produced at a rate of tens of millions per quarter. Complex photonics at volume already exists; it just spins inside a hard drive.

Keysight sells the rulers. CEO Satish Dhanasekaran pointed to the growing pile of MSAs, from LPO to XPO to Open CPX, and showcased a public 448G-class demo with Lumentum, NTT and Coherent. Whichever standard wins, someone has to test it.

The subtext nobody printed on a badge
Optica billed the forum as a response to a volatile geopolitical landscape, in which export controls, investment restrictions, and localization mandates are treated as routine. GenusTechnik's Michael Lebby promised an era of abundance in the 2030s, "as opposed to scarcity today," and handed his supply-chain panel a 90-day outage scenario to chew on. A shortage is physics plus policy, and neither yields quickly. Money does not compress a fab schedule. XPU roadmaps are already set for roughly the next two years, which pushes the earliest optics inside the package to late 2028 or 2029.

Takeaways from Málaga
Everything exciting inside the package is a 2028 story. The next two years belong to whoever can actually ship. Watch the gallium-arsenide alternatives, the smaller laser suppliers who could fill gaps in CW and pump lasers, and, at year-end, whether LightCounting or Hurlston turns out to be right.

In optics, the roadmap is cheap, and the supply is dear. Plan accordingly.