Lumentum says it has no AI optical parts left to sell for roughly the next three years. CEO Michael Hurlston told Bloomberg on October 9, 2026, that the company's optoelectronic capacity is sold out to nearly 2029. Six months earlier his message was that 2028 would be gone within two quarters. The window has moved out by about a year.
This is a hardware story, but it explains something every builder feels: why the AI buildout is limited by far more than GPUs. This post covers what Lumentum makes, what the reported numbers say, what is still unconfirmed, and what it means for data-center plans.
TL;DR
| Question | Answer |
|---|---|
| Who said it? | Michael Hurlston, CEO of Lumentum, in a Bloomberg interview |
| When? | October 9, 2026 (some outlets date it October 8) |
| What was said? | Optoelectronic capacity is sold out to nearly 2029 |
| What changed? | In April 2026 the stated horizon was 2028; earlier it was the end of 2027 |
| Which parts? | Indium phosphide devices: EML chips, pump lasers, parts of 1.6T optical transceivers |
| How big is the gap? | Reported: EML supply more than 30% short; Lumentum 25 to 30% behind demand at full output |
| Who is buying? | Hyperscalers building AI data centers |
| Is it confirmed by filings? | We found no filing with the 2029 wording; it rests on the interview |
What Lumentum makes, in plain language
A modern AI cluster is thousands of GPUs that must talk to each other fast. Inside a rack, copper often works. Between racks, rows and buildings, the data travels as light. Light needs lasers, and the lasers need chips that can emit it cleanly at high speed.
Lumentum is one of the companies that makes those lasers and optical components. The sold-out items reported by the press are mostly built on indium phosphide, a compound semiconductor that, unlike silicon, can emit light efficiently. Three product groups come up:
- EML chips (electro-absorption modulated lasers). These sit inside optical transceivers, the pluggable modules at the end of a data-center cable. A faster transceiver, such as the 1.6T generation that is only starting to ramp, needs more of the fastest lasers.
- Pump lasers. These power optical amplifiers that boost signals over longer links, including links that connect data centers.
- Optical circuit switches. These steer light paths without converting to electrical signals; they are part of the broader optical product line.
Why can't a supplier just add a line? Compound-semiconductor fabs are not the same as silicon fabs. The epitaxy, yield and testing steps are slow to scale.
![]()
The timeline: 2027, then 2028, now 2029
The reason this item matters is the direction of travel. Earlier statements, as reported:
- Earlier in 2026: Lumentum said it was sold out at least until the end of 2027.
- April 2026: He said it would take only two quarters to sell out all of 2028 capacity. He also explained what the phrase means: non-cancelable agreements.
- October 9, 2026: Sold out to nearly 2029, according to Bloomberg and outlets repeating it. Some headlines say "early 2029."
Note the nuance in the April statement. It was a forecast of when 2028 would fill, not a claim that 2028 was already sold. The October claim is a stronger one: the horizon itself has now reached the end of the decade's first third.
The numbers reported
Coverage of the interview and the company's earlier results gives these figures. We could not open the Bloomberg article ourselves, so they come from outlets summarizing it, such as Startup Fortune. Treat them as reported figures.
- EML chip supply is more than 30% short of customer demand.
- Hurlston said Lumentum runs 25% to 30% behind the demand placed on it, even at full output.
- The pump laser business is sold out even after capacity expansion.
- Fiscal Q4 2026 revenue, reported in August, was $1.01 billion, up 109% year over year and 24% from the prior quarter.
- Lumentum joined the S&P 500 in March and the Nasdaq-100 in May.
Other outlets give different product-specific shortfall breakdowns that do not agree, so we do not rely on them.

Nvidia's $2 billion and why it matters
In March 2026, Nvidia invested $2 billion in Lumentum, reported to include multibillion-dollar purchase agreements for lasers and optical components, with money aimed at research, capacity expansion and US manufacturing. Rival Coherent received a similar investment around the same time. The logic is simple: if links are short, GPUs sit idle. A GPU maker that funds its suppliers is buying schedule certainty.
This pattern echoes what we have tracked elsewhere. Our post on the October 2026 AI chip supply chain shows TSMC records, Samsung trimming phone output and GlobalFoundries interposers all pointing at the same thing: AI demand is pulling capacity away from other products. Lumentum is the optical layer of that same story.
Why this is more than a stock story
Share prices are not our subject. The reader-relevant point is that the AI stack has a long chain of scarce parts, and each link has a different lead time:
| Layer | Typical constraint |
|---|---|
| GPUs and accelerators | Advanced packaging, HBM memory |
| Power and sites | Grid connections, local opposition |
| Networking light | Indium phosphide lasers, transceivers |
| Cooling and buildings | Construction, water, permits |
We have covered several of those layers. On power, see Google's 890 MW nuclear uprate deal and the San Francisco data-center moratorium. On scale, see China's 24 GW of compute versus 56 GW in the US. On policy money, see the Genesis Mission industry pledges. Light is simply one more bottleneck on the list.
What it could mean for people who build with AI
This is the part to be careful with. The reporting does not say that any named cluster is delayed. It does support several cautious inferences:
- Expect continued tightness in cluster supply. If optical links are scarce, the fastest build-outs favor buyers with long-term contracts. Smaller operators may wait or pay more.
- Compute prices are unlikely to fall quickly from optics alone. Compound-semiconductor capacity is slow to add, so relief is likely slow. This is an inference, not a reported forecast.
- Alternatives will get attention. Co-packaged optics, shorter copper links and different cluster layouts all reduce light-link demand. None is a quick fix.
- Local AI is unaffected in the short run. A deskside box has no long-haul optics. See our comparison of the Gigabyte GB300 workstation and DGX Spark.

What is claimed and what is verified
Claimed: Sold out to nearly 2029; EML supply more than 30% short; demand outrunning forecast. These come from an executive talking to a news outlet.
Supported by other sources: The earlier 2027 and 2028 horizons are documented in earlier coverage and call transcripts. The Nvidia investment and the August revenue figures were reported by multiple outlets.
Not confirmed: An exact contractual end date, the share of demand unmet in each product, and whether the 2029 horizon appears in an SEC filing. "Sold out" is also an executive's framing. A supplier may sell out by choice, by pricing contracts years ahead, or by rationing. It does not by itself reveal margins or customer names.
How to check: read the next earnings call transcript and the company's investor materials for the stated visibility horizon, and compare it against what Bloomberg quoted. Compare also Coherent's disclosures, since it makes similar parts.
What to watch next
Three signals will show whether the tightness is easing or deepening. First, the visibility horizon stated on the next earnings call: if it moves past 2029, demand is still outrunning capacity. Second, whether Coherent and other laser makers describe the same shortage, since a single supplier's claim is weaker evidence than a pattern across several. Third, whether transceiver makers talk about delayed 1.6T shipments, which would be the first visible sign that laser supply is limiting finished network hardware rather than just filling order books.
For builders, the practical step is to ask cloud and hardware vendors about network lead times, not only GPU availability. A cluster quote that names GPUs but is vague on optics deserves a follow-up question.
Risks and counterpoints
Two cautions apply. First, a bubble reading: some analysts have questioned whether AI spending might outrun real demand. Long-term sold-out claims can look different if buyers cut back, although non-cancelable terms limit that risk for the supplier. Second, supply responds to price. Higher prices bring new entrants and expansions, which could end shortages sooner than a supplier would like to say.
Related reading
- AI chip supply chain, October 2026
- China 24 GW compute versus US 56 GW
- Google's 890 MW nuclear uprates for AI data centers
- San Francisco data-center moratorium
- Genesis Mission industry pledges
- Nvidia DGX Spark price and memory cut
Sources
- Bloomberg: Nvidia-Backed Lumentum Sees Opto-Parts Capacity Sold Out to 2029 (October 9, 2026; paywalled, not read in full)
- Startup Fortune: Lumentum says its AI optical parts are sold out all the way to 2029
- Lumentum investor relations for earnings materials
Figures reflect public reporting as of October 9, 2026 and may change when Lumentum next reports. This is not investment advice.
