Photonics is the AI infrastructure trade now - and 2027 is when it gets decided

13 August 2026 · 12 min read

The short version: as at the close on 12 August 2026, optical and networking is the highest-scoring of the six AI silicon segments we track - median market sentiment score of 82 against 66 for the board - and it holds five of the nine highest scores on a 34-name board. It has been above the board median for eight consecutive sessions. That is the measurement. The forward view, which is judgement rather than measurement, is that 2027 is the year the value inside photonics moves - away from the module that plugs into a switch faceplate and toward the two things nobody can quickly build more of: the light source, and the packaging capacity to put optics next to silicon.

Everything below separates those two things explicitly. Where a number comes from our own data it says so; where a view is an opinion it is labelled as one.

Why photonics is load-bearing in AI infrastructure

An AI cluster is thousands of accelerators that must behave like a single machine. Every training step synchronises state across all of them, and an accelerator waiting on data is the most expensive idle object in the building. Performance is therefore set as much by the fabric as by the compute.

Copper cannot carry that fabric past a certain size. Signal degradation rises with data rate and distance, so at the speeds current accelerators need, passive copper is good for a couple of metres. Inside a rack, fine. The moment a workload spans racks, rows or buildings, the link has to be light.

That is the whole structural argument, and it is worth stating in its strongest form: optics is not a component of the AI buildout, it is the constraint that decides how large a single coherent cluster can be. You cannot buy your way out of a fabric bottleneck with more GPUs - adding accelerators makes the synchronisation problem worse.

What our board actually shows

Median market sentiment score and median five-day price move by segment, as at the close on 12 August 2026. The score is built from price and volume only - momentum, relative strength, trend, volume, volatility and drawdown - so it describes behaviour, not opinion, and can be rebuilt from a daily bar series by anyone who wants to check it.

Two columns are doing the work here, and they say different things.

The score column says optical leads. Lumentum at 87 and Arista at 86 are the two highest readings on the entire board; Coherent at 83, Credo at 81 and Ciena at 79 take three more of the top nine. The only optical name outside that cluster is Marvell at 63, and we will come back to it because it is the most interesting number in the segment.

The spread column says something less comfortable for a bullish story. Optical's 24-point internal spread is the joint-tightest of the six, which means the names are moving together - and a segment whose constituents move together is being traded as a theme, not as six businesses. Neoclouds, by contrast, span 50 points, which is what genuine discrimination looks like.

The climb, session by session

Median score for the six optical names against the median for all 34, every session since 27 July:

The segment has closed above the board median on every session from 3 August onward - eight in a row - after sitting below it for most of the preceding week. Measured from the 29 July low, where five of the six scored in the teens or below, the price moves to 12 August are large: Coherent +60.2%, Lumentum +54.8%, Credo +51.1%, Arista +33.3%, Marvell +32.9%, Ciena +30.8%.

Two weeks is not a trend, and the board as a whole rose over the same fortnight. What distinguishes optical is not that it rose but that it has stayed above the median every session since the turn, which momentum alone does not usually manage.

This is not the same rebound as the rest of the board

The distinction that matters is how far each group is from its own high, because that is what separates a bounce from a re-rating. Percentage of 52-week high, as at 12 August:

• Arista Networks (ANET) - 98.0%
• Credo Technology (CRDO) - 86.9%
• Lumentum (LITE) - 85.9%
• Coherent (COHR) - 80.8%
• Ciena (CIEN) - 67.8%
• Marvell Technology (MRVL) - 65.8%

The group averages 80.9% of its 52-week high. The six neoclouds, over the identical window and after a larger price move, average 65%. Arista at 98% is effectively at its high; nothing in the neocloud group is close to that.

The drawdown component of the score reflects it directly. It is negative for four of the six optical names, but only mildly - Lumentum -0.67 and Credo -0.36 against a bound of 4.80 - and Arista's is positive at +2.79, one of the few positive drawdown readings anywhere on the board. That is a different shape of recovery from a group climbing out of a hole.

The number that argues against the theme: Marvell at 63

Marvell is the second-largest company in the segment at roughly $190 billion, it sells optical interconnect silicon into exactly this buildout, and it scores 19 points below its own segment's median and ranks 20th of 34. The components explain why without any narrative at all:

• Momentum 8.60 of a possible 12.00 - the only optical name not effectively at the ceiling
• Relative strength 0.75 of 9.60 - it barely outpaced the cap-weighted complex, which rose 2.56% over five sessions. Credo's reading is 9.60 of 9.60.
• Volume -1.54 - turnover below its own normal, while Lumentum posted +6.99
• Drawdown -4.33 against a bound of 4.80 - close to as negative as that input goes

Read plainly: the market bought optical exposure in August and did not buy Marvell with it. We have written about this name separately, and the score has no view on why - it reads price and volume, and it does not know what was announced about anything. But it is a useful corrective to "photonics is working". Two of the six largest names in the segment are at two-thirds of their 52-week highs, and one of them is the pure optical-silicon play.

Where the value sits in 2027 - the view

Everything above is measurement. What follows is judgement, and should be read as such: it is not derived from the score, the score has no forward-looking component, and it will be wrong in places.

1. The bottleneck migrates from the module to the light source

The pluggable transceiver is what most people mean by "optics", and it is the layer with the most competition and the most Chinese supply. The layer that cannot be scaled quickly is the one that makes the light: indium phosphide lasers, the substrates they are grown on, and the fabs that process them. Lead times there are measured in years, not quarters.

The clearest evidence that the industry agrees is that Nvidia put $4 billion into it - $2 billion each into Coherent and Lumentum in March 2026, alongside multi-billion-dollar purchase commitments and future capacity access, with Lumentum committing to a new fabrication facility and Coherent expanding its US footprint. A company with Nvidia's balance sheet does not buy equity in its suppliers to secure a component it could simply order. It does it when the component is the constraint.

The August 2026 results from both companies are consistent with that. Lumentum reported pump-laser shipments up more than 80% year over year and described them as effectively sold out. Coherent said it had doubled internal indium phosphide output a quarter ahead of plan and expects to more than double it again by the end of calendar 2027. Sold-out capacity in the least substitutable layer of a supply chain is where pricing power lives - and a supplier racing to double the same capacity twice is telling you where it thinks the constraint will still be in eighteen months.

2. Co-packaged optics is real in 2027 - and it will be additive, not substitutional

Co-packaged optics moves the optical engine next to the switch silicon instead of into a module at the faceplate: shorter electrical path, much lower power per bit. Broadcom's 51.2T Bailly is in the market with a claimed ~70% power reduction against conventional transceivers, and Nvidia has begun shipping its next-generation Spectrum-X CPO switch to selected partners, built with TSMC's COUPE packaging. TrendForce puts the volume ramp in 2027-2028.

The widely held version of this is that CPO displaces pluggables and the module makers get disintermediated. Our view is that this is roughly two years early, and here is the mechanical reason.

CPO does not compete with pluggables for demand in 2027. It competes with accelerators for 2.5D/3D advanced packaging capacity - the same queue, at the same foundries, that produces the AI chips themselves. When a constrained resource is contested by a switch and by the highest-margin processor in the industry, the switch waits. TrendForce names exactly this alongside optical-engine yield and silicon photonics wafer capacity as the binding constraints on the ramp.

Meanwhile pluggable demand is not softening, and the best evidence is not a market forecast but what the suppliers said in the same week this article was written. On 12 August 2026 Coherent reported fiscal Q4 revenue of $2.05 billion and guided fiscal Q1 2027 to $2.2-2.4 billion, describing the 1.6T ramp as pulling in ahead of schedule through 2027. Lumentum said its own 1.6T ramp was accelerating and its pump lasers were effectively sold out. LightCounting's published expectation is that co-packaged optics in scale-up networks deploys across 2026-2027 and only reaches high volume in 2028. A product category whose current generation is supply-constrained is not one about to be cannibalised by a technology queueing behind Nvidia for packaging slots.

The strongest argument against us comes from the same call. Coherent's management said there was absolutely no push-out of CPO demand and expects co-packaged optics revenue to begin contributing in its fiscal second quarter, driven by ultra-high-power continuous-wave laser production ramping in Texas and Sweden. That is a supplier with sight of the order book saying the ramp is on schedule, and it is better evidence than our reasoning about packaging queues. We would note only that "CPO revenue begins contributing" and "CPO displaces pluggables" are separated by several years of volume, and that the second is the claim the equities are being priced on.

The call, stated so it can be marked: through 2027 CPO takes a visible but minority share of new switch ports, pluggable module revenue still grows, and the substitution argument becomes financially material in 2028-2029 rather than next year. If CPO exceeds roughly a quarter of new AI switch ports during calendar 2027, we were wrong.

3. The most under-appreciated line is optical circuit switching

Transceivers make a link faster. Optical circuit switches change the topology - reconfiguring which accelerators are connected to which, in light, without electrical conversion. That is a structurally different product with a structurally different margin, and it is arriving now. In their August 2026 results Lumentum guided to its first triple-digit-million OCS revenue quarter, and Coherent raised its own estimate of the addressable market to above $4 billion, expecting revenue growth across fiscal 2027. Cignal AI, forecasting independently, puts the market above $2.5 billion by 2029.

Three things make this interesting rather than merely new. It is sold largely to a handful of hyperscalers, so revenue arrives in lumps big enough to move a mid-cap's quarter. The two credible suppliers use different physics - Lumentum MEMS-based, Coherent liquid-crystal-on-silicon - so the winner is not yet determined by scale. And unlike a transceiver, an OCS is not made obsolete by the next port-speed generation, because it is not carrying the signal, it is routing it. In a segment where every other product line has a two-year technology clock, that is a materially different asset.

4. What actually breaks the thesis

Not a demand miss. The realistic failure mode for photonics equities in 2027 is that everything the bulls expect happens and the multiple compresses anyway, because the segment enters 2027 having already re-rated on precisely this argument. Lumentum was up 123% year to date going into its August results and rose again on them. It now carries a $72.6 billion market capitalisation against Coherent's $69.6 billion - on roughly half the quarterly revenue ($1.01 billion against $2.05 billion in the June quarter). That is not a valuation call, which we do not publish. It is an observation that the market has already decided which of the two is the purer expression of this thesis, and that the easy part of the re-rating is behind rather than ahead.

The second risk is concentration. Nvidia is simultaneously customer, shareholder and, through its own CPO switches, competitor to parts of the chain it has invested in. That is a good position for Nvidia. It is a less good position for a supplier negotiating its next contract.

The leading photonics companies, and what each actually sells

"Photonics" spans several layers that behave differently. Live scores are as at 12 August 2026 and will have moved.

Broadcom is on the board but not in the optical segment - it is classified with AI chips, because the great majority of its business is not optics. It matters here anyway: it is the most advanced merchant supplier of co-packaged switching, and its score of 52 is a reminder that a company can lead a technology transition without that transition being large enough to move its stock.

Outside the 34 names we track, the photonics universe extends to smaller pure-plays such as Applied Optoelectronics and POET Technologies, and to the private and non-US supply chain - InP substrate suppliers, Chinese module manufacturers, and the foundries doing silicon photonics wafers. We do not price those, so we do not comment on them.

The crowd noticed exactly one of these companies

We count cashtag posts on X ourselves, one row per complete New York day. Optical's share of the board's posts went from 4.2% on 3 August to 9.9% on 12 August, on a day the whole board carried 32,161 posts - which reads like the theme arriving in the retail conversation.

Look at the two days either side of the move and the sequence is unflattering. On 11 August, the day Lumentum reported after the close, Lumentum alone drew 1,592 posts while four of the other five optical names got quieter - Coherent 543 to 302, Marvell 352 to 229, Credo 186 to 144, and Ciena on 30 posts out of 23,227 across the board. The conversation was about one company, not about optics.

On 12 August, the session the prices actually moved, the attention broadened - and Lumentum's own count fell:

• Coherent 302 → 996 posts
• Ciena 30 → 93
• Marvell 229 → 423
• Credo 144 → 233
• Arista 101 → 121
• Lumentum 1,592 → 1,311

That is attention arriving with the move rather than ahead of it, which is the ordinary case. We measured the general question across the whole board and found post volume has a rank correlation of -0.00 with the direction of the next day's move - the working is here. Mention counts measure how much a name is being discussed, and carry no direction at all.

The narrower and more defensible reading is about the theme's visibility. A week ago this segment was outperforming almost unnoticed, at 6.4% of posts for the week of 3-7 August against a median five-day move nearly twice that of AI chips. It is no longer unnoticed. Whether that matters at all is a separate question, and our own answer to it is no.

One caveat on these figures, since it applies to any number we publish: the crawl refetches its whole window on every run, so a completed day's count can be revised after we first see it. The 11 August totals quoted above moved slightly between our first read and this one. That is the mechanism working, not failing - but it is why we date every count rather than describing it as "the latest".

Markers to judge 2027 against

Forecasts are cheap. These are the specific, checkable things that would confirm or break the view above:

1. CPO share of new AI switch ports during calendar 2027. Above roughly a quarter and the substitution case is running ahead of our expectation; in the low single digits and the pluggable ramp has further to run than consensus assumes.
2. Advanced packaging allocation. Whether foundries expand 2.5D/3D capacity fast enough that CPO stops competing with accelerators for it. This is the single most load-bearing variable and it is a capacity question, not a demand one.
3. InP capacity landing early. Coherent has said it will more than double indium phosphide output again by the end of calendar 2027, and Lumentum is building a new fab. Those are the direct test of the laser constraint. If capacity arrives ahead of schedule, the pricing power at that layer is shorter-lived than current multiples imply - a good operational outcome that is a poor equity one.
4. Whether 1.6T stays supply-constrained. Sold-out lasers and an accelerating ramp are what pricing power looks like at this layer. The quarter the suppliers stop saying "sold out" is the quarter module pricing turns.
5. OCS revenue becoming recurring rather than lumpy. A second and third consecutive triple-digit-million quarter at either supplier would make this a business line rather than a set of large orders.
6. Whether the segment's internal spread widens. Our own data answers this weekly. A 24-point spread says the market is buying a theme. Photonics becoming a real sector means that number gets larger, not smaller.

The limits of all this

• The score is price-only and backward-looking. It contains no forecast, no news and no social data. Nothing in the 2027 section above comes out of it, and nothing in it validates the 2027 section.
• Two weeks is not a trend. Our sentiment history began on 7 August 2026, with earlier sessions reconstructed from daily bars, so the series here is short by construction.
• Market moves and published results are not the same claim. Lumentum and Coherent both reported around 11-12 August 2026, and both rose sharply. We can source what was reported and we can measure what the prices did. We are not going to assert which caused which - most financial coverage of a move is written after it and about it.
• Third-party forecasts are forecasts. The TrendForce, LightCounting and Cignal AI figures cited here are cited as what those firms published, not as facts about 2027.
• None of this is a recommendation, and it takes no account of anyone's circumstances.

The optical and networking segment page carries the live scores and history for all six names, and the daily board has all 34, rebuilt every session.

Frequently asked questions

Sources

• ChipSentiment - optical and networking segment - the six constituents, live prices and cap-weighted sentiment history
• ChipSentiment - live board - prices, one-day and five-day moves and scores for all 34 names, quoted as at the close on 12 August 2026
• ChipSentiment - how the sentiment score is calculated - the six components, their weights and their bounds
• ChipSentiment - X mentions board - daily cashtag post counts, counted by this site against the X API; figures quoted for 11 and 12 August 2026
• ChipSentiment - do X mentions predict stock moves? - the -0.00 rank correlation between post volume and next-day direction
• optics.org - Nvidia backs Lumentum and Coherent with $4bn cash investment - $2bn into each company, March 2026, purchase commitments, new Lumentum fab and Coherent US expansion
• TrendForce - Nvidia and Broadcom begin volume ramp of CPO switches - Spectrum-X CPO shipping to select partners, Broadcom 51.2T Bailly and its ~70% power reduction, 2027-2028 volume ramp, optical engine yield, silicon photonics and 2.5D/3D packaging contention
• 24/7 Wall St - Lumentum and Coherent optics earnings, 12 August 2026 - Lumentum fiscal Q4 revenue $1.01bn, pump lasers 'effectively sold out', first expected triple-digit OCS revenue quarter, fiscal Q1 2027 guidance, 123% year-to-date gain
• Investing.com - Coherent fiscal Q4 2026 earnings call transcript, 12 August 2026 - revenue $2.05bn, fiscal Q1 2027 guidance $2.2-2.4bn, 1.6T ramp pulled in, indium phosphide capacity doubled a quarter early and to more than double again by end of calendar 2027, OCS addressable market above $4bn, 'absolutely no push-out of CPO demand'
• Cignal AI - optical circuit switching market to exceed $2.5bn in 2029 - OCS market sizing
• LightCounting - scale-up networks in AI clusters - AI optical transceiver market growth and expected timing of LPO and CPO deployment in scale-up networks

Figures are taken from the public filings and the reporting linked above.

Stocks mentioned

More from ChipSentiment

← All articles