Bubble Wrap: A Cautious Guide to Photonics Valuations
By Oliver Campbell │ ICM HPQC Fund │ October 2026
“Written by humans, please don’t blame the robots for our typos”
After some fantastic news for Salience Labs, raising USD125m in its Series B, we were struck by some of the contrary responses on social media. One sticks in mind, “Oh Gosh! So many optical companies. Again”; a clear reference to the fibre optics companies of the 1999-2001 bubble.
We understand the reaction. Seen from the outside, or through a purely thematic lens, the proliferation of photonics companies, funding announcements and “AI-ready” press releases is bewildering. There are some bubbly ‘ouroboros’ similarities today, albeit more in AI software than in stack hardware.
This note is intended as a guide for those dazzled by the blooming of the sector, and perhaps new to it: how we sort the wheat from the chaff, and the ten checks we run before we allocate.
So many photonics companies? Not for us
There are many photonics companies because photonics is critical to the next generation of compute. Data movement and power per bit are now binding limits on AI clusters, and every layer of the stack that touches them has attracted founders and capital. That is why the sector looks crowded. It only looks that way to a sector or thematic investor, who counts logos rather than assess commercial potential.
From a bottom-up investment perspective however, that changes the count. We sort every company by its actual relevance to a part of the compute stack (co-packaged optics and optical I/O, optical circuit switching, laser sources, packaging and test), and then by the company’s own ability to scale: can it reach volume manufacturing, qualify with a chip vendor or hyperscaler, and last long enough to get there? Apply that filter and the field narrows quickly. There are only a few photonics companies at Series A to C worth allocating to. For us, certainly fewer than ten.
You only get that view if you are ‘in the trenches’. We are in Taiwan, where the foundry, packaging and test ecosystem determines what can actually be manufactured; in the Valley, where the buyers and architects sit; and in the labs and accelerators across Europe, where much of the underlying science and many of the earliest companies emerge. A press release will not tell you which start-up has a qualified process and which has a demo. These factors are critical in assessing commercial potential, and it is why a crowded-looking sector becomes a short list for us.
Most commentary that treats photonics as one undifferentiated bubble comes from people who have not done that sorting. They are comparing today’s market with the fibre optics boom of 1999, when almost anything with a laser in it was funded and most of it failed. But today’s market has different layers, economics and buyers. The ten checks below are the ones we apply.
Ten checks before we allocate
We group them into three questions: is the demand real and when does it arrive, is this the right company, and is the price right and the investment safe.
A. Is the demand real, and when does it arrive?
Check 1: Timing. We worry that we may be early. Earlier this year Pat Gelsinger, the ex-Intel CEO, said1, “I proudly declared that the death of copper was upon us, and everything would switch to optics, but I was just two decades too early. The GPU is sitting there sucking on a straw for more compute. Everyone wants their clusters to be bigger, but the physics of copper is limiting how big they can get”. It seems highly unlikely, but what if he is still two decades too early, copper keeps getting extended, and design wins slip by years?
How we weigh it: We treat timing as a question of when rather than whether, and we are investing in this window of time where compute is seeing enormous physical bottlenecks up and down the stack when. Gelsinger’s own quote highlights the physical constraint: the physics of copper limits how big clusters can get, and that hits harder with each accelerator generation. Photonic adoption does not need copper to die, only to run out of road at the margin. Adoption is already arriving in stages: Nvidia has launched co-packaged optics switches, and optical circuit switching has run in hyperscale networks for years. We then test the financials. We favour companies with enough cash to withstand delays, such as Salience following its Series B. If copper keeps stretching and stretching, they may book revenue later than planned, rather than not at all.
Check 2: Demand extrapolation. The 2000 parallel is also about extrapolation. Fibre was overbuilt because traffic forecasts were extrapolated into perpetuity, while the build was funded with debt and vendor financing. Today’s version is AI capex extrapolation, with circularity worries (chip vendors investing in their own customers) and debt-funded data centres. History doesn’t repeat, but it rhymes.
How we weigh it: That risk hits the whole AI stack, so we ask where a company sits when spending slows and what is bottlenecking the build. Fibre supply multiplied faster than demand as each strand carried more traffic. Today’s compute build is bottlenecked by power, grid connections, advanced packaging (among other constraints), none of which can be scaled at the speed of the 1999 fibre build. Interconnect is a small share of cluster bill of materials (BOM) cost and hyperscaler capex, but it determines how well the far more expensive accelerators and other parts are used. Photonics capex can reduce unit cash operating costs (and also depreciation) while boosting compute capacity, so the payback case is clear. We look for companies whose proposition is unit-cost reduction. We expect to be better defended in a capex digestion phase than businesses selling additional raw capacity, such as more GPUs.
Check 3: Efficiency eroding demand. Model and architecture efficiency could reduce the bandwidth needed per unit of useful compute, shrinking the market that the bottleneck argument assumes.
How we weigh it: We ask whether a company’s case rests on a “total addressable market” (TAM) narrative or on reducing the cost and energy required per bit, and we back the latter. So far, efficiency gains have been absorbed by larger models and larger clusters rather than reducing unit demand. If efficiency does dominate, growth may slow in the near term, but the ratio of data movement to compute tends to worsen as compute improves without technologies like photonics. Compute expansion reveals infrastructure bottlenecks rather than hiding them. (Click here for our comments on Jevon’s Paradox in ESG: Excel, Surveys and Gutfeel).
B. Is this the right company?
Check 4: Crowded sockets. Lightmatter, Ayar, Celestial, Xscape and others are all chasing roughly the same optical I/O socket, and the largest chip companies are spreading bets across multiple players. Most of these companies cannot win the same socket, and there will likely be a shakeout as standardisation matures.
How we weigh it: This is where our sorting does the most work. We map each company to its place in the stack: optical I/O chiplets, optical circuit switching, laser sources and packaging are different points with different competitors (Salience, for example, addresses optical switching rather than the XPU I/O socket). Within a layer we ask which companies can survive, which means who builds to emerging standards such as UCIe and NVLink Fusion and who has a credible path to volume. We expect a shakeout and think it is healthy. Consolidation is a likely exit route, with Marvell/Celestial providing a template. The failure mode differs from 1999: a failed photonics start-up strands R&D spend, not fibre in the ground.
Check 5: Technology and reliability. Co-packaged optics move the laser and optical engine inside the package, so a failure can take out far more than a failed pluggable module. Yield, thermals and field serviceability are why hyperscalers have been cautious, and they are more likely to slow adoption than any financing issue.
How we weigh it: This is probably the most credible technical objection. We look at whether the architecture addresses these concerns: external, replaceable light sources keep the least reliable component serviceable, and optical circuit switching sits outside the compute package altogether. Reliability is a moat for design wins, so we weigh channel checks, manufacturing partnerships and qualification milestones far above lab demonstrations.
Check 6: (Powerful) customer concentration. The customer base is concentrated, and customers could still bring their technology in-house: Nvidia, Broadcom and the hyperscalers are the gatekeepers, and Nvidia could build its own optical interconnects or license the technology. While it is theoretically possible, the (un)likelihood of building in-house is something we addressed in a previous article, “Competition, collaboration or collusion: implications for the stack”, where acquisition still feels like the preferred method to integrate photonic technologies into the stack.
How we weigh it: We read what the gatekeepers actually do, not only what they could do. Nvidia, AMD, Intel Capital, MediaTek and Alchip have all backed or partnered with photonics start-ups (Ayar has joined NVLink Fusion); that is not how a company behaves when it plans to build in-house. Nvidia’s $2B investments into Coherent and Lumentum, alongside its purchase commitments, suggests an instinct to secure external optical supply rather than replicate it. AMD, Broadcom, Marvell and the hyperscalers’ custom-silicon teams also have every incentive to keep the optical layer multi-vendor. We therefore look for strategic investors or design partners already at the table. For a high-quality start-up, having a few well-capitalised acquirers is a feature of the exit as much as a risk to revenue.
Check 7: Value capture and the incumbents. Pluggables, linear-drive optics and further copper improvements keep moving the power and cost goalposts, and the incumbents (Broadcom, Marvell, Coherent, Lumentum) are well placed to absorb the best ideas. Start-ups could prove the technology and still end up as low-margined component suppliers or become tuck-in acquisitions.
How we weigh it: We assume that for most it will be an acquisition, and we price accordingly. That’s a typical venture outcome, not a failure: Marvell’s purchase of Celestial AI followed exactly this path. Each extension of copper and pluggables can cost more power and money per bit, which strengthens the case for the photonic alternative and can raise what incumbents will pay to own it. We expect value to concentrate in the few companies that reach volume manufacturing and a design win, not to spread across every start-up. See our article on Fishing with rods, not nets.
C. Is the price right, and is the investment safe?
Check 8: Step-ups coming steep and fast. Ayar Labs raised $500M2 at a $3.75B valuation, after a $155M round at over $1bn roughly 16 months earlier. Lightmatter’s Series D was a 4X increase in valuation on a 2X increase in funding3. That looks like a narrative re-rating, not a revenue or profit one, even if the future looks brighter than before.
How we weigh it: We separate re-rating from de-risking by asking what the step-up is paying for. Ayar’s $500M is earmarked for volume manufacturing and test capacity, after a decade of R&D, with Nvidia and AMD in the round and a place in Nvidia’s NVLink Fusion ecosystem. Those are de-risking events, not just a re-rating. We also anchor to real transactions: Marvell paid $3.25B upfront for Celestial AI4, with earn-outs taking the potential total to $5.5B that only pay if revenue materialises. Salience follows the same pattern: a 64-port all-optical switch demonstrated, then capital to manufacture it. We want step-ups that track technical and commercial milestones. Revenue still has to follow.
Check 9: Circular financing. We look hard at who is funding whom. The buyers are somewhat different today, but there are echoes of the past. In 2000 many customers were debt-funded new telco-style “CLEC” companies (table below) with little to no cash flow and heavily indebted balance sheets. Today they are hyperscalers funding capex largely from operating cash flow and with still healthy balance sheets, for the most part.
There are some similarities in the co-dependence of circular / vendor financing back then and today: Back then to maintain massive revenue growth, major hardware providers like Lucent Technologies, Nortel Networks, and Cisco Systems effectively functioned as unregulated commercial banks. They extended enormous lines of credit to CLEC (Competitive Local Exchange Carrier) startups to buy their switching gear and fiber components.
The vendor would then record immediate revenue, while the CLECs booked a massive long-term liability. When the CLECs failed to generate real customer revenue, large percentages of these vendor loan portfolios went uncollected, destroying the hardware giants alongside the carriers.
(Alcatel) Lucent working capital and implied vendor financing in the 1990s
There are similarities with the relationships between fabless companies, hyperscalers and AI companies today, although we aren’t at a non-performing loans (NPL)/default stage yet.
How we weigh it: The Lucent charts show where the 1999-2001 stress sat: in vendor receivables and working capital, on balance sheets that funded customers with no cash flow. Today’s buyers are still profitable, largely investment-grade companies, but capex is absorbing most of hyperscalers’ operating cash flow and debt and equity issuance has picked up. This is why we watch this risk particularly closely. The circularity sits higher up, where GPU vendors back their own customers. In photonics, strategic investors buy equity or capacity rather than extending customer credit, so there is no loan book to go uncollected, and start-up losses fall on equity investors. We continue to watch vendor financing, off-balance-sheet leases and hyperscaler credit spreads, and we stay far away from AI software.
Check 10: The public-market anchor. We use listed incumbents with real revenue ramps as an anchor for private valuations. But those incumbents have also re-rated on the AI trade, so the anchor may be correlated with the thing being tested, rather than providing an independent reference point.
How we weigh it: We use the anchor as evidence that real revenue exists in this layer, not as proof that private valuations are right. If AI sentiment turns, listed optics will fall alongside private marks. The difference is what sits underneath: listed incumbents have revenue, margins and, in Nvidia’s case, purchase commitments. A de-rating may compress their multiples rather than erasing the business, whereas in 2000 revenue itself collapsed as customers and counterparties failed. We also look beyond start-up revenue lines. Some anticipated revenues may never come to fruition, but their value can be found in lower unit opex at their acquirers. Those efficiency gains could provide a large buffer against valuation pressure.
So hubble bubble then?
So is it a bubble? Parts of the market will probably behave like one. Valuations are running ahead of revenue in areas, too many companies are chasing overlapping sockets, and the timing of copper’s decline remains uncertain. What differs from 1999 is where the risk sits. Then, leverage lived with customers and vendors, and the losses were stranded in fibre in the ground. Today the exposure is mostly concentrated in equity, the buyers themselves are still cash-generative, and demand is pulled by the known physical constraints of copper, and the need to move data at tolerable power and cost per bit, rather than by a traffic forecast.
That does not make every photonics company a good investment. For those dazzled by the blooming of the sector, the lesson of the ten checks is that the number of photonics companies is the wrong number to look at. First sort by relevance to the stack, then by ability to scale, and then assess the timing, technology, customer and financing risks together. The list worth allocating to at Series A to C is short: for us, fewer than ten.
We expect a shakeout, with most of the value likely to accrue to the few that reach volume manufacturing and a design win. The test is unit economics, not narrative. We watch cost and energy per bit, the conversion of design wins into revenue, and hyperscaler cash flow and credit. If those fundamentals hold, “Oh Gosh! So many optical companies. Again” will look less like a repeat of 1999 and more like the early build-out of a genuine new layer of the compute stack, with a small number of companies emerging as winners.
Oliver Campbell
Investment Principal of the ICM HPQC Fund
MAS Licensed Representative, ICM Global Funds Pte Ltd
October 2026
Sources:
[5] Microsoft FY2026 10-K and annual report (SEC EDGAR), Alphabet Q2 2026 10-Q and 22 July 8-K (SEC EDGAR), Amazon Q2 2026 results release (30 July 2026) and 10-Q, Meta Q2 2026 results release (29 July 2026) and 10-Q, Oracle FY2026 10-K (year ended 31 May 2026).
Important Note:
The information in this article should not be considered an offer or solicitation to deal in the ICM HPQC Fund (Registration number T22VC0112B SF003) or ICM HPQC Fund 2 (Registration number T22VC0112B SF007) (the “Sub-funds”). The information is provided on a general basis for informational purposes only and is not to be relied upon as investment, legal, tax, or other advice. It does not take into account the investment objectives, financial situation, or particular needs of any specific investor. Investors should seek relevant professional advice before making any investment decision. The information presented has been obtained from sources believed to be reliable, but no representation or warranty is given or may be implied that it is accurate or complete. The Investment Manager reserves the right to amend the information contained herein at any time, without notice. Investments in the Sub-funds are subject to investment risks, including the possible loss of the principal amount invested. All forms of investments carry risks, including the risk of losing all of the invested amount. Investors should read the prospectus before deciding whether to acquire the units in the Sub-funds. The value of investments and the income derived therefrom may fall or rise. Past performance is not indicative of future performance. This document is intended solely for institutional investors and accredited investors as defined under the Securities and Futures Act 2001 of Singapore. The whole or any part of this work may not be reproduced, copied or transmitted or any of its contents disclosed to third parties without ICM Global Fund’s express written consent. This advertisement or publication has not been reviewed by the Monetary Authority of Singapore.
ICM HPQC Fund and ICM HPQC Fund 2 are registered Sub-Funds of the ICMGF VCC (the VCC), a variable capital company incorporated in the Republic of Singapore. The assets and liabilities of ICM HPQC Fund and ICM HPQC Fund 2 are segregated from other Sub-Funds of the VCC, in accordance with Section 29 of the VCC Act.
Get to know us better:
Comments