The session was led by Prashanta Kharel (Technology Strategist at Flexcompute, PhD in Physics from Yale, nonlinear/quantum optics background) with live demos run by Bohan Zhang. Kharel’s framing carries extra weight here: the harness demonstrated in the seminar is the same one described in Flexcompute’s own paper, “Autonomous agentic design for photonics” (Kharel, Khavasi, Chen & Hughes, arXiv:2606.00915) — a paper that was already sitting in our own related-work tracking for DSR-CRAG before this seminar happened. More on that below.

What PhotonForge and Tidy3D actually are

For readers newer to the stack: Tidy3D is Flexcompute's cloud, GPU-native multiphysics solver (FDTD, MODE, heat, charge, RF), with adjoint-based inverse design built in. PhotonForge sits a layer up — it's the circuit/schematic/layout environment that wraps Tidy3D with foundry PDK support (8+ processes spanning silicon, silicon nitride, and thin-film lithium niobate), parametric cell (PCell) construction, routing, and design-rule/LVS (Layout Versus Schematic) checking. Together they cover the full stack from "electromagnetic field in a waveguide" to "GDSII ready for a foundry." What turns that stack from a set of tools a human operates into a set of tools an LLM agent operates is a combination of Python APIs, an MCP interface into Flexcompute's documentation and examples (Flexcompute calls their MCP-based copilot FlexAgent, though it's worth noting the seminar itself didn't name the orchestration system on screen), and a library of curated reference designs the agent can draw on instead of starting from a blank file.

The framing question: can AI already design a full chip?

Kharel opened with a distinction worth keeping in mind whenever this topic comes up: designing a PIC and building a product are not the same task. Tape-out, measurement, iteration on measured data — all of that sits downstream of “design” and each step multiplies the complexity. Even design alone, restricted to the PIC, involves translating system-level specs into component-level targets, choosing or designing the components, running the trade-offs between them, and getting everything through DRC before it can go to a foundry.

He used the “jaggedness” framework — AI capability advancing unevenly, strong in some sub-tasks and weak in adjacent ones, gradually filling in — to describe where photonics sits today: past “helps me find papers” and “helps me write simulation scripts,” into “optimizes a single component,” and now edging into “coordinates a design across a whole chip given a human specification.” The remaining gaps he named directly: agent-to-agent coordination, tool access across old and new simulation tooling, cross-run memory (how do you transfer what one design run learned to the next agent that picks up the work?), and — repeatedly flagged as the hardest one — visual and geometric understanding, since so much of photonic design is “does this layout actually mean what I think it means?”.

Under the hood: the harness

The architecture shown is a blackboard-style setup: an orchestrator agent reads the human’s specification, plans out what kind of sub-system the chip needs, and spins up specialist sub-agents — a system designer, a passive-component designer, an active-component designer, a layout/verification agent — each targeting whatever the orchestrator has allocated to it. The orchestrator monitors the sub-agents as they work and can trade off targets between them on the fly (accept a slightly lossier grating coupler if the system-level budget still closes, for instance) — which is closer to how a human photonics team actually negotiates a design than to a single-objective optimizer.

Each specialist runs its own inner loop, and this is the part with the most engineering texture to it. The human supplies a program.md (the target, the metric, the constraints, the rules of engagement — written once and reused, not re-prompted every time) and a design.py (the only file the agent is allowed to edit, holding the initial geometry and the evaluation function). From there the loop is: edit geometry → verify (preview, DRC) → simulate (Tidy3D) → compare to the running best → keep or discard → repeat, with every experiment written to a persistent journal that becomes the agent’s long-term memory for that run. It’s worth being precise about what this isn’t: it isn’t gradient-descent inverse design with extra steps. The agent can change topology entirely between iterations — inverse-design one section, hand-sweep a taper elsewhere because that’s more physically intuitive there — which is a materially different search process than adjoint optimization, even though it can call adjoint optimization as one of its tools.

Demo 1: a four-port waveguide crossing, 30 iterations

The first live demo pointed a fresh agent at a broadband waveguide crossing on the SiEPIC PDK, optimizing for worst-case through-loss across the 1500–1600 nm window — scored at whichever wavelength in that window is worst, not the average, so there’s no credit for a device that’s excellent at 1550 nm and poor at the edges — footprint capped, and told to run 10–30 iterations unsupervised. It started from the “naive thing” — two bare strips crossing at right angles, no taper, no mode shaping — and iterated from there, at one point needing a human nudge to run early sweeps in 2D before committing to expensive 3D simulation (a rule that then got folded back into the harness rather than re-prompted each time).

The convergence data is worth quoting precisely because Flexcompute was explicit about not overstating it: at a loop-standard mesh of 12 steps per wavelength the design showed 0.046 dB worst-case in-band loss; re-run at a converged 20–24 steps/wavelength, the honest number settles at 0.067 dB, with 0.077 dB worst-case under a ±20 nm etch-bias sweep. That’s roughly an 18× improvement over the naive baseline, in a locked 20×20 µm footprint, across 30 FDTD experiments — and the mesh-convergence caveat is logged in the journal rather than left for someone to discover later. It’s a small detail, but it’s the kind of self-auditing habit that matters a lot more than the headline dB figure.

Upgrade to Paid

Read more

EDITORIAL NOTE

Business of Photonics runs on a weekly cadence, but the industry does not take August off, and neither do we. Even in a holiday period, when trading desks thin out and inboxes empty, the underlying business of photonics keeps moving — and this month it moved in a way worth a subscriber’s full attention.

On July 16, 2026, IPG Photonics — the company that effectively invented the high-power fiber laser industry — announced a binding offer to acquire Lumibird Medical for up to €350 million. Three weeks later, on August 4, it reported second-quarter results that told a more complicated story: an earnings beat wrapped around a $17.6 million impairment charge, a 21% year-over-year drop in net income, and a stock that still finished the week up double digits.

Taken together, the acquisition and the earnings print are the same story told twice. IPG is a 25-year-old industrial laser pioneer trying to out-manufacture a structural problem — Chinese competitors have commoditized the mid-power fiber laser business that IPG built — by buying its way into higher-margin, regulation-protected end markets. Lumibird Medical is the clearest evidence yet of how far that pivot has gone, and how much conviction management has in it. This edition is dedicated entirely to understanding both events as one continuous strategic move.

As always, the free section below gives the essential facts. The paid deep-dive goes into the mechanics of the deal, the financial picture behind it, and what both signal about where IPG — and, by extension, a meaningful slice of the industrial photonics sector — is headed.

Subscribe now

Share Precision with Light

Free Section — The Story at a Glance

The Deal

On July 16, 2026, IPG Photonics Corporation (NASDAQ: IPGP) announced a binding offer to acquire Lumibird Medical, the ophthalmology-focused medical laser division of France’s Lumibird SA, for €300 million in cash on a cash-free, debt-free basis, plus a contingent earnout of up to €50 million tied to 2026–2027 performance — a total consideration of up to €350 million, funded entirely from cash on hand. The transaction is expected to close in the fourth quarter of 2026, pending consultation with Lumibird Medical’s French works council and customary closing conditions.

Lumibird Medical designs and manufactures diagnostic and therapeutic laser systems for ophthalmology — treating cataracts, glaucoma, dry eye, and age-related macular degeneration — under three established brands: Quantel Medical, Ellex, and Optotek Medical. Headquartered in France, the company operates three major global facilities with more than 450 employees. For fiscal year 2025, Lumibird Medical reported revenue of €112.2 million and EBITDA of €24.1 million — a 21.5% margin.

Why It Matters

IPG's existing medical business is built around urology lasers. Lumibird Medical's strength is ophthalmology. Combined, IPG expects the deal to roughly triple its medical revenue to approximately $204 million on a pro forma basis, lift Advanced Solutions from about 16% to 26% of total pro forma sales, and add roughly $1 billion to IPG's total addressable market. Management has stated the acquisition is expected to be accretive to gross margin, EBITDA, and adjusted earnings per share.

The Earnings Backdrop

Three weeks after the acquisition announcement, IPG reported second-quarter 2026 revenue of $278.6 million, up 11% year-over-year and its third consecutive quarter of double-digit growth. Adjusted EBITDA rose 54% to $48.5 million, and adjusted EPS of $0.58 beat consensus estimates of $0.39 by a wide margin. GAAP net income, however, fell 21% year-over-year to $5.2 million — weighed down by a $17.6 million impairment charge recorded in the quarter. The stock rallied more than 14% in the days following the report, driven primarily by the scale of the adjusted earnings beat and continued confidence in the medical growth story.

Analyst Reaction

Wall Street’s response to the two announcements has been consistently positive, if not uniform. Needham upgraded IPG to Buy from Hold with a $110 price target shortly after the acquisition news. Stifel reiterated its Buy rating at $125. Roth Capital raised its target to $151. By mid-August, Citi trimmed its target to $137 from $150 while maintaining a Buy rating, characterizing a post-earnings pullback as unwarranted given the company’s exposure to rising semiconductor-driven demand. The consensus rating across nine analysts remains Buy, with a mean price target implying substantial upside from current levels.

The in-depth analysis continues for paid subscribers.

★ PAID · Deal mechanics, the strategic logic behind IPG’s pivot away from commoditized industrial lasers, a close read of the Q2 financials, market reaction, three named risks, and the EWC / Precision with Light scope note.

★ Paid Section — The Deep Dive

Subscribe now

Read more

Read more

Editorial Note

What EPIC’s monthly New Product Release webinars do well is compress the photonics industry into a single hour. It does not look like a conference agenda, a trade show floor — just five companies, each given ten minutes to show what they have actually built, in front of an audience of peers who know the physics well enough to ask the hard questions.

The July 7, 2026 edition was notable for a reason that became apparent only when you’ve looked at the five presentations together. Arden Photonics measured specialty optical fiber with nanometre-level precision. EPIGAP OSA delivered high-power broadband LEDs for hyperspectral sensing. Lightnovo miniaturised Raman spectroscopy into a handheld instrument backed by peer-reviewed photonics science. Keysight Technologies closed the simulation gap between photonic circuit design and system-level electrical validation. And Wave Photonics demonstrated a live platform connecting PIC designers to foundries, EDA tools, testing, and packaging in a single coordinated workflow.

Taken individually, each is a product announcement. Taken together, they describe something more useful: a photonics value chain that is maturing simultaneously at every layer — from the fiber that carries the light, to the emitter that generates it, to the spectrometer that analyses it, to the software that designs the chips that route it, to the platform that manages the manufacturing journey from design intent to packaged die. That simultaneity is the editorial observation that justifies covering all five in a single article.

This publication covers the event in two tiers. The free section below provides a clear description of each company and product for readers who want orientation. The paid section goes inside the technical and commercial substance of each presentation, including what was shown in the live platform demo and what it means for the broader photonic design tool ecosystem.

Hosted by Ivan Nikitski, Photonics Technology Expert at EPIC.

Upgrade to paid

Share Precision with Light

Event Overview

Arden Photonics (UK) · Fiber Geometry Measurement

Arden Photonics, which describes itself as an optical fiber measurement specialist, presented the FGC-GH — a new fiber geometry measurement system launched on June 29, 2026 and designed specifically for the growing category of specialty fibers: hollow-core fiber (HCF), multi-core fiber (MCF), polarisation-maintaining fiber (PMF), and active fibers. Presented by Alistair Robinson, Product Manager, the FGC-GH addresses the gap left by conventional fiber geometry instruments, which were designed for standard single and multimode fiber and do not have the optical resolution or software architecture to handle NANF (Nested Anti-resonant Nodeless Fiber) and DNANF (Double-Nested Anti-resonant Nodeless Fiber) structures.

The system accepts fiber cladding diameters from 80 µm to 250 µm and delivers sub-40 nm measurement repeatability across core diameter, core-to-cladding concentricity, and capillary diameter. A plug-in software architecture allows dedicated measurement routines to be loaded for each fiber type, producing automatic, user-independent results without requiring the instrument to be returned for reconfiguration.

EPIGAP OSA Photonics (Germany) · High Power NIR + SWIR LEDs

EPIGAP OSA Photonics GmbH, formed in 2023 through the merger of two Berlin-based semiconductor companies with roots in the research group that produced the first LEDs in East Germany in 1952, presented the OCI-460 product family under the theme of High Power LEDs for Hyperspectral Imaging. Presented by Franziska Nieder, the OCI-460 covers 700–1700 nm in an industry-standard 3535 SMD package (3.5 × 3.5 mm footprint) with up to 1 W output power and a 125° viewing angle. The company’s broader portfolio spans UV to SWIR from 225 to 2300 nm, with SMD, Chip-on-Board, and TO packaging options.

The application case presented was food quality control — specifically, non-destructive determination of sugar content in potatoes using NIR absorption spectroscopy, with the glucose and fructose absorption peaks clearly resolved in the 1100–1450 nm spectral window. The core commercial argument: replace halogen lamps in hyperspectral systems with targeted LEDs to eliminate the need for additional filters, reduce heat generation, extend operating lifetime, and simplify PCB assembly.

Keysight Technologies (USA) · End-to-End EOE Simulation

Philippe Bouchilloux, Business Development Principal at Keysight, presented the company’s ADS 2026 advancement for AI data centre and optical interconnect design: a unified Electrical-Optical-Electrical (EOE) simulation capability that allows engineers to simulate the complete signal path — from the transmitter’s electrical domain through the optical and photonic circuit to the receiver’s electrical domain — within a single design environment.

The context was clear: as optical interconnects push toward 800 Gbps and 1.6 Tbps per lane, the engineering teams designing them can no longer afford to run electrical and optical simulations separately and manually stitch the results together. The fragmented workflow — separate mode solvers, 3D simulators, circuit simulators, and system simulation tools — introduces errors, delays, and blind spots at exactly the interfaces where the most complex physical effects occur. ADS 2026’s EOE capability, demonstrated through the ADS System Designer for Ethernet, closes this gap with cross-domain simulation, multi-wavelength WDM analysis, and full-duplex optical channel modelling in a single coherent workflow.

Lightnovo (Denmark) · Miniaturised Raman Spectroscopy

Konstantinos Stergiou, Head of Engineering and Manufacturing at Lightnovo ApS, presented under the company’s mission statement: Raman for All — democratising high-end Raman spectroscopy for research, industry, defence, and field deployment. Lightnovo’s product portfolio spans from the miniRaman, which at 112 × 39 × 34 mm and 400 g is among the world’s smallest research-grade Raman spectrometers, through the RG and RG PRO benchtop instruments, to Raman microscopes at both miniature and full laboratory scale.

The technical differentiators presented were a built-in reference channel — a patented approach that self-calibrates the instrument at every acquisition by continuously correcting for laser wavelength drift, mode hops, and temperature variations — and a demagnification spectrometer architecture (described as “light compression”) that achieves high optical throughput in a compact form factor while maintaining high spectral resolution. These innovations are grounded in peer-reviewed science: the core miniaturisation strategy was published in Nature Communications in 2024. Excitation wavelengths span 405 nm, 532 nm, 638 nm, 785 nm, and 830 nm, with tailored configurations for laser power, spectral range, and resolution.

Wave Photonics (UK) · PDK and MPW Management Platform

Matthew Anderson, Co-founder and CSO of Wave Photonics, presented the company’s PDK Development and Management Platform with a live demonstration — the only live product demo of the session. The platform positions itself as the interface layer between PIC designers and the full photonics supply chain: foundries, EDA tools, testing, and packaging. It manages three kit types — Process Design Kits (PDKs), Assembly Design Kits (ADKs), and Test Design Kits (TDKs) — and delivers them through integrations with major EDA environments.

The live demo showed the platform’s PDK Hub (currently listing Cardiff Fab, PhotonFab with 9 PDKs, and Wave Photonics’ own PDKs, with TFLN (Thin-film Lithium Niobate) and two SiN processes as coming soon), the MPW Runs Hub with active Ge-on-Si (Germanium-on-Silicon) fabrication runs from PhotonFab, automated DRC checking running in the background against the fabrication process, and a foundry-facing approval workflow with GDS/OAS approval and rejection actions, timeline tracking, and email notifications. Partners span the full stack: Siemens, Cadence, Synopsys, LUCEDA Photonics, GDS Factory, and Latitude Design Systems for EDA integration (with Keysight coming soon), Cornerstone for fabrication (with AMF, SiPhonIC Technologies, TFLN, and SiN processes coming), ALTER and PhiX for packaging, and CSA Catapult and Wave Photonics for testing.

Paid Section — In-Depth Analysis

Upgrade to Paid

Share Precision with Light

Read more

On April 22, 2026, Marvell Technology announced the acquisition of Polariton Technologies, a high-speed photonics company spun out of ETH Zurich in 2019. The deal — financial terms undisclosed — added a plasmonics-based electro-optic modulator platform to Marvell’s optical portfolio. It was the company’s third photonics-adjacent acquisition in five months.

The first was XConn Technologies, acquired for approximately $540 million in January, filling the PCIe and CXL switching layer. The second was Celestial AI, acquired for $3.5 billion in February, for its optical interconnect between chiplets technology. The Polariton acquisition is the most technically revealing of the three. XConn and Celestial AI extend Marvell’s position in the interconnect stack at the system and chiplet levels. Polariton reaches into device physics — into the material layer where the actual modulation of light happens. That is a different kind of bet, and understanding it requires understanding why silicon’s native modulation mechanism is approaching a scaling ceiling toward 3.2T (Terabits per second (Tbps)) data rates and beyond.

XConn and Celestial AI extend Marvell's position in the interconnect stack. Polariton reaches into device physics — into the material layer where the actual modulation of light happens. That is a different kind of bet.

Upgrade to paid

The physics ceiling on silicon modulation

Silicon photonics modulators almost universally rely on the plasma dispersion effect. When free carriers are injected into or depleted from a silicon waveguide region under applied bias, the real and imaginary parts of the refractive index change. That refractive index change phase-shifts the optical signal, enabling data encoding in a Mach-Zehnder interferometer or a ring resonator configuration. The plasma dispersion effect is efficient enough to have built an industry — but it carries fundamental constraints that become binding as data rates scale.

The Vπ·L figure of merit — the product of the half-wave voltage and the device length required to achieve π phase shift — is constrained by silicon’s relatively weak carrier-dependent refractive index change. State-of-the-art depletion-mode silicon modulators achieve Vπ·L values of roughly 1–2 V·cm. Reducing this further requires aggressive lateral junction engineering with tighter implant profiles — which increases optical absorption from the doped regions and degrades the extinction ratio. More fundamentally, the carrier-based mechanism is not intrinsically fast. Depletion-mode MZMs reach 50–70 GHz with careful RF electrode design; pushing beyond this requires geometry trade-offs that compromise fabrication uniformity.

Energy per bit is where the constraint becomes financial rather than merely technical. A depletion-mode silicon MZM driving a PAM4 signal at 100 GBaud consumes roughly 1–3 pJ/bit at the modulator alone, before accounting for the driver, transimpedance amplifier, DSP, and thermal management. The energy-per-bit roadmap to sub-100 fJ/bit — where large-scale AI interconnect economics need to go — cannot be reached with silicon plasma dispersion alone. At 400G per lane, the commercially demonstrated frontier in 2026, these limits are already clearly in view.

What plasmonics offers — and what it costs

Polariton’s approach exploits surface plasmon polaritons — hybrid electromagnetic modes that exist at the interface between a metal and a dielectric. When confined at this interface, the electromagnetic field can be compressed to dimensions well below the diffraction limit of light, achieving very strong interaction between the optical mode and an electro-optic material embedded in the metallic slot gap. The specific architecture uses organic electro-optic materials — molecules with extremely high Pockels effect coefficients — embedded in a plasmonic slot waveguide.

The Pockels effect is a fundamentally different modulation mechanism from plasma dispersion: a direct, linear change in refractive index in response to an applied electric field, with no free carrier injection or depletion. It is intrinsically fast — the response is electronic rather than carrier-driven — and operates in a very short device length. The on-paper performance figures are compelling: Vπ·L products below 0.1 V·cm, bandwidth exceeding 200 GHz, and energy per bit approaching femtojoule scale. Device footprints measured in micrometres rather than millimetres. These results have been peer-reviewed and published by the ETH Zurich group over several years.

The engineering costs are real, however. Plasmonic waveguides carry very high propagation losses — metal absorbs light significantly — and the devices work because they are very short. Integration with low-loss silicon waveguides requires careful mode transition structures. Organic EO (Electro-Optic) materials have historically had elevated-temperature reliability concerns relevant to data centre environments. And there is no plasmonic silicon photonics PDK in production at any major foundry today. Polariton’s technology is a research platform approaching commercial qualification. This is precisely why Marvell acquired it rather than licensing it: bringing the team and IP in-house gives Marvell the ability to co-develop the PDK, qualify the process alongside its DSP and SerDes roadmap, and control the integration architecture across the package boundary.

Marvell’s five-month acquisition sequence

Read sequentially, the three acquisitions form a vertical integration strategy moving from the package boundary inward to the device level. XConn fills the electrical switching layer. Celestial AI fills the chiplet-to-chiplet optical I/O layer — Marvell’s $3.5 billion bet that the critical bottleneck is the electrical I/O between chiplets within a package. Polariton fills the modulation device physics layer — the question of what modulator technology can reach 3.2T and beyond with acceptable energy per bit once silicon plasma dispersion approaches its ceiling.

Together, the three acquisitions assemble a connected proprietary stack from switching fabric through optical engine to modulator device. For fabless photonic design companies that relied on Marvell’s DSP and SerDes as commodity components, the acquisition sequence signals that Marvell intends to compete across the optical connectivity value chain rather than within it. That is a structural shift in how the interconnect ecosystem is organised.

The analysis continues for paid subscribers.

★ PAID · What follows — device physics depth, tool architecture implications, three resolution scenarios, and the EWC strategic read — is for paid subscribers only.

Upgrade to paid

Read more

Edition #002. Business of Photonics is a two-tier weekly publication from Engineering World Company, the team behind Precision with Light. The free briefing maps the five most relevant developments across silicon photonics, AI interconnect, photonic EDA, and adjacent sectors — with a candid note on what each means for hardware and dedicated software dev…

Read more

On March 26, 2026, GlobalFoundries filed simultaneous lawsuits against Tower Semiconductor at the U.S. International Trade Commission and the U.S. District Court for the Western District of Texas. The complaints allege that Tower has infringed eleven of GF’s manufacturing process patents — covering analog, radio-frequency, and silicon photonics fabrication technologies — and seek both an import ban on the allegedly infringing chips and recovery of lost profits.

Tower’s response was brief and firm: it would defend itself vigorously in court.

On its face, this looks like a standard inter-company IP (Intellectual Property) dispute of the kind that punctuates semiconductor history at regular intervals. It is not. To understand why, you have to look at when it was filed, who filed it, who received it, and what both companies have been racing to build for the past eighteen months. When you do, a different picture emerges: the first open engagement in what will likely become a sustained contest over who controls the process IP substrate of the silicon photonics supply chain — a supply chain that is, in 2026, moving from prototype-grade to industrial scale at extraordinary speed.

"The first open engagement in what will likely become a sustained contest over who controls the process IP substrate of the silicon photonics supply chain."— Business of Photonics #001

Upgrade to Paid

Two specialty foundries on a collision course

GlobalFoundries and Tower Semiconductor occupy adjacent terrain in the foundry landscape. Neither competes at the leading edge — they are not building 2nm logic. Both have built their competitive identities around specialty process technologies: RF (Radio-frequency), analog mixed-signal, silicon germanium, and, increasingly, silicon photonics. For years, this adjacency was commercially comfortable. The SiPho market was small enough that two modestly-sized specialty foundries could coexist without meaningful conflict. That period ended sometime in 2024.

Tower entered 2026 as a company in the middle of a dramatic SiPho expansion. Its silicon photonics revenue reached approximately $52 million in Q3 2025 alone — roughly 70% growth year-on-year — with full-year 2025 revenue targeting $220 million and a Q4 annualised run rate exceeding $320 million. To support this, Tower committed $650 million to triple its SiPho manufacturing capacity by mid-2026, operating 200mm fabs across California, Texas, and Israel, plus a 300mm facility in Japan. By early 2026, Tower’s stock had risen 67% year-to-date, bringing its valuation to approximately $24 billion.

GlobalFoundries entered 2026 having made arguably the most consequential single strategic move in the SiPho foundry space: the November 2025 acquisition of Advanced Micro Foundry (AMF), the Singapore-based pure-play silicon photonics fab widely considered the most capable independent SiPho manufacturer outside the United States. The acquisition gave GF a claim to being the largest dedicated SiPho foundry globally, and positioned its GF Fotonix PDK as the de facto platform for fabless photonics companies seeking a China-free supply chain — a procurement requirement that has become quietly mandatory for U.S. hyperscaler sourcing teams navigating export control and supply chain resilience mandates.

These are two companies growing at uncommon speed, pursuing the same customers, in a market that analysts are projecting to reach $1 billion at GF’s own foundry level within two years. The conditions for conflict were structural. What the March 2026 filing represents is not a surprise but a prepared move — executed at a calculated moment in a race that both companies had been running for over a year.

The analysis continues for paid subscribers.

★ PAID · What follows — process patent mechanics, the dual-venue legal strategy, timing analysis, supply chain exposure, three resolution scenarios, and the EWC strategic read — is for paid subscribers only.

Upgrade to paid

Read more

Welcome to the inaugural edition. Business of Photonics is now a two-tier weekly publication from Engineering World Company, the team behind Precision with Light. Each week, the free briefing maps the five most relevant developments across the silicon photonics, AI interconnect, and photonic EDA landscape — with a candid note on what each means for hard…

Read more

Defence Optics Swiss Photonics Precision Manufacturing

01 — Context

A Note on Swissphotonics Lunchchats

Before the analysis, a word about the source — because the context matters.

Swissphotonics is Switzerland’s national photonics network, and the Lunch Chats it organises are something rare in the photonics industry: genuinely informal, technically honest, community-driven gatherings where practitioners from Swiss photonics companies present actual work to actual peers, without the performance marketing layer that accompanies trade conferences. They happen weekly at noon, over Zoom, and they are open to everyone globally. I attend regularly, and I find them consistently among the most useful single hours in my photonics week — not for headline announcements, but for the quality of what gets said when a presenter is talking to colleagues rather than a press room.

The June 23, 2026 Lunch Chat featured Dr. Martin Forrer, CTO and Senior VP Business Development at FISBA AG, presenting on the company’s positioning in the defence precision optics market under the title: Go Beyond the Visible.

Upgrade to paid

Share Precision with Light

02 — Company

Seventy Years From Textiles to the F-35 Cockpit

Since 1957, FISBA has evolved from a family-led heritage rooted in St. Gallen’s textile legacy into a globally recognised innovator in optics. The same city, the same precision manufacturing culture, the same obsession with tolerances and quality management — the textile mills of St. Gallen produced some of the finest embroidered fabric in Europe for centuries. FISBA applied that same culture to glass.

Dr. Martin Forrer, who presented at the Lunch Chat, brings a profile that is characteristic of the best senior technologists in European deeptech: a physics degree from ETH Zürich, doctoral research at the Institute for Applied Physics at the University of Bern, and an IMD Business School Master of Technology Enterprise (2006) — deep physics training combined with formal business education. He has led R&D at FISBA for 13 years and assumed the combined CTO and Senior VP Business Development role in March 2025.

03 — Track Record

The F-35 Reference: What Serial Production at Scale Actually Means

For more than twelve years, FISBA has been delivering up to 2,000 imaging lenses per year for the F-35 programme — both helmet lenses and flight lenses. The specific components include F/D1.2 and F/D1.3 objectives with aspherical elements and active coma compensation.

“For over a decade, FISBA produces parts for the F-35 programme — up to 2,000 imaging lenses per year, including F/D1.2 and F/D1.3 objectives with aspherical elements and active coma compensation.”

— Dr. Martin Forrer, FISBA at Swissphotonics Lunch Chat presentation, 23 June 2026

Two thousand precision aspheric lens assemblies per year, every year, for twelve years. That is a cumulative delivered quantity exceeding 24,000 units, each meeting Lockheed Martin’s qualification standards for a programme where the cost of a defective component is measured in mission outcomes. The F-35 helmet alone costs approximately half a million dollars per unit — and the imaging performance of the helmet display system depends directly on the optical quality and thermal compensation behaviour of FISBA’s components.

The analysis below — the commodity bifurcation thesis, the VIS/SWIR strategic reading, and the Swiss-EU programme exclusion — is available to paid subscribers of Precision with Light.

Subscribe to Precision with Light →

Share Precision with Light

Read more

Share Precision with Light

Upgrade to Paid

One Date, Two Press Releases, One Industry

June 16, 2026 produced two photonics announcements that, filed separately in most trade coverage, look like routine business news. Read together, they describe the same underlying problem attacked from opposite ends of the value chain.

Keysight Technologies announced it had completed the acquisition of VPIphotonics, adding system-level simulation to its photonic design automation portfolio and enabling optical and electrical engineers to advance designs from component to complete link within a single environment. On the same day, Coherent Corp signed a letter of intent to receive up to $50 million in direct CHIPS Act funding from the US Department of Commerce to expand its 6-inch Indium Phosphide manufacturing facility in Sherman, Texas — timed to a groundbreaking ceremony attended by NVIDIA, federal officials, and Coherent leadership.

One story is about software. The other is about wafers. Both are about the same constraint: the AI infrastructure buildout is moving faster than the tools and the factories that photonics has historically relied on, and the companies best positioned right now are the ones racing to remove whichever bottleneck sits closest to their own business.

Keysight’s Move: Completing the Stack, Not Just Adding a Feature

To understand why the VPIphotonics acquisition matters more than a routine bolt-on, it helps to know what Keysight already owned. RSoft handles device-level simulation for waveguides, gratings, modulators, and laser sources. Photonic Designer delivers comprehensive circuit-level design and verification for photonic integrated circuits. What was missing was the system level — the ability to simulate a complete optical link, end to end, electrical to optical and back to electrical, and predict bit error rate performance before a single wafer is committed.

VPI Optical Link in Keysight ADS enables simulation of the full transceiver path in a single analysis — from electrical to optical and back to electrical — allowing engineers to predict how the whole link will perform, including bit error rate. That capability closes the gap. Keysight’s photonic design automation portfolio now reaches from device physics to full system behavior, in one connected workflow rather than three separate tools requiring manual handoffs between them.

The reasoning in Keysight’s own statement is worth quoting because it states the industry’s real problem plainly: as complexity rises and timelines tighten, engineering teams need simulation that carries a design across every domain without the manual handoffs that slow development and introduce error. Nilesh Kamdar, General Manager of Keysight EDA, was specific about why this matters now rather than five years ago: photonics design complexity continues to increase for customers working at speeds greater than 1 THz, and having a complete suite of tools addressing device-to-system challenges is now crucial.

That single phrase — “speeds greater than 1 THz” — is the tell. This acquisition is not aimed at today’s 800G and 1.6T transceiver designs. It is aimed at the generation after next, where the manual handoff between a device-level tool and a system-level tool becomes not just inconvenient but a source of design errors that nobody can afford at the cost of a modern AI infrastructure tape-out.

Share Precision with Light

Upgrade to Paid

Read more

Returning to the Product Page

This is the inaugural edition of a section that existed under a previous name and is now returning under the Precision with Light brand. Photonics Precision Technologies is where the publication focuses on something that gets less attention than it deserves in the photonics media landscape: the actual products. Not just the …

Read more

The Announcement That Changes the AI Infrastructure Map

This morning, from Taipei at COMPUTEX — timed deliberately alongside Jensen Huang’s keynote — Lightmatter announced it has joined the NVIDIA NVLink Fusion ecosystem.

Read that sentence carefully, because it matters more than it appears. NVLink is not Spectrum-X. NVLink is not a data center fabric switch. NVLink is the scale-up interconnect that binds GPUs to each other directly — the copper backbone of NVIDIA’s GB200 NVL72 systems delivering over one petabit per second of bandwidth within a single rack.

Until today, NVLink has been entirely copper. Through this collaboration, Lightmatter will deliver Co-Packaged Optics (CPO) and Near-Packaged Optics (NPO) products compatible with NVIDIA’s optical and SerDes technologies. That is photonics inside the scale-up interconnect — not just the scale-out fabric. The boundary between electrical and optical in AI computing infrastructure has just moved.

Lightmatter’s groundbreaking Passage® platform — the world’s first 3D-stacked silicon photonics engine — and Guide® — the industry’s first VLSP™ light engine — connect thousands to millions of processors. The Passage M1000, their current reference platform, sits as an active 3D interposer directly beneath the GPU — 4,000mm², 1,024 serial channels at 56 Gbps each, delivering over 57 terabits of photonic bandwidth. The commercial product timeline is late 2027 announcement, 2028 availability.

The 650 Group analyst reading (which appears in the Lightmatter Press Release above) of this is worth quoting directly: “By making its Passage 3D photonic roadmap compatible with NVIDIA’s high-speed interconnect, Lightmatter is significantly expanding the addressable market for its CPO products. This partnership provides a validated blueprint for hyperscalers to overcome traditional I/O bottlenecks and scale AI clusters to the levels required for next-generation intelligence.”

That is the business statement. We will come back to what it actually means after the second story.

Subscribe now

Share Precision with Light

The Other Side of the World: A$100M Sensing Startup Raises A$25M

On May 15, Arkeus — an Australian sensing company — raised A$25 million in a Series A funding round to accelerate the development and global deployment of its AI-powered sensing systems for autonomous platforms. The round was led by QIC Ventures, with participation from R+ VC, Folklore Ventures and DYNE Ventures, alongside continued support from Main Sequence, Salus Ventures and Beaten Zone Venture Partners.

Arkeus builds AI-powered sensing systems that act as the “eyes and brain” of autonomous platforms, enabling machines to see, understand and act in real time — even in degraded, denied or disrupted environments. Arkeus produce and deploy wide-area search, intelligence, surveillance and reconnaissance capabilities across air, land and maritime domains.

The valuation — A$100 million, implied by the Forbes coverage — makes this a pre-revenue or early-revenue defence technology company being priced on roadmap and market positioning rather than current cash flows. The investor mix is instructive: QIC Ventures is a Queensland government-linked vehicle, Main Sequence is CSIRO’s deep-tech fund, and Beaten Zone is a specialist defence venture firm. This is sovereign capability investment, not pure commercial speculation.

The capital will be used to accelerate global expansion, establish advanced manufacturing capability in Queensland and the United States, and deliver Arkeus systems into operational use faster across allied defence markets. “Allied defence markets” means Five Eyes — Australia, the US, the UK, Canada, New Zealand — which is the commercial aperture that makes an Australian defence startup internationally viable rather than domestically constrained.

The hyperspectral imaging angle, covered in the Forbes article, is the technology that connects this to the broader photonics story. Hyperspectral sensors capture dozens to hundreds of narrow spectral bands simultaneously — not the three broad colour channels of a camera, but a full spectral fingerprint of every pixel in the image. The combination of hyperspectral imaging with AI inference at the edge — onboard the drone, in real time, without a communications link back to a human analyst — is precisely what Arkeus is building.

Subscribe now

Share Precision with Light

Read more

The Acquisition Nobody Saw Coming — But Everyone Should Have

Last week, Light Conversion — a laser company headquartered in Vilnius, Lithuania, founded in 1992 as a direct spinout of research led by Professor Algis Petras Piskarskas at Vilnius University — announced the acquisition of Class 5 Photonics GmbH, a Hamburg-based ultrafast laser manufacturer that itself originated as a spinout from DESY, Germany’s national particle accelerator centre, and GSI, the heavy-ion research facility in Darmstadt.

No acquisition price was disclosed. Both companies are privately held. The industry press covered it briefly, noted the complementary technologies, and moved on.

This deal deserves a close reading — not because of its size, but because of what it reveals about the only business model that has consistently produced durable, technically credible companies in the ultrafast laser segment. The cooperation between the Lithuanian and German laser manufacturers is set to strengthen their market position as suppliers of advanced ultrafast technologies, connecting two companies shaped by similar scientific backgrounds and technology-driven approaches to business.

“Technology-driven approaches to business.” That phrase, from Light Conversion’s own press release, is either a platitude or a precise description of a specific strategic philosophy. In this case it is the latter — and understanding what it actually means separates the companies that have lasted in ultrafast photonics from the ones that have not.

Subscribe now

Share Precision with Light

What Each Company Built

Light Conversion’s flagship product family is PHAROS — a femtosecond laser platform that has become the workhorse of scientific research groups worldwide — alongside TOPAS, an optical parametric amplifier that converts the PHAROS output to tunable wavelengths across the visible and infrared. Light Conversion was among the first companies to commercialise OPCPA-based technologies, introducing OPCPA (optical parametric chirped-pulse amplifiers) concepts into its OPA platforms in the early 2000s before launching complete OPCPA systems in 2014. The SYLOS systems built for the ELI-ALPS attosecond science facility in Hungary — some of the most powerful few-cycle laser systems ever constructed — came from Light Conversion in partnership with Ekspla.

Class 5 Photonics came from a different starting point but arrived at adjacent territory. After developing high-power OPCPA systems for extreme-ultraviolet synchrotron and free-electron laser applications, the founding team established the company in 2014 in Hamburg. Over more than a decade, Class 5 Photonics built expertise in the most powerful tunable OPCPA systems, ultrafast mid-infrared lasers and coherent EUV sources, including systems developed for attosecond science, ultrafast spectroscopy, and semiconductor-related EUV applications.

The overlap in OPCPA expertise is the obvious rationale for the deal. The non-obvious part is the Class 5 phrase that deserves to be read twice: “semiconductor-related EUV applications.” We will return to that.

The Nobel Prize Business Thesis

Awarded the Nobel Prize in Physics in 2018, chirped-pulse amplification has fundamentally transformed the field of ultrafast laser science and accelerated the broader adoption of these technologies in both research and industry over the past decade. Several years later, attosecond pulse generation was also recognised with the Nobel Prize, further increasing the demand for more powerful few-cycle lasers operating at longer wavelengths.

Two Nobel Prizes in ultrafast laser physics within five years. The 2018 prize — to Donna Strickland and Gérard Mourou for chirped-pulse amplification — validated and commercialised an entire technology family. The 2023 prize — to Pierre Agostini, Ferenc Krausz, and Anne L’Huillier for attosecond pulse generation — pointed to the next frontier and, in doing so, created demand for more capable systems.

In this context, OPCPA technology, driven by emerging 100-watt-class laser systems, has become an important approach for applications including attosecond science, high-field physics, particle acceleration, ultrafast spectroscopy, and bio-imaging.

This is the business logic hiding in plain sight: the Nobel Committee, with unusual consistency, has been validating the technology roadmap that Light Conversion and Class 5 Photonics have been building along for thirty years. That is either remarkable prescience or confirmation that both companies have been genuinely at the frontier — and the distinction matters enormously for how you evaluate the business.

Read more

Subscribe now

Share Precision with Light

The Scaling Wall Nobody Had Fully Named

Programmable integrated photonic circuits — the silicon photonic meshes of Mach-Zehnder interferometers that implement arbitrary optical matrix transformations in hardware — have been one of the defining technological stories of this publication since the founding series. The Capmany group at iTEAM, Universitat Politècnica de València, and their iPronics spinout introduced the Photonic Field Programmable Photonic Gate Array (PFPGA) concept to this corpus. The non-uniform hexagonal mesh paper — the most recent addition to the platform’s silicon photonics research stack — demonstrated that defect cells and Vernier engineering could extend the spectral and temporal programmability of these meshes beyond what uniform architectures allow.

Both of those contributions addressed what the chip can do. The paper published to arXiv on May 19, 2026 — from Pablo Martínez-Carrasco Romero, Andrés Macho-Ortiz, José Roberto Rausell-Campo, Francisco Javier Fraile-Peláez, and José Capmany — addresses something more fundamental: how much energy it costs to do it, and whether that cost is irreducible or merely the consequence of a mathematical choice nobody had thought to question.

The answer, it turns out, is the latter. And the route to a better answer runs through geometry that Euler identified in the eighteenth century.

The Problem: Phase Shifters Are Eating the Power Budget

Every programmable photonic mesh — whether hexagonal, square, triangular, or feedforward — is built from the same elementary unit: a tunable interferometric building block (BB) consisting of two 50:50 beam splitters and one or more phase shifters. The beam splitters are passive and consume no power. The phase shifters are active — in silicon photonics, they are typically thermo-optic elements that heat a section of waveguide to change its refractive index, or plasma dispersion modulators that inject or deplete carriers. Both mechanisms consume electrical power proportional to the phase shift they apply.

For a small mesh — a 4×4 unitary, perhaps 12 building blocks — the aggregate phase-shifter power is manageable. For a 20×20 unitary, which requires approximately 400 building blocks in a Clements decomposition, the aggregate becomes significant. For the 64×64 switches and the large-scale optical neural network accelerators on the roadmaps of multiple groups, the cumulative power of thousands of simultaneously actuated phase shifters is becoming a hard constraint on chip-level power density and system-level thermal management.

The field’s response to this problem has followed two directions. The first is device-level: improve the efficiency of individual phase shifters — lower the Vπ·L product of electro-optic modulators, improve the thermal isolation of thermo-optic heaters, explore emerging phase-change materials like GST that retain their state without continuous power. The second is system-level: prune less relevant building blocks from the mesh (not every BB contributes equally to a given transformation), or redistribute phase allocation across the mesh using information-theoretic principles to reduce average per-BB phase requirements.

Both approaches have produced meaningful improvements. But both share a common assumption that the paper challenges: that for a given building block implementing a specific 2×2 unitary transformation, the energy cost of that transformation is fixed once the transformation is specified.

The paper’s central insight is that this assumption is wrong.

Read more

The Number Everyone Reported, and the Story Behind It

On May 18, 2026, POET Technologies (NASDAQ: POET) closed a $400 million registered direct offering to a single institutional investor — 19,047,620 common shares and an equal number of warrants at a combined price of $21.00, a premium over POET’s NASDAQ closing price of $20.57 on May 14. The warrants are exercisable at $26.25 for three years — a 25% premium over the offering price — which means the anonymous institutional investor structured the deal with a clear view on where POET’s share price is heading.

Most coverage stopped there. The $400 million headline is large enough to generate attention without requiring further analysis.

But the more interesting story closed four days earlier, and almost nobody led with it.

On May 14 — the same day that became the reference price for the offering — POET announced a supply and development agreement with Lumilens, a Silicon Valley photonics startup founded by veterans of Cisco, Juniper Networks, Meta, Lumentum, and Coherent, which is building next-generation photonic interconnects for AI infrastructure. The agreement includes an initial purchase order of $50 million and a framework for cumulative purchases of up to $500 million. The $400 million raise and the Lumilens agreement are not coincidences occurring in the same week. They are two components of the same strategic move.

Share

What POET’s Optical Interposer Actually Does

Before the strategic analysis, the technology deserves a precise description — because “optical interposer” is a term that gets used loosely in the industry and POET’s implementation is specific.

POET’s photonic integration solutions are based on the POET Optical Interposer — a novel, patented platform that allows the seamless integration of electronic and photonic devices into a single chip using advanced wafer-level semiconductor manufacturing techniques. POET’s Optical Interposer-based products are lower cost, consume less power than comparable products, are smaller in size and are readily scalable to high production volumes.

The key word in that description is “wafer-level.” Most silicon photonic transceivers today are assembled from multiple discrete components — laser dies, modulator chips, photodetector arrays — bonded together with active alignment under microscopy, a process that is accurate but slow and expensive at production volumes. POET’s approach moves the assembly to the wafer level: all components are placed and bonded at the wafer scale before dicing, using the same batch processing infrastructure that makes semiconductor manufacturing economical. The result is a chip-scale optical engine that integrates the laser, modulator, and detector in a single package, without active alignment.

POET’s product portfolio addresses the full data rate roadmap: 800G, 1.6T and above optical engines and optical modules for AI clusters and hyperscale data centers. Beyond transceiver products, POET has designed novel light source products for chip-to-chip data communication within and between AI servers — what the industry calls “near-package optics” or “in-package optics” — targeting the next frontier of AI interconnect where the optical interface moves from the module cage to the chip package itself.

Multiple customer relationships with major manufacturing partners are already in place: Mitsubishi Electric, LuxshareTech, LITEON, Foxconn Interconnect Technology, and ADTRAN. These are not pilot programme relationships. They are production-oriented partnerships with companies that manufacture at scale.

Share

The Lumilens Deal: Why It Matters More Than the $400M

Lumilens is a company that most readers of this publication will not have encountered before this week. It deserves a close reading, because its positioning reveals precisely what POET is being funded to build.

Read more

The Problem Seldom Talked About Enough

Every photonic integrated circuit of any complexity has them. Wherever two waveguide routes need to cross on the same layer of a chip — and in any non-trivial PIC routing topology, they will — there is a waveguide crossing. It is the photonic equivalent of a road intersection: two optical highways meeting at a point, each needing to continue on its way without depositing traffic into the other lane.

The problem is deceptively simple to state and genuinely difficult to solve. At a waveguide crossing, four things happen simultaneously and all of them are bad. Power from Waveguide A leaks into Waveguide B — crosstalk, which corrupts the signal. Power from both waveguides radiates into the substrate — insertion loss, which wastes optical power. The crossing introduces a small but non-negligible reflection back toward the source. And all three effects are wavelength-dependent, meaning a crossing optimised at 1550nm performs differently at 1310nm or at 640nm.

In a small photonic circuit with a handful of crossings, these effects are manageable. In a large-scale PIC — a quantum photonic processor with hundreds of Mach-Zehnder interferometers, a co-packaged optics switch with dozens of wavelength channels, a programmable photonic mesh with a full hexagonal routing fabric — crossings accumulate. A crossing with −0.5 dB insertion loss, unremarkable in isolation, contributes −50 dB of cumulative loss across a cascade of one hundred crossings. At that scale, crossing performance is not a second-order concern. It is a fundamental constraint on how large and how functional a photonic integrated circuit can be.

A paper published in Journal of Lightwave Technology in March 2026 — from Haoran Wang, Yan Fan, Liu Li, Ziyang Xiong, Zhichao Ye, Hao Deng, Shihua Chen, Tong Lin, Junpeng Lu, and Zhenhua Ni — addresses the waveguide crossing problem on the silicon nitride platform with a combination of inverse design methodology and fabrication-robust optimisation that sets new performance benchmarks across two separate wavelength bands simultaneously. The results are, by any reasonable measure, remarkable.

Why Silicon Nitride — Again

Regular readers of this publication will recognise Si₃N₄ (Silicon Nitride) as a recurring protagonist. The quantum photonic processor corpus — the QuiX Quantum 8×8 and 20-mode processors — is built on silicon nitride specifically because two-photon absorption eliminates silicon-on-insulator as a viable platform for single-photon quantum circuits. The Nature Communications inverse design paper from October 2025, already in the platform’s research corpus, demonstrated 1,200× footprint reductions in Si₃N₄ devices. The photonic tensor processor paper covered in the first standalone Research & Technology post used a Si₃N₄ microcomb as its light source.

The reason silicon nitride keeps appearing is structural. It is the platform that spans the widest useful wavelength range of any CMOS-compatible photonic material — from approximately 400nm in the visible through the entire telecom band (O through L, 1260–1625nm) and into the short-wave infrared beyond 2µm. Silicon-on-insulator becomes absorbing below approximately 1100nm, ruling it out for visible applications. Lithium niobate on insulator has excellent electro-optic properties but higher propagation loss and more demanding fabrication. Silicon nitride has no two-photon absorption at any wavelength relevant to classical or quantum photonics, propagation losses approaching 0.1 dB/m in the best demonstrated platforms, and full compatibility with standard CMOS photolithography.

That last point is what makes the Wang et al. paper’s result meaningful beyond its benchmark numbers. The crossing device was fabricated with standard photolithography at a 200nm minimum feature size — not electron-beam lithography, not deep-UV immersion at cutting-edge nodes, but a process accessible at any reasonably equipped Si₃N₄ foundry. Record performance with accessible fabrication is the combination that converts a laboratory demonstration into a component that can appear in production PICs.

Read more

What the Founding Series Actually Was

Four posts. Four months. One argument built in stages.

Post 1 established the existence proof: six research papers, each describing a different instance of the same abstract design problem, collectively making the case that a generative inverse design platform for photonics was not just possible but necessary.

Post 2 made the technical case for inverse design as a categorical capability shift — not a faster version of the same workflow, but a workflow that did not previously exist.

Post 3 situated the platform at the intersection of three converging forces: the data center bandwidth crisis driving co-packaged optics deployment at scale, the programmable photonic mesh enabling software-defined silicon photonic hardware, and the quantum photonic processor scaling from 8 to 20 modes and beyond.

Post 4 addressed the hardest question directly: why should you trust an AI-generated photonic design? The answer required being honest about where black-box machine learning fails in physical design contexts, and specific about how physics-informed neural networks with solver-verified outputs change that calculus.

Together they built a single argument: the photonics industry is at an inflection point where the design software layer has become the rate-limiting constraint, and physics-informed generative AI is the only credible architecture for removing that constraint without sacrificing the physical rigour that fabrication demands.

That argument is now complete. This post is about the platform it justifies.

Subscribe now

The Architecture, Stated Simply

Precision with Light is a four-layer platform. Each layer has a specific job, and the layers are deliberately separated so that each can be improved independently without breaking the others.

Layer 1 — The Intent Layer translates what you want into what the physics allows. A natural language design brief — “I need a hollow-core fiber for gas sensing at 1550nm with less than 0.01 dB/m loss and an overlap factor above 0.95” — enters the DSR-CRAG (DSR-CRAG: A Double-Stream Corrective Retrieval-Augmented Generation Framework for Countless Intent Detection) system, which retrieves the relevant physical constraints from the database, checks the request against fabrication design rules, and produces a structured physics constraint document before any geometry is generated. The constraint check happens first. Always.

Layer 2 — The Generative Engine synthesises geometries from constraints. Physics-Informed Neural Networks have internalised Maxwell’s equations — they cannot produce an effective index above the core material’s refractive index, because doing so would increase the physics loss term during training. Conditional Wasserstein GANs explore the feasible design space and produce candidate geometries that satisfy the constraint document from Layer 1. The engine operates in milliseconds per candidate. A complete design exploration that would take weeks of FEM (Finite Element Method) simulation takes minutes.

Layer 3 — The Simulation Bridge verifies every AI-generated geometry against an independent solver before it leaves the platform. Lumerical FDTD and FDE (Finite Difference Eigenmode) for full three-dimensional device characterisation. COMSOL Multiphysics for fiber cross-section mode solving and thermal analysis. Tidy3D — Flexcompute’s cloud-native GPU-accelerated FDTD — as the preferred backend for silicon photonic components, accessible via API without local installation. Every design that reaches a user carries a solver-verified certificate: the simulated result matched the AI prediction within the specified fidelity threshold. If it did not, the platform flags the discrepancy and requests additional training data in that region of design space.

Layer 4 — The Fabrication Export converts verified designs into files that foundries and fabrication facilities can use directly. GDSII layouts for silicon photonics MPW runs at AIM Photonics, IMEC iSiPP50G, and GlobalFoundries GF Fotonix. STL meshes for two-photon polymerisation printing. Fiber draw tower specifications for specialty fiber manufacturers. OpenEPDA-compatible PDK ingestion — the standardised photonic PDK format developed through the JePPIX consortium — so that foundry design rule updates propagate automatically into the constraint database without manual re-entry.

The platform does not produce interesting geometries. It produces geometries that can be built, in specific materials, on specific process nodes, with specific fabrication tolerances, verified by independent simulation.

Read more

The Wall at 1000°C

This publication has spent considerable time on photonic crystal fibers — their hollow-core variants guiding light through air with near-zero loss, their anti-resonant designs achieving 0.00088 dB/km confinement loss, their gas-filled configurations sensing trace molecules at parts-per-billion concentrations. All of that work has one thing in common: it is built on silica glass.

Silica is the foundation of modern fiber optics for good reasons. It is transparent across a wide wavelength range, it can be drawn into fiber with extraordinary geometric precision, and it is cheap. But silica has a hard ceiling: above approximately 1000°C, the glass softens, the dopants that define the refractive index profile diffuse, and the fiber’s optical and mechanical properties degrade irreversibly. A standard Fiber Bragg Grating inscribed in a germanosilicate fiber — the workhorse of structural health monitoring, strain measurement, and industrial temperature sensing — begins losing its grating structure well below that ceiling.

For most sensing applications, 1000°C is more than sufficient. For the environments where sensing matters most — jet turbine combustion chambers at 1400°C, industrial furnace interiors at 1600°C, advanced materials processing at 1800°C, aerospace propulsion monitoring approaching 2000°C — silica fiber is simply not an option.

The excellent optical transparency, thermal and chemical stability, mechanical robustness, and high melting temperature (~2040°C) of single-crystal sapphire fibers make them a strong candidate for sensing applications in high-temperature environments. Sapphire — aluminium oxide, Al₂O₃, in its single-crystal form — is the material that picks up where silica fails. Its melting point is nearly 1000°C higher than silica’s. It is chemically inert in oxidising and reducing atmospheres. It does not soften, it does not devitrify, and it does not lose its crystalline structure at temperatures that would destroy any glass-based alternative.

The challenge is that sapphire is not silica. You cannot draw it into fiber using a conventional draw tower. You cannot dope it to create a refractive index contrast. You cannot inscribe a Bragg grating using the UV photosensitivity that works effortlessly in germanosilicate glass. Working with sapphire fiber requires a completely different fabrication toolkit — and the paper published to arXiv on May 5, 2026, from a team spanning Chinese institutions and the Oxford Nanophotonics group, represents the most complete solution to that fabrication challenge demonstrated to date.

What Makes Sapphire Fiber Different — and Difficult

Before examining the paper’s specific contribution, it is worth understanding precisely why sapphire fiber sensing has been difficult to commercialise despite its obvious material advantages.

A conventional silica fiber achieves single-mode guidance through a simple refractive index step: the germanium-doped core has a slightly higher index than the pure silica cladding, and total internal reflection confines the light. The index contrast is small — typically 0.3–0.8% — which is sufficient because the core diameter is small and the guidance is tight.

Sapphire fiber, as grown by the laser-heated pedestal growth (LHPG) or edge-defined film-fed growth (EFG) methods, has no cladding. It is a pure sapphire rod, typically 75–250µm in diameter, with no index contrast structure. Light propagates through it, but in a highly multimode fashion — dozens or hundreds of spatial modes coexist, each with different propagation velocities, making coherent sensing signals difficult to interpret. The fiber acts more like a light pipe than a waveguide in the photonics sense.

Previous approaches to creating single-mode or few-mode guidance in sapphire fiber have used femtosecond laser direct writing to create a depressed cladding waveguide: a ring of laser-modified tracks surrounding a central sapphire core, where the modification reduces the local refractive index and creates the index contrast needed for guidance. The approach works — single-mode guidance has been demonstrated, and temperature sensors based on Bragg gratings inscribed within these waveguides have been characterised up to temperatures approaching 1500°C. But depressed cladding fabrication is slow: each track in the ring must be written individually, the ring must be dense enough to confine the mode adequately, and the total number of laser passes required is large.

This is precisely the problem the new paper addresses.

Read more

Same Industry, Same Day, Opposite Market Reactions

Yesterday, May 6, 2026, two of the most important publicly traded photonics companies in the world reported their quarterly results simultaneously. The contrast could not have been more instructive.

Lumentum delivered an exceptional third quarter, with revenue growing 90% year over year to a record $808 million. The stock ticked higher in after-hours trading. IPG Photonics reported revenue of $265.5 million, a 17% year-over-year increase that exceeded analyst expectations — and its stock plummeted 25% in pre-market trading.

One company nearly doubled its revenue. The other grew 17% and beat guidance. The market rewarded one and punished the other severely. Understanding why requires reading both reports carefully — because together they tell you almost everything you need to know about where photonics value is migrating in 2026, and which part of the industry the financial markets are currently willing to pay a premium for.

Lumentum: The AI Photonics Boom in Full Visibility

Lumentum’s numbers need to be read in context to appreciate their significance. Top line growth was primarily driven by the transceiver business and laser chips, while non-GAAP operating margin expanded by over 2,100 basis points year-over-year. Revenue doubling is impressive. Margins expanding by 21 percentage points simultaneously is extraordinary — it means the growth is not being bought with discounting or volume dilution.

The company’s wafer fab capacity in Japan remains at a premium and is fully allocated to meet demand, and Lumentum is shipping twice the number of laser chips compared with the same quarter last year. The constraint is not demand. It is supply. Lumentum cannot build fast enough.

The forward guidance makes the trajectory explicit: for Q4, management expects net revenues to come in between $960 million and $1.01 billion, with the $985 million midpoint representing another new all-time quarterly revenue record. A company that reported $425 million in revenue twelve months ago is guiding to nearly $1 billion in a single quarter. That is the pace of the AI photonics buildout, expressed in quarterly earnings.

CEO Michael Hurlston’s comment in the earnings call is worth marking: “As our key growth drivers of co-packaged optics and optical circuit switches begin to kick in, we would expect further increases in earnings power.” The 90% growth quarter is not, in management’s view, the peak. CPO (Co-Packaged Optics) and OCS (Optical Circuit Switching) — the two product lines that readers of this publication have been tracking since the first Business of Photonics issue in March — have not yet contributed meaningfully to Lumentum’s revenue. They are the next wave. The current wave is transceivers and laser chips.

Supply constraints on critical components are still limiting shipments relative to customer demand — the company is supply-constrained across EMLs (electro-absorption modulated lasers), pump lasers, and transceivers simultaneously. This is why the NVIDIA $2 billion investment, which added directly to Lumentum’s cash balance in Q3, was not charity. It was a capacity reservation. The company is on track to achieve more than 50% growth in EML units by the December quarter of 2026 versus the December quarter of 2025.

IPG Photonics: Strong Operations, Wrong Market Timing

IPG’s story is more nuanced and, frankly, more instructive for understanding the complexity of being a diversified photonics company in the current environment.

Industrial Solutions revenue grew 21% year over year to $227.6 million, representing 86% of total revenue, with notable gains in welding, cutting, marking, and cleaning applications. The core fiber laser business is genuinely healthy. Stationary storage for data centers is emerging as a key driver, requiring higher-power lasers and real-time monitoring for thicker bus bars, which aligns with IPG’s core technological strengths.

So why the 25% share stock price drop? Three factors converged simultaneously and the market chose to price all three harshly at once.

Investor concerns over elevated valuations, tariff headwinds, and a significant litigation settlement overshadowed the operational achievements. The market reaction suggests investors are reassessing the company’s premium valuation — with a P/E (Price to Earnings) ratio of 131 significantly above industry norms — and growing concerned about margin pressures from tariffs.

The P/E of 131 is the central issue. IPG has been priced as a high-growth AI-adjacent photonics company, but its revenue mix tells a more industrial story: 86% industrial cutting, welding, and marking lasers, with the AI-adjacent Advanced Solutions segment actually declining 5% year over year. A $13.5 million charge for settlement of patent litigation with TRUMPF and higher tariffs contributed to an operating loss of $7.7 million and reduced net income to $1.6 million.

The tariff headwind is real and structural: tariffs remain a persistent headwind, impacting gross margins by approximately 150 basis points, with management working on pricing and cost offsets. For a company with significant manufacturing in Germany and sales exposure across China, North America, and Europe, tariff dynamics compress margins in ways that are difficult to offset quickly. The TRUMPF settlement removes a litigation overhang, but the $13.5 million charge in a quarter of $1.6 million net income is the kind of one-time event that makes quarterly earnings almost unreadable.

The signal to retain from IPG’s results: the industrial fiber laser market is growing at 17% and book-to-bill remains above 1. Battery manufacturing for EVs and stationary storage is genuinely strong. Defense (the Crossbow high-energy laser system, the Lockheed Martin order) is an emerging revenue line. IPG is not a broken business. It is a premium-valued industrial laser company caught between a market that re-rated it as an AI play and quarterly results that do not yet fully support that re-rating.

Read more

The Most Expensive Millimetre in Photonics

If you have been following this publication from its earliest posts, you know that the platform we are building has two conceptual pillars. The first is the world of specialty optical fibers — large mode area designs, hollow-core photonic crystal fibers, anti-resonant waveguides. The second is the world of silicon photonics — nanoscale waveguides, ring modulators, programmable meshes, quantum photonic processors on chip. These two worlds were covered separately in the founding series because they are physically and technologically distinct domains.

But they are not isolated. Every time a silicon photonic chip needs to send or receive light from the outside world — from a laser source, from a sensing environment, from a quantum network, from a data center fiber link — it must interface with an optical fiber. And that interface, that single millimetre where light must transfer from a 10-micron fiber mode to a 300-nanometer silicon waveguide mode, is one of the most persistently difficult engineering problems in integrated photonics.

It is the problem that a paper published April 15th 2026 in AIP Advances, from Xinchao Zhou, Tzu-Han Chang, Eric Liu, Saivirinchi Prabandhakavi, and Chen-Lung Hung, addresses with a design that is notable for combining high performance with genuine fabrication simplicity. The title — “On-chip optical fiber-to-nanophotonic waveguide adiabatic coupler” — is technically precise to the point of being deliberately understated. What it describes is a solution to a problem that has constrained the practical deployment of nanophotonic systems for over two decades.

Why the Interface Is So Hard

To understand what makes this paper’s contribution meaningful, it helps to appreciate the scale of the problem it is solving.

A standard single-mode optical fiber — the kind that carries data across continents and connects data center racks — has a mode field diameter of approximately 10 micrometres at 1550nm wavelength. The optical field is relatively large, slowly varying, and well-matched to the gaussian beam profile of most free-space optics.

A silicon nanophotonic waveguide — the kind that routes light on a silicon photonic chip — has a cross-section of roughly 450nm × 220nm. The mode is tightly confined, with dimensions smaller than the wavelength of the light it carries, and the refractive index contrast between silicon (n ≈ 3.47) and its oxide cladding (n ≈ 1.44) is enormous by optical standards.

The mode field diameter mismatch between these two structures is approximately 30:1 in linear dimension, or roughly 1000:1 in area. Attempting to couple light directly from fiber to chip at an abrupt interface loses the vast majority of the optical power to radiation. This is not an imperfection in the manufacturing — it is a fundamental consequence of trying to couple two structures with modes of completely different spatial scales.

Three engineering approaches have been developed over the decades to manage this mismatch. End-fiber coupling, diffraction grating-based coupling, and adiabatic coupling each operate on different physical principles and involve different practical trade-offs. Grating couplers are by far the most commonly used — they are fabricated directly on the chip surface, require only vertical illumination from an optical fiber held above the chip, and are straightforward to integrate into standard PDKs. But grating couplers have limited optical bandwidth and are naturally out-of-plane, which creates both wavelength selectivity constraints and packaging complexity for edge-connected systems.

Adiabatic coupling is the physically elegant alternative. In adiabatic mode transfer, single-mode fiber-waveguide coupling efficiencies as high as 97% are achievable. Efficient coupling is obtained for a wide range of device geometries which are either singly-clamped on a chip or attached to the fiber, demonstrating a promising approach for integrated nanophotonic circuits, quantum optical and nanoscale sensing applications.

The word “adiabatic” here is borrowed from thermodynamics but applied to optics: a transformation is adiabatic if it happens slowly enough that the system remains in its fundamental mode throughout. In an adiabatic coupler, both the fiber and the waveguide are tapered — one getting narrower, the other getting wider — over a sufficient length that the optical mode transitions smoothly from one structure to the other without exciting higher-order modes or radiation modes. No abrupt interface. No mode mismatch loss. Just a gentle, gradual handover of the optical field.

The challenge with adiabatic couplers has always been fabrication. The problem of fiber-to-chip coupling is difficult because the fundamental mode of an optical fiber is roughly 10 µm in diameter, and the dimensions of the fundamental mode of a high-index-contrast waveguide are often less than 1 µm across. Earlier demonstrations of adiabatic coupling required either substrate releasing — a delicate microfabrication step that suspends the waveguide in air to prevent light leakage into the substrate — or cladding the tapered fiber tip in a higher-index polymer material to prevent the field from leaking before it reaches the chip waveguide. Both approaches add fabrication complexity and reduce yield.

Read more