POET Technologies ($POET) and Sivers Semiconductors ($2DG.F) stock after Marvell ($MRVL) cancelled the Celestial AI co-packaged-optics (CPO) orders — what it means for each stock, the CPO laser supply chain ($LITE Lumentum, $COHR Coherent), and the AI photonics trade, with a live supply-chain map.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
TL;DR. On April 27, 2026 Marvell ($MRVL) cancelled all purchase orders to POET Technologies ($POET) tied to its Celestial AI subsidiary, citing a confidentiality breach after POET's CFO disclosed PO timing on a podcast. $POET stock crashed ~47% on the news. The orders ran through POET's Optical Interposer, which was being paired with Sivers Semiconductors ($2DG.F) DFB lasers for AI co-packaged optics (CPO) deployments. This post explains what each company actually contributes to the CPO stack, why $POET stock got hit so much harder than $2DG.F, who else builds external light sources (ELS) for AI photonics — Lumentum ($LITE), Coherent ($COHR), Furukawa, Ayar Labs, Lightmatter — and the realistic 2026 outcome for each name. Live prices for the five tickers above are at the top of the page.
On April 27, 2026 the photonics corner of the AI trade got an ugly surprise. Marvell ($MRVL) reportedly canceled all purchase orders tied to its Celestial AI subsidiary — the very orders POET Technologies ($POET) had just spent the past week celebrating in a CFO interview. The stated reason was a confidentiality breach: $POET's CFO went on the record about purchase order timing and shipment plans, and $MRVL pulled the plug. $POET shares slumped ~47% intraday. The Celestial AI program — the showcase AI fabric customer for the joint $POET / Sivers Semiconductors ($2DG.F) external light source (ELS) module — is the chunk of revenue that disappeared.
For Sivers Semiconductors ($2DG.F), this is bad news in the short term — $POET was one of the most visible go-to-market channels for Sivers' indium phosphide (InP) DFB lasers into hyperscale AI workloads. But it is also, for anyone who actually understands what $2DG.F ships into the rack, very much not an existential event. The lasers are still the lasers. The bottleneck is still the bottleneck. The customer just changed.
This post walks through, at a technician's level of detail, what $2DG.F actually makes, how those parts get integrated, what the $POET deal would have done, and why upstream and downstream players in the silicon photonics & optical interconnects trend still have very strong reasons to source from $2DG.F. If you came here for the $POET stock impact specifically, jump to "What this means for POET Technologies stock" below.
$2DG.F's photonics business sells continuous-wave (CW) Distributed Feedback (DFB) lasers — and increasingly, arrays of them on a single chip — fabricated in the InP (indium phosphide) material system.
A few specifics that matter:
These are not commodity parts. The grating is etched into the III-V epi stack to set the wavelength, and yields drop sharply when you start packing 8, 16, or 32 emitters onto a single bar with tight wavelength spacing. $2DG.F's edge is in the epitaxy and grating fab — they have qualified capacity at WIN Semiconductors ($3105.TWO) as a fab partner, which is what lets them ship at the volumes hyperscalers need.
To understand why the customer can change without the laser changing, you need to picture the actual stack. A modern Co-Packaged Optics (CPO) switch or AI accelerator looks roughly like this, from light source to fiber:
The key thing to internalize: the laser is not the optical engine, and the optical engine is not the laser. They are decoupled by design — that is the whole point of "External" Light Source. $2DG.F makes the part that emits the photons. $POET, Lightmatter, Ayar Labs, $MRVL, $AVGO, $INTC, and every other photonics IC company makes the parts that modulate, route, multiplex, and detect those photons. They need each other.
The September 2025 partnership had a clean division of labor:
Combined, the two were producing a "plug-and-play" light engine sub-system that could be sold into transceiver makers and CPO programs without the customer needing to learn III-V packaging. $MRVL/Celestial AI was the showcase customer for the AI cluster use case (1.6 Tbps+ optical fabric stitching together accelerator pods).
When $MRVL yanks the orders, what gets removed from the picture is the integration partner's revenue and the showcase socket. What does not get removed is the underlying physics. $MRVL still has Celestial AI's photonic fabric program. That program still needs InP CW lasers as light sources. There are not many companies on Earth that can ship qualified, high-power, low-FIT CW DFB arrays at the volumes a hyperscale AI fabric requires — Lumentum ($LITE), Coherent ($COHR), Furukawa ($5801.T), and $2DG.F are essentially the merchant short list.
$POET stock is the directly exposed name. The Celestial AI socket was the showcase customer underwriting POET's optical interposer ramp narrative — the thing that justified the run from sub-$2 in 2024 to the post-partnership highs. Three things to internalize on $POET specifically:
For $POET stock, the realistic 2026 case is: support around the prior trading range while new customer announcements grind through. The bear case is a dilutive raise into weakness. The bull case is a quiet re-engagement with $MRVL once NDAs get patched. None of those are priced in cleanly yet — that's why $POET stock news in the weeks after April 27, 2026 has been jumpy.
Three reasons, in order of importance:
If $MRVL continues building Celestial AI's photonic fabric, it has to source CW DFB arrays from somebody. The cleanest paths forward:
In all three cases, the laser BOM line is unchanged. $POET's revenue is what's at risk; $2DG.F's photon-emitting silicon is still in the bill.
Even if every $MRVL/Celestial dollar evaporated, $2DG.F's photonics roadmap has multiple legs that don't share customer concentration with $POET:
These programs have nothing to do with $MRVL's Celestial AI POs. Losing one channel hurts visibility but does not change the structural demand for InP CW lasers across the broader CPO/ELS rollout, which is why the silicon photonics & optical interconnects trend and the underlying compound semiconductor epitaxy category are the right places to track this thesis from.
This is the part the market is mispricing today. There is no shortage of clever photonic interposer designs — $POET, Lightmatter, Ranovus, $AVGO, $INTC, and TSMC's ($TSM) own photonic SOIC roadmap are all credible options. There is a shortage of qualified, hyperscale-grade III-V CW laser capacity. InP epi reactors are slow to bring online, the recipes are jealously guarded, and reliability qualification cycles can take 12–18 months per customer. When the integration layer reshuffles, the laser supplier sitting at the bottleneck doesn't get cut — it gets a new shipping address.
That asymmetry is exactly why @aleabitoreddit and other photonics analysts have been pointing out that $POET trades at ~9–11× the market cap of its laser supplier ($2DG.F). The market priced in the integration story; the laser bottleneck story has been undervalued. Today's news doesn't break the laser thesis — if anything, it forces investors to look at where the actual scarcity sits.
These are the risks that are activated today by the $POET–$MRVL event, separate from sympathy-driven price action. For each, what specifically goes wrong, what to watch for confirmation, and what the realistic financial impact looks like.
The mechanics: $2DG.F's photonics revenue ramp was modeled around $POET's prototype-to-production schedule (early customer demos in H1 2026, production by year-end 2026). If $MRVL's cancellation forces $POET to either rebuild the $MRVL relationship under new terms or qualify replacement customers ($2317.TW, $002475.SZ, others), each path adds 3–9 months of delay.
Even if $2DG.F's BOM share is unchanged, the perception of single-customer fragility now sits on the stock. Photonics revenue today is unevenly distributed — Ayar Labs has historically been the largest single line item, and $POET-channel revenue was supposed to diversify that. Losing the visible $POET ramp temporarily increases concentration on Ayar Labs at exactly the moment investors are scrutinizing it.
This is the genuinely fragile part of the story, and it's not about the technology. $2DG.F has run negative free cash flow consistently and has a history of equity raises at unfavorable prices. A revenue slip combined with continued opex growth means the next financing window arrives sooner than the prior plan assumed.
A non-trivial part of $2DG.F's retail bid was momentum carried over from $POET coverage. That correlation cuts both ways: when $POET-related catalysts evaporate, $2DG.F-the-stock takes the hit even when $2DG.F-the-business doesn't.
These are the slower, structural risks. None of them were triggered by the $POET news — but they are the ones that, if they materialized, would mean the entire laser-bottleneck thesis is wrong, not just delayed. Listed in rough order of probability.
What would have to be true: Linear Pluggable Optics (LPO) and improved pluggables scale cleanly through 1.6T and even 3.2T per port, delivering acceptable power-per-bit without requiring the move to co-packaged optics. Hyperscalers (Microsoft ($MSFT), $META, Google ($GOOGL), Amazon ($AMZN), $NVDA-aligned cloud providers) defer their CPO rollouts and keep buying pluggables.
Why it could happen: LPO has real momentum. Recent demonstrations of 1.6T LPO at sub-15 pJ/bit are surprisingly close to the early CPO power budgets. Hyperscalers strongly prefer the field-replaceable form factor of pluggables, and switching infrastructure has long depreciation cycles.
Why it probably doesn't: at 3.2 Tbps per port and beyond, the SerDes electrical reach budget gets brutal. Copper between the ASIC and the faceplate at those rates is borderline non-functional. CPO becomes a physics requirement, not a marketing preference. The question is when, not whether.
What would have to be true: III-V-on-Si hybrid integration (wafer-scale InP-to-Si bonding), quantum-dot lasers grown directly on Si, or other monolithic-laser approaches scale to hyperscale qualification volume by ~2027–2028. The "external" in External Light Source becomes optional. CPO designs increasingly integrate the laser into the same die or the same interposer as the modulators, removing the standalone ELS module entirely.
Why it could happen: real, well-funded research programs exist. $INTC's hybrid silicon photonics platform (with bonded III-V), Tower Semiconductor's ($TSEM) PH18 photonic platform, IMEC's silicon-photonics roadmap, Scintil Photonics' integrated lasers, and academic work on Si-based QD lasers are all credible. $NVDA's photonic engine designs publicly hint at moving toward more integration over time. Eliminating the ELS module is a natural cost and reliability optimization once the underlying integration is yield-clean.
Why it probably doesn't (yet): the thermal argument for keeping lasers external next to a 700W ASIC is strong, and yield/reliability for hybrid-bonded high-power CW lasers at hyperscale qualification (sub-1-FIT) remains unproven at volume. Quantum-dot Si lasers don't yet hit the wavelength stability and output power that ELS specs require. Even if integration wins long-term, the transition window covers all of the 2026–2029 ramp where $2DG.F is positioned.
What would have to be true: One of the big incumbent laser vendors decides hyperscale CPO ELS is a strategic priority and dedicates serious manufacturing and qualification resources to it. They out-spec $2DG.F on output power or wavelength stability, undercut on price by leveraging their volume, or — most painful — win sole-source qualification at a hyperscaler that $2DG.F was targeting.
Why it could happen: $LITE has publicly positioned itself as the market reference for high-power 1310nm CPO lasers. $COHR has the materials breadth and hyperscale customer relationships. Either could decide to crowd out the smaller specialists once the market is large enough to matter for them. Hyperscalers also prefer dual sourcing, but if the second source can't keep pace on yield or qualification, they'll consolidate.
Why it probably doesn't kill $2DG.F: hyperscalers explicitly want diversified suppliers in critical bottleneck components — single-sourcing CPO lasers from one vendor is a concentration risk they actively manage against. $2DG.F's $3105.TWO qualified capacity is real and shipping. Even in a $LITE-dominated market, "second supplier" status at a hyperscale customer is a viable, profitable business — the question is whether the multiple compresses to reflect commodity-tier economics.
What would have to be true: an InP wafer or epitaxy-precursor supply disruption (rare-earth-style export controls, a critical fab incident, or trade restrictions affecting $3105.TWO or other Taiwan-based fab partners) prevents $2DG.F from shipping even when demand is there.
Why it could happen: III-V semiconductor materials (InP, GaAs) have concentrated global supply. $3105.TWO is in Taiwan, with all the associated cross-strait risk. US export controls on advanced photonics components into China could cap one of $2DG.F's growth markets (the $0877.HK / $ENA.V Asian-hyperscaler channel).
Why it's a tail risk, not a base case: this is a generic risk that affects every photonics player, not $2DG.F specifically. It's worth pricing in but it doesn't differentiate the investment thesis.
A clean confirmation that $MRVL really did cancel for the reason stated (NDA breach), no quiet re-engagement, no replacement integrator stepping in, AND a parallel slowdown in Ayar Labs / $JBL / $0877.HK programs would be needed to actually break the thesis from $POET-style customer events alone. That's a stack of low-probability events that all have to fire. The realistic 2026 outcome is messier and slower than the bull case wanted, but well short of broken.
The April 2026 cancellation forces investors to look at the broader photonics stocks universe that benefits from the AI co-packaged optics ramp, not just $POET. The realistic public-equity menu, in rough order of how exposed each is to the same theme:
The bigger investing question this event surfaces: the photonics layer of the AI infrastructure stack is still the most under-diversified part of the hyperscale capex picture. Five or six public names carry a disproportionate share of the CPO ramp narrative. The $MRVL/$POET cancellation is a reminder that single-customer concentration risk is real even in companies that look like picks-and-shovels plays.
If you're holding or watching $2DG.F, the practical workflow is:
The $POET–$MRVL cancellation is genuinely bad for $POET and uncomfortable for $2DG.F in the near term. It removes a high-profile channel and bruises the narrative. But the lasers themselves — high-power, wavelength-stable, low-FIT InP DFB arrays — are exactly the kind of part that gets re-routed rather than replaced when an integration partner falls out.
$MRVL still needs photons. Hyperscalers still need photons. The integrator can change; the photon source is harder to swap. That's the part of the story today's news doesn't touch.
Disclosure: this post is technical analysis, not investment advice. Photonics ramps are notoriously bumpy, small-cap photonics names are notoriously volatile, and customer concentration in this space cuts both ways.
According to POET Technologies' April 27, 2026 press release, $MRVL terminated the Celestial AI-related POs after a confidentiality breach. $POET's CFO discussed purchase order timing, shipment schedules and Marvell-specific deal mechanics in a podcast interview, which Marvell determined was a violation of the NDA covering the relationship. The technology itself was not the problem; the disclosure was.
Co-packaged optics is a data-center networking architecture where the optical engine (modulators, photodetectors, multiplexers) sits inside the same package as the switch ASIC or AI accelerator, instead of inside a pluggable transceiver at the faceplate. This shortens the high-speed electrical path from inches to millimeters, cuts power per bit by roughly 2–4×, and enables 1.6 Tbps+ per port at the densities AI fabrics need. The laser that produces the light stays external (in an external light source / ELS module) so it isn't cooked by the 700 W+ ASIC next to it.
The merchant short list for high-power, hyperscale-qualified continuous-wave DFB lasers in CPO ELS modules is: $LITE (Lumentum), $COHR (Coherent), Furukawa ($5801.T), and $2DG.F (Sivers Semiconductors). The integration partners around them — $POET (Optical Interposer), Ayar Labs (TeraPHY), Lightmatter (Passage), $AVGO (TH5/TH6 CPO), $INTC (silicon photonics), $MRVL (Celestial AI Photonic Fabric) — all need lasers from one of those four merchant suppliers, plus increasingly hyperscaler-internal sources.
This post is not investment advice. The structural argument is that the laser bottleneck for AI CPO is genuine and $2DG.F is one of four merchant suppliers that can fill it at hyperscale qualification. The near-term argument is that the $POET cancellation slips visibility into 2026 revenue and increases concentration risk perception until a new diversifying customer is announced. Both are true at once. Pull the full Sivers Semiconductors ($2DG.F) supply chain graph and the silicon photonics & optical interconnects trend on Macroplane before deciding.
The next two catalysts to watch are (1) POET's next quarterly update — whether the H1 2026 prototype milestone is reaffirmed or re-guided — and (2) any new hyperscale or Tier-1 transceiver customer announcement that backfills the Celestial AI socket. A dilutive raise into weakness is the genuine bear case; a quiet re-engagement with $MRVL once NDAs get patched is the genuine bull case. The technology itself (the wafer-level Optical Interposer™ with passive alignment) is real and qualified by other customers.
$MRVL closed the $3.25B acquisition of Celestial AI on February 2, 2026. The acquisition brings Marvell the Photonic Fabric — Celestial AI's optical interconnect technology for scale-up AI clusters — and the engineering team behind it. The Photonic Fabric still needs high-power InP CW lasers as light sources regardless of which integrator $MRVL chooses for the optical interposer layer. That is why the laser-supplier bottleneck argument survives even when individual integration partners get reshuffled.
The April 2026 Marvell–POET cancellation
Marvell's acquisition of Celestial AI
The September 2025 POET–Sivers partnership
Sivers' diversifying customer programs
Market data
Image credits
On Macroplane
According to POET Technologies' April 27, 2026 press release, $MRVL terminated the Celestial AI-related POs after a confidentiality breach. $POET's CFO discussed purchase order timing, shipment schedules and Marvell-specific deal mechanics in a podcast interview, which Marvell determined was a violation of the NDA covering the relationship. The technology itself was not the problem; the disclosure was.
Co-packaged optics is a data-center networking architecture where the optical engine (modulators, photodetectors, multiplexers) sits inside the same package as the switch ASIC or AI accelerator, instead of inside a pluggable transceiver at the faceplate. This shortens the high-speed electrical path from inches to millimeters, cuts power per bit by roughly 2–4×, and enables 1.6 Tbps+ per port at the densities AI fabrics need. The laser that produces the light stays external (in an external light source / ELS module) so it isn't cooked by the 700 W+ ASIC next to it.
The merchant short list for high-power, hyperscale-qualified continuous-wave DFB lasers in CPO ELS modules is: $LITE (Lumentum), $COHR (Coherent), Furukawa ($5801.T), and $2DG.F (Sivers Semiconductors). The integration partners around them — $POET (Optical Interposer), Ayar Labs (TeraPHY), Lightmatter (Passage), $AVGO (TH5/TH6 CPO), $INTC (silicon photonics), $MRVL (Celestial AI Photonic Fabric) — all need lasers from one of those four merchant suppliers, plus increasingly hyperscaler-internal sources.
This post is not investment advice. The structural argument is that the laser bottleneck for AI CPO is genuine and $2DG.F is one of four merchant suppliers that can fill it at hyperscale qualification. The near-term argument is that the $POET cancellation slips visibility into 2026 revenue and increases concentration risk perception until a new diversifying customer is announced. Both are true at once. Pull the full Sivers Semiconductors ($2DG.F) supply chain graph and the silicon photonics & optical interconnects trend on Macroplane before deciding.
The next two catalysts to watch are (1) POET's next quarterly update — whether the H1 2026 prototype milestone is reaffirmed or re-guided — and (2) any new hyperscale or Tier-1 transceiver customer announcement that backfills the Celestial AI socket. A dilutive raise into weakness is the genuine bear case; a quiet re-engagement with $MRVL once NDAs get patched is the genuine bull case. The technology itself (the wafer-level Optical Interposer™ with passive alignment) is real and qualified by other customers.
$MRVL closed the $3.25B acquisition of Celestial AI on February 2, 2026. The acquisition brings Marvell the Photonic Fabric — Celestial AI's optical interconnect technology for scale-up AI clusters — and the engineering team behind it. The Photonic Fabric still needs high-power InP CW lasers as light sources regardless of which integrator $MRVL chooses for the optical interposer layer. That is why the laser-supplier bottleneck argument survives even when individual integration partners get reshuffled.