To The Moon

$Applied Optoelectronics(AAOI)$ 

$AXT Inc(AXTI)$

$Marvell Technology(MRVL)$ 

$Corning(GLW)$

$Ciena(CIEN)$

On March 12, 2026, three separate optical-related MSAs dropped on the same day. That does not happen by coincidence. It means the entire industry has converged on the same conclusion: the physical layer of AI infrastructure needs to change, and it needs to change now. Call it the opening shot of the Optical War.

This article breaks down what each MSA actually standardizes and which companies are positioned to benefit.

First, What Is an MSA?

MSA stands for Multi-Source Agreement. It is an industry consortium where multiple companies align in advance on component specs, interfaces, mechanical dimensions, and management protocols so that no single vendor ends up with a monopoly on a given interface. Think of it as a binding agreement to keep products interoperable across vendors, rather than locking customers into one proprietary ecosystem.

The key difference from official standards bodies like JEDEC or IEEE is speed. Standards bodies pull in hundreds of companies and take years to reach consensus. An MSA gets a small group of core players in a room, defines the interface fast, and ships products. This is how SFP, QSFP, and OSFP became the de facto industry standard for optical module form factors. The MSA route worked then, and the industry is betting it works again now.

Three in one day means the industry is in a hurry.

Three MSAs, Three Different Layers

OCI MSA (Optical Compute Interconnect)

Founding members: AMD, Broadcom, Meta, Microsoft, NVIDIA, OpenAI.

OCI standardizes short-reach optical links inside the AI scale-up domain, specifically the common optical PHY connecting GPU to GPU and GPU to scale-up switch. It is NRZ plus WDM: GEN1 at 200 Gbps per direction, GEN2 at 400 Gbps per direction (800 Gbps per fiber), with a long-term roadmap beyond 3.2 Tbps. It does not touch the protocol layer. Whether you are running UALink (AMD and Broadcom) or NVLink (NVIDIA), OCI only unifies the physical layer underneath. It supports pluggable, on-board optics, and CPO.

The membership structure is the real story. Three hardware vendors and three hyperscalers sitting at the same table is a departure from the vendor-driven standards bodies of the past. The buyers are at the table this time.

The NRZ choice is also worth noting. PAM4 dominates scale-out, but OCI starts with NRZ. NRZ has simpler signal processing and lower power draw. For short-reach, you can hit adequate signal quality without a complex DSP. Bandwidth comes from adding WDM wavelengths rather than pushing per-channel speed. The design philosophy is power efficiency and bandwidth density at the same time, which maps directly onto the CPO direction.

Open CPX MSA (Open Co-packaging)

Founding members: Ciena, Coherent, Marvell, Molex, Samtec, TeraHop.

If OCI defines the optical PHY spec, Open CPX standardizes the physical form factor of the optical engine that implements that spec. It covers the socket, connector, thermal management, electrical pinout, and optical interface between the ASIC and the CPO/NPO optical engine. Put simply: OCI says what optical signal to use; Open CPX says how to physically plug that optical engine next to the ASIC.

The LightCounting CEO has projected that annual CPO/NPO port shipments will exceed 100 million units within five years. Getting that volume onto a multi-vendor model requires physical compatibility at the optical engine level.

XPO MSA (eXtra-dense Pluggable Optics)

Led by Arista Networks, with founding members including Lightmatter, Eoptolink, and TeraHop. More than 40 optical module vendors have joined.

XPO defines a next-generation pluggable form factor that goes beyond the limits of current OSFP. The target is 12.8 Tbps per module and 204.8 Tbps per OCP rack unit, four times the front-panel density of OSFP. An integrated liquid-cooled cold plate handles over 400W of cooling per module. A single XPO module covers DR, FR, LR, SR, ZR, ZR+, and coherent-lite. One XPO replaces eight OSFP modules. Arista claims that in a 400MW AI data center, XPO reduces switch rack count by 75%. Volume production is targeted for 2027.

Why Now: The Copper Wall Is Coming

In early March, Broadcom CEO Hock Tan said this on the earnings call:

"We will push copper in scale-up as long as we can." 

"CPO is coming. Not this year, probably not next year either." 

"Silicon photonics becoming meaningful is not far off, but not yet."

The market reacted immediately. Optical-related stocks dropped more than 10% across the board. The read was simple: copper wins, optical transition delayed. Some called it the end of the optical thesis.

Then, one week later, Broadcom co-founded the OCI MSA. At OFC 2026, the company unveiled a CPO-integrated Tomahawk 6 switch, the industry's first 400G/lane optical DSP, its own 400G EML laser, and a VCSEL-based NPO. The company that said it would push copper showed up with a full optical stack.

Is that a contradiction? No. The market misread what Tan actually said.

The correct reading: scale-out already runs on optical. Scale-up will also move to optical eventually. Just not yet. 

The key phrase is "eventually." He was talking about timing, not direction. His "not yet" horizon runs to roughly 2028, meaning copper still has a real advantage in scale-up on latency, power, and cost for the next two years or so.

Why 2028? Because that is when copper's physical limits hit critical mass simultaneously on three fronts.

First, reach. Above 800 Gbps, copper cable maxes out at a few meters, and holding acceptable error rates without retimers becomes difficult. This is why NVL72's copper connections are confined to a single rack. Multi-rack scale-up requires optical.

Second, power. Copper networking already consumes more than half of data center network power. As GPU counts scale up, that burden becomes unmanageable.

Third, bandwidth density. There is a hard physical ceiling on how much I/O bandwidth you can pull off an ASIC shoreline using copper. At some point, the chip gets faster but the copper I/O cannot keep up.

All three of these constraints converge around 2027 to 2028. That is also when the industry is targeting CPO volume production. Tan's "not yet" was not a denial of the destination. It was an observation that the inflection point has not arrived. The OCI MSA is about locking in the standard before it does.

"Copper versus optical" is not a binary choice. It is a sequence. Broadcom extracts margin from copper today while simultaneously building the standard and the product stack for when optical becomes necessary in two to three years. The optical-related stocks selloffs after Tan's comments reflected the market collapsing a time sequence into a simple "copper wins" headline.

Why NVIDIA Joined an Open Consortium

The obvious question: NVLink already exists, so why does NVIDIA join an open optical consortium?

OCI does not replace NVLink. It builds the optical physical layer that NVLink runs on. Protocol and PHY are different layers. NVIDIA opened NVLink Fusion in May 2025, extending the NVLink ecosystem to external custom silicon partners. In March of the same year, the company announced Spectrum-X Photonics and a CPO switch, pulling optical directly into the network core. For NVLink to reach further and pack more tightly, it needs a standardized optical PHY and a multi-vendor supply chain underneath it.

At the same time, NVIDIA has committed $4B to the optical supply chain. That is $2B each to Lumentum and Coherent, with multi-year preferred purchase agreements, plus participation in Ayar Labs' $500M Series E. The play is straightforward: use standardization to bring down component costs and spread supply risk, while keeping NVLink itself proprietary. Open PHY does not mean open protocol. Only the optical layer gets opened up; the protocol lock-in stays intact.

Who Benefits: Breaking Down the Value Chain

To think clearly about optical transition beneficiaries, you have to decompose the value chain: laser source, modulator and receiver (PIC), optical DSP, optical engine and module assembly, connectors and packaging, and the system layer that integrates all of it with the ASIC. The three MSAs each standardize a different layer, so the beneficiaries split accordingly.

System Players: ASIC and Optics Under One Roof

The structurally strongest position belongs to companies that hold both the switch or compute ASIC and meaningful optical technology.

Broadcom is an OCI MSA founding member, and at OFC 2026 it showed up with a CPO-integrated Tomahawk 6 switch, a 3nm 400G/lane optical DSP (Taurus), its own 400G EML, and a VCSEL-based NPO. The vertical portfolio runs from switch ASIC to optical engine to DSP to laser. OCI's silicon-centric model maps directly onto the CPO architecture Broadcom has been building for years. Running copper for current revenue while preparing the full optical stack is a strategy to hold both the present and the future at the same time.

Marvell has pulled up to the same tier. It is not in OCI, but it is an Open CPX founding member. The Celestial AI acquisition closed in February 2026 for $3.25B, folding in Photonic Fabric-based optical interconnect. According to the company, Celestial AI's technology delivers more than twice the power efficiency of copper, nanosecond-level latency, and thermal stability next to multi-kilowatt XPUs. Combine that with Marvell's existing PAM4 optical DSP and UALink scale-up switch roadmap, and you get DSP plus switch plus optical fabric from a single vendor. At OFC 2026, Marvell led with a 1.6T PAM4 optical DSP and a multi-rack optical scale-up Photonic Fabric platform.

These two are the clearest system-level beneficiaries of the optical era. They are writing the standards and shipping the products that comply with those standards at the same time.

Laser Sources: Whoever Controls the Light Controls the Supply Chain

Every optical connection, whether CPO, pluggable, or XPO, starts with a laser. The number of companies that can actually manufacture high-performance InP-based EML and CW lasers at volume is very small.

Lumentum owns an upstream InP wafer fab, manufactures EML, CW, and ELS lasers, and runs downstream OCS and cloud optical modules. It is the most vertically integrated pure-play optical company. The MEMS mirror technology behind Google's Palomar OCS, used in TPU clusters from v4 through v7, comes from Lumentum. Add the NVIDIA $2B investment and multi-year preferred purchase agreement.

Coherent (formerly II-VI) has InP lasers and LCoS-based OCS. OCS pilots are currently running with seven customer accounts. Also $2B from NVIDIA with a multi-year purchase contract. Both companies are expanding U.S. manufacturing capacity. In a CPO world, these two companies control whether laser supply can actually keep up with demand.

A multi-vendor ecosystem means more total volume, and that structurally expands the TAM for laser sources.

CPO and NPO Optical Engines: The Direct Standard Beneficiaries

Ciena carries a traditional networking reputation, but after acquiring Nubis Communications, it launched the Vesta 200 6.4T CPX, entering the scale-up optical market with a 200G/lane CPO engine. The linear-drive architecture without retimers reportedly cuts power consumption by up to 70%. Since Ciena does not make its own ASICs, it is a structural beneficiary that grows as the CPO ecosystem grows, rather than a company driving that ecosystem.

Lightmatter builds Passage, a silicon photonics interposer that sits underneath the processor package. That is architecturally different from conventional CPO, which extracts bandwidth from the package edge. Passage is edgeless I/O. The company has demonstrated over 60 Tbps of inter-package optical bandwidth per chip and 16-wavelength bidirectional DWDM.

Ayar Labs takes the approach of replacing electrical SerDes with a TeraPHY optical I/O chiplet. UCIe-compatible. The target is 144 Tbps bidirectional bandwidth per chip, roughly five times what Rubin GPU carries today. Series E was $500M with NVIDIA participation. Manufacturing runs on GlobalFoundries' silicon photonics platform.

What these startups share is that their technology can slot in as compliant product on top of the standards the three MSAs define. Broader standards adoption opens up larger opportunity. The risk is that they are still in volume production validation.

Pluggable Modules: Volume War, Present and Future

Chinese vendors already dominate the pluggable module market. 

InnoLight and Eoptolink are estimated to have captured roughly 60% of NVIDIA's 800G volume. Eoptolink is also a founding member of the XPO MSA and is demonstrating a 12.8 Tbps XPO module at OFC 2026.

On the U.S. side, Applied Optoelectronics (AAOI) is worth watching. It has been a Tier 2 player in data center optical modules for a long time, but the 800G and 1.6T transition has pushed the company to broaden its portfolio into VCSEL and EML-based modules. Lower geopolitical risk relative to Chinese vendors is a supply chain diversification argument that U.S. hyperscalers take seriously.

OCI's support for multi-vendor pluggables and XPO's definition of the next-generation pluggable form factor create a clear near-term tailwind. The long-term question is different: as CPO adoption accelerates, the pluggable share of the market shrinks. The ability to transition into CPO will determine whether this is a growth story or a sunset story.

Upstream: Materials and Devices

At the base of the value chain sit the materials and device companies that make optical components possible. They are not at the MSA table, but they benefit structurally as total optical volume grows.

AXT (AXTI) is a primary supplier of InP (indium phosphide) substrates. InP wafers are the raw material for the EML and CW lasers that Lumentum and Coherent manufacture. If laser source demand explodes in the CPO era, InP substrate demand follows.

Corning (GLW) is synonymous with optical fiber and cable. Scale-out already consumes Corning fiber at scale, including a $6B fiber agreement with Meta. If scale-up converts to optical, that adds intra-rack fiber demand on top of what already exists. The stock sold off after Tan's comments, but Corning management made clear that the company's growth strategy does not depend on scale-up optical revenue. The scale-out story alone is sufficient.

On the device side, Lightwave Logic (LWLG) is developing electro-optic polymer-based modulators. The pitch is wider bandwidth and lower drive voltage compared to conventional silicon photonics modulators like MZM and MRM. Volume manufacturing is not there yet, and commercial deployment is still a distance away. The technology direction is interesting, but there is a gap between here and production.

POET Technologies takes an Optical Interposer approach, integrating laser, modulator, and receiver onto a single chip. The concept points in the same direction as Lightmatter's Passage, but the current situation is very different. POET is pre-revenue. The Foxconn 1.6T pluggable module production timeline is 2026 to 2027, but there is no shipping track record to validate it. POET does not appear in any of the three MSAs. The technology direction is credible, but two prerequisites for being part of this wave, demonstrated volume production and a presence in the standard ecosystem, are both absent. Whether Foxconn production gets to scale and whether POET enters an MSA ecosystem will determine whether this story develops.

Enablers: Foundry, Test, and Packaging

Someone has to actually fabricate optical chips. The importance of foundries that can manufacture silicon photonics PICs at volume is rising fast.

Tower Semiconductor (TSEM) is one of the leading silicon photonics foundries. It holds SiPho and SiGe BiCMOS platforms. In November 2025, it announced the extension of 300mm wafer bonding technology to 3D-IC for CPO applications, along with a $300M investment plan to double SiPho manufacturing capacity by end of 2025 and triple it by mid-2026. At OFC 2026, Tower is leading with a DWDM laser source developed with Scintil Photonics and a CPO platform.

GlobalFoundries holds a 300mm SiPho platform (Fotonix) and counts Ayar Labs among its key customers. 

TSMC is approaching CPO from the packaging angle through COUPE (Compact Universal Photonics Engine), with expected application in NVIDIA Quantum-X switches and the next-generation Rubin platform. SiPho capacity at these foundries could be the real bottleneck for the 2026 optical ramp. Industry estimates have high-end optical chip production capacity growing more than 80% year over year in 2026, but that still leaves a 5 to 15% supply shortfall relative to demand.

Test and packaging matter too. Aehr Test Systems (AEHR) builds wafer-level burn-in and test equipment. As silicon photonics device reliability verification demand scales up, Aehr is a direct beneficiary. Reliability verification is the production gating function for CPO, which raises the strategic importance of test equipment.

Keysight Technologies (KEYS) and FormFactor (FORM) are positioned to benefit in optical interconnect test and measurement. Keysight covers high-speed optical signal analyzers and BER testers; FormFactor provides wafer-level optical probing.

BE Semiconductor (BESI) makes die bonding and packaging equipment. As the process of 3D-stacking PIC and EIC in CPO expands, BESI benefits from the incremental volume.

Conclusion

The question of whether the optical era is coming has been settled. The race to lock in standards and supply chains is already running. Three MSAs launching on the same day is the starting gun.

The question now is which companies get to the standard and the supply chain first.

@Daily_Discussion  @Tiger_comments  @TigerStars  @TigerObserver  @TigerPM