Sivers and SemiNex Launch $3.4M InP Program for AI Data Centers

Sivers Semiconductors and SemiNex launched a $3.4 million program that puts a specialized photonics partnership into the google news cycle. The companies are developing indium phosphide light sources for AI data center interconnects. Their real challenge is turning higher optical power into reliable, manufacturable products.

The program targets external lasers, multi-wavelength distributed feedback laser arrays, and semiconductor optical amplifiers. These components support optical links connecting accelerators, switches, and other computing hardware. Customer sampling and early production ramps are targeted for the second half of 2027.

That timeline makes the announcement more than an immediate product launch. It is a manufacturing and engineering commitment aimed at architectures still moving toward commercial deployment. Larger suppliers, including GlobalFoundries and established optical component manufacturers, are chasing the same infrastructure transition.

The central contest is between optical performance and production reality. Sivers and SemiNex need more power, more wavelengths, and better thermal behavior without sacrificing efficiency or dependable manufacturing. A successful development program would strengthen their position in AI data center optics, but the announcement alone does not establish volume demand.

The $3.4M Program Behind the Google News Headline

Sivers and SemiNex are treating the light source as a system-level constraint, not another interchangeable optical component.

The companies announced the program on August 13, 2026. Sivers operates photonics development and manufacturing activities from Glasgow, Scotland, while SemiNex is based near Boston, Massachusetts. The initial program value is approximately $3.4 million.

According to the program announcement, the work covers three related product areas. Each addresses a different limitation emerging as AI systems require denser optical connectivity.

The first area is high-power external laser sources for co-packaged optics. Co-packaged optics, or CPO, places optical engines close to a processor or network switch. This shortens demanding electrical connections but changes how designers supply, distribute, and manage laser light.

The second area involves high-channel-count distributed feedback laser arrays. A DFB laser produces light at a tightly controlled wavelength. Combining several wavelengths allows a fiber to carry multiple channels through wavelength-division multiplexing.

The third area is semiconductor optical amplifier gain stages. An SOA is a chip that increases an optical signal’s strength. The program aims to use those amplifiers to extend link reach as systems add channels and increase throughput.

These elements address one shared problem. Optical architectures require enough usable power at every wavelength after accounting for splitting, coupling, modulation, and transmission losses. More channels create more opportunities for power to disappear before a signal reaches its destination.

The companies say their development work will concentrate on the design and processing of indium phosphide, usually shortened to InP. This compound semiconductor is widely used for lasers and optical amplifiers at communications wavelengths. Silicon handles many electronic and photonic functions well, but it does not efficiently generate laser light.

Sivers brings experience in precision DFB lasers, arrays, and wafer processing. Its InP100 platform uses four-inch wafers and a library of qualified process modules. The company says it supports wafer-level coating and testing, which can reduce later assembly work.

SemiNex brings high-power InP lasers, amplifier designs, and epitaxial engineering. Epitaxy is the controlled growth of semiconductor layers that determine a device’s electrical and optical behavior. Small changes in these layers can affect output power, efficiency, wavelength, and heat tolerance.

This division of expertise explains why the collaboration matters. The program is not simply matching a finished Sivers product with a finished SemiNex product. It combines design and process work around a set of demanding light-source requirements.

Sivers CEO Vickram Vathulya described advanced light sources as being as important as compute, memory, and silicon photonics inside AI systems. SemiNex CEO Ronald Moore was more direct about the engineering constraint.

“The optical layer is becoming the constraint on how far AI infrastructure can scale,” Moore said in the announcement. He argued that existing devices cannot meet future requirements simply by being driven harder.

That claim remains a company position, not an independently verified conclusion. Still, it identifies the program’s technical thesis. Optical scaling increasingly depends on the properties of the light source itself.

The google news headline captures the funding amount, but the development schedule deserves equal attention. Sampling and early production are not expected until the second half of 2027. Customers must then qualify the devices inside complete optical systems.

That gap creates the article’s main tension. Sivers and SemiNex have defined a plausible technical target, but market relevance depends on execution beyond a laboratory result.

Why AI Data Center Optics Need More Than Faster Transceivers

AI clusters are forcing optical links deeper into computing systems, where power, heat, and reliability become tightly connected.

Traditional data center networks place pluggable optical transceivers at a switch’s faceplate. Electrical traces carry signals from the switching silicon to those modules. This familiar design offers replaceability and separates sensitive optical components from the hottest chips.

However, electrical connections become harder to manage as speeds and switch bandwidth rise. Longer copper traces consume more energy and suffer greater signal loss. Engineers can compensate with additional signal-processing circuitry, but that adds power and heat.

CPO reduces those electrical distances by moving optical conversion closer to the switching chip. The approach can improve bandwidth density and energy efficiency. It also places optical hardware in a harsher thermal and operational environment.

The laser does not always have to sit inside that package. An external laser source can generate light elsewhere and deliver it through fiber. This arrangement separates the light-producing component from the hottest silicon while allowing several optical engines to share a source.

Yet an external source needs enough power to survive distribution losses. It must also supply stable wavelengths across operating temperatures. Failure behavior, serviceability, coupling loss, and redundancy all affect whether the architecture works outside a demonstration.

Sivers and SemiNex are targeting that collection of requirements. Their proposed external sources would serve both scale-up and scale-out systems.

Scale-up links connect computing resources inside a tightly coupled system. These links prioritize low latency and dense communication among accelerators. Scale-out links connect larger groups of systems across racks or clusters, often over longer distances.

Both architectures use optical bandwidth differently. Scale-up environments can demand dense connections over relatively short reaches. Scale-out fabrics cover broader physical topologies and introduce different link budgets.

A link budget tracks gains and losses across an optical path. Every connector, waveguide, coupler, splitter, and modulator affects the amount of light available at the receiver. Designers need sufficient margin to handle manufacturing variation and component aging.

Adding wavelengths raises total capacity without adding a separate fiber for every channel. However, each wavelength needs usable power and stable spectral behavior. A multi-wavelength array must therefore control several emitters without allowing heat or fabrication variation to undermine performance.

SemiNex has promoted O-band amplifiers with saturation output exceeding one watt. O-band refers to the wavelength region around 1,310 nanometers, commonly used for short-reach data communications. The company also claims power-conversion efficiency above 25 percent for its latest amplifier approach.

Those figures appeared in a sponsored InP photonics webinar, not an independent comparative test. They show the engineering direction, but buyers will require device data under their own operating conditions.

Sivers separately advertises DFB outputs up to 100 milliwatts across the O-band. It positions those products for 800-gigabit, 1.6-terabit, CPO, and optical I/O applications.

Combining a laser array with amplifier stages offers several possible design paths. Amplifiers can increase total optical power or compensate for distribution losses. They can also introduce noise, heat, and additional control requirements.

That tradeoff is why raw output power cannot determine the program’s success. Wall-plug efficiency matters because it measures optical output against electrical input. Thermal stability matters because wavelengths and output levels can shift as temperatures change.

Reliability matters even more in a hyperscale deployment. Replacing a failed part inside a densely integrated system costs more than replacing a conventional pluggable module. External light sources can improve serviceability, but only if the architecture isolates failures effectively.

The transition therefore pressures optical component suppliers, module makers, foundries, and system designers simultaneously. Each group must decide where the laser belongs, how much power it should provide, and who owns system qualification.

This makes AI data center optics a broader coordination problem. It is not enough for one component to post a higher laboratory measurement. The complete link must perform predictably across thousands of systems.

InP Light Sources Are Becoming a Competitive Bottleneck

The Sivers-SemiNex InP program enters a crowded race to control the light source connecting increasingly expensive compute.

Sivers is already building a network of partnerships around this opportunity. It has worked with Ayar Labs on optical I/O and has announced projects involving Jabil, O-Net Technologies, Enablence Technologies, POET Technologies, and GlobalFoundries.

Each relationship addresses a different part of the optical stack. That breadth can expand Sivers’ addressable market. It also shows how much coordination stands between a laser design and a deployed data center system.

Ayar Labs develops in-package optical I/O for computing systems. In late 2024, the companies discussed preparing Sivers laser arrays for manufacturing at scale. Ayar Labs said it intended to use development funding and product pre-purchases to support that preparation.

At the time, Sivers expected shipment volumes to ramp over an 18-to-24-month horizon. That forecast was forward-looking and tied to another company’s product schedule. The current SemiNex program introduces a separate 2027 target rather than confirming that earlier deployment.

Jabil represents a more conventional transceiver route. In 2026, the companies announced work on a 1.6-terabit pluggable optical transceiver using Sivers DFB technology. Pluggable modules remain important because operators understand their deployment and maintenance model.

The Jabil collaboration puts Sivers on both sides of an architectural transition. It can supply lasers for established pluggables while pursuing external sources and CPO.

That flexibility reduces dependence on one deployment model. However, it also requires the company to manage several development paths and partner schedules. Engineering resources, manufacturing capacity, and customer qualification can become limiting factors.

GlobalFoundries adds another competitive reference point. Its silicon photonics strategy combines manufacturing scale with advanced optical engines. Sivers announced that its laser arrays would support selected GlobalFoundries platforms targeting AI infrastructure.

The GlobalFoundries project demonstrates that laser suppliers do not compete only through standalone specifications. They compete for placement within foundry, packaging, and module ecosystems.

Large photonics vendors also have established manufacturing operations, customer relationships, and qualification histories. Companies such as Lumentum, Coherent, and Broadcom participate across portions of the optical supply chain. Their scale creates a demanding comparison for smaller suppliers.

Sivers and SemiNex are pursuing differentiation through specialized InP design, higher output, and customization. That can appeal to customers whose architecture exceeds the limits of standard components. Customization can also lengthen development cycles and fragment production.

The primary contest is therefore performance versus manufacturing readiness. A customer may value a laser with more power or wavelengths. It will still require repeatable wafers, acceptable yields, stable packaging, and predictable supply.

Sivers says its platform allows on-wafer coating and testing. Testing devices before wafer separation can identify failures earlier and reduce later packaging costs. The benefit depends on actual yield, throughput, and correlation with packaged performance.

SemiNex describes a hybrid manufacturing model and has announced cleanroom expansion. It says the investment includes automated testing and inspection for volume production of known-good die. Known-good die are chips tested before being integrated into more expensive assemblies.

These capabilities fit the needs of AI data center optics. They do not yet establish that the joint program can meet hyperscaler volumes. Neither company disclosed a named customer, committed unit quantity, or production purchase order for this project.

That omission is understandable at an early development stage. Customers often keep optical architectures confidential while comparing suppliers. Still, it limits what readers should infer from the $3.4 million value.

The figure signals funded work and a defined commercial relationship. It does not represent product revenue from a completed deployment. It also does not disclose how the value is divided between engineering services, prototypes, equipment, or future components.

This distinction can disappear as the announcement travels through google news and investor discussions. The headline sounds like immediate entry into AI data centers. The underlying program is a development pathway with sampling nearly a year away.

What the Announcement Does Not Prove

The program validates a development agenda, but it does not validate product yield, customer adoption, or data center economics.

The first uncertainty concerns technical integration. A high-power laser or amplifier can perform well as an individual device while creating problems inside a complete optical system. Coupling, packaging, controls, and cooling can change the result.

Higher optical power can improve link margin. It can also increase electrical consumption and heat generation. If efficiency does not rise with output, the source may transfer a networking constraint into the facility’s power budget.

Multi-wavelength operation creates another test. Customers need stable power across every channel, not just a strong average. Uneven output can force additional calibration or reduce usable capacity.

Thermal drift can change both wavelength and power. That behavior matters near processors and switches that operate under variable workloads. The companies must show stability across realistic temperature cycles and operating lifetimes.

The second uncertainty concerns manufacturing yield. InP fabrication uses different materials and processes from mainstream silicon manufacturing. Defects or small process variations can affect optical performance, especially across complex arrays.

Sivers’ wafer-level methods are designed to improve screening and production economics. SemiNex’s process expertise is intended to raise device performance. The partnership must demonstrate that those strengths remain compatible at repeatable production volumes.

The third uncertainty is packaging. Optical alignment demands precision, and packaging can represent a substantial portion of component complexity. Arrays may reduce individual alignment steps, but they require consistent spacing and wavelength behavior across several emitters.

External light-source architectures also need connectors, fibers, monitoring, and redundancy. A source with impressive chip-level performance can lose its advantage through an expensive or fragile assembly.

The fourth uncertainty concerns demand timing. AI infrastructure spending is high, but architecture transitions do not occur uniformly. Operators can continue deploying pluggable optics while testing CPO in selected systems.

CPO adoption depends on switch designs, packaging standards, service models, and system economics. Different customers can choose different approaches. Some deployments may use external lasers, while others retain lasers closer to optical engines.

The Sivers-SemiNex program targets several configurations, which gives it flexibility. However, supporting multiple designs can dilute standardization benefits. Customers may request unique wavelength, power, and packaging combinations.

The fifth uncertainty is competitive response. Larger suppliers can improve existing devices, acquire specialized technology, or bundle optical sources with other components. Foundries can also expand partnerships with multiple laser vendors.

Sivers reported a large opportunity pipeline in 2026, but a pipeline is not contracted revenue. Opportunities can change in value, timing, or probability. Readers should separate market engagement from recognized sales.

The company’s 2025 annual report said total revenue rose 40 percent to SEK 307 million. It also reported 17 percent growth in its photonics business. Those figures show operational momentum before the SemiNex announcement.

However, the same report described continuing work toward production ramps. The business still depends on customers moving from development activity into repeatable product orders.

A further issue is concentration. AI optical programs often involve a small number of large buyers. Winning one design can materially affect a supplier, while a delay or redesign can have the opposite effect.

No independent customer has publicly confirmed adoption of the joint light sources. There is also no disclosed performance comparison against competing production devices. The companies have provided targets, technical reasoning, and a schedule.

That is enough to treat the announcement as credible development news. It is not enough to call the optical bottleneck solved.

The most useful reading of the google news event is narrower. Two InP specialists are pooling resources around a specific systems problem. Their success will be measured through samples, qualification, and orders rather than announcement language.

Three Signals Will Decide Whether the Program Matters

The next evidence must move from engineering intent to customer-visible execution.

The first signal is the delivery of customer samples. Sivers and SemiNex target sampling and early production ramps during the second half of 2027. Before that window, investors and customers should look for prototype milestones or updated qualification schedules.

Samples would show that the companies converted their joint design work into physical devices. They would not prove production readiness. Still, a named sampling milestone would strengthen the case that the program remains on schedule.

A delay without a clear technical explanation would weaken that case. Photonics programs often encounter packaging, yield, or thermal issues after early device results. Schedule changes can reveal where the real bottleneck sits.

The second signal is third-party technical validation. Useful evidence would include measured power per wavelength, wall-plug efficiency, thermal performance, lifetime testing, and array uniformity. Results should describe operating conditions and packaging.

Customer qualification would carry more weight than a company presentation. A system partner naming the devices inside an optical engine would connect component claims to a real architecture.

Independent testing could also clarify whether SemiNex’s high-power amplifier approach improves total link economics. More output matters only when the complete source remains efficient, stable, and serviceable.

The third signal is a commercial commitment. A production purchase order, disclosed customer program, or repeatable product revenue would separate the partnership from early development work.

The amount matters less than the structure. A multi-year order or capacity reservation would suggest that a customer expects continuing demand. A one-time engineering payment would provide weaker evidence of deployment.

Sivers’ other partnerships provide useful checkpoints. Progress with Ayar Labs, Jabil, GlobalFoundries, O-Net, Enablence, or POET can show whether its laser platform works across architectures.

Success in pluggable modules would not automatically validate CPO. However, it could demonstrate manufacturing quality and customer confidence. Those capabilities transfer more readily than a particular module design.

The same logic applies to SemiNex. Its claimed amplifier performance needs to appear in qualified assemblies. Cleanroom expansion becomes strategically important only when equipment supports stable output and customer deliveries.

Readers should also watch whether the companies narrow their product target. Early programs often cover several possible devices before customer feedback identifies the strongest configuration.

A sharper specification could indicate progress rather than retreat. It would show that system requirements are becoming concrete. Constantly expanding the scope without samples would signal less discipline.

The program arrives as compute vendors and network suppliers search for ways around electrical interconnect limits. That demand creates a real opening for specialized photonics companies. It also attracts rivals with deeper manufacturing and sales resources.

Sivers and SemiNex do not need to dominate every optical architecture. They need to secure a defensible role in the light-source layer. That requires performance customers cannot obtain easily elsewhere and production they can trust.

The $3.4 million commitment gives the teams resources and a public schedule. It also makes future progress easier to judge. By the second half of 2027, the market should expect more than another partnership announcement.

The google news headline is therefore the starting point, not the result. Watch for sample delivery first, third-party system data second, and a production commitment third.

If all three appear, the partnership will have converted an optical scaling thesis into a commercial position. If they do not, higher power will remain a compelling specification without proven data center adoption.

For engineers and infrastructure buyers, the immediate question is practical: where will optical power become the limiting factor in the next system design? Track link budgets, thermal assumptions, and service requirements before choosing an architecture. Those details will reveal whether the Sivers-SemiNex approach addresses a real deployment constraint or only a promising laboratory target.