GlobalFoundries Fotonix Was Early. Now Silicon Photonics Is Becoming AI Infrastructure. GF Fotonix Silicon photonics Co-packaged optics Optical I/O Nvidia networking Lightmatter Marvell Celestial TSMC COUPE AI factories

Section 1  ·  Historical frameWhat the 2022 article got right

The 2022 SemiAnalysis piece, GlobalFoundries Fotonix, The Leading Silicon Photonics Foundry For Co-packaged Optics And Processing, made a simple but unfashionable argument.^[1] Silicon photonics would reshape semiconductor design because I/O demand was rising faster than electrical I/O efficiency. Page 2 framed the data-center network bandwidth scaling problem and the gap between switch bandwidth growth and electrical I/O scaling. Page 3 walked through GF's monolithic electro-optic integration concept across modulators, germanium photodetectors, waveguides, fiber couplers, laser attach, and CMOS plus photonic integration. Pages 6 and 7 connected that to Nvidia's long-reach signalling and photonics-link diagrams, with page 9 showing co-packaged photonics with optical engines near GPUs and switches.^[1]

The 2022 piece positioned GlobalFoundries Fotonix as a leading silicon photonics foundry platform and highlighted partnerships with Nvidia, Broadcom, Marvell, Cisco, Macom, Ayar Labs, Lightmatter, PsiQuantum, Ranovus, Xanadu, Ansys, Cadence, and Synopsys. It made an argument that often gets missed in foundry comparisons: GF's advantage was not only manufacturing capability, but the ecosystem around PDKs, simulation, packaging, fiber attach, and customer enablement. Page 14 walked through packaging features such as copper pillars, on-die laser attach, and V-groove fiber attach. Page 17 explained why PDKs, simulation, and the EDA partners around Cadence, Synopsys, and Ansys actually matter for silicon photonics manufacturing.^[1]

Underneath all of that, the strategic argument was that GF was differentiated because it could serve silicon photonics, RF, SiGe, SOI, and specialty processes rather than only commodity mature-node logic. The article also called out something that has aged unusually well: Nvidia would eventually need co-packaged optics as AI systems scaled and electrical I/O power became harder to manage. Four years later, that part of the thesis is no longer a forecast. It is a product roadmap.

Section 2  ·  Bottleneck shiftFrom compute to communication

The AI race is usually described as a compute race. That description is incomplete. AI factories need massive data movement between GPUs, switches, memory pools, racks, and data centers. Larger models require more accelerators. More accelerators require more networking. More networking means more I/O power. Copper links face reach, power, heat, and signal-integrity limits at the bandwidths AI clusters now demand. Optical links can move more data over longer distances with better power efficiency, and silicon photonics lets optical communication be manufactured using semiconductor-like processes.

The next bottleneck is not only how fast GPUs calculate. It is how efficiently the AI factory can move data.

That shift in framing matters for foundry strategy. If AI infrastructure were only GPUs, then advanced logic would be the only foundry conversation that mattered. But once you add memory, packaging, networking, optics, power, and software, the foundry conversation gets wider. Several specialty platforms become strategic in their own right. Silicon photonics is one of them.

Section 3  ·  The wedgeWhy GF Fotonix matters

GF's wedge is not leading-edge AI logic. It is specialty manufacturing for optical I/O and electro-optical integration. The Fotonix announcement in 2022 framed it as a single platform that brings together photonics, RF and CMOS, modulators, photodetectors, waveguides, couplers, laser attach, packaging, and PDK support, with customer quotes from Cisco on custom silicon photonics, Nvidia on optical interconnect, Marvell on SiGe, Ayar Labs on PDK and process, PsiQuantum on quantum photonics, and Ranovus on co-packaged optics for AI.^[2] GF's March 2026 photonics and advanced packaging investor webinar updates the roadmap with 1.6T and 3.2T photonics portfolio examples, pluggables, co-packaged optics, optical compute, and quantum interconnect, alongside 300mm and 200mm manufacturing across New York and Singapore.^[3]

In silicon photonics, the product is not just the wafer process. The product is the process plus PDK plus packaging plus fiber attach plus simulation ecosystem.

That phrasing is the practical translation of the 2022 thesis. A leading silicon photonics foundry has to make the wafer process, the PDK that designers actually use, the packaging that turns a wafer into a working module, the fiber attach that gets light in and out, and the simulation ecosystem that allows complex photonic systems to be designed reliably. GF is one of the small set of foundries trying to do all of those things at once. Most pure-play foundries do not even attempt it.

Section 4  ·  Demand validatorNvidia validated the market

The strongest validation of the GF Fotonix thesis is not another GF press release. It is Nvidia productising silicon photonics for AI factories. Nvidia's silicon photonics page describes a co-packaged optics strategy in which CPO switches replace pluggable transceivers with silicon photonics integrated onto the same package as the switch ASIC. Nvidia claims approximately 5x better power efficiency and approximately 10x higher network resiliency versus pluggable transceivers, and positions Quantum-X InfiniBand Photonics and Spectrum-X Ethernet Photonics as platforms for massive-scale AI infrastructure.^[6]

The strongest validation of the GF Fotonix thesis is not another GF press release. It is Nvidia productising silicon photonics for AI factories.

The product claims are Nvidia's. They should be read as Nvidia's framing, not as endorsed forecasts. The strategic point holds either way. Once Nvidia ships co-packaged optics inside switch platforms designed for million-accelerator-class AI infrastructure, the question of whether optical I/O matters for AI is settled. The remaining questions are about who manufactures it, where it is packaged, and how quickly the ecosystem can scale.

Section 5  ·  From platform to capacityGF expanded with Advanced Micro Foundry

In 2025, GlobalFoundries acquired Advanced Micro Foundry (AMF), a Singapore-based silicon photonics foundry, to accelerate its silicon photonics global leadership and expand its AI infrastructure portfolio. GF says the acquisition expands its silicon photonics technology portfolio, capacity, and R&D in Singapore, complements its US capabilities, and supports communications, computing, LiDAR, sensing, AI data centers, and other applications on a 200mm platform with plans to scale toward 300mm as demand grows. GF describes the resulting company as the largest pure-play silicon photonics foundry by revenue.^[4]

The interpretation is straightforward. The 2022 version of this story was GF has a strong Fotonix platform. The 2026 version is GF is building a global silicon photonics manufacturing footprint. That includes a New York anchor, a Singapore expansion through AMF and its A*STAR R&D ties, and a 200mm-to-300mm scaling path. None of that guarantees market share in a contested category. All of it is the kind of capacity story that has to be in place before a foundry can credibly serve AI-scale customers.

Section 6  ·  PackagingThe New York Advanced Packaging and Photonics Center

Photonics is not only a wafer-process problem. It is a packaging problem. GF's January 2025 announcement of a New York Advanced Packaging and Photonics Center frames the program as supporting advanced packaging and test for AI, automotive, aerospace and defense, communications, and other markets, with capabilities across silicon photonics packaging, Trusted Foundry flows, wafer-to-wafer bonding, 3D and heterogeneous integration, assembly, and testing. GF expects approximately US$575M of investment plus approximately US$186M of R&D over more than 10 years.^[5]

The hard packaging problems are not exotic in name. They are exotic in volume engineering. Optical coupling has to be aligned within tight tolerances. Fiber attach has to be reliable. Laser attach has to be thermally stable. Electrical-to-optical interfaces have to behave under real workloads. Thermal management has to keep photonic devices in their operating window. Reliability testing has to qualify modules across data-center duty cycles. Advanced packaging has to combine all of the above with logic, memory, and networking in a single coherent module. Trusted supply-chain flows have to be available for customers with security requirements.

Silicon photonics does not become infrastructure until it can be packaged, tested, and manufactured reliably.

Section 7  ·  Into the packageLightmatter shows optical I/O moving into the AI package

Lightmatter's March 2025 Passage M1000 announcement framed the product as a 3D photonic superchip for next-generation XPUs and switches. Lightmatter claimed approximately 114 Tbps of total optical bandwidth, described a multi-reticle active photonic interposer larger than approximately 4,000 mm², and positioned the system to connect thousands of GPUs in a single domain. Lightmatter said Passage M1000 uses GF Fotonix as its underlying silicon photonics process.^[7]

The strategic read is consistent with the 2022 thesis. Silicon photonics is moving from the edge of the rack toward the package, the interposer, the switch, and the accelerator complex. Lightmatter is one example, but the pattern is wider. The Passage M1000 claim is a company claim. The structural direction it represents is shared across multiple platforms in the same window.

Silicon photonics is moving from the edge of the rack toward the package, the interposer, the switch, and the accelerator complex.

Section 8  ·  ConsolidationMarvell, Celestial AI, and AMD

Marvell announced in December 2025 that it would acquire Celestial AI, citing the value of Celestial AI's Photonic Fabric for package, system, and rack-level optical I/O, and framing the deal around the argument that copper must give way to optics within racks, systems, and even packages. Marvell described Celestial's first chiplet as delivering 16 Tb/s of bandwidth and positioned the combined offering as a multi-rack scale-up fabric for next-generation data centers.^[8] AMD acquired Enosemi in 2025 to expand its co-packaged optics offerings for AI systems, signalling that optical I/O is now strategic to AI accelerator companies as well as to networking companies.^[9]

Big semiconductor companies are not treating optics as a science project anymore. They are buying it.

That is the consolidation signal. When networking incumbents and accelerator vendors both decide to internalize optical I/O capability through acquisition, the category is past its science-project phase. It is now a strategic asset. That changes how downstream manufacturing decisions get made and tightens the timeline on the photonics foundry conversation.

Section 9  ·  The strategic competitorTSMC is also building photonics

GF may be one of the strongest pure-play silicon photonics foundries, but TSMC controls the most important advanced logic and advanced packaging ecosystem. TSMC has published work on COUPE, the Compact Universal Photonic Engine, as a silicon-photonics integration platform aimed at HPC applications.^[10] Read alongside TSMC's CoWoS, InFO, SoIC, and broader 3DFabric platforms, the picture is that TSMC will not concede photonics to specialty foundries quietly. Where photonics is tightly coupled to leading-edge AI logic and advanced packaging, TSMC is a natural integration partner.

GF's leadership is real, but not uncontested.

Section 10  ·  EcosystemThe optical AI infrastructure map

Section 11  ·  The proof pointsWhat GF must prove

GF has a strong position. The market is not guaranteed. The honest version of the story has a clear list of things GF still has to demonstrate at AI infrastructure scale.

The question is not whether silicon photonics matters. The question is who can manufacture it at AI infrastructure scale.

Section 12  ·  EvidenceEvidence ledger

Section 13  ·  Risk registerRisks and limitations

This essay is an analysis of public disclosures and historical context. It is not investment advice. The honest risks against the read above run in several directions, and they are listed here so the argument can be stress-tested.

Section 14  ·  Bottom lineBottom line

Section 15  ·  DefinitionsGlossary

Section 16  ·  MethodSources and method notes