The Other Leading Edge.Original analysisNot investment advice
GlobalFoundries did not win the 2nm race. It stopped running that race. But AI is expanding the definition of leading edge. The future is not only smaller transistors. It is optical data movement, low-power edge intelligence, RF connectivity, power efficiency, automotive reliability, and heterogeneous integration.
In semiconductors, “leading edge” usually means one thing. 2 nm. Gate-all-around. EUV. Transistor density. TSMC, Samsung, Intel. By that definition, GlobalFoundries left the race years ago. It stopped chasing 7 nm. It stopped trying to match TSMC at the most advanced logic node. It chose a different path. That made the easy conclusion tempting: GF is no longer leading edge.
The uploaded 2021 SemiAnalysis piece argued the opposite. It said GF was still leading edge, but not in the narrow Moore’s Law sense. GF was pushing advanced manufacturing in silicon photonics, FD-SOI, GaN-on-silicon, 3D integration, heterogeneous integration, RF, and other specialty platforms.1
In 2026 that argument looks more important. AI does not only need smaller logic transistors. It needs optical interconnect. It needs power efficiency. It needs memory movement. It needs RF connectivity. It needs sensors. It needs automotive-grade reliability. It needs edge devices that can run intelligence under strict power budgets.
That is not one frontier. It is many frontiers. And GF is trying to own the other leading edge.
The correct claim is not that GlobalFoundries is a leading-edge foundry like TSMC. The correct claim is that GlobalFoundries is leading edge in specialty manufacturing. In an AI world, that matters more than people think.
I. The 2021 thesis
In June 2021, Dylan Patel published a SemiAnalysis piece arguing that “leading edge” should not only mean the smallest logic node. GlobalFoundries had walked away from the 7 nm race, but it had not walked away from advanced manufacturing. The piece highlighted FD-SOI, 22FDX, 12FDX, silicon photonics, GaN-on-silicon, silicon nitride, RF platforms, 3D integration, and heterogeneous integration as places where GF was genuinely advanced. It also discussed Lightmatter’s photonic compute stack as an early example of what GF’s 45CLO and 90WG silicon photonics processes could enable when combined with GF logic.1
I revisited that piece because the broader framing has aged into the dominant question of AI infrastructure. Density alone no longer describes the frontier.
GlobalFoundries left the Moore’s Law logic race, but it did not leave the leading edge. It shifted toward specialty frontiers where the bottleneck is not just transistor density: photonics, RF, power, FD-SOI, and heterogeneous integration.
II. GF is not fighting TSMC’s war
The cleanest way to read GF is to take it on its own terms. It is not trying to beat TSMC at 2 nm. If the question is who manufactures the most advanced GPU logic, the answer is not GF. If the question is who manufactures the specialty technologies around AI, automotive, RF, photonics, power, and edge intelligence, GF becomes much more interesting.
Density only
- Frontier · smallest logic node, EUV, GAA transistors.
- Winners · TSMC at 3/2 nm, Intel 18A/14A, Samsung.
- Test · can you print a smaller transistor?
- Limit · misses memory, photonics, RF, power, automotive.
Many frontiers
- Frontiers · logic + memory + photonics + RF + power + automotive + integration.
- Winners · TSMC for logic, GF for selected specialties, others elsewhere.
- Test · can you move, power, sense, connect, and integrate intelligence?
- Limit · needs a wider definition of “advanced.”
GF is not trying to replace TSMC. It is trying to be essential in the layers that pure logic-node narratives miss.
III. Photonics is the centre of the story
AI systems are data-movement machines. GPUs need HBM. Racks need switches. Switches need optics. AI clusters need lower power per bit. Copper does not disappear overnight, but it starts to hurt at the highest bandwidths. GF’s silicon photonics platform is positioned for exactly that pressure. The company describes a monolithic electro-optical integration approach on 300 mm wafers with PDK support for photonics and RF-CMOS co-design, plus packaging hooks that include passive fiber attach and 2.5D integration.2
The 2021 SemiAnalysis piece argued that GF’s 45CLO process could integrate active and passive photonic components, and that co-packaged optics would push optics closer to the switch ASIC as bandwidth and power constraints grew.1 That direction has only intensified since.
GF wants to be part of the AI factory’s movement layer, not the GPU layer.
IV. SCALE turns photonics into an AI scale-up product
In 2026, GF gave its photonics strategy a sharper edge. The Q1 2026 results materials describe the launch of SCALE, the Silicon Photonics Co-packaged Advanced Light Engine, tailored to Optical Compute Interconnect (OCI) MSA requirements for modern AI scale-up architectures.3 Industry coverage of SCALE highlights support for CWDM and DWDM transmission and notes 8λ and 16λ bidirectional DWDM demonstrations on the platform.4
Scale-up is what makes many accelerators behave like one larger system. It needs high bandwidth, low latency, low power, reliable interconnect, and dense physical packaging. SCALE is GF’s attempt to be the foundry layer for the optical engine that lives inside that architecture.
AI factories need optical scale-up. GF wants to be the foundry layer for that optical engine.
V. AMF makes this a scale strategy
Silicon photonics is not only a technical bet. It is a manufacturing-scale bet. GF’s acquisition of Advanced Micro Foundry (AMF) in Singapore expanded its silicon photonics technology portfolio, production capacity, and R&D. GF positions itself as the largest pure-play silicon photonics foundry by revenue post-deal and frames the move around AI data centers, quantum, and optical connectivity. Reuters covered the acquisition with similar framing.56
A startup can demonstrate a photonic chip. AI infrastructure needs process control, PDKs, packaging, test, yield, reliability, fiber attach, customer trust, and global supply. AMF gives GF more of those operational levers in one place.
GF wants to be the photonics foundry for the AI infrastructure era.
VI. The U.S. trusted-manufacturing layer
The other pillar of GF’s 2026 strategy is U.S. capacity. GF announced a $16B U.S. investment plan across New York and Vermont, tied to AI-driven demand for power-efficient and high-bandwidth semiconductors. The plan specifically calls out silicon photonics, 22FDX, GaN-based power, 3D heterogeneous integration, and advanced packaging, and includes the New York Advanced Packaging and Photonics Center.7
In a world of AI geopolitics, where a chip is made can matter almost as much as what the chip does.
VII. FD-SOI becomes an edge-intelligence platform
FD-SOI puts an insulating layer under the transistor channel to reduce leakage and improve power behaviour. It supports body biasing, which lets designers tune performance and power. It is well-suited to low-power, cost-sensitive, always-on, RF-adjacent, and automotive/industrial systems.
GF’s AutoPro 150 announcement is the clearest 2026 signal. GF says the platform offers Auto Grade 1-ready embedded MRAM on its FDX platform, with endurance up to 500k cycles, sub-10 ns read speed, and operation up to 150°C, targeting automotive SoCs, software-defined vehicles, ADAS, and emerging physical-AI systems.8
Cars are becoming rolling compute platforms. Factories are becoming sensor networks. Robots and drones need efficient local intelligence. Industrial systems need secure, reliable, low-power chips. Physical AI does not always need 2 nm. It often needs the right specialty process.
FD-SOI is not dead. It is becoming a platform for low-power, reliable edge intelligence.
VIII. RF-SOI keeps GF inside the wireless stack
Wireless performance is not just about the modem. It is also RF front-end modules, switches, filters, power amplifiers, antenna tuning, radio platforms, process technology, and thermal/power behaviour.
Reuters reported that Soitec will supply Soitec’s 300 mm RF-SOI wafers to GF for the 9SW radio platform used in future 5G and Wi-Fi chips, with Soitec framing the move around 5G-Advanced and 6G needs for higher performance, better energy efficiency, and compactness.9
The modem gets the attention, but the RF platform decides how well the device actually talks to the world.
IX. GaN and power matter because AI needs electricity
AI data centers and physical AI systems create power pressure. Power electronics matter for data-center power delivery, conversion, EVs, defense systems, RF power, industrial systems, renewable energy, and robotics. GF’s U.S. investment plan calls out GaN-based power technologies as one of its five focus areas, alongside silicon photonics, 22FDX, 3D heterogeneous integration, and advanced packaging.7
This is the unglamorous half of the AI infrastructure story. The same factories that need optical interconnect also need GaN-class power devices to feed them.
The AI factory is an electrical machine before it is an intelligence machine.
X. Lightmatter was the early proof of heterogeneous integration
The 2021 SemiAnalysis piece used Lightmatter as a concrete example of what GF’s specialty platforms could enable. Lightmatter combined GF’s 90WG silicon photonics with GF logic to build a photonic compute stack, with 3D integration concepts that brought SRAM close to compute and reduced data movement. The point was not one startup, it was the integration pattern: logic plus photonics plus SRAM plus packaging.1
Five years later, that pattern looks like the early template for what AI infrastructure photonics is becoming.
XI. The new definition of leading edge
The old definition was tidy. Smallest node. Highest transistor density. Moore’s Law. EUV. Logic scaling. The AI-era definition is messier and more honest.
XII. Where TSMC still wins
This article is not pretending GF is the new TSMC. TSMC dominates the logic frontier, runs the largest advanced packaging operation in the industry, and manufactures the GPUs and AI ASICs at the centre of the current cycle. Its 2025 annual report frames AI/HPC as a primary driver of advanced process demand and packaging investment.11 ASML’s 2025 strategic report adds that AI requires leading-edge processors and a significant increase in DRAM compared with traditional compute, both of which lean on TSMC and the high-end memory makers.10
Logic frontier dominance
- Most advanced node · 3 nm shipping, 2 nm, A16, A14 in the pipeline.
- Advanced packaging at scale · CoWoS family for AI and HPC.
- Customer trust · neutral, deep ecosystem.
- Capex muscle · among the largest in the industry.
TSMC dominates the logic frontier. GF is betting that AI creates more than one frontier.
Quick terms
- Leading edge
- Most advanced manufacturing frontier, traditionally measured by logic-node scaling.
- Specialty foundry
- Foundry focused on differentiated non-leading-logic platforms such as RF, photonics, power, and embedded memory.
- FD-SOI
- Fully depleted silicon-on-insulator, useful for low-power and RF-adjacent applications.
- RF-SOI
- Silicon-on-insulator technology used for radio-frequency front-end chips.
- Silicon photonics
- Using silicon manufacturing methods to build optical data-movement components.
- CPO
- Co-packaged optics, moving optical engines close to compute or switch ASICs.
- SCALE
- GlobalFoundries’ Silicon Photonics Co-packaged Advanced Light Engine.
- eMRAM
- Embedded magnetoresistive memory.
- GaN
- Gallium nitride, useful for power and RF applications.
- Heterogeneous integration
- Combining different chip technologies into one system.
- 2.5D / 3D integration
- Packaging methods that integrate multiple dies closely.
- PDK
- Process design kit used by chip designers.
- Optical interconnect
- Using light to move data.
- Physical AI
- AI deployed into robots, vehicles, factories, sensors, and other real-world systems.
XIII. What could break the thesis
A serious piece needs counterarguments. The specialty-edge thesis has plausible failure modes.
- Value capture. Even if silicon photonics grows quickly, the highest value may accrue to system companies, hyperscalers, and advanced logic leaders rather than the foundry.
- Crowded photonics field. TSMC COUPE, Intel OCI, Broadcom CPO, Ayar Labs, Celestial AI, and Lightmatter are all moving around optical interconnect.
- Margin profile. Specialty foundry markets can be lower margin than leading-edge logic, and can be cyclical.
- FD-SOI ceiling. FD-SOI may stay important without ever becoming a large market.
- RF-SOI and GaN competition. Both are technically attractive and have multiple credible suppliers.
- Automotive timing. Auto demand is slow, qualification-heavy, and sensitive to OEM roadmaps.
- AMF integration. Singapore-to-300 mm scaling depends on operational execution.
- SCALE adoption. Customer wins, not specs, will determine whether SCALE matters.
- Pricing power. Specialty markets can be important without being highly profitable.
- Geopolitics cuts both ways. Trusted-manufacturing positioning helps in some scenarios and hurts in others.
The strongest bear case is simple: GF has good specialty technology, but the highest value still accrues elsewhere.
XIV. What to watch
Working checklist, not a prediction. The signals below would move first if the specialty-edge thesis is paying off.
- SCALE customer announcements.
- OCI MSA adoption beyond demonstration silicon.
- AMF integration progress in Singapore.
- 200 mm to 300 mm silicon photonics scale-up.
- Silicon photonics revenue growth at GF.
- AI data-center optical interconnect demand.
- Co-packaged optics deployments outside pilots.
- New York Advanced Packaging & Photonics Center milestones.
- Renesas / automotive FDX design wins.
- AutoPro 150 eMRAM qualification ramps.
- RF-SOI 9SW design wins.
- 5G-Advanced and 6G RF platform demand.
- GaN power and RF traction.
- Defense and aerospace trusted-foundry demand.
- Margin profile of specialty platforms.
- TSMC / Intel / GF competition in photonics.
- Lightmatter and other photonic-compute ecosystem adoption.
XV. The other leading edge
GlobalFoundries did not win the 2 nm race. It stopped running that race. But AI is expanding the definition of leading edge. The future is not only smaller transistors. It is optical data movement, low-power edge intelligence, RF connectivity, power efficiency, automotive reliability, embedded memory, heterogeneous integration, and trusted manufacturing.
That does not make GF the next TSMC. It makes GF something different: the foundry for the other leading edge.
1 Patel, D. (Jun 2021). GlobalFoundries Is A Leading-Edge Foundry Despite Claims Otherwise. SemiAnalysis. Historical anchor for the specialty-leading-edge framing, including FD-SOI, silicon photonics (90WG, 45CLO), GaN-on-silicon, silicon nitride, 3D and heterogeneous integration, and the Lightmatter photonic compute example. Used as inspiration only. No content, structure, or charts reproduced.
2 GlobalFoundries. Silicon Photonics. 300 mm silicon photonics manufacturing, monolithic electro-optical integration, photonics and RF-CMOS co-design with PDK support, and packaging-related framing including 2.5D integration and passive fiber attach.
3 GlobalFoundries (2026). First quarter 2026 financial results. Launch of SCALE (Silicon Photonics Co-packaged Advanced Light Engine), Optical Compute Interconnect MSA framing, OFC 2026 ecosystem context, and Renesas / FDX commentary.
4 Signal Integrity Journal. GlobalFoundries accelerates adoption of co-packaged optics with SCALE. SCALE details including CWDM/DWDM transmission support, 8λ and 16λ bidirectional DWDM demonstrations, AI scale-up framing, and copper interconnect limitations.
5 GlobalFoundries (Nov 2025). GlobalFoundries acquires Advanced Micro Foundry. AMF acquisition in Singapore, expansion of silicon photonics technology portfolio and capacity, framing as the largest pure-play silicon photonics foundry by revenue, plans to scale AMF’s 200 mm platform toward 300 mm, and AI/quantum/optical connectivity context.
6 Reuters (Nov 2025). GlobalFoundries buys Singapore’s Advanced Micro Foundry. Independent reporting on the AMF acquisition, AI data centers and quantum computing framing, and GF’s silicon photonics expansion.
7 GlobalFoundries. $16B U.S. investment plan. New York and Vermont sites, focus on silicon photonics, 22FDX, GaN-based power, 3D heterogeneous integration, and advanced packaging, with the New York Advanced Packaging and Photonics Center as a dedicated photonics packaging facility.
8 GlobalFoundries. AutoPro 150 eMRAM on enhanced FDX. Auto Grade 1-ready embedded MRAM, endurance up to 500k cycles, sub-10 ns read speed, operation up to 150°C, targeting automotive SoCs, software-defined vehicles, ADAS, and physical-AI systems.
9 Reuters (Dec 2024). Soitec to supply RF-SOI wafers to GlobalFoundries 9SW. Soitec 300 mm RF-SOI wafers for the GF 9SW radio platform, with 5G-Advanced and 6G context emphasising higher performance, energy efficiency, and compactness.
10 ASML (2025). 2025 Annual Report, strategic report section. AI requires leading-edge, high-performance processor chips and a significant increase in DRAM compared with traditional compute architectures.
11 TSMC. 2025 Annual Report. Robust AI-related demand, advanced packaging and 3D stacking investment, CoWoS context, and specialty-technology considerations.
- When AI Runs Out of Copper. Companion essay on optical I/O and the broader photonics race across Intel, Nvidia, Broadcom, TSMC, and Ayar Labs.
- The Inference Efficiency War. Companion essay on Qualcomm AI200/AI250 and cost-per-token inference infrastructure.
- The Custom Silicon Flywheel. Why hyperscalers turn their biggest workloads into chips.
- Nvidia’s Earnings Quality Test. AI capex, customer concentration, and the durability of Nvidia’s revenue.
- The AI Memory Tax. AI servers repricing DRAM, NAND, and consumer electronics.
- The AI Memory Wall. DRAM, HBM, packaging, and semicap as the new center of computing.
- The Boring Back-End Boom. Mature nodes, wirebonding, and packaging becoming strategic again.
- The Density Illusion. Why Moore’s Law became a system problem.
- Nvidia Built the AI Factory Anyway. Vertical system integration as the new moat.
- The Modem-to-Antenna War. Apple unbundling Qualcomm’s modem-RF stack.
- MediaTek and the Fragmented Compute War. A neutral fabless platform in a bifurcated compute world.
- The Dry Resist War. Patterning as a strategic process technology for AI-era chipmaking.
- The AI Field Manual. Reference layer for the AI stack: hardware, memory, models, agents, safety, economics.
This is Essay No. 024. The topics: intelligence, AI, systems, knowledge, and the questions underneath the questions everyone else is asking. If you read this far and disagreed with any part of it, write to me. I read everything.