The Real AI Bottleneck Isn’t Compute — It’s the Optical Supply Chain
$13 billion in six months signals a shift from chips to connections
In the span of six months, more than $13 billion flowed into a part of the AI stack that, until recently, barely registered outside specialist circles. Not into GPUs. Not into memory. Into the connections between them.
As companies like NVIDIA, Marvell Technology and Credo Technology Group move aggressively into optical technologies, a new constraint is emerging — one that has less to do with compute and more to do with whether the industry can physically move data at the scale AI now demands.
This is not a technology transition. It is a supply chain constraint revealing itself.
From compute to connectivity
For the better part of a decade, AI scaling has been framed as a compute problem: faster GPUs, more memory, larger clusters.
That framing is now breaking down.
Modern AI systems are no longer defined by individual chips, but by how effectively thousands of them operate as a single system. Training clusters and inference fabrics depend on synchronization, latency and bandwidth across massive networks of accelerators.
Performance is no longer bounded by compute alone. It is bounded by the ability to move data. And at scale, that is where the system begins to fail.
Copper interconnects — the physical layer of today’s data centers — degrade under the weight of AI workloads. Power consumption rises, heat becomes unmanageable and bandwidth density plateaus. At hyperscale, these are not incremental inefficiencies. They are structural limits.
The industry’s response is optical.
The quiet capital shift
Between late 2025 and early 2026, that response turned into a coordinated capital movement.
- Marvell Technology acquired Celestial AI
- NVIDIA committed $2B each to Lumentum and Coherent, alongside long-term purchase commitments
- Credo Technology Group moved to acquire DustPhotonics
- Molex acquired Teramount
- Ayar Labs raised $500M to accelerate co-packaged optics
- GlobalFoundries expanded into silicon photonics manufacturing capacity
Individually, these are strategic moves.
Together, they form a pattern.
Capital is moving simultaneously across every layer required to make optical interconnects work at scale.
Design.
Lasers.
Packaging.
Manufacturing.
This is not venture exploration. This is infrastructure being secured under time pressure.
The misconception: silicon photonics alone
The dominant narrative frames silicon photonics as the solution — the technology that will replace copper and unlock the next phase of AI scaling.
That narrative is incomplete.
Silicon photonics excels at routing light. It does not efficiently generate it.
Light sources — lasers — depend on III-V materials such as indium phosphide (InP). In practice, every scalable optical system is hybrid: silicon for integration, III-V materials for emission and advanced packaging to bind the system together.
“Silicon photonics is not the answer on its own. You need silicon photonics plus III-V materials such as InP to generate the light. And if you want to scale, you have to scale every link in the supply chain — otherwise it becomes a dead-end street.”
Martijn Heck
Professor of Heterogeneous Integration, Eindhoven University of Technology
The implication is straightforward, and often overlooked: The bottleneck is not a component. It is the system.
Where the constraint is forming
That system is already under strain.
Demand for optical components is accelerating faster than production capacity can follow. Suppliers are signaling multi-year visibility into full order books. Large buyers are no longer just purchasing components — they are securing future capacity.
This changes how the market functions.
- Access becomes priority-based
- Capacity is pre-allocated
- Pricing becomes secondary to availability
For smaller players or late entrants, the implication is structural: access to the next generation of AI infrastructure may simply not be guaranteed.
At the same time, constraints are emerging across multiple layers:
- Materials → InP wafer supply remains limited
- Manufacturing → photonics fabs are not designed for hyperscaler-scale volumes
- Packaging → fiber-to-chip coupling and co-packaged optics remain complex
- Geopolitics → critical inputs such as indium are increasingly regulated
Individually, each constraint is manageable. Together, they form a bottleneck that is difficult to resolve quickly.
The shift to manufacturable systems
The recent deal activity also reveals a deeper shift in how value is assigned.
Historically, photonics has been evaluated on technical merit: better performance, lower latency, novel architectures.
That lens is no longer sufficient.
- Marvell Technology acquired Celestial AI before meaningful revenue
- Ayar Labs is deploying capital into production and test infrastructure
- Molex moved to close a packaging gap through Teramount
The pattern is consistent: Value is shifting from design to the ability to deliver complete, manufacturable systems at scale.
In other words, the constraint is no longer innovation. It is execution.
AI as a physical system
The broader implication is that AI is entering a fundamentally different phase.
For years, progress was defined by improvements in computation. More powerful chips enabled larger models, which in turn enabled new applications.
That loop is now constrained by something more fundamental.
AI systems require:
- physical connectivity
- energy transport
- materials
- manufacturing capacity
These are not software problems. They are industrial ones. AI is no longer just a compute problem. It is a physical system with real-world limits.
The emerging hierarchy of scarcity
The next phase of AI infrastructure will not be defined by demand, but by which constraints bind first.
Three stand out:
- Laser supply (III-V materials)
Without scalable light sources, optical systems cannot expand - Manufacturing and packaging capacity
Integration remains complex and difficult to scale - Strategic access to suppliers
Capacity is increasingly locked in by large buyers
Layered on top of this is a fourth dimension: geopolitics.
If indium phosphide and its upstream materials become strategically constrained, the optical stack begins to resemble other critical supply chains — where control, not capability, defines outcomes.
Conclusion
Six months ago, photonics was a niche domain. Today, it is the most aggressively funded bottleneck in AI infrastructure.
The narrative has shifted — from chips to connections, from compute to interconnect, from design to supply chain. The constraint is no longer theoretical.
It is already being secured.
The next decade of AI will not be defined by who builds the fastest chips, but by who controls the pathways through which data — and light — can move.
Caption:
AI scaling is no longer constrained by compute, but by the physical limits of moving data. Optical interconnects shift the bottleneck from chips to the supply chain that enables light.
Credit:
Altair Media (Illustration)
