The Layer That Decides AI: Why Connectivity — Not Compute — Is the Real Bottleneck
At OFC 2026, one overlooked company reveals where power in the AI economy is shifting
The dominant narrative in artificial intelligence remains focused on compute. Faster GPUs, larger clusters, more advanced models. It is a story driven by scale and visibility—by the companies that design chips and train systems.
But as that scale increases, a different constraint is emerging. Less visible, less discussed, but increasingly decisive.
The bottleneck is no longer how much data can be processed. It is how that data moves.
At OFC 2026 in Los Angeles, this shift is difficult to ignore. The industry’s transition toward 1.6 terabit per second interconnects and the growing emphasis on extreme energy efficiency point to a deeper structural problem. Moving data between systems—quickly, reliably and without excessive power consumption—is becoming the limiting factor in AI deployment.
“Optics is no longer ‘behind the scenes’ — it is the connecting infrastructure that enables performance, efficiency and scalable growth in the AI-driven economy.”
Tetsuya Hayashi, General Chair OFC 2026 (Sumitomo Electric), OFC News Release 2026
The overlooked layer
For years, improvements in AI performance were largely a function of better chips. That equation is changing. As clusters grow, the complexity of connecting thousands of processors begins to outweigh the gains from incremental compute improvements.
This is not a marginal issue. It is a structural one.
Electrical interconnects struggle under the combined pressure of bandwidth, latency and heat. Optical technologies—long positioned as an enabling layer—are now moving into the critical path. Without them, scaling AI systems becomes inefficient, expensive and ultimately constrained.
Yet even within photonics, one layer remains largely invisible.
The gatekeeper of scale
PHIX Photonics Assembly does not build AI models. It does not design processors. It does not operate data centers.
Instead, it operates at the point where systems either work—or fail.
Its role is deceptively simple: connecting optical fibers to photonic chips with extreme precision. In practice, this is one of the hardest challenges in the entire stack. Misalignment at microscopic scale results in signal loss, inefficiency and system instability. At AI scale, those losses compound rapidly.
This is the layer that determines whether theoretical performance can be translated into real-world deployment.
The relevance of that position becomes clearer when looking at who depends on it.
Where innovation meets reality
One of the more telling signals at OFC 2026 is not found in product announcements, but in partnerships. U.S.-based Lightmatter, widely positioned as a next-generation AI hardware challenger, has selected PHIX as a key integration partner.
The implication is straightforward. Advanced photonic chips alone are not enough. Without the ability to reliably connect them into larger systems, their performance remains theoretical.
“The collaboration with pioneers like Lightmatter shows that the bottleneck in AI is no longer just compute — it is connectivity. We deliver the scalable assembly needed to get that compute power out of the chip and into the network.”
David van Duinen, Business Development Manager, PHIX Photonics Assembly, PHIX Corporate News 2026
The statement reflects a broader industry shift. The limiting factor is no longer located inside the chip, but between chips.
From specialization to strategy
What makes PHIX particularly notable is not only its technical role, but its position within a broader ecosystem.
The Netherlands has quietly developed a concentration of capabilities across the photonics value chain—from chip design to system integration and critically packaging. This is not a coincidence. It reflects a long-term focus on a layer of technology that sits beneath mainstream visibility, but underpins system performance.
That positioning is now moving from niche to strategic.
Scaling the invisible
Historically, Europe has struggled to translate research leadership into industrial scale. The launch of the Dutch Photonic Chip Pilot Line in Eindhoven suggests a shift in that pattern.
By bringing manufacturing closer to research, the initiative aims to accelerate the transition from prototype to production—particularly for complex photonic systems.
“By bringing research and production closer together, we strengthen the position of the Netherlands in the European semiconductor landscape and help companies scale high-end photonic technology faster.”
Tjark Tjin-A-Tsoi, CEO TNO, TNO Press Release, March 9, 2026
For companies operating in the packaging layer, this matters. Scale is not just about volume. It is about consistency, reliability and integration into global supply chains.
The infrastructure phase of AI
The broader implication of OFC 2026 is that artificial intelligence is entering a new phase—one defined less by algorithmic breakthroughs and more by physical constraints.
Energy consumption, bandwidth limitations and system-level integration are becoming first-order challenges. In that context, control over infrastructure layers becomes a source of leverage.
Compute defines capability. Connectivity defines scalability.
This distinction reframes how competitive advantage is built.
A quiet position of influence
The Netherlands is unlikely to dominate the AI headlines. It does not produce the largest models or the most visible hardware platforms.
But it is positioning itself within a layer of the stack that those systems cannot function without.
PHIX is a case in point. Not because it represents scale, but because it represents dependency.
AI systems can be designed anywhere. But without the ability to connect them efficiently, they cannot scale.
And as the industry moves deeper into this infrastructure phase, it is increasingly clear where that constraint—and therefore that leverage—resides.
Credit:
Image: AI-generated visualisation of photonic chip assembly (Altair Media)
Caption:
Precision photonic packaging enables the high-speed optical interconnects required for next-generation AI infrastructure.
