Hyperscalers, semiconductor companies, data-center operators and energy providers are racing to build the physical foundations of AI: clusters, accelerators, interconnects, optical links, cooling systems and power infrastructure. 

But modern AI cannot scale while it’s constrained by the energy required to inefficiently move data between memory, accelerators, interconnects and nodes. This isn’t just about chip power — it’s got as much to do with what physically connects them. 

Copper simply cannot keep up with the bandwidth, speed and energy demands of increasingly distributed AI infrastructure which has driven accelerated adoption of and investment in photonics  

Photonics is now deployed in AI infrastructure due to its ability to move data at high bandwidth and with greater efficiency than purely electrical approaches which is critical in data centers as clusters scale up, across and out. 

But treating photonics solely as a connectivity technology vastly underestimates the technology, and the opportunity. 

The data movement bottleneck

The AI infrastructure conversation is still heavily focused on accelerators: GPUs, XPUs, ASICs and other specialized processors. 

But performance is increasingly shaped by what happens between compute elements: memory movement, interconnect overhead, synchronization, conversion and data transfer. 

These workloads move vast amounts of data through memory hierarchies and across distributed systems before useful work can be completed. Every movement adds latency, every conversion consumes energy and every synchronization step impacts performance. 

Simply put, the burden of data movement on processors and energy grids is unsustainable with AI’s growth trajectory 

The US Department of Energy has reported that data centers could consume approximately 9% of total US electricity by 2030, while Goldman Sachs project US data center power demand to more than double to 66 GW in 2027 from 31 GW in 2025. This energy demand is a binding and extremely expensive constraint — with PwC projecting that data center expenditure will rise from around $800 billion per year in 2026 to $1.8 trillion per year in 2050. 

The implications are clear: if the US is going to lead the next era of AI, it cannot rely only on building more data centers and deploying more accelerators. It needs coordinated collaboration and sovereign architectures that make every watt, every bit and every link do more useful work. 

The copper wall and the limits of conventional systems

For decades, electronic computing advanced by making transistors smaller, faster and more efficient. But AI workloads are pushing systems into a different set of barriers. 

Copper is constrained by reach, speed, supply and power while memory bandwidth and performance has been outpaced by processor demands. Data movement increasingly dominates system cost and, as clusters scale, so the overhead of moving data between compute elements has become as much of a concern as the computation itself. 

Photonics helps move data with high bandwidth and lower energy over distances – hence the growing adoption of optical interconnects into AI infrastructure. 

But transport is only one part of the equation. 

If data is moved optically, converted back into electronics, processed, stored, moved again and converted again, the system still pays a large data movement and conversion cost. 

At Optalysys we’ve asked ourselves if optical infrastructure can do more than simply carry data between electronic processing points. Whether photonics has progressed enough to play a part in computation itself? 

The answer is yes. 

Introducing photonic Compute-in-Transit

The solution we are developing is photonic Compute-in-Transit. 

Rather than transporting data between memory and processors before useful work happens, we’re adding compute capability within the optical data path itself. 

In simple terms: compute happens as data moves. 

This doesn’t mean photonic systems replace electronic systems — it means playing to each domain’s strengths for all-round smarter systems.  

Electronics remains essential for control, programmability, memory and general-purpose processing while integrated photonics enables a new class of high-throughput, structured operations that align naturally with signal propagation. 

When useful work happens directly in the optical path, data movement stops being pure overhead; latency decreases, energy efficiency improves and infrastructure can scale more sustainably. 

Why this matters for US AI infrastructure

The US has a unique concentration of AI infrastructure leadership: hyperscalers, semiconductor companies, networking leaders, frontier AI labs, national laboratories and data center developers all increasingly focused on solving a physical infrastructure problem. 

The next wave of AI systems will require more bandwidth, more interconnect density, more power and highly efficient movement of data across increasingly distributed architectures. 

This wave is where photonics can move from enabling technology to foundational compute architecture. 

Today, communication and computation are often treated as separate processes. We move the data, then compute on it — we transport, then transform. Compute-in-Transit changes that equation. 

The future scalability, security and sovereignty of AI depend on architectures that move and process data more intelligently. Compute-in-Transit is one such architecture. 

Photonics is already a major part of the AI story, but it’s approaching a new frontier. If the US is to maintain its leading role in that story, we must recognize that photonics should compute, not just connect. 

---

At Optalysys we’re pioneering smarter data movement for AI through Compute-in-Transit Get in touch with us to find out how we can bring efficiency gains to your use case →