Britain’s chance to lead in heterogeneous AI infrastructure

There won't be a sole capability, chip or machine that defines the data centre of 2035 – the future of AI systems will blend architectural innovations at every layer of the stack.

techUK's recent Digital Infrastructure 2035 report highlights an important change taking place across computing: future infrastructure combining classical CPUs and GPUs with specialised accelerators, neuromorphic systems, quantum processors and photonic technologies.

The semiconductor industry is moving away from a world in which almost every workload benefited from the same underlying trend: smaller transistors delivering predictable improvements in performance and efficiency. As those gains become harder to achieve and AI places unprecedented pressure on power, memory and communication, computing is entering a new, heterogenous realm.

The next infrastructure era will therefore not be defined by one technology replacing all the others, but by putting the right type of computation in the right place. Here, Britain has a chance to lead.

Different bottlenecks require different architectures

GPUs have transformed artificial intelligence because they are exceptionally good at highly parallel computation, but the performance of an AI system depends on far more than the arithmetic capability of the processor: information must also travel between memory, accelerators, servers and racks.

This means that as clusters expand, what happens between the processors is becoming as critical to performance as what happens on them.

This is one reason optical technologies are being rapidly adopted throughout AI networking infrastructure - photonics provides the bandwidth necessary to move enormous quantities of information efficiently.

At Optalysys we have asked ourselves whether optical infrastructure can contribute more than bandwidth alone. When data is already travelling as light, why wait for an inefficient conversion back to digital and pass through a separate processor before any useful operation can take place?

Computing without competing

This is the premise for photonic Compute-in-Transit – using silicon photonics to perform operations with light, directly in the data path, to tackle the memory, power and copper walls facing AI scaling, latency and throughput.

Critically, this isn't intended as an alternative to XPUs, but a complementary technology that adds new compute capability where there wasn't any before, for the workloads that would benefit most from it.

Adding compute into the optical path delivers more useful work for every transmitted bit for ultimate system efficiency in a heterogeneous architecture.

This approach is substantiated in the Digital Infrastructure 2035 report, with techUK predicting that the data centres of 2035 will combine classical silicon, neuromorphic chips, quantum processors and optical interconnects and processors while an intelligent orchestration layer directs each workload to the compute best suited to it.

The report also highlights how photons are particularly attractive in the context of the energy and latency limitations confronting modern computing and identifies the progression from photonics being predominantly used for connectivity towards architectures capable of performing computation directly in light.

This creates an opportunity for Britain's semiconductor strategy

The UK is unlikely to compete with the scale of semiconductor manufacturing investment in the United States, Taiwan or South Korea.

That should not be the only measure of success.

Global technology leadership can also come from owning strategically valuable pieces of an international supply chain.

Britain already has strong capabilities in semiconductor design, photonics, optical communications, advanced materials and computing research.

The architectural transition now underway creates an opportunity to turn some of those strengths into infrastructure technologies used around the world.

That requires a different ambition from attempting to replicate an entire supply chain domestically.

It means identifying the parts of the emerging stack where Britain has genuine differentiation and ensuring that UK companies can commercialise them before global architectures consolidate.

Partnerships matter now more than ever

Heterogeneous computing also means collaboration becomes imperative - no company or sector will build this future alone.

Hardware vendors, hyperscalers, networking and interconnect companies, Government and regulators and deep tech innovators must all work together to design and deploy the evolving systems AI demands.

Adding compute capability, whether that's in rack, in memory or the data path requires integrations with the systems that are already being developed rather than whole stack overhauls.

The next stage of UK compute strategy isn't a race to find a single national champion capable of replacing every incumbent technology, but an opportunity to build strategically important layers in the global computing system.

This window won't remain open forever

Architecture transitions open up new positions in supply chains, and computing is facing its biggest transition in decades.

Britain has the research base and technical strengths required to meaningfully participate in this transition, but we must translate that expertise into products, partnerships and companies – and do so quickly.

At Optalysys, we echo techUK's calls for the government to publish a Digital Infrastructure Strategy and an interim update to the UK Compute Roadmap, with the addition of dedicated sections on neuromorphic computing, photonics and quantum computing.

The strategic decisions that determine who will own the next generation of compute technologies versus who will be consuming them are being made now. Government and industry must work together to ensure the UK capitalises on its strengths and play a leading part in the next era of digital infrastructure.