Event Round-up: Is UK data centre infrastructure ready for next-generation compute?
Event overview
Moore's Law is running out of road. The next major gains in computing performance will not come from smaller transistors alone, but from new materials, novel architectures, and convergence with photonics, quantum and neuromorphic technologies. This session explores how the UK’s data centre infrastructure can prepare for this post-silicon era.
Published in June 2026, the government’s AI Hardware Plan includes a £120m Innovation Programme and is supporting UK firms to develop novel chip architectures and bring new forms of compute closer to deployment. As this ecosystem develops, the infrastructure around it needs to be ready to support and scale these technologies. Data centres are where that shift becomes concrete. A chip breakthrough stays in the lab until it can be housed, powered, cooled and integrated at scale, and post-silicon chips bring different power draw and thermal profiles to today's GPUs.
Building on techUK's recently published Digital Infrastructure report, this session explored how heterogeneous computing systems could reshape the UK’s infrastructure requirements over the next decade. It convened a small group of chip makers, chip designers, data centre operators and hyperscalers to explore the practical implications and existing bottlenecks while identifying the tangible next steps UK government and industry can take to prepare for the post-silicon era.
Below is a summary of some of the key points from the discussion:
Next-gen chips – what's coming and what will it unlock?
The bottleneck has moved to memory and data movement. Core counts have risen sharply while memory per socket has barely moved. Gains will come from bringing memory and compute closer together and from faster interconnects, with optical technologies such as co-packaged optics and hollow-core fibre a major opportunity.
GPUs will run alongside emerging technologies over the next five years, not be replaced by them. GPUs solve one class of problem very well, but many workloads still run better on CPUs, and many organisations struggle to feed the GPUs they already have, so bringing compute to the data is an underused approach. Adoption of novel compute is low, and photonic, neuromorphic and quantum paradigms may be ten or more years from wide use. Early adopters are those who know exactly what they want to run, such as defence and financial services, but the UK should prepare now.
Hardware is outpacing software and supply chains. Software takes longer to mature than new chips, and bringing a new processor to market is a huge undertaking. Wafer capacity, memory and lead times are all under pressure, which makes diversity of supply a resilience issue as well as a technical one.
Software, orchestration and open standards are the battleground. An agnostic software layer, workload managers that know where each job should run, and clear migration paths are essential for heterogeneous systems. Open interfaces matter because technologies are maturing at different speeds in different countries.
Preparing our infrastructure for next-gen compute
Power is the biggest constraint. Grid connection queues of up to eight years and high energy costs are pushing training and large-scale inference to markets such as the Nordics. Latency-sensitive workloads stay in and around London, and there are sovereign workloads that are legally obliged to stay in the UK. Better use of existing power through digitisation and flexibility, reform of the connection queue, and faster self-build options are all needed.
Workloads will sit across a continuum. Hyperscale, colocation, edge and on-premises will all have a role. Large-scale training needs vast facilities, but fine-tuning frontier models for UK use cases such as defence and healthcare can happen in smaller, modular sites, which may suit the UK better.
Building for uncertainty is hard. Facilities are built for decades but chips change every few years, and demand is high but unpredictable. Contracts are also shortening, so operators struggle to know what to design for and when.
Signals on future chip requirements are rarely shared. Roadmaps and pricing are commercially sensitive, and expertise is not joined up, so the same problems are solved many times over. Participants called for a trusted forum with clear ground rules, with techUK suggested as a natural convenor alongside government.
The AI Hardware Plan needs a clearer route to deployment. Funding is seen as small for the scale of the challenge, and there is little clarity on procurement or on whether operators will host unproven chips. Participants want defined requirements, testing and validation, and clear signals on which use cases will drive demand over the next five years.
What can be upgraded, and what must be built?
Retrofit is possible but costly and often partial. Putting new high-power chips into existing buildings was likened to "Ferrari GPUs with Victorian plumbing". Power distribution, cooling, floor loading and switchgear all need upgrading, so dense heterogeneous systems will often need new facilities.
Reuse brownfield and industrial sites. Investors are favouring sites with existing power and connectivity, though many are not ready to scale. Former power stations, industrial land, telecoms exchanges and office space were all suggested, rather than building in the countryside.
Fix grid connection, planning and skills. Faster connections, lower friction and quicker planning would unlock the most capacity, since time is the biggest cost for invested capital. A national standard that fast-tracks compliant developments was proposed, alongside investment in electricians, specialist contractors and engineers who can run mixed-architecture systems.
Set a clear national direction. Participants pointed to Japan's grand challenge approach and the EU's AI gigafactories as models the UK lacks. Backing specific use cases such as defence and the NHS, and investing in testbeds that smaller companies can use, would create the demand signals industry needs.
Design for adaptation, not optimisation. Optimising tightly for one chip suits a few very large players but not most of the industry, where investors need 10–50-year horizons. Facilities should be designed for "exit" from the outset, so components, cooling and power systems can be swapped as chips change.
Wider activity
This event formed part of a year-long initiative exploring how the UK can lead on the development and deployment of Frontier Compute technologies, the next phase of compute set to enable wide-ranging innovations and applications through the convergence of advanced hardware, software, and infrastructure. Findings from this activity will inform policy paper to be published in 2027 featuring policy recommendations for industry and government outlining the strategy, regulation, infrastructure, and investment required for UK leadership in the next frontier of compute.
This focus brings together experts, policymakers, and industry leaders to explore how the UK can build on its world-class research and industrial strengths to drive future growth, productivity, and global leadership in frontier compute technologies, such as quantum, photonics and neuromorphic compute.
If you would like to get involved in our frontier compute work and help shape upcoming activity, please reach out to Ella Shuter at Ella.Shuter@techuk.org to arrange an intro chat, or find out more here.
