POWER
Meeting the engineering challenges of AI workloads: Spiky power, accelerated lifecycles, and upstream
By Martin Brannigan, operations director, TES T
he worldwide explosion of AI compute workloads is radically altering how the data centre industry approaches data centre power system design, operation and lifecycle management.
What were once relatively predictable environments are giving way to far more complex conditions, as the era of predominant cloud and colocation infrastructure decisively comes to an end. In 2027, AI is projected to account for 28 per cent of the overall data centre market, up from just 13 per cent, fundamentally reshaping assumptions that have underpinned facility power architectures for more than a decade. At the core of the issue is power management and the difference between supporting AI workloads and more traditional cloud or colocation environments. AI is replacing steadier, average-driven workloads with hyper-dynamic compute cycles, defined by steep spikes and unpredictable troughs. This in turn is creating a new generation of increasingly complex power and thermal systems built to accommodate volatility and sustained high-density compute, rather than optimising for steady-state efficiency. The shift introduces engineering challenges that extend well beyond load variability alone. Compressed project timelines and accelerated infrastructure refresh cycles are compounding the stress on power systems, forcing engineers to reconcile 20–25 year electrical asset lifespans with compute platforms that may be refreshed every three to five years. The result is a growing mismatch between digital innovation cycles and the physical realities of power infrastructure.
In response to these pressures data centre builders are increasingly turning to prefabricated modular solutions within their powertrains as a way to accelerate construction, scale operations across multiple sites, and introduce greater flexibility into maintenance and upgrade strategies.
18 JUNE 2026 | ELECTRONICS FOR ENGINEERS
AI workloads and their impact on power systems
AI workloads can be broadly divided into two discrete categories: training and inference. Training, the development and refinement of AI models, is compute-intensive and power-hungry, but relatively latency- insensitive; it can be located remotely where power availability and cost are more favourable. Inference, which runs trained models in production, powers every user- facing AI application and is far more latency dependent.
By 2030, inference is expected to account for approximately 80 per cent of all AI workloads globally, making efforts to host it among the most impactful digital infrastructure overhauls of the decade.
Unlike cloud or colocation workloads, both AI training and inference processes are characterised by rapid, high-magnitude fluctuations in power demand. High-density GPU clusters can spin up from near-idle to peak load in seconds, and back down again just as fast. These spiky load profiles are very different from more conventional data centre activities which tend to exhibit predictable, steady-state or slowly varying loads.
While data centre power systems have long been designed to handle some degree of dynamism, AI-driven fluctuations are a whole new kind of challenge. Repeated load cycling can lead to premature ageing of critical power components and thermal time constants within transformers, UPS systems and switchgear often lag far behind GPU ramp rates, increasing mechanical and thermal stress, thereby reducing effective asset life if systems are not explicitly designed to accommodate this behaviour.
Holistic resilience and the move to modularity
Addressing these issues requires a shift from component-level optimisation to a holistic, system-level design philosophy. Oversizing infrastructure is no longer sufficient or economically viable on its own. Instead, resilience must be engineered through coordinated interaction across the entire power chain.
Data centre design is evolving in response to the stresses of AI workloads and accelerated delivery timelines, as AI companies race to deploy capacity at unprecedented speed. The industry
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50
