Bottom Line
The Agentic AI transition is a triple power-demand cycle where gas turbines, optical networking, and the efficiency paradox operate simultaneously — more compute efficiency actually amplifies absolute power demand.
Gas Turbines, Optical Networking, and How Efficiency Amplifies Absolute Demand
The Agentic AI transition is a triple power-demand cycle where gas turbines, optical networking, and the efficiency paradox operate simultaneously — more compute efficiency actually amplifies absolute power demand.
Sub-agent parallelism is exploding token and memory I/O 10-100x, copper interconnect physics are forcing an optical transition, and Jevons Paradox is converting efficiency gains into higher total power consumption.
Beneficiaries — GE Vernova (80GW gas turbine backlog), Coherent (CPO commercial year one), Fabrinet and Tower Semi (optical component manufacturing), NVIDIA and Neocloud. Pressure — legacy grid infrastructure, copper-based interconnect suppliers.
Quarterly GE Vernova gas turbine backlog and Coherent CPO revenue share — both rising simultaneously confirms the triple-chain thesis remains intact.
This section is the logical foundation of the report. It decomposes in four stages why agentic workloads are not simply a 'compute demand × N' model but rather reshape power, networking, and infrastructure simultaneously.
Chat-era workloads were relatively sparse inference patterns of hundreds to thousands of tokens per user session. In the agentic environment, a single user runs dozens to hundreds of sub-agents in parallel, each independently performing tool calls, context processing, reflection, and replanning. Per-user token consumption is estimated at 10–100× chat levels, which non-linearly increases cluster-level compute, memory, and network loads.
IEA quantifies this: accelerated server power grows at 30% annually in the Base Case, while conventional servers grow 9%. -specific compute expands more than 2× faster than data center averages, and this gap is the substance of the investable 'AI infrastructure premium.'
The core constraint of parallel sub-agent operation is not FLOPs but memory I/O. Each agent holds an independent KV cache and context state, and synchronization, retrieval, and cross-reference between them depend directly on inter-GPU bandwidth. Swapping between traditional storage tiers (HDD/eSSD) and memory severely degrades inference latency, driving HBM capacity and bandwidth spec upgrades and scale-up domain expansion.
NVIDIA's Rubin-generation NVLink extends to 14.4 Tbps per GPU, but copper-based scale-up is limited to 2 meters, confining the scale-up domain to 1–2 racks. Agentic clusters require low-latency coherent domains of thousands to tens of thousands of GPUs — physically unreachable via copper.
NVIDIA responds with Quantum-X InfiniBand in 2026 H1 and Spectrum-X Ethernet CPO in 2026 H2 commercial launches. CPO-based systems reduce per-port power to ~9 W and cut electrical loss to ~4 dB. Dell'Oro forecasts 2026 as the initial volume-ramp year for CPO across both InfiniBand and Ethernet switches.
CPO improves per-port power by 3.5× but simultaneously expands the total port count and scale-up domain per cluster. As the unit economics of agentic workloads improve, more organizations are incentivized to run more agents, amplifying rather than dampening demand. Absolute data center power demand re-rates upward along with the efficiency curve.
This is the classic Jevons Paradox observed repeatedly in economic history. Just as 19th-century steam-engine efficiency improvements increased rather than decreased coal consumption, AI compute efficiency gains produce net increases — not decreases — in total power consumption. IEA reflects this in its projection: 415 TWh (2024) → 945 TWh (2030), with the US and China accounting for ~80% of the global increase.
Sign in with Google to unlock the full body of every free report instantly.
Sign in with GoogleThis site runs on ads — the tier system rewards community contributions.
This report is provided for informational purposes only and does not constitute a recommendation to buy or sell any financial instrument. Investment decisions should be made based on your own judgment and responsibility. The analysis and opinions contained herein are based on information available at the time of writing and are subject to change.