During COMPUTEX 2026 and Nvidia GTC Taipei, energy once again dominated the AI data center conversation — only this time the question was not whether enough electricity existed, but whether it could arrive on time, arrive clean, and sustain 24/7 carbon-free operations.
DIGITIMES observed that as power-supply and grid-interconnection bottlenecks tightened, AI data center development was shifting toward on-site, behind-the-meter generation. Gas turbines remained the incumbent option, but solid oxide fuel cells, or SOFCs, were rapidly emerging as an alternative. South Korea's Doosan Fuel Cell and Delta Electronics were expected to enter mass production in 2026 and 2027, respectively, following the trail blazed by US-based Bloom Energy.
Energy is Nvidia's foundation — and its next frontier
Nvidia CEO Jensen Huang described an AI five-layer cake framework composed of energy, chips, infrastructure, models, and applications, with energy as the base. Every token, in his framing, traces back to energy converted into computing power. As the industry moves deeper into agentic AI, future gigawatt-scale AI factory investments are projected to reach US$80 billion to US$100 billion, with competition increasingly defined by tokens per watt and throughput per megawatt.
Nvidia's sustainability report put it plainly: "Energy is the Limiter." On the efficiency front, the company has made substantial progress — under the same megawatt of power consumption, the Vera Rubin NVL72 delivers 10 times more tokens than the GB200 NVL72, shifting the conversation from raw compute output to AI yield per unit of power.
Nvidia also met its goal of running on 100% renewable energy across its own offices and data centers in 2025. The harder challenge, however, lies upstream. Moving from Scope 2 to Scope 3 accountability, supply chain emissions in 2025 accounted for roughly 87% of Nvidia's total — with wafer manufacturing, advanced packaging, and servers as the main pressure points. Whether Nvidia ultimately requires low-carbon or carbon-free energy from its suppliers could become an important issue for the broader AI industry.
That pressure is already rippling across the supply chain. When cloud service providers commit to carbon-free energy, two things happen: Scope 3 emissions intensity from GPU usage falls, and Nvidia and its suppliers face growing pressure to disclose their own emissions and raise their share of clean power.
Carbon-free sources span solar, wind, hydro, geothermal, nuclear, green-hydrogen fuel cells, and battery storage. Nvidia has also invested in fusion startup Commonwealth Fusion Systems and helped build a digital twin of its SPARC fusion device — a move DIGITIMES characterized as a medium- to long-term bet on solving AI-era power constraints at the source.
The grid can't keep up — and that's opening the door for SOFCs
The problem for AI data centers has shifted. Rising electricity demand is only part of it; the deeper issue is whether grid buildout and interconnection can keep pace with computing capacity. Inference demand and the proliferation of gigawatt-scale data centers are expected to push global data center electricity consumption to between 1,000TWh and 1,350TWh by 2030. Yet more than 2,500GW of renewable energy, storage, and large-load projects remain stuck in interconnection queues.
That gap is fueling demand for behind-the-meter onsite generation. Gas turbines are the default today, but lead times stretching up to five years and rising costs have made SOFCs an increasingly attractive alternative — one that can be deployed within months and paired with battery energy storage systems. The global SOFC market is expected to exceed US$10 billion by 2030.
The efficiency case for SOFCs is compelling. Generation efficiency can reach 60%, against 31% to 36% for gas turbines. They produce direct current compatible with 800VDC architectures, carry a lower environmental footprint, and while they currently run on natural gas, a transition to green hydrogen could bring them to near-zero emissions.
A three-tier market takes shape
DIGITIMES divided the SOFC supplier landscape into three technology maturity tiers: US-based Bloom Energy at the top; the technology-transfer ecosystem anchored by UK-based Ceres and its licensees, Doosan Fuel Cell and Delta Electronics; and Japan's Mitsubishi Heavy Industries, which remains at the megawatt-scale demonstration stage.
Bloom Energy is already operating at gigawatt scale and is targeting 2GW to 5GW of capacity. The company generated more than US$2 billion in 2025, with annual growth of 37.3%, and projects 2026 revenue of US$3.4 billion to US$3.8 billion. Oracle's Project Jupiter will deploy Bloom Energy fuel cells for 2.45GW of capacity, displacing gas turbines and diesel generators.
Within the Ceres ecosystem, Doosan Fuel Cell's megawatt-scale factory is now operational and set to begin commercial SOFC sales in early 2026. Delta Electronics currently runs a 110kW proof-of-concept site at the Taiwan Power Research Institute. Its Tainan plant is expected to reach mass production by end-2026, while its new Guanyin plant in Taoyuan is planned to scale to gigawatt size — with the first phase entering production in 2027 to 2028 and the second phase expanding in 2028. Delta's SOFC generation cost is estimated at NT$4.2 (US$0.13) to NT$4.8 per kWh, covering 20 years of capex amortization, maintenance, and natural gas costs.
Gas turbines remain the most common onsite generation choice for AI data centers, but their limitations — long lead times, lower efficiency, noise, and carbon output — are becoming harder to ignore. SOFCs, offering higher efficiency, faster deployment, cleaner output, and native compatibility with high-voltage direct current systems, have moved from niche consideration to leading behind-the-meter solution. For the top suppliers, scaling capacity is now the priority.
Article translated by Jingyue Hsiao and edited by Jerry Chen
