The bottleneck for AI data centres is power, not GPUs

When you can get power decides where you can build. CAYNMARS chose fuel cells for time, not for unit cost.

Problem

Power is already a structural bottleneck

Electricity demand from AI data centres is rising quickly, while building new transmission takes years and domestic grid connection queues keep lengthening. Sites and capital can be ready and the project still cannot break ground for lack of power.

Alternatives

Few sources can actually be built today

Source	Can you choose it today?
Grid connection	Queues make the timing hard to commit to
Gas turbines	Order books at the major OEMs are already full
SMR	Commercialisation still needs more time
Fuel cells	Fastest to build, using existing gas infrastructure

For an operator that needs power now, electricity several years from now is not an option.

Regulation

Toward bringing your own power — regulation has set the direction

Since 2024 the rules for large electricity demand have been moving the same way in Korea and abroad. The following are facts confirmed in published legislation and regulatory filings.

2024.06

Korea

The Distributed Energy Act takes effect. Parties intending to consume 10MW or more are subject to a power system impact assessment.

2026.05

Korea

The AI Data Centre Industry Promotion Act passes the National Assembly, expected to take effect in February 2027. It includes an exemption from the power system impact assessment for AI data centres built or expanded outside the capital region below a certain scale. The detailed threshold is to be set by enforcement decree.

2026.06

USA

FERC issues Federal Power Act Section 206 orders to all six grid operators, directing them to justify or reform their interconnection rules for large and co-located load.

2026.08

USA

PJM files a large-load framework with FERC, defining a single site with peak demand of 50MW or more as a large load and exempting those that secure qualifying new capacity from curtailment exposure. The proposed start of application is June 2027. Pending at FERC

Ongoing

Asia

As data centre demand presses against available grid capacity, delayed connections and rejected applications are being reported.

On-site generation and grid flexibility are moving from an option to a condition.
That is why CAYNMARS designs the power architecture before the facility.

And CAYNMARS NEO OS^™ is what we built to carry the cost and quality responsibility that follows the moment power is generated on site. See what the software does →

Confirmed as of September 2026; subject to change as proceedings and legislation advance. Reviewed quarterly.

Technology roadmap

Capability	2026	2027
Verification native	Verification roadmap as a product function	First integration with an external specialist body
Regulatory native	Domestic rules engine	Response to the AI data centre act
AI-load native	Profiles from published literature	Per-mode model v1
Finance · PF outputs	—	Technical advisory output format

These are our technology development targets and may change with external conditions. Investment, construction and completion milestones are not included in this table. See the capability structure →

Economics

Today’s cost, and the direction of two curves

On a total-cost basis fuel cell power is more expensive than grid electricity today. We state that plainly. Industrial tariffs rise, while fuel cell unit cost falls as the technology matures. Where those two lines meet is the decision point for this business.

One rising grid tariff line, three on-site LCOE paths by decline speed. Where each crosses the tariff line is the golden cross. Switch scenarios above and move across the chart. No amounts appear anywhere — the vertical axis carries no scale.

Disclosure principle

Cost, margin, plant capacity and model assumptions are not published. Detailed materials are provided after an NDA.

See the reference project See Technology