Can a 100% Behind-the-Meter Power System Support an AI Data Centre?
A Japan–Malaysia stress test of islanded power: gas-only versus hybrid solar, 4-hour battery storage and firm generation.
Research note · 24 October 2026 · Zaixing Wang
Executive view
A 100% behind-the-meter (BTM) artificial intelligence data centre is not necessarily a 100% renewable facility. BTM describes the electrical and commercial boundary. In this research note, it means meeting the full 100 MW continuous campus load from an on-site, island-capable power system without relying on energy imports from the public grid. Reliability is a constraint; annual energy balance by itself is not proof of firm supply.
Using Japan and Malaysia as contrasting APAC markets, this screening analysis reaches three conclusions:
- Solar and short-duration batteries are useful, but are not a substitute for dispatchable firm generation in a fully islanded, around-the-clock 100 MW facility. Adequacy must be assessed against multi-day low-solar periods, outages and reserve requirements.
- A hybrid can be economically comparable to gas-only generation, especially with high gas prices or a carbon cost, while reducing fuel dependence. In the reference four-hour BESS case, the screening LCOE is approximately US$147/MWh in Japan and US$86/MWh in Malaysia, almost level with the corresponding gas-only baseline.
- Local engineering CAPEX matters more than importing a US cost curve. Malaysia has a materially lower assumed solar-plus-storage cost than Japan, although installed PV CAPEX estimates have important project-scope and market-year differences.
These are indicative, assumption-driven screening outputs, not a dispatch-verified claim of zero loss of load or an investment-grade feasibility study.
01 — Set the technical boundary first
The reference facility is a flat 100 MW net electrical demand supplied for 8,760 hours a year, equal to 876 GWh annually. We compare two illustrative locations: Chiba, Japan, and Johor, Malaysia. The model represents a 20-year operational horizon with on-site firm gas generation retained in every feasible scenario. No electricity export revenue or grid-import revenue is included.
| Modelling input | Japan — Chiba | Malaysia — Johor |
|---|---|---|
| Firm AIDC demand | 100 MW | 100 MW |
| Electricity needed | 876 GWh/year | 876 GWh/year |
| Evaluation horizon | 20 years | 20 years |
| PV installed CAPEX | US$950/kWdc | US$450/kWdc |
| PV annual generation | 1,250 kWh/kWdc | 1,450 kWh/kWdc |
| PV effective utilisation | 85% | 80% |
| BESS 4h installed CAPEX to MV | US$110/kWh | US$80/kWh |
| Natural gas | US$14/MMBtu | US$7/MMBtu |
| Gas electrical heat rate | 8 MMBtu/MWh | 8 MMBtu/MWh |
| Discount rate | 7% | 8% |
All figures are assumptions for an illustrative 2026 screening, not verified project quotations. Installed BESS prices refer to the modelled AC/DC scope up to the medium-voltage interface and exclude land, site-specific high-voltage interconnection and owner-specific development costs unless included separately.
02 — Compare feasible configurations, not technologies in isolation
| Scenario | Gas firm plant | Solar PV | BESS | What it tests |
|---|---|---|---|---|
| G — Firm generation | Included | — | — | Cost and reliability reference |
| H1 — Balanced hybrid | Included | 150 MWdc | 60 MW / 240 MWh (4h) | Economically restrained renewable integration |
| H2 — High-solar hybrid | Included | 300 MWdc | 100 MW / 400 MWh (4h) | Higher PV penetration and more curtailment exposure |
Why keep gas? The load does not disappear at sunset, and a 4-hour battery is not a four-day energy source. For islanded sites, firm generation, redundancy, reserve margin and credible outage analysis cannot be skipped. Solar offsets fuel; batteries shift hours and help dynamic performance.
In H1, gross annual PV output is about 187.5 GWh in Japan and 217.5 GWh in Malaysia. Applying illustrative utilisation produces about 159 GWh and 174 GWh, or 18% and 20% of the campus’s annual load. The remaining electricity is expected from dispatchable generation, subject to losses and actual hourly dispatch.
03 — System CAPEX and first-pass economics
| Scenario | Japan total CAPEX | Malaysia total CAPEX |
|---|---|---|
| Gas-only reference | US$210 million | US$165 million |
| H1: 150 MWdc PV + 60 MW / 240 MWh BESS + gas | US$378.9 million | US$251.7 million |
| H2: 300 MWdc PV + 100 MW / 400 MWh BESS + gas | US$539 million | US$332 million |
The H1 incremental CAPEX is transparent: in Japan, 150,000 kWdc × US$950 = US$142.5 million of PV plus 240,000 kWh × US$110 = US$26.4 million of battery. Adding the US$210 million gas reference yields US$378.9 million. In Malaysia, PV adds US$67.5 million, BESS adds US$19.2 million, and the US$165 million gas reference yields US$251.7 million.
The model’s published screening LCOE comparison for the gas reference and H1 is:
| Metric | Japan | Malaysia |
|---|---|---|
| Gas-only LCOE | US$146.8/MWh | US$85.8/MWh |
| H1 hybrid LCOE | US$147.4/MWh | US$86.2/MWh |
| Hybrid premium | +US$0.6/MWh | +US$0.4/MWh |
Interpretation: the balanced hybrid is nearly cost-neutral under the chosen fuel and capital-cost assumptions. Its potential value lies in fuel hedging, electricity self-supply resilience and decarbonisation. It would be incorrect to conclude that PV+BESS is always cheaper or that the model establishes a data-centre-grade reliability guarantee. The narrow LCOE differences are sensitive to gas price, solar output, generator part-load efficiency, BESS operating assumptions and asset replacement costs.
04 — Storage duration changes the effective US$/kWh
Battery storage includes both power-dependent components (PCS, MV transformer, controls) and energy-dependent components (cells and containers). A single four-hour unit cost should not be reused unchanged for every duration.
The screening model uses:
BESS CAPEX = (power cost in US$/kW × BESS MW × 1,000) + (energy cost in US$/kWh × BESS MWh × 1,000)
| Parameter | Japan | Malaysia |
|---|---|---|
| Power-related component | US$120/kW | US$80/kW |
| Energy-related component | US$80/kWh | US$60/kWh |
| 2-hour installed system | US$140/kWh | US$100/kWh |
| 3-hour installed system | US$120/kWh | US$86.7/kWh |
| 4-hour installed system | US$110/kWh | US$80/kWh |
| 6-hour installed system | US$100/kWh | US$73.3/kWh |
| 8-hour installed system | US$95/kWh | US$70/kWh |
These values are model assumptions, not manufacturer quotations or a forecast that capacity can be added indefinitely at the energy-only marginal price. Augmentation, ageing, fire-safety design, warranty limitations and site-specific civil works should be modelled separately.
05 — How realistic are the PV cost assumptions?
For Malaysia, the US$450/kWdc base case represents competitive, large-scale engineering execution. It is not presented as a measured national average. Public project awards help set context: Solarvest reported approximately RM685.4 million of EPC contracts for 345.9 MWp of Corporate Green Power Programme capacity in its 2025 annual report (around RM1.98/Wp, contract-scope dependent), while the same company reported a RM1.06 billion award for a 690 MWp solar project in April 2026 (around RM1.54/Wp). EPC contract prices and total owner’s all-in installed CAPEX are not interchangeable. The model’s US$450/kWdc therefore warrants site-specific land, grid and EPC checks.
Japan’s US$950/kWdc assumption is likewise an optimised large-project scenario, not the historic industry mean. METI’s committee data summarised a ¥250,000/kW average for ground-mounted business PV projects installed in 2024, highlighting the importance of Japan’s soft and construction costs. Without evidence from site-specific EPC pricing, the modelled Japan PV price should be stress-tested upward.
As a global cross-check, IRENA reports US$691/kWdc as the global weighted-average total installed cost of utility PV commissioned in 2024. Country composition, project size and system boundary make it a directional benchmark rather than a direct Japanese or Malaysian quote.
06 — Reliability is the gating constraint
A constant 100 MW IT and facility load places a different requirement on a plant than annual PV generation forecasts:
- A positive annual energy balance does not establish hourly supply adequacy.
- A 4-hour BESS charged from solar cannot bridge several consecutive low-irradiance days without enough dispatchable power or extreme excess PV.
- Gas generator N-1 or N-2 design, maintenance overlaps, fuel security, black-start and island-mode transient response must be tested.
- Auxiliary loads, medium-voltage transformer losses, usable state-of-charge, battery augmentation and UPS interfaces must be reflected in the electrical boundary.
- Zero unserved energy in one historical weather year is not the same as a data-centre service-level guarantee.
A bankable next iteration requires chronological 8,760-hour or sub-hourly dispatch using location-specific irradiance and temperature, chronological forced outages, fuel cost sensitivity, at least one prolonged cloud/storm scenario, and explicit reliability metrics such as loss-of-load expectation and expected energy not served.
Conclusion — pursue firm hybrid power, not a false binary
The strategic question is not simply whether renewables can outbid a gas turbine on their generation LCOE. A fully islanded AI data centre purchases availability at every hour. Under plausible APAC assumptions, a gas-backed PV+BESS hybrid can add renewable supply without a substantial screening LCOE premium, but that outcome is fragile to gas fuel pricing and installed costs. Malaysia merits particular attention given its comparatively low PV and BESS cost assumptions; Japan requires stronger focus on construction costs, land, fuel exposure and reliability design.
Bottom line: evaluate PV + appropriately sized 4-hour BESS + firm dispatchable generation as an integrated behind-the-meter supply architecture. Treat pure solar-plus-short-duration-battery supply for a continuously loaded, fully islanded AIDC as unproven until adequacy modelling demonstrates otherwise.
Source notes and reproducibility
- IRENA — Renewable Power Generation Costs in 2024 — global installed PV and storage cost context.
- Japan METI — FY2025 Procurement Price Committee report — Japanese historic business solar CAPEX inputs.
- Solarvest — Annual Report 2025 — Malaysian EPC awards and capacities.
- Solarvest — April 2026 large-scale solar EPC announcement — Malaysian large-project benchmarking.
- IEA-PVPS — Malaysia country profile — installed-market context.
Scope: This is an independent scenario comparison based on explicitly disclosed modelling assumptions. No hourly dispatch file, site-specific EPC quotations or full financial workbook is embedded in this post. Numerical results should not be interpreted as a completed feasibility study, warranty or verified equipment offer.