Malaysia is building power for AI. Make each megawatt answer more.
PUE shows how well a building delivers power to its chips. Nothing yet shows how much of that power ends up in answers. Croutons measures it and cuts the waste.
What a megawatt is worth in Johor
One megawatt of facility power, for one year. Estimates.In a hall where power is the limit, one megawatt’s electricity costs about RM4.4 million a year, while the GPU time it runs sells for about RM50–167 million. The answers a capped megawatt can’t serve are worth roughly 10 to 40 times its electricity bill.
| Per MW, per year | Value | Worked out as |
|---|---|---|
| Electricity | RM4.4m | 8,760 MWh at RM0.50/kWh |
| GPU time, 30% of list price | RM50m | About 2.4m GPU-hours at about RM21 |
| GPU time, full list price | RM167m | About 2.4m GPU-hours at about RM69 |
Assumptions: 1 MW of facility power at PUE 1.4 leaves about 714 kW for IT, enough for about 397 NVIDIA GB200 GPUs at 1.8 kW each at the wall (our estimate, including a share of host and switching). GPUs busy 70% of hours. Electricity at RM0.50/kWh, rounded from the ultra-high-voltage tariff (47.09–51.18 sen/kWh from 1 Jul 2025, before fuel surcharge), with the full megawatt drawn all year, which overstates the bill. GPU time priced from 30% of the Malaysian GB200 list price (Nov 2026) up to full list, about RM21–69 per GPU-hour; large customers pay less than list. Croutons, November 2026.
Where things stand.
- 8.39 GW
- Projected peak demand from Peninsular Malaysia’s data centres by 2030. Johor alone is 4.42 GW.
- PETRA, via Bernama, Oct 2026
Building a data centre in Malaysia? Read the guide · See the research and data
Add answers per megawatt to the standard.
Report energy and carbon per answer next to PUE, WUE and CUE. Every operator already has what it takes: GPU telemetry and request logs.
At a PUE of 1.4, each megawatt of GPU load cut frees 1.4 MW of grid capacity and avoids about 9,100 tonnes of CO₂e a year on the Peninsular grid, at the Energy Commission’s 0.740 kg CO₂e/kWh (2024, provisional).
- The grid
- Less load for the same AI output, at the meter where it counts.
- Operators
- More paid work from each megawatt they were allocated.
- Regulators
- A number for the work itself, not only the building.
- Tenants
- A lower cost per answer, on chips that are hard to add.
We measured it.
A controlled benchmark on a production-class AI inference stack, across four workload types common in enterprise AI: support chat, questions over long documents, multi-step agents and long-form summaries. Accelerator energy was read from the hardware and joined to engine and request logs, and every request was labelled so each finding could be checked.
| Finding | Result |
|---|---|
| Energy follows busy time | Idle draw, plus a fixed draw while the accelerator is busy, plus a small cost per token, explains 99.8% of measured energy. Counting tokens alone explains 58%. |
| Utilisation decides efficiency | The same hardware used 3.3 times less energy per request at peak load than at low load. |
| Memory, not compute, is the limit | Working memory for active requests was above 90% full a third of the time and reached 100%; requests were paused and recomputed while compute still had room. |
| Where the energy went | Long summaries were 7% of requests and 37% of energy. Agent tasks were 64% of requests and 38% of energy. |
| Avoidable work | 4.7% to 15.7% of accelerator energy went to work that did not need doing: identical retried calls, shared documents placed mid-prompt, and repeated questions. |
Source: Croutons controlled benchmark, November 2026. Method and data available to operators and regulators under NDA.
What that means for a 20 MW AI hall.
Applied to a reference 20 MW AI inference hall at 70% average load, 70% of IT energy in accelerators, PUE 1.4 and WUE 2.2 m³/MWh. Carbon uses the MITI data-centre guideline’s figure of 0.758 t CO₂/MWh (Peninsular, 2021), not the Energy Commission’s current 0.740 kg CO₂e/kWh (2024, provisional). Every result scales with these assumptions.
| Indicator | Value |
|---|---|
| Facility energy per AI answer | 0.17 Wh |
| Carbon per 1,000 answers | 129 g CO₂ |
| Avoidable facility energy a year | 20,300 to 67,900 GJ |
| Carbon avoided a year | 4,300 to 14,300 t CO₂ |
| Water avoided a year | 8,900 to 29,700 m³ |
| Accelerator capacity freed | 0.7 to 2.2 MW |
Source: Croutons benchmark results, November 2026, scaled to the reference hall above. PUE, WUE and CUE stay at 1.40, 2.20 and 1.06 either way. The saving shows in absolute energy, carbon and water, and in energy per answer, which is why we propose reporting it.
Written as EECA energy saving measures.
A hall this size is far above the Energy Efficiency and Conservation Act 2024 threshold of 21,600 GJ a year and files a yearly Energy Efficiency and Conservation report. Croutons findings are written as energy saving measures for that report.
| Measure | Facility energy (GJ/yr) | Carbon (t CO₂/yr) |
|---|---|---|
| ESM 1. Remove identical retried calls | 9,086 | 1,913 |
| ESM 2. Place shared documents first so earlier work is reused | 11,250 | 2,369 |
| ESM 3. Answer repeated questions from a cache | up to 47,594 | up to 10,021 |
Source: Croutons benchmark results, November 2026, scaled to the reference hall; carbon at the MITI guideline’s 0.758 t CO₂/MWh. Each measure counts only after real requests are replayed with the change and answer quality holds. Repeat rates vary by service, so ESM 3 is an upper bound.
Get your hall’s baseline.
A one-week pilot reads the request logs, engine metrics and accelerator telemetry you already export, on your own machine, and returns your hall’s energy per answer, MITI indicators and EECA energy saving measures. Nothing is installed in production and no prompts leave the site.