New researchMeasuring AI work in Malaysian data centres

More answers per megawatt
from your H100s

Croutons measures the energy and GPU memory behind every AI answer, finds the workloads that waste capacity, and proves each fix by replaying real prompts before it ships.

Request a briefing Scan your AI logs

croutons scan --demo

Worked for 4s

8,066 model calls cost $130 between 26 October and 1 November: about $574 a month at this rate. One workflow, research-agent, is 56% of it.

Two H100s used 80 kWh. Utilization read 67%, but SM activity was 0.27. When requests queued, the cache was nearly full: memory is the limit, so smaller requests buy the headroom.

Request logs
Engine metrics
GPU telemetry

What to fix first9 findings, largest firstOpen

What to fix first

Retries, loops and oversized contextDemo

8,066 calls · 2 × H100 · 26 Oct to 1 Nov

577 calls resend a prompt within 2 minutes or after a failure
Cap retries per request and back off on failure.
measuredsaves $36.96
7 of 84 agent runs made 21 or more model calls
Set a call budget per run and stop loops that repeat a tool.
measuredsaves $30.08
339 calls send 8,000+ tokens to get under 2% of that back
Retrieve only the passages the job needs.
estimatesaves $13.31
GPU utilization read over 90% in 5,017 minutes
In 40% of them SM activity was under 0.3. Track SM and tensor activity before buying more GPUs.
measuredin 40%

Demo data, made up: the output of croutons scan --demo. Your logs stay on your machine.

A chip serves more when less of its work is wasted.

Llama
DeepSeek
Qwen
Mistral
Gemma

The Croutons mascot, a chrome robot cat, between open models and the accelerators that run them

H100
H200
B200
GB200
MI300X

Reads telemetry from deployments running these models and accelerators. Nothing sits in the request path.

Croutons finds the waste and proves the fix

Every token a model reads costs compute and holds GPU memory. At 128K tokens, that memory alone can outgrow the GPU (OSDI 2026).

Energy, memory and cost per answer

From the logs and telemetry you already have. Spend by model and workflow, and what the GPUs drew while they served it.

croutons scan ./logs --gpu dcgm.csv

Spend by workflow, and what the GPUs drew while they served it.

Workflow	Calls	Cost	Share
research-agent	1,066	$72.31	56%
contract-review	280	$24.50	19%
ticket-triage	2,800	$22.29	17%
support-chat	1,820	$10.86	8.4%

2 × H100 · 80 kWh used · 238 W average per GPU

Demo data, made up: the output of croutons scan --demo.

croutons scan ./logs

Nine findings. Start with the calls that repeat.

577 calls resend a prompt within 2 minutes or after a failure
Cap retries per request and back off on failure.
measuredsaves $36.96
7 of 84 agent runs made 21 or more model calls
Set a call budget per run and stop loops that repeat a tool.
measuredsaves $30.08
339 calls send 8,000+ tokens to get under 2% of that back
Retrieve only the passages the job needs.
estimatesaves $13.31

Demo data, made up: the output of croutons scan --demo. Findings can overlap, so the savings do not add up.

croutons scan --model config.json

How many requests each instance can hold at once, today and at half the tokens.

Workflow	Tokens a request	At once	At half the tokens
doc-qa	11,394	39	79
ticket-summary	1,925	234	469

Llama-3.1-8B-Instruct · 80 GB GPU · the cache holds about 452,037 tokens per instance

Demo data, made up: the output of croutons scan --demo.

Losses multiply. The cheapest one to cut is at the bottom.

A token saved at the request also saves its share of every loss above it. At a PUE of 1.4, each watt a GPU spends on a token the answer never needed costs 1.4 watts at the meter.

Power from the grid

Power reaching the serversLost to cooling and power conversion · facility teams

Useful GPU workLost to idle and underused GPUs · schedulers and clouds

Answers deliveredLost to wasted tokens · Croutons

Widths are illustrative, not measured. PUE 1.4 as an example. Croutons, November 2026.

Facility tools stop at the rack.
GPU tools stop at the chip.
AI logs stop at the API call.
Croutons starts at the request, the one layer nobody else checks.

Beside your stack, not in it

Croutons reads; it does not sit in the request path.

Reads what you already export

Request logs, engine metrics, GPU telemetry and facility data. Gateway logs, OpenTelemetry spans, vLLM and DCGM metrics all work as they are.

How the scan works

Runs on your own machine

Nothing is installed in production and no prompts leave the site.

How data is handled

Proved on your traffic

Start with a scan of your own logs. It shows where your compute goes. Every fix after that is measured against it, for cost and for answer accuracy.

npm install -g @croutons/cli

Then run croutons scan ./logs, or croutons scan --demo to try it on a made-up week.

One command, on logs you already keep

JSON, JSON Lines and CSV. Your logs are never uploaded.

Read the docs

Teams paying for AI by the token: A smaller bill for the same answers.
Teams running their own GPUs: More users on the cluster you already own.
GPU clouds: More tenant work on the capacity you’ve built.
Public-sector AI: More from finite national compute.

An open method, on public data

How to verify AI work and measure output per allocated megawatt, with every source dated.

See all

Method paper

Measuring AI Work in Malaysian Data Centres

An open, privacy-preserving method to verify AI use in Malaysian data centres and measure output per allocated megawatt: AI workload share, answers per MWh, and energy, water and carbon per answer, mapped to MITI, EECA 2024 and ISO/IEC 30134.

Public data

Malaysia Data Centre Research: Climate, Cooling and Public Data

Ten years of hourly weather for 11 Malaysian data-centre districts, which figures conflict and which to use, what is not public, and the full DESAC and EECA terms. Sourced and dated.

Guide

Building AI Data Centres in Malaysia: Power, Water, Rules and Land

Where AI data centres can be built in Malaysia in 2027–2030 and what decides it: grid power, water allocations, approvals, MITI efficiency targets, land prices and flood risk, with sources.

Questions

Do you need access to our systems?

Not to start. The scan runs on your machine, on log samples you choose.

How much will it save?

That depends on your traffic, so we don’t quote a number up front. The scan gives you your own baseline first.

Does sending less hurt accuracy?

It can if it’s done carelessly. Accuracy is measured on the same questions as cost, and a change that loses accuracy doesn’t ship.

Is our data safe?

Samples are used only for the agreed work and kept only as long as needed. Personal information is never sold.

Before you buy another GPU, see what yours are doing.

Request a briefing Scan your AI logs

More answers per megawatt
from your H100s

A chip serves more when less of its work is wasted.

Croutons finds the waste and proves the fix

Energy, memory and cost per answer

The workloads that waste capacity

Replayed on real prompts before it ships

Headroom in the operator’s own terms

Losses multiply. The cheapest one to cut is at the bottom.

Beside your stack, not in it

Reads what you already export

Runs on your own machine

Proved on your traffic

One command, on logs you already keep

An open method, on public data

Measuring AI Work in Malaysian Data Centres

Malaysia Data Centre Research: Climate, Cooling and Public Data

Building AI Data Centres in Malaysia: Power, Water, Rules and Land

Questions

Before you buy another GPU, see what yours are doing.

More answers per megawatt from your H100s

A chip serves more when less of its work is wasted.

Croutons reads telemetry from deployments running open models on your own accelerators

Energy, memory and cost per answer

The workloads that waste capacity

Replayed on real prompts before it ships

Headroom in the operator’s own terms

Reads what you already export

Runs on your own machine

One command, on logs you already keep

Measuring AI Work in Malaysian Data Centres

Malaysia Data Centre Research: Climate, Cooling and Public Data

Building AI Data Centres in Malaysia: Power, Water, Rules and Land

Questions

Before you buy another GPU, see what yours are doing.

More answers per megawatt
from your H100s