Only pay for GPUs that work.

We raise GPU utilization across inference and training, put each workload on the right chip and show the cost per model. Read-only to start, and we never touch your models or code.

smalBlu Console›GPU workloads

GPU workloads

$1,389/mo found
GPU fleet utilizationLive
Autostop idle g5.xlarge, 19:00 to 07:00$332/mo
AUTOSTOP_IDLE_GPUConfidence85%
Approve
Move nightly training to Spot capacity$664/mo
SPOT_SWAP_TRAININGConfidence75%
Approve
Downshift llama-3-8B to distilled variant$393/mo
MODEL_SWAPConfidence55%
Approve
  • Up to50%lower cloud bill
  • Up to30%better performance
  •  Day oneyour full GPU picture, read-only

What we do

Every GPU, working for its cost.

GPUs are the most expensive thing in your cloud, and often the least used. We fix both sides: how much you use, and what you pay for it.

  1. 1

    Keep every GPU busy

    GPU utilization tracked on every chip. Idle GPUs, notebooks and endpoints switched off automatically, and capacity reservations that never get used, found.

    smalBlu Console›GPU workloads

    11 GPUs idle since Friday night

    Notebooks and dev endpoints

    $8,400/mo
    notebook-ml-03, idle 62 hoursSwitch off
    endpoint-dev-02, idle 41 hoursSwitch off
    training-sweep-7, finished FridayRelease
    Reserved and never used: 4 GPUsFound
  2. 2

    Slice and pack workloads

    Small jobs share a GPU instead of each holding a whole one. Workloads are packed onto fewer GPUs, with memory and compute sized to each job.

    smalBlu Console›GPU sharing

    4 inference jobs, 1 GPU

    Before: 4 GPUs, each mostly waiting

    3 GPUs freed
    search-rank, 18% of one GPUShared
    support-bot, 26%Shared
    vision-qa, 21%Shared
    doc-parse, 14%Shared
  3. 3

    The right capacity, the right chip

    Autoscaling and EKS node pools sized for inference and training, capacity reservations for big runs, and the most cost-effective chip for every workload.

    smalBlu Console›Capacity

    Cheaper chips, same speed

    Simulated on 30 days of your jobs

    3 ready
    Move inference to a newer GPU type27% less
    Training on spot, with checkpoints62% less
    Reserve capacity for the big run31% less
  4. 4

    Know what every model costs

    Cost per model, team and project, in one view. GPU and AI API spend side by side, drift from your planned setup flagged, and budgets that alert you before they are breached.

    smalBlu Console›Model costs

    What each model costs

    This month, chips and API tokens

    4 models
    support-bot$18,400
    search-rank$9,200
    vision-qa$4,100
    Cost per million tokens$0.42

What we do

Every GPU, working for its cost.

GPUs are the most expensive thing in your cloud, and often the least used. We fix both sides: how much you use, and what you pay for it.

Keep every GPU busy

Idle GPUs, notebooks and endpoints switched off automatically.

  • Experiment capacity off overnight and at weekends
  • Reserved GPUs that never get used, found
  • Scale to zero when nothing is running

Slice and pack workloads

Small jobs share a GPU, instead of each holding a whole one.

  • GPU slicing and sharing, where your chips support it
  • Workloads packed onto fewer GPUs
  • Memory and compute sized to each job

The right capacity, the right chip

The most cost-effective chip for every workload.

  • Newer, cheaper instance types
  • AWS’s own AI chips, where your model supports them
  • The right mix of reserved, spot and on-demand
  • Capacity that follows demand, reserved for big training runs

Know what every model costs

Cost per model, team and project, in one view.

  • GPU and AI API spend side by side
  • Cost per million tokens, self-hosted or not
  • Budgets with alerts before they are breached

What we do

Every GPU, working for its cost.

GPUs are the most expensive thing in your cloud, and often the least used. We fix both sides: how much you use, and what you pay for it.

1 of 4

Keep every GPU busy

GPU utilization tracked on every chip. Idle GPUs, notebooks and endpoints switched off automatically, and capacity reservations that never get used, found.

smalBlu Console›GPU workloads

11 GPUs idle since Friday night

Notebooks and dev endpoints

$8,400/mo
notebook-ml-03, idle 62 hoursSwitch off
endpoint-dev-02, idle 41 hoursSwitch off
training-sweep-7, finished FridayRelease
Reserved and never used: 4 GPUsFound

2 of 4

Slice and pack workloads

Small jobs share a GPU instead of each holding a whole one. Workloads are packed onto fewer GPUs, with memory and compute sized to each job.

smalBlu Console›GPU sharing

4 inference jobs, 1 GPU

Before: 4 GPUs, each mostly waiting

3 GPUs freed
search-rank, 18% of one GPUShared
support-bot, 26%Shared
vision-qa, 21%Shared
doc-parse, 14%Shared

3 of 4

The right capacity, the right chip

Autoscaling and EKS node pools sized for inference and training, capacity reservations for big runs, and the most cost-effective chip for every workload.

smalBlu Console›Capacity

Cheaper chips, same speed

Simulated on 30 days of your jobs

3 ready
Move inference to a newer GPU type27% less
Training on spot, with checkpoints62% less
Reserve capacity for the big run31% less

4 of 4

Know what every model costs

Cost per model, team and project, in one view. GPU and AI API spend side by side, drift from your planned setup flagged, and budgets that alert you before they are breached.

smalBlu Console›Model costs

What each model costs

This month, chips and API tokens

4 models
support-bot$18,400
search-rank$9,200
vision-qa$4,100
Cost per million tokens$0.42

What we do

Every GPU, working for its cost.

GPUs are the most expensive thing in your cloud, and often the least used. We fix both sides: how much you use, and what you pay for it.

1

Keep every GPU busy

GPU utilization tracked on every chip. Idle GPUs, notebooks and endpoints switched off automatically, and capacity reservations that never get used, found.

smalBlu Console›GPU workloads

11 GPUs idle since Friday night

Notebooks and dev endpoints

$8,400/mo
notebook-ml-03, idle 62 hoursSwitch off
endpoint-dev-02, idle 41 hoursSwitch off
training-sweep-7, finished FridayRelease
Reserved and never used: 4 GPUsFound
2

Slice and pack workloads

Small jobs share a GPU instead of each holding a whole one. Workloads are packed onto fewer GPUs, with memory and compute sized to each job.

smalBlu Console›GPU sharing

4 inference jobs, 1 GPU

Before: 4 GPUs, each mostly waiting

3 GPUs freed
search-rank, 18% of one GPUShared
support-bot, 26%Shared
vision-qa, 21%Shared
doc-parse, 14%Shared
3

The right capacity, the right chip

Autoscaling and EKS node pools sized for inference and training, capacity reservations for big runs, and the most cost-effective chip for every workload.

smalBlu Console›Capacity

Cheaper chips, same speed

Simulated on 30 days of your jobs

3 ready
Move inference to a newer GPU type27% less
Training on spot, with checkpoints62% less
Reserve capacity for the big run31% less
4

Know what every model costs

Cost per model, team and project, in one view. GPU and AI API spend side by side, drift from your planned setup flagged, and budgets that alert you before they are breached.

smalBlu Console›Model costs

What each model costs

This month, chips and API tokens

4 models
support-bot$18,400
search-rank$9,200
vision-qa$4,100
Cost per million tokens$0.42

What we do

Every GPU, working for its cost.

GPUs are the most expensive thing in your cloud, and often the least used. We fix both sides: how much you use, and what you pay for it.

  1. 01

    Keep every GPU busy

    GPU utilization tracked on every chip. Idle GPUs, notebooks and endpoints switched off automatically, and capacity reservations that never get used, found.

  2. 02

    Slice and pack workloads

    Small jobs share a GPU instead of each holding a whole one. Workloads are packed onto fewer GPUs, with memory and compute sized to each job.

  3. 03

    The right capacity, the right chip

    Autoscaling and EKS node pools sized for inference and training, capacity reservations for big runs, and the most cost-effective chip for every workload.

  4. 04

    Know what every model costs

    Cost per model, team and project, in one view. GPU and AI API spend side by side, drift from your planned setup flagged, and budgets that alert you before they are breached.

How it works

Find. Simulate. Apply. Keep it off.

  1. 1

    Find

    Every GPU across your servers, clusters and AI services in one view, with what it costs and how much of it you use.

  2. 2

    Simulate

    Every change is tested against your real workloads first: what it saves, and what it does to speed.

  3. 3

    Apply

    Capacity, instance types, scaling and reservations change after your approval. Your models and code stay as they are.

  4. 4

    Keep it off

    New idle GPUs and drift are caught as they appear, and the cost per model is tracked week by week.

What we measure

Busy is not the same as working.

We show reserved, busy and working side by side, for NVIDIA GPUs and AWS’s own AI chips alike, so no single flattering number hides the waste.

MeasureWhat it meansWhy it matters
Paid hoursGPU and AI chip hours on your bill.Visible from day one, on every account.
ReservedThe share of paid capacity assigned to a job.Shows how much you hold, not how much you use.
BusyThe share of time any work ran on the chip.The number most dashboards stop at.
WorkingThe share of the chip doing useful work.The honest number. A busy GPU can be mostly waiting.
Memory usedPeak and typical memory on each chip.Decides which jobs can share a GPU.
Cost per modelEvery cost tied to one model: chips, storage, network and API tokens.The number your CFO asks for.

What you see

Your whole GPU fleet, on one screen.

What you run, what it costs and how much of it is working.

smalBlu Console›GPU workloads

GPU workloads

$1,389/mo found
GPU fleet utilizationLive
Autostop idle g5.xlarge, 19:00 to 07:00$332/mo
AUTOSTOP_IDLE_GPUConfidence85%
Approve
Move nightly training to Spot capacity$664/mo
SPOT_SWAP_TRAININGConfidence75%
Approve
Downshift llama-3-8B to distilled variant$393/mo
MODEL_SWAPConfidence55%
Approve
1234
  1. 1

    Every GPU, in one place

    Across servers, clusters and AI services.

  2. 2

    What it costs

    Hour by hour, and per model.

  3. 3

    How much is working

    Reserved, busy and working, side by side.

  4. 4

    What to change next

    Each change with its saving, ready to approve.

Your GPU fleet, hour by hour

An idle GPU costs the same as a busy one.

Each row is a GPU. Each square is an hour of the day. Switch the view to see what changes.

BusyIdle, and you payStopped, and you do not

Clear lines

What we change, and what we never touch.

We change where and how your workloads run. Never the models, and never the code.

  • We change capacityInstance types, scaling, slicing, scheduling and reservations, after your approval.
  • We never touch your models or codeWhen a newer serving engine or a different chip would pay off, we show you the numbers and your team decides.
  • Read-only to startSee your full GPU picture from day one. If you don’t collect GPU metrics yet, we help you switch them on.
  • Every change on recordWhat changed, who approved it, and what it saved.
Up to50%

lower cloud bill

Jobs that leave your team’s list.

Switch off idle notebooksDone by smalBlu
Slice GPUs for small inference jobsDone by smalBlu
Move training to spot capacityDone by smalBlu
Right-size 12 GPU instancesDone by smalBlu
Reserve capacity for the big runDone by smalBlu
Show what each model costsDone by smalBlu

Common questions

What teams ask us about AI Infrastructure.

Will you touch our models or code?

No. smalBlu changes where and how your workloads run, never the models or the code themselves.

Do we need to change anything to start?

No. Connect read-only and see your GPU picture right away: every GPU, what it costs and how much of it you use. If you don’t collect GPU metrics yet, we help you switch them on.

Which chips do you support?

NVIDIA GPUs on AWS, and AWS’s own AI chips, Trainium and Inferentia. We measure usage the same way for all of them.

Kubernetes or plain servers?

Both. EKS clusters and node pools, plain EC2 servers, auto scaling groups and SageMaker endpoints, for inference and training alike.

Is GPU slicing safe for production?

We only share a GPU where the chip and the workload allow it, and we simulate the effect first. Jobs that need a whole GPU keep one.

Can you track our AI API spend too?

Yes. Bedrock and other AI API spend sits beside your GPU costs, so every model gets one cost, self-hosted or not.

200 OK. Your cloud, sorted.

Get a free review of your cloud and see what smalBlu would run, fix and improve in your first month.

Get your free cloud review