Google Colab Pro vs Pro+ (2026): Compute Units, GPU, RAM, Runtime Limits & Which Plan to Choose

Google Colab—also known as Google Colaboratory—is a browser-based notebook platform from Google that lets you run Python code on cloud-hosted computing resources.

It is widely used for:

• Python programming;
• machine learning;
• artificial intelligence;
• PyTorch;
• TensorFlow;
• JAX;
• data science;
• computer vision;
• generative AI;
• large language model experiments;
• GPU-accelerated workloads.

One of Colab’s biggest advantages is that you can use cloud CPU, GPU and TPU resources without manually building and maintaining a local AI workstation.

However, choosing between:

• Colab Free
• Colab Pro
• Colab Pro+
• Pay As You Go

can be confusing.

The most important thing to understand is:

Paying for Colab does not mean you are buying a dedicated GPU.

Paid Colab generally gives you better access to compute resources, premium accelerators, higher-memory configurations and Compute Units—but hardware allocation remains dynamic and subject to availability. Google explicitly states that GPU and TPU types available in Colab vary over time.

Quick Answer: Colab Free vs Pro vs Pro+

Google states that free Colab notebooks can generally run for up to 12 hours, depending on availability and usage patterns. Colab Pro, Pro+ and Pay As You Go provide increased compute availability based on Compute Unit balance. Pro+ can support continuous code execution for up to 24 hours when sufficient Compute Units are available.

What Is Google Colab?

Google Colab is a hosted Jupyter notebook environment.

A notebook normally uses the:

.ipynb

file format.

Your code runs on a remote runtime rather than directly inside your browser.

A simplified architecture is:

Your Browser
    ↓
Google Colab Interface
    ↓
Cloud Runtime / Virtual Machine
    ↓
CPU + System RAM + Optional GPU/TPU
    ↓
Temporary Runtime Storage

Google states that the code is executed inside a virtual machine associated with your session and that those virtual machines are deleted after inactivity and have service-enforced maximum lifetimes.

This leads to one of the most important Colab rules:

Treat the runtime as temporary.

Do not keep the only copy of:

• trained models;
• datasets;
• checkpoints;
• results;
• logs;
• configuration files;

inside the temporary Colab runtime.

What Can Google Colab Be Used For?

Common workloads include:

• learning Python;
• data analysis;
• NumPy;
• Pandas;
• machine learning;
• neural-network training;
• PyTorch;
• TensorFlow;
• JAX;
• image generation;
• computer vision;
• natural-language processing;
• LLM experimentation;
• embeddings;
• inference;
• data preprocessing;
• scientific computing;
• educational notebooks;
• research;
• prototyping.

Colab is particularly attractive when you occasionally need GPU acceleration but do not want to purchase and maintain a dedicated GPU workstation.

What Is Colab Pro?

Colab Pro is a paid subscription tier that provides improved access to Colab compute resources compared with the free version.

Typical benefits can include:

• paid Compute Units;
• improved access to GPUs;
• premium accelerator access where available;
• high-system-memory configurations subject to availability;
• longer runtimes;
• fewer free-tier restrictions.

However:

Colab Pro does not guarantee a particular GPU.

It also does not turn Colab into a permanently reserved VM.

What Is Colab Pro+?

Colab Pro+ is the higher paid consumer subscription tier.

The most important practical difference is background execution.

Google states that Pro+ users can continue executing notebooks after closing the browser tab, provided Compute Units remain available.

Google also currently states that Pro+ supports continuous code execution for up to 24 hours, subject to sufficient Compute Units.

This can be especially useful for:

• long model-training jobs;
• overnight experiments;
• data preprocessing;
• batch inference;
• hyperparameter testing;
• long-running research workloads.

But:

Pro+ is still not a permanent 24×7 cloud server.

Does Colab Pro or Pro+ Guarantee an NVIDIA A100?

No.

This is one of the most important points to understand before purchasing.

Paying for Colab Pro or Pro+ does not mean you are purchasing permanent access to:

• NVIDIA A100;
• NVIDIA L4;
• NVIDIA T4;
• or any other exact GPU.

Google says the types of GPU and TPU available in Colab change over time and that premium GPUs on paid plans are still subject to availability.

Therefore:

Colab subscription = improved access to pooled compute

not:

Colab subscription = dedicated A100 reservation

If your workload absolutely requires the same hardware every time, a dedicated cloud GPU VM is more appropriate.

What GPUs Can Google Colab Provide?

The exact hardware can change.

Users may encounter accelerators such as:

• NVIDIA T4;
• NVIDIA L4;
• NVIDIA A100;
• other supported GPU types;
• TPUs.

Do not rely on a fixed public list forever.

Google explicitly states that GPU/TPU types available in Colab vary over time.

How to Check Which GPU You Actually Received

Run:

!nvidia-smi

This usually shows:

• GPU model;
• VRAM;
• driver version;
• CUDA compatibility;
• GPU utilization;
• GPU processes;
• GPU memory usage.

Always check the actual runtime.

Do not assume the GPU from your subscription name.

Check GPU Using PyTorch

Use:

import torch

print("CUDA available:", torch.cuda.is_available())

if torch.cuda.is_available():
    print("GPU:", torch.cuda.get_device_name(0))

If:

CUDA available: False

then PyTorch currently does not see an accessible CUDA GPU.

Check GPU VRAM

Use:

import torch

if torch.cuda.is_available():
    props = torch.cuda.get_device_properties(0)
    print("GPU:", props.name)
    print("VRAM:", round(props.total_memory / 1024**3, 2), "GB")

VRAM becomes especially important for:

• large language models;
• large image-generation models;
• big batch sizes;
• long context windows;
• large neural networks;
• high-resolution training.

What Are Colab Compute Units?

Compute Units are the paid resource credits used by eligible Colab compute.

A very common misunderstanding is:

1 Compute Unit = 1 hour

That is not correct.

A Compute Unit is not a fixed time unit.

Different resource types consume Compute Units at different rates.

A conceptual model is:

Compute Unit usage
≈
Resource consumption rate × time

Therefore:

100 Compute Units

does not automatically mean:

100 GPU hours

A premium GPU may consume Compute Units faster than a lower-cost resource.

What Can Consume Compute Units?

Depending on the runtime and current Colab offering, Compute Units can be used by resources such as:

• GPUs;
• premium GPUs;
• TPUs;
• high-system-memory runtimes;
• longer-running paid compute;
• Pro+ background execution.

The exact consumption rate can change.

Therefore:

Always check the current Compute Unit rate displayed inside Colab before starting an expensive workload.

Can You Buy Compute Units Without Pro?

Google supports Pay As You Go Compute Units.

Google’s Colab documentation says anyone can purchase additional Compute Units through Pay As You Go without necessarily holding a Pro subscription.

This can make sense if you:

• only need GPU occasionally;
• do not want a monthly subscription;
• have irregular AI workloads.

What Happens When Compute Units Run Out?

When the available paid Compute Units are exhausted, paid-resource access can become restricted.

You may experience:

• inability to select premium GPU;
• reduced GPU availability;
• high-memory restrictions;
• runtime termination;
• inability to continue long paid sessions.

Google states that backend termination can occur if available Compute Units are exhausted on Pro, Pro+ or Pay As You Go.

Google Colab Pricing

Colab pricing can vary by:

• country;
• currency;
• taxes;
• promotional offers;
• account;
• subscription availability;
• Pay As You Go pricing changes.

Google’s paid-service terms explicitly allow prices and usage limits to change over time.

For that reason, this article should not hard-code a monthly price that may become wrong.

Before purchasing:

1. open your Colab account;
2. select the paid-plan or Compute Unit purchase screen;
3. check the current local price;
4. check current Compute Unit allocation;
5. check applicable taxes.

That is more reliable than using an old blog post or screenshot.

System RAM vs GPU VRAM

These are completely different resources.

System RAM

Used by:

• Python processes;
• Pandas;
• NumPy;
• dataset preprocessing;
• CPU-side operations;
• decompression;
• file caching.

GPU VRAM

Used by:

• model weights;
• CUDA tensors;
• gradients;
• activations;
• training;
• inference;
• GPU operations.

Therefore:

More system RAM does not automatically give you more GPU VRAM.

You could have:

System RAM: 50 GB
GPU VRAM:   16 GB

and still receive:

CUDA out of memory

because the GPU memory is full.

Check System RAM in Colab

Run:

!free -h

Or use Python:

import psutil

ram = psutil.virtual_memory()

print("Total RAM:", round(ram.total / 1024**3, 2), "GB")
print("Available RAM:", round(ram.available / 1024**3, 2), "GB")
print("Used:", ram.percent, "%")

Do not assume every Colab session receives exactly the same RAM.

Google states that free users access standard system-memory profiles and paid users can access high-system-memory machines depending on availability and Compute Unit balance.

High RAM Does Not Mean High VRAM

This deserves separate emphasis.

Suppose:

System RAM: 52 GB
GPU VRAM:   16 GB

A model requiring:

22 GB GPU VRAM

will still not fit simply because the Colab runtime has 52 GB system RAM.

The bottleneck is GPU memory.

Check CPU Configuration

Run:

!lscpu

or:

import os

print("Logical CPUs:", os.cpu_count())

This is useful because some workloads are CPU-bound rather than GPU-bound.

Check Temporary Disk Space

Run:

!df -h

The common runtime working location is:

/content/

Treat this as temporary storage.

If the runtime is deleted, files stored only there can disappear.

How Long Can Colab Run?

Google currently says:

Free Colab notebooks can run for up to 12 hours, depending on availability and usage patterns.

In general, managed Colab runtimes remain subject to runtime limits.

Paid plans offer increased compute availability based on Compute Unit balance.

Pro+

Google specifically states that Colab Pro+ can support continuous code execution for up to 24 hours when sufficient Compute Units remain.

That does not guarantee that every workload always reaches 24 hours.

What Is Background Execution?

Background execution is a Pro+ feature.

It allows eligible notebook code to continue running even after the browser tab is closed, while Compute Units remain available.

Useful scenarios include:

• overnight training;
• batch processing;
• long preprocessing jobs;
• inference;
• hyperparameter experiments.

This is one of the strongest reasons to choose Pro+ over Pro.

Can I Close My Browser While Colab Is Running?

Free

Do not rely on it.

Pro

Do not assume closing the browser will provide background execution.

Pro+

Background execution is specifically available, subject to Compute Units and runtime conditions.

Can Google Colab Run 24×7?

No, not as a normal managed Colab runtime.

Colab is primarily an interactive notebook service.

Even Pro+ continuous execution is limited and policy-dependent.

If you require:

• permanent processes;
• stable IP;
• long-running APIs;
• persistent server applications;
• dedicated GPU;
• predictable infrastructure;

use an appropriate cloud VM, container platform or dedicated server instead.

Can Colab Be Used as a Production Server?

Generally, normal managed Colab should not be treated as a production application server.

It is better suited for:

• experimentation;
• notebooks;
• research;
• learning;
• prototyping;
• temporary compute.

Google’s paid-service terms also impose usage restrictions and resource limits that can vary over time.

Why Does Colab Disconnect?

Possible reasons include:

• idle timeout;
• maximum runtime limit;
• exhausted Compute Units;
• resource pressure;
• VM recycling;
• browser/network interruption;
• runtime crash;
• excessive system RAM usage;
• GPU failure;
• policy limits.

A professional workflow must therefore be designed to survive disconnections.

Save Checkpoints During Long Training

Do not wait until training finishes before saving.

Example in PyTorch:

torch.save({
    'epoch': epoch,
    'model_state_dict': model.state_dict(),
    'optimizer_state_dict': optimizer.state_dict()
}, '/content/drive/MyDrive/checkpoint.pt')

This allows you to resume after a disconnected runtime.

For long GPU jobs, checkpointing can save:

• time;
• Compute Units;
• training progress.

Mount Google Drive

Use:

from google.colab import drive

drive.mount('/content/drive')

Files are commonly available under:

/content/drive/MyDrive/

Google Drive is useful for:

• checkpoints;
• trained models;
• datasets;
• results;
• logs;
• notebook outputs.

Do Not Treat Google Drive as Local SSD

Mounted Google Drive can be slower than local runtime storage.

Processing thousands of tiny files directly from Drive may become inefficient.

A useful workflow is:

Google Drive
    ↓
Copy archive to /content/
    ↓
Extract locally
    ↓
Train/process
    ↓
Save results back to Drive

Example:

!cp /content/drive/MyDrive/dataset.zip /content/
!unzip /content/dataset.zip -d /content/dataset/

Then process from:

/content/dataset/

Google Drive Input/Output Error

A possible error is:

OSError: [Errno 5] Input/output error

Possible causes include:

• excessive small-file access;
• temporary Drive mount issue;
• storage quota;
• remote-storage limitations;
• unstable runtime.

Possible approaches include:

• remount Drive;
• reduce small-file operations;
• archive datasets;
• copy data locally before processing.

Google Colab GPU Not Available

If:

torch.cuda.is_available()

returns:

False

check:

Runtime → Change runtime type

Then select an available GPU accelerator.

Reconnect and run:

!nvidia-smi

Then:

import torch

print(torch.cuda.is_available())

Possible reasons GPU is unavailable include:

• GPU runtime not selected;
• GPU resources temporarily unavailable;
• account/resource limits;
• exhausted Compute Units;
• high demand;
• package/environment problem.

CUDA Out of Memory Error

A very common error is:

CUDA out of memory

This normally means the GPU does not have enough available VRAM for the requested operation.

Possible fixes include:

• reduce batch size;
• reduce image resolution;
• reduce sequence length;
• use a smaller model;
• use mixed precision;
• use gradient accumulation;
• use gradient checkpointing;
• free unused tensors;
• restart the runtime where appropriate;
• use a GPU with more VRAM if available.

Check GPU Memory Usage

Run:

!nvidia-smi

For PyTorch:

print(torch.cuda.memory_summary())

Why Restarting Runtime Can Help CUDA OOM

Python may release objects while the CUDA allocator retains reserved memory for reuse.

After many model loads or experiments, VRAM may become fragmented or heavily reserved.

A runtime restart can clear GPU state.

But restart also deletes temporary runtime data.

Save important files first.

Colab System RAM Exhausted

Symptoms can include:

• notebook freezing;
• runtime crashing;
• process killed;
• kernel restart;
• dataset-loading errors.

Possible causes:

• loading huge datasets all at once;
• very large Pandas DataFrames;
• multiple copies of arrays;
• memory leaks;
• decompression;
• excessive preprocessing.

Possible solutions:

• process data in batches;
• stream datasets;
• read CSV files in chunks;
• delete unused objects.

Example:

del large_object

import gc
gc.collect()

Why Is GPU Utilization Low?

A GPU runtime does not guarantee that your code is actually using the GPU.

Check:

!nvidia-smi

If utilization remains near 0%, possible causes include:

• model still on CPU;
• tensors still on CPU;
• slow data loader;
• CPU preprocessing bottleneck;
• batch size too small;
• frequent CPU↔GPU transfers;
• workload not suitable for GPU.

PyTorch: Confirm Model Is on GPU

Example:

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = model.to(device)

And move input tensors:

inputs = inputs.to(device)

If the model is on GPU but the data remains on CPU, errors or performance problems can occur.

Mixed Precision to Reduce VRAM Usage

Mixed precision can significantly reduce GPU-memory requirements for supported workloads.

Typical PyTorch workflows use:

torch.autocast

and appropriate gradient-scaling techniques.

It can help:

• reduce VRAM;
• improve throughput;
• allow larger models or batches.

But numerical behaviour should be validated for the workload.

Gradient Accumulation

Suppose you want an effective batch size of:

64

but only:

8

fits in GPU memory.

You can accumulate gradients across multiple mini-batches.

Conceptually:

8 samples × 8 accumulation steps
≈ effective batch size 64

This can reduce peak VRAM consumption.

Gradient Checkpointing

Gradient checkpointing saves GPU memory by recomputing some intermediate activations during backward propagation.

Trade-off:

lower memory usage

in exchange for:

more computation

Useful for large neural networks.

Is Colab Good for LLM Training?

It depends on model size.

Colab can be excellent for:

• small models;
• fine-tuning;
• LoRA;
• QLoRA;
• inference;
• experimentation;
• educational work.

It can be unsuitable for:

• very large models;
• multi-GPU training;
• long distributed jobs;
• production workloads;
• workloads requiring guaranteed hardware.

Is Colab Good for Stable Diffusion or Image Generation?

Yes, depending on:

• available GPU;
• VRAM;
• model;
• resolution;
• batch size;
• platform rules.

But runtime availability and hardware assignment can vary.

Is Colab Pro Worth It?

Colab Pro can make sense if:

• you use Colab regularly;
• free GPU access is too restrictive;
• you need paid Compute Units;
• you want better access to premium resources;
• you occasionally need high system RAM.

It may not be worthwhile if:

• you rarely use GPU;
• your work runs locally;
• you need guaranteed exact hardware;
• you need a permanent server.

Is Colab Pro+ Worth It?

Pro+ is most useful when background execution and longer continuous workloads materially affect your work.

Choose Pro+ when you frequently run:

• long training jobs;
• overnight experiments;
• batch inference;
• long preprocessing jobs.

Google’s current documentation confirms background execution and up to 24-hour continuous code execution when sufficient Compute Units remain.

Colab Pro vs Pay As You Go

Pro

Better for:

• recurring Colab usage;
• users who want a monthly subscription;
• regular access to paid compute.

Pay As You Go

Better for:

• occasional GPU users;
• irregular workloads;
• users who only need extra Compute Units sometimes.

Because pricing and CU allocations change, compare the current purchase screen before choosing.

Colab vs Local GPU PC

Colab Advantages

• no GPU purchase;
• no hardware maintenance;
• browser access;
• ready Python environment;
• quick experiments;
• flexible cloud resources.

Local GPU Advantages

• predictable hardware;
• no session timeout;
• permanent storage;
• complete environment control;
• no Compute Unit charges;
• offline work.

Colab vs Dedicated Cloud GPU VM

Colab

Best for:

• notebooks;
• experiments;
• intermittent workloads;
• students;
• research;
• prototyping.

Dedicated GPU VM

Best for:

• guaranteed GPU;
• stable infrastructure;
• persistent services;
• production;
• APIs;
• custom drivers;
• exact hardware requirements;
• long-running workloads.

Colab Local Runtime

Colab also supports connecting the Colab interface to a local runtime.

Google warns that connecting a local runtime gives notebook code access to your local machine, including the ability to read, write or delete files. Therefore, only connect notebooks you trust.

This option can be useful if you like the Colab interface but want code to run on your own GPU workstation.

Important Security Warning

A notebook can execute arbitrary code.

Be careful with notebooks downloaded from:

• GitHub;
• forums;
• Discord;
• Telegram;
• random websites.

A malicious notebook could potentially:

• access mounted Google Drive data;
• steal credentials entered in the notebook;
• upload private files;
• delete data;
• execute unwanted commands.

Never blindly click:

Run all

on an untrusted notebook.

Secrets and API Keys

Avoid hard-coding API keys directly into notebooks that will be shared.

Bad:

API_KEY = "my-secret-key"

If you share the notebook, that key may become exposed.

Use secure secret-management mechanisms where possible.

Google Colab Troubleshooting Checklist

When a Colab workload fails, check:

GPU

!nvidia-smi

CUDA

import torch
print(torch.cuda.is_available())

VRAM

print(torch.cuda.memory_summary())

System RAM

!free -h

CPU

!lscpu

Disk

!df -h

PyTorch Version

import torch
print(torch.__version__)

CUDA Runtime Information

print(torch.version.cuda)

Common Mistakes

Mistake 1: Assuming Pro Guarantees A100

It does not.

Mistake 2: Thinking 100 Compute Units = 100 Hours

Compute Unit consumption depends on the resource.

Mistake 3: Confusing RAM With VRAM

More system RAM does not increase GPU memory.

Mistake 4: Treating /content as Permanent Storage

It can disappear when the runtime ends.

Mistake 5: Not Saving Checkpoints

A disconnect can destroy hours of training progress.

Mistake 6: Selecting the Most Expensive GPU for Every Task

CPU workloads do not need GPU.

Mistake 7: Running Large Datasets Directly From Drive

Local runtime storage can be faster for intensive small-file processing.

Mistake 8: Assuming Paid Colab Has Unlimited Runtime

It does not.

Mistake 9: Using Colab as a Permanent Production Server

Colab is primarily an interactive notebook service.

Mistake 10: Trusting Random Notebooks

Notebook code can be dangerous.

Frequently Asked Questions

1. What is Google Colab?

Google Colab is a Google-hosted Jupyter notebook environment for Python and computational workloads.

2. Is Google Colab free?

Yes, a free tier exists, with resource limits and dynamic availability.

3. What is Colab Pro?

A paid subscription that provides improved access to compute resources and Compute Units.

4. What is Colab Pro+?

A higher paid tier that includes features such as background execution and longer continuous execution capability.

5. Does Colab Pro guarantee a GPU?

No.

6. Does Colab Pro+ guarantee an A100?

No.

7. Which GPUs does Colab provide?

Available GPU types change over time and depend on availability.

8. What are Compute Units?

Paid resource credits used by eligible Colab compute resources.

9. Is one Compute Unit one hour?

No.

10. Can Compute Units run out?

Yes.

11. What happens when they run out?

Paid-resource access can become restricted and runtimes may terminate.

12. Can I buy Compute Units without Pro?

Google supports Pay As You Go Compute Units.

13. How long can free Colab run?

Google says up to 12 hours depending on availability and usage patterns.

14. How long can Pro+ run?

Google currently says continuous code execution can run up to 24 hours when sufficient Compute Units remain.

15. Can Pro+ run when the browser is closed?

Yes, background execution is a Pro+ feature while eligible resources/Compute Units remain available.

16. Is Colab a permanent server?

No.

17. Can Colab run 24×7?

Do not rely on normal managed Colab for permanent 24×7 service.

18. What is CUDA out of memory?

The requested GPU operation does not fit into available VRAM.

19. Will high RAM fix CUDA OOM?

No. System RAM and GPU VRAM are separate.

20. How do I check the GPU?

Run:

!nvidia-smi

21. How do I check Colab RAM?

Run:

!free -h

22. How do I save work permanently?

Save notebooks and important outputs to persistent storage such as Google Drive.

23. Can I train LLMs on Colab?

Yes for many small and medium experiments, but hardware and runtime limits can restrict larger models.

24. Is Colab Pro worth it?

It can be worthwhile for regular users who need better paid compute access.

25. Is Pro+ worth it?

It is particularly attractive for users who need background execution and longer-running jobs.

Final Recommendation

Choose Colab Free when:

• you are learning;
• GPU usage is occasional;
• interruptions do not matter;
• workloads are small.

Choose Colab Pro when:

• you use Colab regularly;
• you want paid Compute Units;
• free resource availability is too restrictive;
• you need better access to premium compute.

Choose Colab Pro+ when:

• background execution matters;
• you frequently run long jobs;
• overnight training is common;
• the larger paid-resource allowance justifies the higher cost.

Choose Pay As You Go when:

• your GPU needs are irregular;
• you do not want a monthly subscription;
• you only occasionally need extra Compute Units.

Choose a dedicated GPU VM or workstation when:

• exact GPU hardware is mandatory;
• you require stable resources;
• you need 24×7 operation;
• production workloads depend on the environment;
• interruptions are unacceptable.

The most important principle is:

Do not purchase Colab Pro or Pro+ expecting a permanently assigned A100 or unlimited GPU runtime.

Google Colab uses dynamically allocated cloud resources, and availability, Compute Unit consumption and hardware types can change.

Technical Disclaimer

Google can change:

• subscription pricing;
• Compute Unit allocations;
• Compute Unit consumption rates;
• GPU/TPU availability;
• runtime limits;
• paid-plan benefits;
• resource restrictions.

Always check the current Colab purchase screen and official FAQ before making a purchase decision or calculating long-term GPU costs.

Google Colab Pro vs Pro+: GPU, Compute Units & Runtime Limits

Compare Google Colab Free, Pro and Pro+ in 2026. Learn about Compute Units, GPUs, RAM vs VRAM, 12/24-hour runtime limits, background execution, CUDA OOM and which Colab plan to choose.

#GoogleColab #Colab #ColabPro #ColabProPlus #ColabFree #ComputeUnits #GPU #CloudGPU #A100 #L4 #T4 #CUDA #CUDAOutOfMemory #VRAM #SystemRAM #MachineLearning #ArtificialIntelligence #DeepLearning #PyTorch #TensorFlow #JAX #Python #DataScience #LLM #GenerativeAI #ModelTraining #GPUTraining #CloudComputing #Jupyter #JupyterNotebook #GoogleDrive #Runtime #RuntimeLimit #BackgroundExecution #HighRAM #GPUComputing #NVIDIAGPU #AITraining #AITools #PythonProgramming #CloudNotebook #ColabGPU #ColabCUDA #ColabRAM #PayAsYouGo #GoogleCloud #MLTraining #BISONKnowledgebase #BISONKB #TechGuide