GPU credits for AI startups

Training and serving models is expensive, and GPUs are the line item that scares most AI founders. Several programs exist specifically to cut that cost: through cloud GPU credits, discounted hardware, or dedicated compute grants.

The programs split into two lanes. The hyperscalers fold GPU capacity into their broader startup credits, while dedicated GPU clouds (Lambda, CoreWeave, RunPod) and serving platforms like Together AI, Replicate, and Groq run startup programs of their own, priced for teams that live on the GPU rather than visit it. NVIDIA Inception sits alongside both, pairing preferred hardware pricing with partner cloud credits and free technical training.

These are the GPU and compute programs we track for AI startups. Each links to the full eligibility and how to apply, and because the programs are independent, joining Inception doesn’t stop you stacking cloud GPU credits on top of it.

GPU credits for AI startups

19 programs

Open to all

NVIDIA Inception

NVIDIA

Free membership with preferred pricing on select NVIDIA hardware and software, free cloud credits from NVIDIA and partners, and free self-paced technical training.

$100k in credits

GPUAI creditsDevTools
Open to all

Google for Startups Cloud Program

Google Cloud

Up to $200,000 USD in Google Cloud credits over two years, up to $350,000 USD for AI-first startups

$200k in credits

CloudAI creditsDatabase
Open to all

Baseten Inference for Startups

Baseten

Baseten startup credits and perks

AI creditsGPU
Open to all

Nebius for Startups

Nebius

Up to $150K in Nebius credits

$150k in credits

GPUCloud
Open to all

Crusoe Cloud for Startups

Crusoe Energy

Crusoe Energy startup credits and perks

GPUCloud
Open to all

Paperspace for Startups

Paperspace (DigitalOcean)

Paperspace (DigitalOcean) startup credits and perks

GPUCloud
Open to all

Vast.ai for Startups

Vast.ai

Vast.ai startup credits and perks

GPUCloud
Open to all

Together Compute for Startups

Together AI

Together AI startup credits and perks

AI creditsGPU
Open to all

Anyscale for Startups

Anyscale

Anyscale startup credits and perks

$20k in credits

AI creditsGPUCloud
Open to all

Baseten AI Startup Program

Baseten

Baseten startup credits and perks

AI creditsGPU
Open to all

CoreWeave Startup Accelerator Program

CoreWeave

CoreWeave startup credits and perks

GPUCloud
Open to all

CoreWeave for Startups

CoreWeave

CoreWeave startup credits and perks

GPUCloud
Open to all

Lambda GPU Cloud for Startups

Lambda

Lambda startup credits and perks

GPUCloud
Open to all

RunPod for Startups

RunPod

RunPod startup credits and perks

$1k in credits

GPUCloud
Open to all

Modal for Startups

Modal

Modal startup credits and perks

AI creditsCloudGPU
Open to all

Groq for Startups

Groq

Groq startup credits and perks

AI creditsGPU
Open to all

Together AI for Startups

Together AI

Together AI startup credits and perks

AI creditsGPU
Open to all

Replicate for Startups

Replicate

Replicate startup credits and perks

AI creditsGPU
Open to all

Vultr for Startups

Vultr

Vultr startup credits and perks

CloudGPU

Common questions.

Straight answers. For the full picture, read the FAQ.

NVIDIA Inception offers hardware discounts and credits, and the major clouds include GPU capacity inside their AI startup tiers. Together they can cover a meaningful share of early training and inference costs.

Yes, joining Inception is free. It unlocks discounts, credits, and technical resources; the specific benefits depend on your stage and program partners.

Take a hyperscaler’s program when your whole stack lives there anyway, since one credit pool then covers training alongside everything else. Dedicated GPU clouds like Lambda, CoreWeave, and RunPod make sense when compute is most of your bill and you want GPU-first pricing. Many teams run both.

Both, but from different providers. Training-oriented GPU clouds discount raw compute, while inference platforms in the catalog, including Together AI, Replicate, Groq, and Modal, run startup programs aimed at serving costs. If your model is already trained, the inference-side programs are the ones to read first.