modal

🇺🇸 USA · Modal

Status: 🟩 COMPLETE 🟦 LIVING Last updated: 2026-06-26 Plain-English tagline: Serverless GPU / CPU compute for AI — write Python locally, deploy to cloud GPUs with zero infrastructure setup. Pay per second of execution. The “Vercel of GPU workloads.”

---

Front-matter facts

Field	Value
Vendor	Modal Labs (San Francisco, USA)
Country / origin	🇺🇸 USA
Recommended for Australian users?	✅ Yes — fully accessible from AUS
Privacy summary	No training on customer data; your code runs in isolated containers
Free tier	US$30/month free compute
Paid tiers	Pay-per-second on top of free tier; Team / Enterprise quoted
First released	2022
Last reviewed	2026-06-26
Official site	https://modal.com

---

What it is

Modal is serverless GPU / CPU compute for AI workloads. You write Python functions locally, decorate them, and Modal runs them on cloud GPUs (Nvidia A100, H100, L40S, etc.) without you managing servers, containers, or infrastructure.

Example workflow:

import modal
 
app = modal.App("my-app")
 
@app.function(gpu="A100")
def generate_image(prompt):
    # runs on cloud A100 GPU when called
    # all dependencies handled automatically
    pass
 
@app.local_entrypoint()
def main():
    result = generate_image.remote("a kookaburra")

Modal handles:

• Container building automatically
• Dependency installation (pip / apt / etc.)
• Cold starts (typically 1-10 seconds)
• Auto-scaling (zero to thousands of concurrent executions)
• Per-second billing

Use cases:

• Run open-source AI models on demand without infrastructure
• Batch processing at scale
• Fine-tuning models with GPU access
• Web apps / APIs with serverless GPU backends
• Background jobs for AI processing

---

What you’d use it for

• Self-hosted model inference on cloud GPUs (Llama, Mistral, Whisper, Stable Diffusion, etc.) without server management
• Fine-tuning models on your data
• Large-scale batch processing of AI workloads
• Building AI products with serverless backend
• Running specific open-source models not available on Together / Fireworks
• Data science / ML research with cloud GPU access

---

How to use from Australia

1. Sign up at modal.com — US$30/month free compute
2. Install: pip install modal
3. Authenticate: modal token new
4. Write Python functions with @app.function() decorators
5. Deploy: modal deploy my_script.py
6. Call from local Python or expose as HTTPS endpoint
7. AUS card accepted

---

What it costs

Free tier

• US$30/month free compute (substantial — covers many small projects)

Per-second pricing

• CPU: very cheap (~US$0.0001/sec for small CPUs)
• GPU T4: ~US$0.00018/sec(US$0.65/hour)
• GPU A100 80GB: ~US$0.001/sec(US$3.60/hour)
• GPU H100 80GB: ~US$0.0024/sec(US$8.60/hour)
• Per-second billing means you only pay when functions actually execute

Storage

• Some persistent volume storage included free
• Additional storage charged per GB-month

Hidden costs

• Long-running idle GPUs can add up if you mis-configure
• Cold starts are fast but real (1-10s); design for them

---

How it compares to alternatives

Aspect	Modal	Lambda Labs	RunPod	CoreWeave	AWS GPU instances
Serverless (no provisioning)	Yes (best)	Limited	Limited	No (rent GPUs)	Limited
Per-second billing	Yes	Per-hour	Per-hour or per-second	Per-hour	Per-second
Auto-scaling	Best	Limited	Limited	Manual	Manual / auto-scaling
GPU types	A100 / H100 / T4 / L40S	Broad selection	Broad selection	Broad enterprise	Broad
Best for	Serverless AI workloads	Cheap rented GPUs	Cheap rented GPUs	Enterprise GPU clusters	AWS-stack

For developers wanting serverless GPU access without managing infrastructure, Modal is the cleanest option. For renting raw GPUs cheaply, RunPod / Lambda Labs.

---

Privacy / data handling

• Code runs in isolated containers per request
• No training on customer code
• US data centres
• For AUS data residency, AWS / Azure / GCP with AUS regions

---

Recent changes

• 2026: H200 / Blackwell GPU support
• 2025: Persistent volumes + sandbox improvements
• 2024: Major adoption among AI developers

---

Gotchas

• Cold starts are 1-10s typically; design async workflows accordingly
• Python-first — for non-Python AI work, less natural
• Per-second billing means watch your idle time — don’t leave functions running idle
• For inference of common models (Llama, Mistral), Together / Fireworks / Groq usually simpler than self-hosting on Modal
• For Bible-Quest-scale projects, Modal is overkill — Vercel / Supabase covers most needs

---

See also

• Lambda Labs 🟥
• RunPod 🟥
• CoreWeave 🟥
• AWS overview 🟩 🟦
• Together AI 🟩 🟦
• Fireworks AI 🟩 🟦
• Replicate 🟩 🟦
• Hugging Face 🟩 🟦
• Nvidia AI 🟩 🟦
• What is the cloud? 🟩 🟦

---

Sources

• Modal — modal.com
• Modal — Pricing
• Modal — Documentation