GPUBank vs Replicate
Replicate is a San Francisco platform, joining Cloudflare since November 2025, that runs public and custom models behind an API and bills GPU time per second. It is a serverless model API, not a GPU cloud: there is no VM, SSH or disk, and its H100 80GB costs $5.49 per GPU-hour ($0.001525 per second) with the form factor not stated. GPUBank rents an H100 SXM instance at $1.85 per GPU-hour, so the two overlap only on inference.
GPUBank
Tallinn, Estonia
H200 SXM
$2.49 / h
- vCPU, RAM, NVMe, IPv4 and traffic included
- Per-second billing after the health check
- Top up in Bitcoin, USDT and 12+ other coins
vCPU, RAM, NVMe, IPv4 and traffic included
Replicate
San Francisco, United States
H200 SXM · serverless
$5.49 / h
- Public models bill only active processing time, setup and idle are free
- One API for 50,000+ models, including proprietary ones sold per output
- Cog packaging, autoscaling deployments and rolling updates
Replicate
Serverless
United States23 points compared21 of their pages readprices last checked on 29 September 2026page updated 2 November 2026
The verdict
Scale-to-zero model APIs at Replicate. Raw GPU instances at a lower hourly rate at GPUBank.
Replicate suits bursty inference: public models bill only active seconds, and the catalogue cites 50,000+ models, some sold per output. For steady work the hourly rate matters more, and GPUBank costs 55 % less on the H200 and 84 % less on the L40S. A month of L40S time is $2,527.20 on Replicate against $396 on GPUBank. Replicate gives no SSH or disk access, puts H200 and most multi-GPU sizes behind committed spend contracts, and publishes no regions.
ReplicateGPUBank 01The real cost
The same GPUs, side by side, per hour.
Every GPU model both of us rent, at each provider's own published price for one card, in US dollars, excluding VAT. Prices last checked on 29 September 2026.
| GPU | Replicate per hour | GPUBank per hour | Difference | 100 h at Replicate | 100 h at GPUBank |
|---|---|---|---|---|---|
 H200 SXM141 GB · their serverless price |
$5.49 | $2.49 | −55 % | $549.00 | $249.00 |
 L40S48 GB · their serverless price |
$3.51 | $0.55 | −84 % | $351.00 | $55.00 |
Replicate bills model runtime per second on managed hardware, not GPU instances. Figures are its per-second price times 3600. H200 is available only with committed spend contracts.
Where Replicate publishes no on-demand price for a model, its lowest published serverless price is used, as labelled on the row.
An inference box, a month, 720 h, on an L40S
Replicate
$2,527.20
GPUBank
$396.00
Saves $2,131.20 at GPUBank (−84 %)
A long-context job, 40 h on an H200
Replicate
$219.60
GPUBank
$99.60
Saves $120.00 at GPUBank (−55 %)
02Their catalogue
Every GPU Replicate lists.
5 GPU models listed by Replicate, each at its lowest public price per GPU-hour. 2 of them are also rented at GPUBank (in blue).
- H200serverless$5.49
- H10080 GB · serverless$5.49
- A10080 GB · serverless$5.04
- L40S48 GB · serverless$3.51
T416 GB · serverless$0.81
03What the price includes
What appears on the bill besides the GPU.
The same H200 SXM session of 2 h 14 m 03 s, line by line, as each provider describes its billing.
Replicate
2 h 14 m 03 s on H200 SXM
- GPU · $5.49 / h
- $12.27
- vCPU and RAM
- Included
- Storage
- Not published
- Egress
- Not published
- Public IPv4
- No IP, API endpoints only
- Billing granularity
- Per second
- Minimum charge
- Not published
GPUBank
2 h 14 m 03 s on H200 SXM
- GPU · $2.49 / h
- $5.56
- vCPU and RAM
- Included
- Storage
- Local NVMe included; persistent disk $0.03 per GB-month
- Egress
- Included
- Public IPv4
- Included
- Billing granularity
- Per second, from the health check
- Minimum charge
- None
While stopped
What keeps running on the bill.
Replicate
Scales to zero; private models bill idle time.
Public models bill active time only. A deployment with min_instances above 0 is billed continuously.
GPUBank
Only the persistent disk, $0.03 per GB-month. Compute, IPv4 and scratch cost nothing.
04Where the GPUs run
1 country against 4, and where they are.
Countries where Replicate says its GPUs run, from its own region list or documentation, against GPUBank's four regions. A pin marks a country, not the exact site.
United States
Paris
Frankfurt
Ashburn
Singapore
Replicate1
- United States
GPUBank4
Paris
Frankfurt- Ashburn
Singapore
Replicate · 1 countryGPUBank · 4 regions
05Paying and signing up
How each one takes your money.
Replicate
Payments and sign-up
Payment methods
Cards
Bank transfer
No cryptoMinimum top-up Not published
To start
- 1GitHub account
- 2Age 18 or over
- 3Payment method
Billing modelPrepaid credit or monthly in arrears
GPUBank
Payments and sign-up
Payment methods
A prepaid USD balance, topped up in crypto. Every GPU-second is drawn from it.
Minimum top-up $25
To start
- 1Email and password
- 2Top up in crypto
- 3Launch a GPU
Billing modelPrepaid balance
06Trust and support
Certificates, uptime and who answers.
| Replicate | GPUBank | |
|---|---|---|
| SOC 2 | Not published | Not published |
| ISO 27001 | Not published | Not published |
| HIPAA | Not published | No |
| GDPR DPA | DPA available | Published DPA, applies automatically |
| Availability target | Enterprise only, level not published | 99.9 % monthly, per region |
| Status page | replicatestatus.com | Public status page |
| Support | Email and Discord; priority for enterprise | Tickets from a $25 balance |
07Their site, our site
What you see before you pay.
Both sites as a visitor sees them, without an account: the Replicate pricing page captured on 29 September 2026, and GPUBank's live board. Cookie banners and chat bubbles are closed; nothing else is edited.
- 1The pricing page opens with per-output prices for example models before the hardware table.
- 2The hardware table lists each SKU with a price per second and per hour, GPU count, CPU count and RAM.
- 3An Additional hardware block lists H200 and multi-GPU sizes, each marked available with committed spend contracts.
- 1Live stock per region and every price, without an account.
- 2Per-hour and per-second prices, with vCPU, RAM, NVMe, IPv4 and traffic in them.
- 3A machine-readable price feed at /prices.json.
08The spec sheet
23 points, read off both sites.
GPUBank aheadReplicate aheadEven
| Criterion | GPUBank | Replicate |
|---|---|---|
| Price and catalogue | ||
| H100 SXM, per GPU-hourBilled as $0.001525 per second of hardware time. | $1.85, all-in | $5.49, H100 80GB, form factor not stated |
| Cheapest GPU-hour | $0.23, RTX 4090 | $0.81, T4 |
| GPU models listed | 10 | 5 |
| What the GPU price includesNot comparable: Replicate sells model runtime, with no disk, IP or instance. | vCPU, RAM, NVMe, IPv4, traffic | CPU and RAM per hardware SKU |
| Billing | ||
| Billing granularity | Per second | Per second |
| Minimum charge | None | Not published |
| Stopped instancePublic models bill active time only. A deployment with min_instances above 0 is billed continuously. | Disk only | Scales to zero; private models bill idle timeAhead |
| Persistent storageNo disks. API prediction data is deleted after one hour. | $0.03 per GB-month | Not published |
| Egress traffic | Included | Not published |
| Public IPv4 | Included | No IP, API endpoints only |
| Billing modelPrepaid credit is non-refundable and valid for 12 months. | Prepaid balance | Prepaid credit or monthly in arrearsAhead |
| Payment and sign-up | ||
| Crypto payments | Bitcoin, USDT, Ethereum and 12+ more | Not published |
| Card payments | No | Credit and debit cards, bank transferAhead |
| Minimum depositAuto reload has a $5 minimum threshold and a $15 minimum reload balance. | $25 | Not published |
| Platform | ||
| API and CLI | REST API, CLI, Python SDK | HTTP API, CLI, 4 official clients, MCP serverAhead |
| Terraform | Official provider | No provider found |
| Serverless GPUs | No | Yes, scale-to-zero model APIAhead |
| Managed Kubernetes | No | No |
| Ready-made templates | 25 | Public model catalogue, 50,000+ modelsAhead |
| Interruptible or spot | Yes, 40 % off | No |
| GPUs per instance | Up to 8 | Up to 8; most multi-GPU sizes need a contract |
| Multi-node interconnect | Not offered | Not published |
| Network and regions | ||
| Regions | 4 | None published, no region choice |
| Countries | France, Germany, US, Singapore | United States (hosting subprocessors) |
| EU regions | Paris, Frankfurt | None |
| Trust and support | ||
| Availability target | 99.9 % monthly | Enterprise only, level not published |
| Compliance reports | None published | No SOC 2 or ISO 27001 found; DPA for enterprise |
| Public status page | Yes | replicatestatus.com, on Cloudflare Status |
| Support | Tickets from a $25 balance | Email and Discord; priority for enterprise |
09The right fit
Different needs. A clear choice.
What Replicate does better
- Public models bill only active processing time, setup and idle are free
- One API for 50,000+ models, including proprietary ones sold per output
- Cog packaging, autoscaling deployments and rolling updates
- Python, Node.js, Go and Swift clients, a CLI and an MCP server
- Cards, debit cards or bank transfer, and billing in arrears
Choose GPUBank if
- H100 SXM at $1.85 per GPU-hour, against $5.49 for Replicate's H100
- SSH, JupyterLab, custom Docker images and persistent disks
- H200 and 8-GPU instances without a committed spend contract
- Interruptible at 40 % off and reserved at 20 % off
- Four named regions, including Paris and Frankfurt, and crypto top-ups
Moving from Replicate takes an afternoon.
Replicate models are standard Docker containers built with Cog, so the same image runs on a GPUBank instance. Moving means replacing Replicate API calls and its autoscaling with your own inference server.
- 1Build your Cog model image and push it to a registry you control.
- 2Launch a GPUBank instance with the matching GPU and run the image with its HTTP server.
- 3Point your application at the new endpoint and keep outputs on a persistent disk.
# 1. a GPU and a disk on GPUBank gpubank launch --gpu h200-sxm --region eu-paris --disk data # 2. copy your data from the old machine rsync -azP old-host:/workspace/ root@<new-ip>:/workspace/ # 3. run the same image or template ssh root@<new-ip> 'cd /workspace && ./start.sh'
A few things you might be wondering.
Per hour, yes. Replicate lists its H100 at $5.49 per GPU-hour, billed as $0.001525 per second, and does not state the form factor. GPUBank charges $1.85 per GPU-hour for an H100 SXM. Replicate can still cost less for rare requests, since public models bill only active time.
No. Replicate runs models packaged with Cog behind an API, with no VM, SSH, Jupyter or persistent disk. GPUBank rents GPU instances with SSH, JupyterLab, custom Docker images and persistent disks.
It depends on the model. Public models bill only active processing time. Private models and deployments also bill setup and idle time, and a deployment with min_instances above 0 is billed continuously.
Replicate's terms list credit card, debit card or bank transfer through Stripe, and no crypto option is published. GPUBank is paid from a USD balance topped up in Bitcoin, USDT, Ethereum and more than a dozen other coins.
Replicate publishes no regions and offers no region choice. Its hosting subprocessors (AWS, CoreWeave, GCP, Fly.io) are listed in the United States. GPUBank runs in Paris, Frankfurt, Ashburn and Singapore.
Without. Every price on this page excludes VAT and sales tax, for Replicate and for GPUBank, so the two columns compare like for like. Tax is added at checkout according to where the buyer is based.
Sources
Every figure about Replicate was read on its own pages or official documentation. Prices last checked on 29 September 2026.
- 01replicate.com/pricingHardware per-second prices, SKUs, specs, per-output model prices, committed-spend gating
- 02replicate.com/Positioning, scale-to-zero, pricing excerpt
- 03replicate.com/docs/topics/billingBilling models, instance lifecycle, what is billed for public/private models and deployments, failed runs, free limits
- 04replicate.com/docs/topics/billing/prepaid-creditPrepaid credit, 1-year validity, non-refundable, auto reload minimums
- 05replicate.com/termsLegal entity and address, payment methods, Stripe, arrears vs prepaid, 18+ eligibility, export controls
- 06replicate.com/privacyGitHub sign-in data, legal entity
- 07replicate.com/signinSign-in method (GitHub only)
- 08replicate.com/enterpriseSLA, DPA, priority support, volume discounts, reserved compute
- 09replicate.com/blog/replicate-cloudflareOwnership
- 10replicate.com/aboutCareers on Cloudflare
- 11replicate.com/docs/topics/deploymentsAutoscaling, hardware choice, monitoring
- 12replicate.com/docs/topics/deployments/create-a-deploymentDeployment API, min/max instances
- 13replicate.com/docs/topics/models/hardwareHardware switching
- 14replicate.com/docs/reference/client-librariesSDKs, MCP
- 15replicate.com/docs/reference/open-sourceCog, official CLI
- 16replicate.com/docs/reference/httpAPI endpoints incl. hardware.list
- 17replicate.com/docs/reference/how-does-replicate-workCold boots
- 18replicate.com/docs/topics/predictions/data-retentionOne-hour deletion of API prediction data
What we could not verify (9)
- H100 form factor (SXM, PCIe or NVL) not published; recorded as model 'other' (80GB). H100 SXM price therefore null.
- A100 80GB form factor (SXM or PCIe) not published; recorded as model 'other'.
- H200 GPU memory and CPU/RAM for the 'Additional hardware' SKUs not listed.
- Data center regions not published; countries set to United States only because every hosting subprocessor listed (AWS, CoreWeave, GCP, Fly.io) is located in the United States.
- Crypto payment, PayPal, minimum first credit purchase and free credit amount not published.
- No SOC 2, ISO 27001 or HIPAA statement found; trust.replicate.com does not resolve and replicate.com/security returns 404.
- Enterprise SLA levels, volume discounts and committed-spend contract terms not published.
- Egress or bandwidth charges not published.
- Founding year not published on the pages read.
In short
GPUBank vs Replicate, in facts.
The essentials of this comparison in one table, for quick readers and AI assistants. Prices last checked on 29 September 2026.
Read as Markdown- H100 SXM, per GPU-hour
- Replicate: $5.49, H100 80GB, form factor not stated
GPUBank: $1.85, all-in - Cheapest GPU-hour
- Replicate: $0.81, T4
GPUBank: $0.23, RTX 4090 - Billing granularity
- Replicate: Per second
GPUBank: Per second - Stopped instance
- Replicate: Scales to zero; private models bill idle time
GPUBank: Disk only - Persistent storage
- Replicate: Not published
GPUBank: $0.03 per GB-month - Egress traffic
- Replicate: Not published
GPUBank: Included - Crypto payments
- Replicate: Not published
GPUBank: Bitcoin, USDT, Ethereum and 12+ more - Card payments
- Replicate: Credit and debit cards, bank transfer
GPUBank: No - Minimum deposit
- Replicate: Not published
GPUBank: $25 - GPU models listed
- Replicate: 5
GPUBank: 10 - Regions
- Replicate: None published, no region choice
GPUBank: 4 - Serverless GPUs
- Replicate: Yes, scale-to-zero model API
GPUBank: No - Availability target
- Replicate: Enterprise only, level not published
GPUBank: 99.9 % monthly - Compliance reports
- Replicate: No SOC 2 or ISO 27001 found; DPA for enterprise
GPUBank: None published - Support
- Replicate: Email and Discord; priority for enterprise
GPUBank: Tickets from a $25 balance - Headquarters
- Replicate: San Francisco, United States
GPUBank: Tallinn, Estonia
How this page was made
Every price and policy of Replicate was read on its own website, documentation or public API on 29 September 2026, with the address of each page listed above. A value we could not read on a primary source is marked "Not published" rather than estimated. GPUBank figures come from this site and its terms. Prices change: check both providers before you buy. Replicate, its logo and its website belong to their owner and are shown only to identify the service compared. GPUBank is not affiliated with Replicate.
