GPUBank vs Baseten
Baseten is a San Francisco inference platform, founded in 2019, that runs models as autoscaled dedicated replicas billed per minute, plus training jobs and per-token Model APIs. It is not a raw VM rental. Its H100 costs $0.11 per minute, or $6.50 per GPU-hour; a GPUBank H100 SXM costs $1.85. Baseten sells autoscaling, scale-to-zero, a 99.9 % SLA and compliance reports; GPUBank sells plain GPU instances with SSH, billed by the second.
GPUBank
Tallinn, Estonia
H100 SXM
$1.85 / 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
Baseten
San Francisco, United States · since 2019
H100 SXM
$6.50 / h
- Autoscaled model replicas that scale to zero when idle
- 99.9 % monthly SLA with service credits of 10 to 30 %
- SOC 2 Type II, ISO 27001, HIPAA and PCI DSS listed
Baseten
Serverless
United States26 points compared21 of their pages readprices last checked on 29 September 2026page updated 2 November 2026
The verdict
Managed inference and compliance at Baseten. Raw GPUs at a fraction of the price at GPUBank.
Baseten fits teams that serve models in production and want autoscaling, scale-to-zero, managed inference engines, a contractual SLA and SOC 2, ISO 27001 or HIPAA reports. You deploy a model package or container, not a machine. For the four GPUs both of us offer, GPUBank costs 58 to 72 % less per GPU-hour, and it bills by the second, not per started minute with cold starts and image builds on the bill. GPUBank gives a full instance with SSH, JupyterLab and persistent disks, paid in crypto; Baseten bills a card or US bank account after the fact.
BasetenGPUBank 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 | Baseten per hour | GPUBank per hour | Difference | 100 h at Baseten | 100 h at GPUBank |
|---|---|---|---|---|---|
 B200 SXM180 GB |
$9.98 | $4.15 | −58 % | $998.00 | $415.00 |
 H200 SXM141 GB |
$7.50 | $2.49 | −67 % | $750.00 | $249.00 |
 H100 SXM80 GB |
$6.50 | $1.85 | −72 % | $650.00 | $185.00 |
 RTX PRO 600096 GB |
$4.00 | $1.15 | −71 % | $400.00 | $115.00 |
Baseten prices are per minute for autoscaled model replicas, not raw VM rentals. Figures here are the per-minute rate x60, per GPU. Cold starts, image builds and idle warm replicas are billed.
A fine-tune, 40 h on an H100 SXM
Baseten
$260.00
GPUBank
$74.00
Saves $186.00 at GPUBank (−72 %)
A long-context job, 40 h on an H200
Baseten
$300.00
GPUBank
$99.60
Saves $200.40 at GPUBank (−67 %)
A full day, 24 h on a B200
Baseten
$239.52
GPUBank
$99.60
Saves $139.92 at GPUBank (−58 %)
02Their catalogue
Every GPU Baseten lists.
9 GPU models listed by Baseten, each at its lowest public price per GPU-hour. 4 of them are also rented at GPUBank (in blue).
- B200180 GB · on demand$9.98
- H200141 GB · on demand$7.50
- H100 SXM80 GB · on demand$6.50
- H100 MIG 40GB40 GB · on demand$3.75
A100 PCIe 80GB80 GB · on demand$4.00
L424 GB · on demand$0.85
A10G24 GB · on demand$1.21
T416 GB · on demand$0.63- RTX PRO 600096 GB · on demand$4.00
At GPUBank
vCPU, RAM, NVMe and traffic included. All prices
03What the price includes
What appears on the bill besides the GPU.
The same H100 SXM session of 2 h 14 m 03 s, line by line, as each provider describes its billing.
Baseten
2 h 14 m 03 s on H100 SXM
- GPU · $6.50 / h
- $14.52
- vCPU and RAM
- Included
- Storage
- Not published
- Egress
- Not published
- Public IPv4
- Not published
- Billing granularity
- Per minute, partial minutes rounded up
- Minimum charge
- Cold starts and image builds billed
GPUBank
2 h 14 m 03 s on H100 SXM
- GPU · $1.85 / h
- $4.13
- 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.
Baseten
Scaled to zero: no charge.
Warm replicas are billed while idle, and the default scale-down delay keeps them billed for 15 minutes.
GPUBank
Only the persistent disk, $0.03 per GB-month. Compute, IPv4 and scratch cost nothing.
04Where the GPUs run
2 countries against 4, and where they are.
Countries where Baseten 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
Baseten2
- United States
- European Union (country not specified)
GPUBank4
Paris
Frankfurt- Ashburn
Singapore
Baseten · 2 countriesGPUBank · 4 regions
05Paying and signing up
How each one takes your money.
Baseten
Payments and sign-up
Payment methods
Cards
Bank transfer
No cryptoMinimum top-up None, postpaid
To start
- 1Email (or Google / GitHub sign-in)
- 2Payment method
- 3Organization verification
Billing modelPostpaid invoice
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.
| Baseten | GPUBank | |
|---|---|---|
| SOC 2 | Type II | Not published |
| ISO 27001 | Listed, report not public | Not published |
| HIPAA | Listed, report not public | No |
| GDPR DPA | Listed, report not public | Published DPA, applies automatically |
| Availability target | 99.9 % monthly, credits up to 40 % | 99.9 % monthly, per region |
| Status page | status.baseten.co | Public status page |
| Support | Email and chat; Slack and Zoom on Pro | 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 Baseten pricing page captured on 29 September 2026, and GPUBank's live board. Cookie banners and chat bubbles are closed; nothing else is edited.
- 1GPU prices are shown per minute by default, with a per-hour toggle, and need no account.
- 2The pricing page lists 7 single-GPU entries; H200, RTX PRO 6000 and multi-GPU sizes appear only in the docs.
- 3The pricing FAQ says you do not pay for idle time, while the billing docs say idle warm replicas are billed.
- 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
26 points, read off both sites.
GPUBank aheadBaseten aheadEven
| Criterion | GPUBank | Baseten |
|---|---|---|
| Price and catalogue | ||
| H100 SXM, per GPU-hourAutoscaled replica price. The form factor is not named; the published specs match H100 SXM. | $1.85, all-in | $6.50 ($0.11 per minute) |
| Cheapest GPU-hour | $0.23, RTX 4090 | $0.63, T4 |
| GPU models listed | 10 | 9, including a 40 GB H100 MIG slice |
| What the GPU price includesStorage, egress and IP prices are not published. | vCPU, RAM, NVMe, IPv4, traffic | vCPU and RAM |
| Billing | ||
| Billing granularity | Per second | Per minute, partial minutes rounded up |
| Minimum chargeFailed boots and image pulls are not billed. | None | Cold starts and image builds billed |
| Stopped instanceWarm replicas are billed while idle, and the default scale-down delay keeps them billed for 15 minutes. | Disk only | Scaled to zero: no chargeAhead |
| Persistent storage | $0.03 per GB-month | Not published |
| Egress traffic | Included | Not published |
| Public IPv4 | Included | Not published |
| Billing model | Prepaid balance | Postpaid invoice |
| Payment and sign-up | ||
| Crypto payments | Bitcoin, USDT, Ethereum and 12+ more | Not listed; card or US bank account |
| Card payments | No | Cards and US bank accountsAhead |
| Minimum deposit | $25 | None, postpaidAhead |
| Platform | ||
| API and CLI | REST API, CLI, Python SDK | APIs, Baseten and Truss CLIs, Python SDKs |
| Terraform | Official provider | No provider found |
| Serverless GPUs | No | Autoscaled replicas, scale to zeroAhead |
| Managed Kubernetes | No | No |
| Ready-made templates | 25 | Model library, count not published |
| Interruptible or spot | Yes, 40 % off | None published |
| GPUs per instance | Up to 8 | Up to 8 per replica, multi-nodeAhead |
| Multi-node interconnect | Not offered | InfiniBand between nodesAhead |
| Network and regions | ||
| RegionsDefault deployments run on a global multi-cloud pool. No data-center cities are published. | 4 | 2 region slugs, us and eu |
| Countries | France, Germany, US, Singapore | US and an unnamed EU country |
| EU regions | Paris, Frankfurt | eu slug, verified organizations only |
| Trust and support | ||
| Availability target | 99.9 % monthly | 99.9 % monthly, credits up to 40 %Ahead |
| Compliance reports | None published | SOC 2 Type II, ISO 27001, HIPAAAhead |
| Public status page | Yes | status.baseten.co |
| Support | Tickets from a $25 balance | Email and chat; Slack and Zoom on ProAhead |
09The right fit
Different needs. A clear choice.
What Baseten does better
- Autoscaled model replicas that scale to zero when idle
- 99.9 % monthly SLA with service credits of 10 to 30 %
- SOC 2 Type II, ISO 27001, HIPAA and PCI DSS listed
- Per-token Model APIs and multi-node InfiniBand deployments
- Postpaid invoices, no deposit needed
Choose GPUBank if
- H100 SXM at $1.85 per GPU-hour, against $6.50
- Per-second billing that starts after the health check
- A full GPU instance with SSH, JupyterLab and persistent disks
- Four named regions to pick from at launch
- Pay in Bitcoin, USDT and 12+ other coins, no card
Moving from Baseten takes an afternoon.
Baseten models are open-source Truss packages or custom Docker images that pull weights from Hugging Face, S3 or GCS, so the same container runs elsewhere. There are no inference volumes to export.
- 1Launch the GPU you need on GPUBank with a persistent disk, and pull your model weights onto it.
- 2Run the same Docker image, or build one from your Truss package, and expose its port.
- 3Point your clients at the new endpoint, then scale the Baseten deployment to zero.
# 1. a GPU and a disk on GPUBank gpubank launch --gpu h100-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.
Yes. Baseten lists its H100 at $0.11 per minute, which is $6.50 per GPU-hour. GPUBank charges $1.85 per GPU-hour for an H100 SXM, with vCPU, RAM, NVMe, IPv4 and traffic included. The two products differ: Baseten runs autoscaled model replicas, GPUBank rents GPU instances.
An inference platform. You deploy a model or container as dedicated replicas that autoscale and can scale to zero, or call per-token Model APIs. There is no general-purpose VM or Jupyter product, and SSH into inference containers must be enabled by support.
Yes, unless the deployment scales to zero. Warm replicas with min_replica 1 or more are billed, and the default scale-down delay of 900 seconds keeps a replica billed for 15 minutes after traffic stops. Cold starts and image builds are billed too.
Baseten lists card or US bank account through its payment processor, billed after use. No crypto or PayPal option is published. GPUBank is paid from a USD balance topped up in Bitcoin, USDT, Ethereum and more than a dozen other coins.
Baseten. It publishes a 99.9 % monthly SLA with service credits of 10, 15 or 30 % of the bill, and lists SOC 2 Type II, ISO 27001, HIPAA and PCI DSS on its Trust Center. GPUBank targets 99.9 % monthly per region in its terms and publishes no SOC 2 or ISO 27001 report.
Without. Every price on this page excludes VAT and sales tax, for Baseten 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 Baseten was read on its own pages or official documentation. Prices last checked on 29 September 2026.
- 01baseten.co/pricing/Plans, per-minute GPU prices on the page, Model API prices, FAQ (credits, idle time, support, discounts)
- 02docs.baseten.co/deployment/resourcesFull GPU instance table with per-minute prices, vCPU, RAM, VRAM, GPU specs
- 03docs.baseten.co/organization/billingPer-minute metering, lifecycle billing, payment methods, credits, invoice cadence, x60 conversion
- 04docs.baseten.co/deployment/autoscaling/overviewScale-to-zero, default scale-down delay 900 s
- 05docs.baseten.co/deployment/regional-deploymentsRegions us and eu, verified-organization requirement
- 06docs.baseten.co/hosting-options/cloudMulti-cloud pool, GPUs never shared, self-service
- 07docs.baseten.co/hosting-options/single-tenantEnterprise custom pricing, private networking
- 08docs.baseten.co/observability/securityData retention, isolation, pentesting
- 09docs.baseten.co/reference/overviewAPIs, CLIs, SDKs
- 10docs.baseten.co/inference/sshSSH into deployments, enabled on request
- 11docs.baseten.co/training/concepts/cacheTraining storage behaviour
- 12docs.baseten.co/inference/model-apis/pricing-and-limitsAccount tiers Basic unverified/verified, rate limits
- 13baseten.co/service-level-agreement/99.9% SLA and service credits
- 14status.baseten.co/90-day uptimes
- 15trust.baseten.coSOC 2, ISO 27001, HIPAA, PCI DSS, FedRAMP 20x, subprocessors
- 16baseten.co/terms-and-conditions/Legal name, San Francisco address, invoice terms
- 17baseten.co/about-us/Founded 2019, founders
- 18baseten.co/deployments/baseten-cloud/Deployment options, compliance claims
What we could not verify (9)
- Funding and investors: not stated on the pages read
- Amount of starting credits for new workspaces: not published
- Storage, egress and IP pricing: not published
- Reserved or committed capacity and volume-discount prices: sales only
- Spot or interruptible tier: none published
- Terraform provider: none found in the docs index
- What organization verification involves (documents, card, company check): not documented
- Exact GPU form factors (H100 SXM vs PCIe, A100 SXM vs PCIe) not named; inferred from published memory bandwidth
- Physical data-center locations: only region slugs us and eu, plus subprocessor locations on the Trust Center (AWS, Google Cloud global; Crusoe US, Ireland; Oracle US, UK, UAE)
In short
GPUBank vs Baseten, 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
- Baseten: $6.50 ($0.11 per minute)
GPUBank: $1.85, all-in - Cheapest GPU-hour
- Baseten: $0.63, T4
GPUBank: $0.23, RTX 4090 - Billing granularity
- Baseten: Per minute, partial minutes rounded up
GPUBank: Per second - Stopped instance
- Baseten: Scaled to zero: no charge
GPUBank: Disk only - Persistent storage
- Baseten: Not published
GPUBank: $0.03 per GB-month - Egress traffic
- Baseten: Not published
GPUBank: Included - Crypto payments
- Baseten: Not listed; card or US bank account
GPUBank: Bitcoin, USDT, Ethereum and 12+ more - Card payments
- Baseten: Cards and US bank accounts
GPUBank: No - Minimum deposit
- Baseten: None, postpaid
GPUBank: $25 - GPU models listed
- Baseten: 9, including a 40 GB H100 MIG slice
GPUBank: 10 - Regions
- Baseten: 2 region slugs, us and eu
GPUBank: 4 - Serverless GPUs
- Baseten: Autoscaled replicas, scale to zero
GPUBank: No - Availability target
- Baseten: 99.9 % monthly, credits up to 40 %
GPUBank: 99.9 % monthly - Compliance reports
- Baseten: SOC 2 Type II, ISO 27001, HIPAA
GPUBank: None published - Support
- Baseten: Email and chat; Slack and Zoom on Pro
GPUBank: Tickets from a $25 balance - Headquarters
- Baseten: San Francisco, United States
GPUBank: Tallinn, Estonia
How this page was made
Every price and policy of Baseten 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. Baseten, its logo and its website belong to their owner and are shown only to identify the service compared. GPUBank is not affiliated with Baseten.
