One line of Python.
Every cloud's best price.
Verlex reads any Python function, picks the hardware it needs, prices it live across 10 clouds, and runs it on the cheapest one. If a provider fails or stocks out, the job restarts on the next cheapest, automatically. No hardware picking. No quotas. No DevOps.
import verlex def train(data): # your existing code, unchanged return results # zero config: Verlex reads train(), picks the hardware it needs, # and runs it on whichever of 10 clouds is cheapest right now results = verlex.cloud(train, data) # or fully automatic: offload only when this machine runs hot verlex.overflow()
Ten clouds, one endpoint. Your job lands on whichever is cheapest at submit time.
Cheap GPUs are everywhere.
Using them is the hard part.
To run one GPU job today, you pick a provider, compare prices, file quota requests, write deploy scripts, install drivers, and babysit the machine until it finishes. Then you do it all again next week, when prices move or your provider runs out of capacity.
Verlex collapses all of that into one Python call. You write the code you were going to write anyway. We handle the cloud, end to end.
No more comparing GPU prices across five dashboards, no more deploy YAML, no more 2am restarts when a box dies. You ship code; the infrastructure just happens.
A marketplace hands you a catalog and a login for every provider.
Verlex hands you the finished job.
Everything you'd have to build.
Stacked into one call.
No trust-us claims here. Each piece below ships with the mechanism that makes it work.
Every job runs an auction. The cheapest cloud wins.
Verlex prices your job across every configured provider, VMs and serverless containers alike, and launches on the cheapest hardware that fits your request. Hardware catalogs refresh continuously, and capacity marked unavailable heals itself, so the comparison is live, not a cached brochure.
We do not own the GPUs. So we have no reason to keep you on an expensive one.
A stockout is our problem, not yours.
When a launch fails, your job does not. Verlex walks a ladder of alternatives, in price order, until your code is running. Failed regions are remembered for five minutes so retries never thrash.
The whole chain runs under one provisioning deadline: your job runs or fails loudly. It never queues forever.
Cheapest provider, primary region
The job launches wherever the live price is lowest.
Nearby regions, closest first
Stocked out? Verlex retries the same cloud next door.
Next cheapest cloud
Still nothing? The job moves down the price-ordered list of providers.
Substitute GPU
Last resort: a comparable GPU across all clouds, unless you pinned one.
Spot prices, without the babysitting.
in this example, service fee included, billed per second. Spot discounts run 60 to 80% depending on provider.
Your machine first. The cloud only when it counts.
Two lines make any machine hybrid. verlex.overflow() watches CPU, RAM and GPU, and the moment your box runs hot, the heavy functions overflow to the cheapest cloud. Everything else keeps running locally, for free.
Only the work that needs the cloud ever leaves your machine. No annotations, no cluster setup, no decorators. To our knowledge, nobody else does this automatically.
The parts you'd never get around to building.
You never pay for idle
Jobs meter in one-second increments from prepaid credits. When your code finishes, the instance is destroyed and the meter stops. There is no machine burning money overnight because someone forgot it.
Small jobs skip the VM entirely
Short jobs route to container lanes on Cloud Run, Fargate, ACI, RunPod and more, with cold starts in tens of seconds instead of full VM boots. Serverless for bursty small jobs, VMs for long runs, clusters for multi-GPU: one auction prices all three lanes.
Up to 8 GPUs in one line
On the Performance plan, ask for eight GPUs and get them on a single machine. No cluster YAML, no placement groups, no capacity reservations. Same one-line call.
One credit pool for the whole team
Corporate accounts share an organization credit pool. Admins allocate credits to members, claw them back when priorities change, and invite teammates by email. Every job draws from one place.
Route. Run. Recover.
One endpoint. The provider never becomes a decision you have to make.
Route
You call Verlex. It prices every configured cloud, VM and serverless, and picks the cheapest hardware that fits. On the Performance plan, it picks the fastest option inside your budget instead.
Run
Your code runs unchanged and meters per second from prepaid credits. A job runs until it finishes or your credits do, with a hard seven-day backstop.
Recover
Your checkpoint directory syncs to object storage while the job runs. If a provider dies or reclaims a spot machine, Verlex restarts from the last checkpoint on the next cheapest capacity.
Stop paying list price to babysit consoles.
Single clouds hand you quotas, consoles, and on-demand list prices. Managed platforms are simpler, but you rent their hardware at their markup. Verlex passes through the live provider price at cost and adds a fixed service fee per GPU-hour. That is the whole model.
A monthly subscription, plus a service fee.
Two simple parts: our monthly subscription, and a service fee on each job on top of the live GPU price. The service fee varies with the hardware you run and your plan, and you see it before anything starts. No commission, no premium hardware tax, no invoice surprises.
We make money on the flat fee, never a percentage of your compute. So when a cheaper GPU appears anywhere in the market, routing you to it costs us nothing, and we always do.
- Cost-first routing across 10 clouds
- Automatic failover across providers and regions
- Spot routing with checkpoint recovery
- Serverless lane for small jobs
- 25 GB storage · pay in USD, BRL, INR, THB, or CAD
- Fast mode: the fastest qualified hardware within your budget
- Clusters: up to 8 GPUs on a single instance
- Warm capacity first, full VM boots only when needed
- 300 GB storage
- Cancel the subscription anytime, keep your credits
| Hardware tier | Examples | Standard | Performance |
|---|---|---|---|
| CPU-only | any vCPU job | $0.05/hr | $0.10/hr |
| Small | T4, L4, A10, RTX 3090 | $0.10 | $0.20 |
| Mid | A100, L40S, RTX 4090 | $0.30 | $0.45 |
| Large | H100, H200, MI300X | $0.40 | $0.60 |
| Flagship | B200, B300, GB200 | $0.50 | $0.75 |
That table is the entire fee schedule: per GPU, per hour, prorated to the second. Tiers follow the hardware's raw TFLOPS, so a new GPU prices itself the day it exists. No percentage markup, no cold-start surcharge, no invoice surprises. The exact price is shown before every job runs.
Every way this could bite you, defused.
Your balance is the ceiling
You load credits before anything runs, and a job can never spend money you have not loaded. When credits run out, the job stops. There is no surprise invoice at the end of the month.
The meter stops when the job does
Billing runs in one-second increments while your code runs. Afterward the instance is destroyed, so there is nothing left to charge you for.
Run exactly as long as you choose
Let a job run until its credits are spent, or set a spending limit. Auto top-up is strictly opt-in, and every job has a hard seven-day backstop no matter what.
Cancel in one click
Performance is a $10 monthly subscription through Stripe. Cancel whenever, keep your remaining credits, and your code still runs anywhere Python runs.
The things you'd ask before trusting us with a job.
How does Verlex make money?
The provider's GPU price passes through at cost. Verlex adds a fixed service fee of $0.05 to $0.75 per GPU-hour depending on the hardware tier and plan. No percentage markup, no hidden margin.
How much cheaper is Verlex than AWS or GCP list prices?
Every job is priced across 10 clouds and routed to the cheapest qualified GPU. Spot discounts at the provider typically run 60-80% off on-demand list. Example: an H100 at $3.50/hr list can run at about $1.05/hr spot plus a $0.40 service fee - about $1.45/hr, roughly 59% off all-in.
What happens if a spot GPU is reclaimed mid-job?
Checkpoints sync to object storage while the job runs. Verlex restarts the job from the last checkpoint on the next cheapest capacity, and the reclaimed time is not charged.
Can I get a surprise bill?
No. Credits are prepaid and jobs meter per second. When credits run out the job stops. Auto top-up is strictly opt-in, and every job has a hard 7-day backstop.
Do I need to change my code to run it on a cloud GPU?
No. results = verlex.cloud(train, gpu="A100") runs your existing Python function on a cloud GPU - two lines including the import.
Which clouds does Verlex route across?
AWS, Google Cloud, Azure, Verda, RunPod, Vast.ai, Hyperstack, JarvisLabs, TensorDock and Lyceum, plus serverless container lanes for small jobs on Beam, Northflank, Novita and Cerebrium.
Stop babysitting clouds.
Ship the function.
One line of Python, prepaid credits, 10 clouds competing for every job. We onboard in small waves to keep routing quality high, so grab your spot.