60-Second Provisioning
From `tscale instances create` to SSH-ready, your GPU is online in under a minute. Cold-start a cluster of 64 H100s in the time it takes to make coffee.
/ INSTANCES
On-demand GPU compute at bare-metal speed
Tscale Instances deliver dedicated, single-tenant GPU servers — H100, H200, A100, L40S, MI300X — ready in under 60 seconds, billed by the second, no per-API markup. The cloud GPU, finally done right.
Per-Second Billing
No 60-minute minimum, no rounding up, no per-API markup. Stop a job at 14:23:07, pay for 14:23:07. The cloud GPU billing experience you’ve been waiting for.
Single-Tenant Hardware
Every instance is a dedicated physical GPU — no noisy neighbours, no shared MIG, no oversubscription. Whatever you run, you get the full FLOPS.
/ INSTANCE CATALOG
The latest GPUs, ready when you are
From the budget-friendly L40S to the new H200 and NVIDIA Blackwell, Tscale’s catalog covers every workload profile. Mix and match across regions — every instance is identical, predictable, and dedicated.
H200 SXM5
NewThe newest Hopper generation — 141GB HBM3e, 4.8TB/s bandwidth, optimised for the largest LLM training runs.
GPU1x H200 SXM5
vCPU26 cores
Memory200 GB
Storage2 TB NVMe
Network200 Gbps
GPU Mem141 GB HBM3e
$4.12 / hour
Launch →
H100 SXM5
PopularThe workhorse of modern AI — 80GB HBM3, NVLink, transformer engine, FP8 precision. The default for serious workloads.
GPU1x H100 SXM5
vCPU26 cores
Memory200 GB
Storage2 TB NVMe
Network200 Gbps
GPU Mem80 GB HBM3
$3.42 / hour
Launch →
A100 80GB
ValueThe proven default — 80GB HBM2e, NVLink, ideal for training and inference at the best dollar per FLOP.
GPU1x A100 80GB
vCPU24 cores
Memory192 GB
Storage1.5 TB NVMe
Network100 Gbps
GPU Mem80 GB HBM2e
$1.92 / hour
Launch →
L40S Ada
InferenceThe inference sweet spot — 48GB GDDR6, AV1 encoder, and tensor cores. Perfect for vLLM at 100+ tok/s.
GPU1x L40S 48GB
vCPU16 cores
Memory128 GB
Storage1 TB NVMe
Network50 Gbps
GPU Mem48 GB GDDR6
$1.18 / hour
Launch →
AMD MI300X
NewAMD’s flagship — 192GB HBM3, ROCm 6.2, and exceptional FP6/FP8 throughput. Run massive models without model sharding.
GPU1x MI300X 192GB
vCPU32 cores
Memory256 GB
Storage2 TB NVMe
Network200 Gbps
GPU Mem192 GB HBM3
$3.88 / hour
Launch →
B200 Blackwell
NewThe next generation — 192GB HBM3e, 8TB/s, 5th-gen tensor cores, FP4/FP6 precision for trillion-parameter training.
GPU1x B200 192GB
vCPU32 cores
Memory256 GB
Storage2 TB NVMe
Network400 Gbps
GPU Mem192 GB HBM3e
$5.20 / hour
Launch →
/ TRANSPARENT PRICING
Pay for what you use, nothing else
Tscale’s pricing model is built for AI workloads that ramp up and down. Per-second billing, no API markup, no data egress fees, no surprise tier-2 storage charges — just the GPU.
- Per-second billing — start at 14:23:07, stop at 14:58:41, pay for exactly 35 minutes 34 seconds.
- No data egress — move a trained model off Tscale without paying per-GB transfer fees.
- Volume discounts — committed-use discounts from 20% to 60% for predictable baseline capacity.
- Reserved instances — 1-year and 3-year terms for stable workloads, with full price transparency.
/ PLATFORM
Everything you need around the GPU
Instances aren’t just bare servers — they’re a complete cloud platform. Storage, networking, snapshots, and integrations come standard with every instance you launch.
Storage
- Local NVMe (up to 8 TB)
- High-performance block storage
- S3-compatible object storage
- Snapshots & clones
- Cross-region replication
Networking
- 200 / 400 Gbps fabric
- NVLink for multi-GPU
- InfiniBand for HPC
- Private VPCs
- Floating IPs & load balancers
Operating Systems
- Ubuntu 22.04 / 24.04 LTS
- Rocky Linux 8 / 9
- Debian 12
- Custom images
- Bring-your-own VM
Pre-installed
- NVIDIA drivers (latest)
- CUDA + cuDNN
- Docker + containerd
- NCCL for multi-GPU
- DCGM exporter
Regions
- Lagos, Nigeria
- West Africa (Q2)
- East Africa (Q3)
- Europe (Q4)
- Sovereign deployments
Access
- Web Console
- Tscale CLI
- REST API (Radar)
- Terraform provider
- SSH & JupyterHub
Performance
From request to SSH
Average cold-start time across all instance types. Hot pools for popular SKUs cut this to under 15 seconds.
Guaranteed network throughput
Every instance gets the full 200/400 Gbps line rate. No oversubscription, no shared uplinks.
vs. hyperscalers
Same NVIDIA H100s, same SXM5 boards, 80% lower bill. No data egress, no per-API markup.
Backed by enterprise SLA
Hot-standby failover for control plane, redundant power & networking, 24/7 SRE on call.
For teams that need more than a VM
Built for Engineers
Standard Linux. Root access. Bring your own kernel modules. No proprietary abstraction layer between you and the silicon.
Learn MoreSecure & Sovereign
Data residency in your jurisdiction. Encryption at rest and in transit. Optional dedicated tenancy for sensitive workloads.
Learn MoreCompose the rest of the stack
Instances are the building blocks. Combine them with managed Slurm, Kubernetes, and inference to run the full AI lifecycle on one platform — without leaving Tscale.