Thunder Compute (YC S24) is live · pricing from $0.35/GPU-hr · last verified 30 Oct 2026 · viability monitor
Thunder Compute (YC S24)
Rent GPU instances and per-second GPU sandboxes for AI training, RL rollouts, and agent workloads from $0.35/hr.
For bursty workloads Thunder Compute is close to a straight arbitrage: a $1.09 A100 80GB and per-second sandbox billing mean evals and rollouts stop paying for idle time. The ceilings are real though — 8 GPUs self-serve, no regional spread, sandboxes gated behind request access. Great for a small team shipping fast; a hard stop if you need multi-region residency or a signed SLA.
Verdict generatedVerified 10h ago · liveness 68/100 · cite: rightaichoice.com/tools/thunder-compute-yc-s24
- ML engineers with bursty training and inference who want published on-demand rates with no spot preemption
- RL and eval teams fanning out hundreds of short-lived GPU sandboxes per second
- Small startups where per-minute billing and zero egress make GPU spend a predictable line
- Developers who want a GPU reachable from terminal, browser console or VS Code
- Teams needing multi-region deployment for latency or data residency
- Enterprise buyers requiring custom SLAs or a formal security review at the on-demand tier
- Multi-node training beyond 8 GPUs in a self-serve deployment
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip Thunder Compute if your workload needs GPUs spread across multiple regions, bare-metal servers, or a multi-node cluster larger than the 8-GPU self-serve ceiling.
Instance disk past the first 100 GB per GPU runs $0.219/GB/mo, and the meter keeps running while your instance is up.
The rate card fits cash-conscious startups, research labs and individual ML engineers: $0.35/GPU-hr for an RTX A6000 up to $3.20/GPU-hr for H100 SXM, billed per minute only while running, with no egress fees. Against Lambda ($3.29 H100), Runpod ($3.99) and CoreWeave ($6.16), it sits at the cheap end of self-serve on-demand. Cheaper options exist only in the spot-market tier, which trades away the never-preempted guarantee. Teams needing 8+ dedicated GPUs move onto reserved capacity through
In short
Thunder Compute (YC S24) — Rent GPU instances and per-second GPU sandboxes for AI training, RL rollouts, and agent workloads from $0.35/hr. Best for ML engineers with bursty training and inference who want published on-demand rates with no spot preemption, RL and eval teams fanning out hundreds of short-lived GPU sandboxes per second, Small startups where per-minute billing and zero egress make GPU spend a predictable line. Plans from $0.35/GPU-hr.
What is Thunder Compute (YC S24)?
Thunder Compute (YC S24) — Rent GPU instances and per-second GPU sandboxes for AI training, RL rollouts, and agent workloads from $0.35/hr. Best for ML engineers with bursty training and inference who want published on-demand rates with no spot preemption, RL and eval teams fanning out hundreds of short-lived GPU sandboxes per second, Small startups where per-minute billing and zero egress make GPU spend a predictable line. Plans from $0.35/GPU-hr.
Thunder Compute sells GPU capacity in two shapes on one fleet. Instances are persistent Ubuntu machines you SSH into for training, CUDA work and inference. Sandboxes are isolated GPU-enabled microVMs spawned from a few lines of Python for RL environments, evals and CI jobs, scaling to zero when the work ends. The audience is ML engineers, RL researchers and small teams who want a GPU without a procurement cycle. The rate card does most of the persuading. RTX A6000 runs $0.35 per GPU-hr, L40 $0.79, A100 80GB $1.09, H100 SXM $3.20, configurable at 1x, 2x, 4x and 8x. Instances bill per minute and only while running. Sandboxes bill per second across GPU, vCPU, memory and storage — the H100 SXM sandbox lists $0.000819/GPU-sec. The first 100 GB of instance disk per GPU is included, snapshots cost $0.05/GB/mo, and egress is $0.00/GB. Tooling covers the ways people actually provision: a `tnr create` / `tnr connect` / `tnr snapshot` CLI, browser console, VS Code extension, REST API, Python SDK, instance templates, and an MCP server whose `get_availability` tool lets coding agents check live GPU stock before launching. Snapshots park a disk between runs instead of leaving a GPU idle. Nothing gets preempted. A June 2026 engineering post details the GPU-over-TCP virtualization layer underneath both products. Against Lambda ($3.29 for H100 on the vendor's Oct 30, 2026 comparison) and Runpod ($3.99), Thunder's $3.20 undercuts the field, and its sandbox second is priced below Daytona and Modal. Where it gives ground is footprint and scale: self-serve tops out at 8 GPUs, and sandboxes go through a request-access flow rather than a card.
Key features
- Per-minute billing on GPU instances, per-second billing on sandboxes
- RTX A6000 $0.35, L40 $0.79, A100 80GB $1.09, H100 SXM $3.20 per GPU-hr
- GPU sandboxes for RL rollouts, evals and CI, spawned via thunder.Sandbox.create in Python
- Sandboxes scale to zero and bill per second for GPU, vCPU, memory and storage
- 1x, 2x, 4x and 8x multi-GPU instance configurations
- Instances and sandboxes are never preempted — run until stopped or TTL expiry
- Snapshot an instance disk at $0.05/GB/mo and restore it later
- Zero data egress fees on instances and sandboxes
- CLI workflow: tnr create, tnr connect, tnr snapshot, tnr status
- VS Code extension for launching and iterating on GPU instances in-editor
- MCP server with a get_availability tool exposing live GPU stock to coding agents
- REST API, Python SDK and instance templates for programmatic provisioning
- Full Ubuntu GPU instances with CUDA, ready to SSH into in minutes
- GPU-over-TCP virtualization layer underlying instances and sandboxes
- CPU-only sandboxes billed on vCPU, memory and storage only
Behind the verdict
I'd reach for Thunder Compute the moment a workload is spiky. Fanning out hundreds of short RL rollouts or eval runs is where per-second sandbox billing actually changes the math — a 15-minute job billed by the second versus a whole GPU-hour is a different budget conversation. Same for a training box you use eight hours a day and snapshot overnight instead of paying for idle GPU time. When to pass: if your roadmap needs multi-region deployments for latency or data-residency reasons, Thunder doesn't offer the regional spread the hyperscalers do. If you're an enterprise buyer who needs a formal security review or a custom SLA at the on-demand tier, that's a conversation, not a checkout. And if you're pushing multi-node training past 8 GPUs self-serve, plan on talking to sales about reserved capacity first. The closest alternative is Runpod for breadth, or Lambda for a more established managed-training story. Thunder's pitch is narrower and cheaper on the same on-demand list price — its H100 at $3.20 sits under Lambda's $3.29 and Runpod's $3.99 on the vendor's Oct 30, 2026 comparison. Modal and Daytona are the sandbox competitors; Thunder's $0.000819/GPU-sec H100 sandbox is the lower number on that same check. In practice, two caveats bite. First, sandboxes require request access, so an agent pipeline that expects instant card-free provisioning will need a sales handoff. Second, availability fluctuates — the `get_availability` MCP tool exists precisely because stock can change before you launch. Budget for the occasional "sold out, pick another config" moment. Where Thunder earns its keep is smaller teams converting an unpredictable GPU line-item into something you can forecast. Per-minute billing, zero egress, and included disk per GPU mean fewer surprise fees than
Limitations
- Thunder Compute rents GPU hardware (RTX A6000, L40, A100 80GB, H100 SXM) rather than hosting AI models, so it provides compute, not managed model inference.
- Instances are billed per minute while sandboxes bill per second for each resource (GPU, memory, vCPU).
- Sandbox access currently requires requesting it through sales before setup.
- Pricing comparisons are on-demand list prices checked Oct 24, 2026 and may change before launch.
as of 2026-10-09
Researching Thunder Compute (YC S24)? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Is Thunder Compute (YC S24) still active?
Thunder Compute (YC S24) is live · pricing from $0.35/GPU-hr · last verified 30 Oct 2026 · viability monitor.
Viability score
How well maintained and how widely used is Thunder Compute (YC S24)? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →What's new in Thunder Compute (YC S24)
Checked 4 days agoAcross the latest 1 update: 1 changelog entry.
What do users say about Thunder Compute (YC S24)?
What people actually say about Thunder Compute (YC S24) — is it worth it?
We ran a structured research pass across product reviews, community discussions, and post-purchase forum threads to surface the patterns vendors won't publish themselves. Below: the recurring strengths, the hidden costs people mention most, and the cohort that consistently regrets adopting this tool.
ResearchedDiscussed by users across 1 source (Hacker News) · researched Oct 21, 2026.
Weighted by the 5 posts each of 1 source contributed.
- +Genuinely aggressive GPU pricing — RTX A6000 at $0.35/hr, A100 80GB at $1.09/hr
- +Per-second billing on sandboxes with scale-to-zero suits RL rollouts and CI evals
- +Neither instances nor sandboxes are ever preempted, unlike spot-priced competitors
- +Zero egress fees removes a classic cloud surprise on data-heavy ML workloads
- +CLI, MCP server and Python SDK fit how agent-driven ML engineers now work
- −No regional spread — a single footprint constrains latency and data-residency needs
- −Self-serve tops out at 8 GPUs; larger runs require a manual conversation
- −Sandbox access is request-gated rather than self-serve by credit card
- −Zero independent reviews means uptime and support claims are unverified publicly
- −Small seed-stage team means support likely competes with engineering time
- • Snapshots billed monthly at $0.05/GB beyond the 100 GB per-GPU free disk
- • Sandbox vCPU, memory and storage billed separately from the GPU second
- • Larger-than-8-GPU runs require a sales conversation, not a self-serve quote
How we verified this
Every listed fact about Thunder Compute (YC S24)— pricing, features, links, status — is re-read from the vendor's own pages by our verification pipeline and each change is logged with a date; the last pass ran on 30 Oct 2026. Pages mix model-assisted synthesis with human review; see how we verify.
Verification history
We have re-verified Thunder Compute (YC S24) 11 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — changed pricing tiers, our verdict, features, who it suits, description
- — changed pricing page, limitations, SSL certificate
- — changed who should skip it, our verdict, who it suits, features, description, summary, pricing tiers
- — changed platforms, limitations, pricing page
- — changed integrations, summary, description, our verdict, who it suits, features, pricing tiers, who should skip it
- — changed pricing page, models, limitations
Showing the 6 most recent of 11 verification passes.
Free to cite with attribution — this page re-verifies continuously.
How much does Thunder Compute (YC S24) cost in 2026?
Thunder Compute (YC S24): Plans from $0.35/GPU-hr. Pricing last verified 30 Oct 2026.
Plans compared
Every published Thunder Compute (YC S24) tier, its price as the vendor publishes it, who it fits, and what it adds over the tier below. The 12-month cost projection follows.
| Plan | Price | Billing | Who it's for |
|---|---|---|---|
| RTX A6000 instance (1x) | $0.35/GPU-hr | usage | — |
| L40 instance (1x) | $0.79/GPU-hr | usage | — |
| A100 80GB instance (1x) | $1.09/GPU-hr | usage | — |
| H100 SXM instance (1x) | $3.20/GPU-hr | usage | — |
| Sandbox — A100 80GB | $1.95/GPU-hr | usage | — |
| Sandbox — H100 SXM | $2.95/GPU-hr | usage | — |
| Add-ons | $0.05/GB/moVendor: “$0.04/vCPU-hr extra; $0.219/GB/mo disk; $0.05/GB/mo” | monthly, per seat | — |
| Reserved / volume | Custom | quote | — |
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Where the pricing makes sense
The company stage and team size where Thunder Compute (YC S24)'s pricing actually pencils out — and where peers do it cheaper.
The rate card fits cash-conscious startups, research labs and individual ML engineers: $0.35/GPU-hr for an RTX A6000 up to $3.20/GPU-hr for H100 SXM, billed per minute only while running, with no egress fees. Against Lambda ($3.29 H100), Runpod ($3.99) and CoreWeave ($6.16), it sits at the cheap end of self-serve on-demand. Cheaper options exist only in the spot-market tier, which trades away the never-preempted guarantee. Teams needing 8+ dedicated GPUs move onto reserved capacity through
Has Thunder Compute (YC S24)'s pricing changed?
No plan price for Thunder Compute (YC S24) has moved since we began logging in Aug 2026; the pricing page is re-read on every verification pass.
Who should pick Thunder Compute (YC S24)?
Pick Thunder Compute (YC S24) if you are ML engineers with bursty training and inference who want published on-demand rates with no spot preemption, RL and eval teams fanning out hundreds of short-lived GPU sandboxes per second, Small startups where per-minute billing and zero egress make GPU spend a predictable line, Developers who want a GPU reachable from terminal, browser console or VS Code.
Look elsewhere if you are Teams needing multi-region deployment for latency or data residency, Enterprise buyers requiring custom SLAs or a formal security review at the on-demand tier, Multi-node training beyond 8 GPUs in a self-serve deployment, Buyers who need bare-metal or non-virtualized GPU servers.
Skip it if Skip Thunder Compute if your workload needs GPUs spread across multiple regions, bare-metal servers, or a multi-node cluster larger than the 8-GPU self-serve ceiling.
GeneratedReal-world workflow fit
Concrete scenarios for the personas Thunder Compute (YC S24) actually fits — and what changes day-one when you adopt it.
Fine-tuning a 7B model overnight on an A100 80GB instance launched with tnr create, snapshotted at the end of each session so the CUDA environment survives between runs.
Outcome: Pays roughly $1.09/GPU-hr for 176 hours a month instead of the same GPU hours on AWS, and resumes the next session from a snapshot instead of reinstalling dependencies.
Fans out hundreds of short-lived GPU sandboxes from a Python script, each with a 900-second TTL, to collect rollouts in parallel, then lets them scale to zero.
Outcome: Per-second billing means ninety-second rollouts cost ninety seconds, and hundreds of parallel environments don't require reserving a standing cluster.
Wires the MCP server's live availability data into an agent so it can check stock by configuration and launch an H100 sandbox on demand with a TTL.
Outcome: The agent provisions exactly the GPU it needs for the task and the sandbox disappears when the TTL expires, avoiding a permanently reserved GPU.
Use cases
- Run fine-tuning jobs on A100 80GB GPUs at $1.09/GPU-hr instead of AWS's on-demand rate.
- Deploy an 8x H100 SXM instance for batch inference with 640 GB of VRAM.
- Launch a Jupyter notebook on an L40 GPU in minutes through the VS Code extension.
- Snapshot your training environment and restore it later without rebuilding dependencies.
- Scale from a single RTX A6000 to a multi-GPU instance without changing your workflow.
- Fan out hundreds of parallel sandboxes for RL rollouts or eval sweeps, billed per second.
- Connect AI coding agents to live GPU availability through the MCP server.
- Track experiments in Weights & Biases while training on Thunder Compute instances.
Setup time & first value
How long it actually takes to get something useful out of Thunder Compute (YC S24) — broken out by persona, not the marketing-page minute.
For an individual developer, first value is a few minutes: install the VS Code, Cursor or Devin Desktop extension, authenticate via the command palette, add a payment method, and connect to an instance. CLI users get there as fast with tnr create followed by tnr connect. Teams wiring sandboxes into an RL or eval harness should budget an afternoon to adapt their rollout code to the Python SDK and
Switching to or from Thunder Compute (YC S24)
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From AWS EC2 / SageMaker: reuse your Docker images and SSH habits; launch the same GPU class as an instance and move training scripts over unchanged.
- →From Lambda or Runpod: re-create your instance configuration on the same A100 or H100 class, move datasets via your existing transfer tooling, and swap the launch command for tnr create.
- →From Paperspace: port notebooks into the VS Code extension workflow and export datasets before cancelling, since instance disk and snapshots are per-run on Thunder Compute.
- →From CoreWeave: start with self-serve instances on the GPU type you already depend on, then move clusters of 8+ GPUs onto reserved capacity via sales.
- ↗To Lambda or Runpod: export your data and recreate the environment from Docker; note that you gain regional options but give up per-minute billing tail savings.
- ↗To a hyperscaler: containerize your workloads and rebuild IAM, networking and storage from scratch, since Thunder Compute's console-level instances and sandboxes have no direct equivalent.
- ↗To a managed inference provider: move model serving off raw GPU instances entirely if you no longer need control of the CUDA stack.
Integrations
Resources & guides
Tutorials & Learning
YouTube returned 6 videos for “Thunder Compute (YC S24)”, and we withheld 6: 6 did not mention Thunder Compute (YC S24). We are showing none, because we could not prove any of them are about Thunder Compute (YC S24).
Official links
What are the best Thunder Compute (YC S24) alternatives?
The closest alternatives to Thunder Compute (YC S24) we list are Paperspace, CoreWeave and RunPod. Ranked by direct product-type match, not category overlap.
| Alternative | Pricing | Starts at | Last verified |
|---|---|---|---|
| Paperspace | Freemium | Pricing not published | — |
| CoreWeave | Paid | Pricing not published | — |
| RunPod | Paid | Pricing not published | — |
Alternatives to Thunder Compute (YC S24)
View all
Paperspace
Paperspace is a cloud GPU platform from DigitalOcean offering per-second-billed NVIDIA H100, A100 and V100 instances, hosted notebooks, and managed ML training
Tools that pair well with Thunder Compute (YC S24)
Common stack mates teams adopt alongside Thunder Compute (YC S24), with the specific reason each pairing earns its keep.
Paperspace
Paperspace is a cloud GPU platform from DigitalOcean offering per-second-billed NVIDIA H100, A100 and V100 instances, hosted notebooks, and managed ML training
CoreWeave
CoreWeave is the AI-native GPU cloud for training, reinforcement learning post-training, and low-latency inference on NVIDIA cluster-scale
RunPod
Rent GPU pods, serverless inference endpoints and multi-node clusters by the second on Runpod's AI developer cloud.
Thunder Compute (YC S24) vs its closest rival
Thunder Compute (YC S24)'s closest head-to-head is with Spider Cloud — Spider Cloud and Thunder Compute solve completely different problems. Choose Spider Cloud if you need real-time web data for AI agents or RAG pipelines — its Rust engine, AI extraction, and browser commands are purpose-built. Choose Thunder Compute if you need cheap, on-demand GPUs for training or inference — its per-minute billing and GPU virtualization undercut traditional clouds. They are complementary, not competing. Verdict generated
Head-to-head comparisons
Thunder Compute Yc S24 vs Spider Cloud
Spider Cloud and Thunder Compute solve completely different problems. Choose Spider Cloud if you need real-time web data for AI agents or RAG pipelines — its Rust engine, AI extraction, and browser commands are purpose-built. Choose Thunder Compute if you need cheap, on-demand GPUs for training or inference — its per-minute billing and GPU virtualization undercut traditional clouds. They are complementary, not competing.
GeneratedThunder Compute Yc S24 vs Voyage Ai
Voyage AI and Thunder Compute serve entirely different layers of the AI stack. Choose Voyage AI if you need domain-tuned embeddings for high-accuracy retrieval in enterprise RAG pipelines, especially in regulated sectors like finance or legal. Choose Thunder Compute if you’re a data scientist or startup seeking dirt-cheap, on-demand GPU compute for training or inference, with per-minute billing and fast provisioning. They are complementary, not direct competitors.
GeneratedThunder Compute Yc S24 vs Temporal Ai
Temporal AI and Thunder Compute address entirely different needs — one is a durable execution platform for orchestrating resilient workflows, the other is cheap GPU compute for training/inference. Choose Temporal if you need fault-tolerant AI agents and long-running processes; choose Thunder Compute if you need affordable, on-demand GPUs with per-minute billing. They are complementary: you could use Thunder Compute for GPU-intensive tasks triggered by Temporal workflows.
GeneratedFrequently asked questions
Categories
Best-of guides
Used Thunder Compute (YC S24)? Help shape our editorial sentiment research.