Crusoe Cloud suits teams that want published GPU rates and free data transfer, since eight accelerators are priced openly by the hour. Expect to call sales for the newest rack-scale parts, for any spot capacity and for Provisioned Throughput, and note that dedicated inference endpoints cost more than raw instances.
What changed
- · Update Crusoe contracts 4.9 GW of AI infrastructure capacity, announces Starcloud space partnership source
- · Update Crusoe closes $300M conventional debt round, total funding reaches $2.77 billion source
- · New feature Crusoe Managed Inference launches with MemoryAlloy, delivering up to 9.9x faster time-to-first-token source
Crusoe Cloud, a Denver GPU provider, rents NVIDIA and AMD accelerators by the hour at eight published rates and says it does not bill for network ingress or egress. The same account adds Managed Kubernetes, Managed Inference and Serverless Fine-Tuning, with the newest parts quoted by sales. Its 2026 funding round was $3.9B or more.
About Crusoe Cloud
Crusoe Cloud is the GPU cloud of Crusoe, a Denver, Colorado company founded in 2018 by Chase Lochmiller and Cully Cavness. Crusoe describes itself as vertically integrated: it sources energy, builds and runs AI data centers, and sells cloud capacity on top. Its own timeline lists a $600M Series D in 2024 alongside general availability of Crusoe Cloud, a $1B+ Series E and the Atero acquisition in 2025, and a $3.9B+ Series F in 2026. It also records the divestment of its bitcoin mining business to NYDIG to focus on AI infrastructure.
The cloud sells GPU instances by the GPU-hour, CPU instances, block, shared and object storage, and Crusoe Managed Kubernetes. On top sit Managed Inference (serverless per-token endpoints, self-serve dedicated deployments, tailored deployments and provisioned throughput) and Serverless Fine-Tuning, launched in 2026. The pricing page lists most NVIDIA generations plus one AMD part with public hourly rates, while the newest rack-scale and AMD parts need a sales conversation. Its pricing FAQ says transfer is free inside a VPC and over the public internet, which removes a line item that hyperscalers bill for data-heavy training and serving.
The serverless catalog is mostly open-weight models. Serverless Inference lists DeepSeek V4 Pro alongside GLM, Gemma and Nemotron families, and Crusoe positions its MemoryAlloy KV cache as the engine behind low time-to-first-token on repeated prefixes. NVIDIA hardware runs through the NVIDIA partnership, and the Crusoe timeline records a 900 MW Abilene, Texas campus announced in 2026 to support Microsoft AI infrastructure and the earlier Stargate site start there, the project tied to OpenAI.
It suits ML engineers and AI platform teams that want published GPU prices, free egress and managed Kubernetes without a hyperscaler contract. Who should look elsewhere: teams that need a published spot rate, a CDN, DDoS protection or domain services will not find them here, and anyone who wants the newest rack-scale capacity must negotiate. CoreWeave is the closest peer, and Atlas Cloud, Pipeshift and machine0 cover smaller or inference-first needs.
Crusoe has not published a changelog, so the 2026 company timeline is the best dated record: Edge Zones and the Spark Factory in Brighton, Colorado, Serverless Fine-Tuning and Self-Serve Inference Deployments, a nuclear partnership with Aalo Atomics, and a 1.0 gigawatt campus in Childress, Texas with Lancium. Prices on this page were read from the vendor in October 2026 and change without notice, so confirm the figure for the instance you plan to rent before committing.
Pricing
Crusoe lists on-demand GPU prices per GPU-hour on its pricing page, read 19 October 2026: B200 $8.50, H200 $5.40, H100 $4.50, A100 SXM $2.30, A100 PCIe $2.00, L40S $1.50, AMD MI300X $3.45 and HGX B300 $9.90. GB200 NVL72 and AMD MI355X are contact-sales only, and the spot column shows contact sales for every GPU, so no spot rate is published. CPU instances cost $0.04 per vCPU-hour (general-purpose) or $0.09 (storage-optimized).
Storage is $0.08 per GiB-month for persistent disks, $0.07 shared, $0.10 container registry and $0.06 object storage, and a Managed Kubernetes cluster costs $0.10 per cluster-hour. Network ingress and egress are free. Managed Inference bills per million tokens, and dedicated Self-Serve Deployments start at $6.30 per GPU-hour on H100.
There is no free tier; compare with CoreWeave for another GPU-first cloud and Hetzner for CPU hosting.
Key Features
- Zero egress fees: Crusoe states it does not charge for network ingress or egress, within a VPC or to and from the public internet, so data-heavy training and serving jobs carry no transfer line item.
- Published hourly GPU prices: Eight accelerator types carry an on-demand rate on the pricing page, while the newest rack-scale and AMD parts are quoted by sales.
- Managed Inference: Shared per-token endpoints for open models, plus dedicated and custom-benchmarked deployments and committed capacity sold in AI Model Units.
- Serverless Fine-Tuning: Fine-tunes the open-weight models on its list, priced per million training tokens and counted as dataset tokens times epochs.
- Managed Kubernetes: Managed clusters at $0.10 per cluster-hour that schedule across GPU and CPU instances.
- Energy-first infrastructure: Crusoe builds its own AI data centers around renewable and low-carbon energy, including a nuclear partnership with Aalo Atomics announced in 2026.
Strengths
- Free network ingress and egress per Crusoe, which keeps large dataset moves and multi-region serving off the bill.
- GPU rates are published for eight accelerators, so a training budget can be drafted without a sales call.
- One vendor covers raw instances, managed Kubernetes, serverless inference and serverless fine-tuning.
- Dedicated Self-Serve Deployments let a team stand up an endpoint for an open or fine-tuned model without a sales engagement.
Weaknesses
- The spot column on the pricing page shows contact sales for every GPU, so there is no published spot rate to plan around.
- The newest rack-scale NVIDIA and AMD capacity needs a sales quote and has no self-serve price.
- Self-serve dedicated deployments cost more per GPU-hour than the matching raw instances because the managed layer is included.
- It is a GPU platform, not a general cloud: no CDN, DDoS protection or domain services are listed.
Frequently Asked Questions
What does Crusoe Cloud sell, and who is it for?
Crusoe Cloud sells GPU and CPU instances, storage, Managed Kubernetes, Managed Inference and Serverless Fine-Tuning. It is aimed at ML engineers and AI platform teams that train or serve open-weight models and want published hourly rates and no network transfer charges. Crusoe is a Denver company that also builds the data centers its cloud runs on.
How much does a Crusoe GPU cost per hour?
On the pricing page read on 19 October 2026, on-demand rates were $4.50 per GPU-hour for H100, $5.40 for H200, $8.50 for B200, $2.30 for A100 SXM, $2.00 for A100 PCIe, $1.50 for L40S, $3.45 for AMD MI300X and $9.90 for HGX B300. GB200 NVL72 and MI355X are quoted by sales, and no spot rate is published. There is no free tier.
Does Crusoe charge for data transfer?
No. Crusoe answers no on its pricing page, for traffic inside a VPC and for traffic to or from the public internet. Storage is still billed separately, and the policy is worded as applying at this time, so read the pricing page again before a very large migration.
How is Crusoe Managed Inference billed?
Serverless Inference is pay-as-you-go per million tokens, split into input, output and cached tokens, with a rate per model on the pricing page. Self-Serve Deployments bill per GPU-hour for a dedicated endpoint. Tailored Deployments and Provisioned Throughput, sold in AI Model Units, are priced through sales.
How does Crusoe compare with CoreWeave?
Both are GPU-first clouds. Crusoe publishes eight hourly GPU rates and says egress is free, while the CoreWeave page on Hokai covers that provider. Crusoe adds serverless fine-tuning and a per-token inference catalog on the same account. Pick the one whose region, GPU generation and contract terms match your job, and request quotes from both for committed capacity.
Top Alternatives
Hokai guides that mention Crusoe Cloud
- Fireworks AI vs Together AI (2026): Managed Platform or Neocloud?: Fireworks AI and Together AI charge the same per token on seven shared open models. So pick by setup: Together rents GPU clusters, Fireworks sells managed ones.