FluidStack

What FluidStack actually is

FluidStack is a GPU cloud that sells managed clusters to AI labs, and it is not a traditional hyperscaler. The company started in London in 2017 as a marketplace that matched idle consumer GPUs with researchers, and over time it pivoted to what it does today: taking third-party datacenter capacity and operating it as a fully managed GPU cloud. In our view the fairest description is "infrastructure broker with a real engineering layer on top". SemiAnalysis reached a similar conclusion in its ClusterMAX 2.0 review, calling FluidStack the hired gun that turns bronze-tier datacenter infrastructure into a gold-tier customer experience.

The scale of the pivot is easy to underestimate. In November 2025 FluidStack announced a $50 billion deal with Anthropic to deliver custom data centers in New York and Texas, and it is separately developing its own US campuses (an Oklahoma site among them). Reported and confirmed since: a $1.5 billion round led by Jane Street at a valuation above $18 billion, plus earlier rounds that had never been publicly disclosed until mid-2026. The company now describes itself as US-headquartered after relocating its global headquarters to New York, though it remains a British-founded operation.

What you can actually buy falls into three buckets. On Demand covers 8 to 4,000+ GPUs billed per hour. Reserved Clusters cover 256 to 10k+ GPUs on terms of 30 days or more with monthly or annual billing and discounted rates. Private Cloud is bespoke: dedicated clusters, custom stacks, region-specific deployments, custom terms. Everything is delivered as managed Slurm or managed Kubernetes on bare metal — you are handed a working cluster, not a login page and a blank VM. This is the shape of product covered in our /category/gpu-cloud comparisons.

Pricing: what you really pay

FluidStack publishes an on-demand rate card, which is more than several competitors do. The rates we pulled from their pricing page: NVIDIA H200 SXM at $2.30 per GPU per hour, H100 SXM at $2.10 per GPU per hour, and A100 80GB SXM at $1.30 per GPU per hour. B200 SXM and GB200 NVL72 are quoted on request. Reserved clusters are also priced on request, with the page promising discounted rates and guaranteed resources in exchange for a commitment of at least 30 days.

Included in those rates: on-node NVMe storage, and no egress or ingress fees. That second point matters more than it sounds — moving datasets in and out is where GPU bills from other providers go feral, and FluidStack's stated policy of zero data movement fees removes an entire class of surprise. Support is bundled too: 24/7 engineering support with a claimed 15-minute response SLA, and fully managed Slurm or Kubernetes with no setup charge.

Two caveats. First, third-party trackers such as GetDeploying record FluidStack as a contracted, quote-per-deal provider with no self-serve signup, which matches our reading of how the larger deals work even though an on-demand rate card exists. In practice, the rate card is the floor of the conversation; reserved and private deals are negotiated. Second, the vendor's own marketing makes claims like "3-5x lower server costs than hyperscalers" and "80% lower costs". We would treat those as directional rather than literal — the honest comparison is against specific instance types on specific hyperscaler regions, and that comparison shifts monthly. Check current GPU pricing against our /benchmarks page before committing.

Our take: for a team that needs double-digit GPU counts for weeks at a time, the published rates are genuinely competitive with the big clouds, and the absence of egress fees is a real differentiator. For small experiments, this is the wrong shop — the on-demand floor is 8 GPUs.

Performance & infrastructure

The most useful third-party evidence we have is SemiAnalysis's hands-on ClusterMAX 2.0 evaluation, which awarded FluidStack a Gold debut rating — the only cloud to debut in that tier. Their testing gives a concrete picture of what customers get. A Slurm cluster arrived with pyxis and MPI support integrated into srun, and two-node nccl-tests landed within the expected range for large message sizes. Kubernetes came with Cilium for networking, a CSI with ReadWriteMany support, node-problem-detector, kube-prometheus-stack, draino, the NVIDIA Network Operator and GPU Operator, all managed through ArgoCD. That is a competent, standard-shaped stack rather than a bespoke snowflake.

The evaluation was equally clear about the rough edges. Health checks were bloated enough to add over a minute to interactive job scheduling. HPC-X was missing from the base image. A vulnerable nvidia-container-toolkit version (NVIDIAScape, CVE-2025-23266) was present at handover, as was an out-of-date GPU Operator chart on the Kubernetes side. Passive node health checks were not enabled. The pattern behind those findings matters: FluidStack's software supply chain security process was reactive, not proactive. But the response pattern matters too — HPC-X was deployed within 24 hours, the CVE was patched within minutes of reporting on Slurm and under an hour on Kubernetes, and correctness fixes to DCGM metrics landed the next day. Issues that would sit in a hyperscaler ticket queue for weeks were fixed in hours.

Infrastructure-wise, FluidStack operates mostly on third-party datacenter capacity. GetDeploying tracks four locations — Canada, the Netherlands, Norway and Sweden — while FluidStack's own about page claims a network spanning six continents with tens of thousands of GPU servers. Its own build programme is US-centric: New York and Texas for Anthropic, an Oklahoma campus in development. For EU data residency the Nordic and Netherlands capacity is the relevant part, and it is genuinely useful for teams that need European training data to stay in Europe. FluidStack claims clusters are fully burned-in and tested to deliver 95%+ of theoretical performance from day one; we have not verified that claim ourselves.

Where it falls short

The aggregated model cuts both ways. FluidStack's strength is that it can stitch together capacity fast; the consequence is that hardware provenance varies between deployments. You are buying an operational layer, not a single homogeneous fleet, and the ClusterMAX findings show that "day one" polish is not guaranteed — the team fixes things quickly, but some things should not need fixing.

Contract friction is real. Reserved clusters start at 256 GPUs, on-demand at 8. There is no free tier, no hobby plan, and no realistic path for a solo developer who wants to fine-tune a 7B model on one GPU at weekend rates. If you want self-serve with a credit card and one GPU, look at the consumer-friendly names on our /benchmarks page instead.

Transparency is mixed. There is a public rate card, which we credit. But there is no public uptime SLA we could find, no published datacenter list, no public documentation library to speak of, and most commercial terms are negotiated. For a company now building gigawatt-scale infrastructure, the public surface is surprisingly thin. And while the engineering responsiveness documented by SemiAnalysis is impressive, it was measured during an evaluation period where a customer had the team's attention — assume that attention scales down with deal size.

Finally, track record. FluidStack is a 2017-founded company whose explosive growth is recent, and a meaningful share of its capacity story is tied to a small number of very large AI customers. For enterprise buyers, the financial and operational due diligence you would do on any fast-growing supplier applies here.

Who should pick this

We would pick FluidStack when a team needs managed Slurm or Kubernetes on 8 or more GPUs, wants transparent per-GPU-hour pricing without egress fees, and values a responsive engineering partner over a self-service console. Fine-tuning runs, mid-scale training, and inference fleets with EU residency needs all fit. It is a strong option for AI labs and platform teams reserving hundreds to thousands of GPUs with committed terms — the reserved cluster product is clearly built for that.

We would not pick it for single-GPU experiments, for teams that need a deep public documentation ecosystem before signing, or for workloads that demand a contractual uptime SLA. Those buyers are better served elsewhere in /category/gpu-cloud.

Key takeaways

• Published on-demand rates — H200 SXM $2.30, H100 SXM $2.10, A100 80GB SXM $1.30 per GPU per hour — with on-node NVMe and no egress or ingress fees.
• Managed Slurm or Kubernetes on bare metal is the product; SemiAnalysis rated the platform Gold in ClusterMAX 2.0 and documented real fixes in hours, not weeks.
• Scale runs from 8 GPUs on-demand to 10k+ reserved; this is not a single-GPU, hobby-friendly cloud and there is no free tier.
• Aggregated third-party capacity means variable hardware and, in one evaluation, missing security patches at handover — diligence on the day-one state is warranted.
• EU capacity (Netherlands, Nordics) plus US builds in New York, Texas and Oklahoma; no public uptime SLA or datacenter list.
• Best fit: AI teams committing to double-digit GPU counts who want a hands-on infrastructure partner rather than a console.

Sources

• https://fluidstack.io/pricing
• https://fluidstack.io/about-us
• https://fluidstack.io/press/fluidstack-selected-by-anthropic-to-deliver-custom-data-centers-in-the-us
• https://getdeploying.com/fluidstack
• https://clustermax.ai/cloudreview/fluidstack
• https://www.cbinsights.com/company/fluidstack
• https://forbes.com/sites/iainmartin/2026/09/03/a-tiny-startup-helping-google-take-on-nvidia-is-now-worth-18-billion