
Neoclouds need their own data platforms because the work that feeds a GPU runs continuously while training runs episodically, and today almost all of that work happens on a hyperscaler. Cleaning, joining, deduplicating, embedding and rebuilding evaluation sets fill the hours between training jobs. A neocloud that sells only the training window leaves that continuous, billable work, and the margin attached to it, with someone else.
CoreWeave’s third-quarter 2026 results, reported November 10, show how the category is moving up the stack. The managed inference platform launched earlier in the year had already booked more than $250 million in ARR, and CFO Nitin Agrawal said its growth was “constrained primarily by the capacity allocated to it.” Seventy percent of new cloud customers in the quarter were already Forge customers. The climb runs from capacity to inference to developer experience. A search of the Q3 press release, the earnings call transcript and the Form 10-Q for “data processing”, “ETL”, “Spark”, “analytics” and “data preparation” returns nothing for any of the five terms.
What Is a Neocloud?
A neocloud is a cloud provider whose infrastructure, software and commercial terms are organized around one job: delivering GPU capacity at scale to teams that train or serve large models. Industry writing also calls it a GPU cloud, an AI cloud or a specialized AI infrastructure provider. No standards body defines the term, so the working definition comes from what these providers actually ship:
- GPU-dense fleets sold in units sized for training and inference, not carved out of a general-purpose compute pool.
- A dedicated accelerator fabric for GPU-to-GPU traffic, because multi-node training is sensitive to how fast gradients move between cards.
- Bare-metal or near-bare-metal access, so provisioned hardware behaves the way its specification sheet says.
- Short provisioning cycles, measured in minutes to days rather than a quota request.
- Mixed commercial terms: on-demand, spot and multi-month or multi-year reserved clusters.
What a neocloud provider sells
A neocloud provider sells depth in one layer, the accelerator and the network around it. A hyperscaler sells breadth: identity, databases, queues, warehouses and GPUs as one product line among hundreds. That is the whole difference, and it explains both why teams rent neocloud GPU capacity and why their data rarely follows it.
What a neocloud provider leaves out
A neocloud provider leaves out the long tail of managed data services. The documentation set usually covers compute, storage, networking and sometimes orchestration, and stops. There is no managed warehouse, no lakehouse, no native catalog and rarely a workflow orchestrator. The narrowness is a deliberate design choice, and it is the reason the data layer is still unclaimed.
Neocloud Companies and What Their Platforms Include
The best-known neocloud companies publish catalogs that look similar at the category level. Read by what each provider’s own documentation says it ships:
| Neocloud company | Compute and orchestration | Storage | Data processing |
|---|---|---|---|
| CoreWeave | Managed Kubernetes, Slurm on Kubernetes | S3-compatible object storage | Not mentioned in Q3 2026 disclosures |
| Nebius | VMs, managed Kubernetes, Slurm | Block, file and S3-compatible object storage, managed Postgres | Managed Service for Apache Spark, in Preview |
| Lambda | On-demand VMs, 1-Click Clusters, preinstalled ML stack | Not detailed in its public-cloud docs | Not in the published catalog |
| Crusoe | GPU VMs, orchestration | Persistent and object storage; local storage is ephemeral | Not in the published catalog |
| Together AI | GPU clusters on Kubernetes or Slurm, inference endpoints | Not detailed in its cluster docs | Not in the published catalog |
Nebius is the one exception in the data-processing column, and its terms of use put the Spark service at Preview stage, with no service level agreement and free of charge during that preview. No primary source puts a revenue figure on it.
Also Read: How to Reduce Cloud Data Platform Costs
Neocloud GPU vs Neocloud Storage: Where the Platform Stops
Neocloud GPU capacity is the product; neocloud storage is there to keep the GPUs fed. Object storage is close to universal across the category, usually S3-compatible and tuned for the large sequential reads a training job makes. Some providers add block volumes, shared file systems or a managed Postgres. What the storage layer does not come with is anything that turns raw data into training data: no catalog, no governance layer, no engine for joins, aggregations or deduplication over that storage.
So the neocloud platform ends at the bucket. Everything that happens to data before it lands there, and after the model’s outputs come back out, happens somewhere else.
The Split Estate Neocloud Customers Live With
Most AI teams that rent from a neocloud today run a split estate. Raw and intermediate data is stored and processed on a hyperscaler, because the warehouse, the lake, the catalog and the existing ETL jobs already live there. GPUs sit on the neocloud, because that is where capacity was cheapest or available soonest.
The boundary between the two costs money and time on every cycle:
- Egress. Moving a training dataset out of a hyperscaler crosses its data-transfer-out meter, and the charge recurs every time the dataset is refreshed. It is billed by the hyperscaler, so it never appears on the GPU contract.
- Duplicate copies. Feeding the GPUs means copying data across, which adds storage, adds egress and raises the question of which copy is current.
- Split governance. Access control, lineage and audit logs live in the hyperscaler’s catalog. Data copied to a second provider either gets its own rules, inherits them through federation, or sits outside both systems. The third option is where compliance gaps appear.
- Engineering hours spent keeping two environments in sync.
The neocloud sees only the training window. The feature building before it and the evaluation work after it are billed to somebody else.
Why Data Preparation Belongs on the Neocloud Platform
Data preparation belongs on the neocloud platform because it is the continuous workload in an AI lifecycle, and continuous work is what a heavily financed fleet needs. Training starts, holds a block of GPUs for days or weeks, and ends. Preparation has no such rhythm. Every new source needs cleaning, joining and partitioning before a GPU touches it. Every model swap re-embeds the retrieval index. Every change to the eval harness rebuilds the evaluation set. That work runs whether or not a training job is in flight.
The financing makes the point sharper. CoreWeave’s net interest expense was $846 million in Q3 2026, against $310 million a year earlier, with $1.05 to $1.15 billion guided for Q4. CEO Mike Intrator said more than 130,000 GPUs come off contract in 2027 and roughly 170,000 more in 2028. A fleet carrying that much debt and roll-off risk, and depreciating the whole time, needs software-margin revenue that shows up whether or not a given training cluster is booked this week. The data layer is the largest piece of that revenue nobody in the Q3 disclosures has claimed, and building it does not require another data center.
The Nebius counter-case
The Nebius counter-case is real but narrower than it first looks. Nebius describes its Spark service as “a fully managed data processing engine that simplifies and accelerates data engineering and machine learning workloads,” framed around “using Spark for data preparation and a specialized framework for training.” That is the layer this article describes, built by a direct peer. It is also in Preview, unpriced, and disclosed in the same shareholder letter in which Nebius dropped ARR as its headline metric and narrowed full-year revenue guidance to the low end of $3.0 to $3.4 billion. One peer is building toward the layer. Nobody has proved it out.
The GPU data-prep counter-case
The GPU data-prep counter-case says the existing fleet can absorb some of this work. Cloudera and NVIDIA announced in August 2026 that Cloudera Data Engineering can run Apache Spark 4.1 ETL and data-preparation pipelines with “up to 4x workload acceleration” on NVIDIA GPUs versus “traditional CPU infrastructure.” NVIDIA’s own cuDF documentation is explicit that unsupported operations and user-defined functions fall back to CPU. Some preparation moves onto GPUs; the rest still needs CPU capacity next to the data. Either way it is continuous, billable work the neocloud does not capture today.
Six Requirements for a Neocloud Data Platform
A data platform built for a neocloud’s own environment comes down to six requirements:
- It runs inside the operator’s environment. If data has to leave for someone else’s cloud to be processed, the argument for building the layer collapses.
- It spans the environments customers already use. Almost nobody doing serious AI work is single-cloud.
- It executes multi-tenant. A dedicated cluster per account kills the margin story.
- It reads existing catalogs in place. Hive Metastore, Unity Catalog and AWS Glue stay where they are; moving governance is the project that stops adoption before it starts.
- It carries enterprise certifications. SOC 2, HIPAA, GDPR and ISO 27001 are what procurement asks for.
- Its pricing does not meter per core, hour or job. A vendor that takes a cut of every compute-hour takes margin straight out of the thing the operator is trying to build margin on.
An Existing Example of the Architecture
All six requirements can be met in one shipping product today, and Yeedu is a clean example. Yeedu is a Spark execution layer, re-engineered for performance while staying compatible with existing Spark code, that runs inside the customer’s own cloud account or on-premises servers rather than as a hosted service. Its stated deployment targets are AWS, Azure, GCP, Oracle Cloud and on-premises servers, on ARM, x86 and GPU runtimes. It supports multi-tenant execution, and its catalog layer reads Hive Metastore, Unity Catalog and AWS Glue directly, with no metadata duplication. It carries ISO 27001, HIPAA, GDPR and SOC 2. Licensing is fixed-price with unlimited usage: no per-core, per-hour or per-job charge.
Nothing here says Yeedu runs on a neocloud, and no partnership exists. The claim is narrower. The architecture a neocloud would need, environment-resident, multi-tenant and catalog-federated, already exists as a product category rather than a hypothetical. Adoption on these terms is a job at a time. A Spark job is pointed at the engine while orchestration, notebooks and the catalog stay where they are, which is also the model a neocloud’s own data layer would have to follow, whoever builds it.
Two problems stay open. Who owns the data once an infrastructure operator runs the processing layer, not just the GPUs, gets harder as the layer gets more capable. And the transition period, with pipelines half on a hyperscaler and half on a neocloud, is exactly the kind of project every team defers.
Conclusion
- A neocloud is a GPU-first cloud provider with a deliberately narrow catalog: compute, storage, networking and sometimes orchestration.
- Neocloud storage keeps GPUs fed, but the neocloud platform stops at the bucket, with no catalog and no processing engine.
- Customers therefore run a split estate, paying egress, duplicate storage and split governance on every cycle.
- Data preparation is the continuous workload in the AI lifecycle, and the Q3 2026 disclosures from neocloud companies leave it unclaimed.
- Nebius’s Preview Spark service and GPU-accelerated ETL are real counter-cases, but neither has turned the layer into revenue yet.
- A neocloud data platform needs six things: environment-resident, multi-cloud, multi-tenant, catalog-federated, certified and not metered per compute-hour.
FAQs
What is a neocloud provider?
A neocloud provider is a cloud company built around selling GPU capacity for AI training and inference, with dense GPU fleets, a dedicated accelerator network, bare-metal access and short provisioning times. Its service catalog is intentionally narrow compared with a hyperscaler’s.
Which companies are neocloud companies?
CoreWeave, Nebius, Lambda, Crusoe and Together AI are among the best-known neocloud companies. Each publishes compute, storage and networking services. Nebius is the only one of the five with a documented data-processing service, a Spark offering in Preview.
How is a neocloud GPU different from a hyperscaler GPU?
A neocloud GPU is the core product, sold on bare or near-bare metal with a dedicated GPU-to-GPU fabric and on-demand, spot or reserved-cluster terms. On a hyperscaler, GPUs are one instance family among many, often quota-constrained in busy regions, sitting next to a far wider set of managed services.
What storage does a neocloud offer?
Neocloud storage is usually S3-compatible object storage tuned for large sequential reads, plus block or file volumes on some providers. It holds datasets and checkpoints but comes without a catalog, governance layer or processing engine.
Why do neoclouds need their own data platforms?
Data preparation runs continuously, while training runs in episodes. Without a data platform, that continuous work stays on a hyperscaler, so customers pay egress and split governance, and the neocloud misses recurring software revenue on hardware it has already financed.
Can data preparation run on neocloud GPUs?
Part of it can. GPU-accelerated Spark runs supported operators on the GPU, and unsupported operations and user-defined functions fall back to CPU. A practical neocloud data platform needs both GPU and CPU execution close to the data.



