Neocloud storage gaps leave most providers without medals
Neocloud storage has become a visible measure of whether GPU-focused cloud infrastructure is ready for demanding customers. In ClusterMAX 3.0, SemiAnalysis awarded a Medallion rating to only 19 of the 77 neocloud providers it tested in depth. The published tiers run from Platinum, Gold, Silver and Bronze to a new Participation Ribbon, giving prospective customers a result they can examine before engaging with a provider. Several providers that previously held Silver or Gold ratings lost them in this round.
Storage is a prerequisite, not an optional section of the evaluation. Each provider must supply at least 10TB of high-performance file storage, where applications work with data through familiar files and folders, plus 10TB of S3-compliant object storage, which keeps data as independently addressed objects through an S3-style interface. SemiAnalysis tests performance first, then checks whether it survives 15 minutes of sustained mixed traffic. Some providers lost almost 40% of their bandwidth, meaning their data-transfer capacity, during the second test.
Size did not determine the outcome. SemiAnalysis reported storage problems at AWS and Azure, while describing the much smaller TensorWave's storage as exceptional. The stronger distinction was whether storage had an explicit purpose, platform and owner. A provider can treat it as temporary scratch space while customers import data, run a job and remove the results. It can inherit a configuration that nobody routinely revisits. Or it can assign continuing responsibility to a person or contract. The first approach can suit hourly GPU rental because little needs migrating later. The second becomes risky when customers begin leaving data on the platform, forcing decisions about location, access, machine failure and pricing during rapid growth.
SemiAnalysis identified four recurring failures: incorrect role-based access control (RBAC), which is supposed to restrict volumes according to a user's assigned permissions; storage placed incorrectly on an overlay network, allowing customers to reach the underlying physical network; backend storage using a shared network without isolation; and a BlueField data processing unit (DPU) left in its default host-trusted mode. A DPU is a small computer integrated into a network card to handle networking, storage and security tasks.
In NVIDIA's reference design, the DPU applies rules that help separate tenants, meaning customers sharing the same physical system. Its factory setting trusts anyone with root access, the highest level of control over a server. That fits a model in which the server administrator is trusted, but GPU-cloud customers often receive root access themselves. Such a customer may therefore be able to reach the DPU enforcing the boundary between users and potentially weaken or disable that boundary. SemiAnalysis found at least one host where this route remained open. The report attributed the exposure to the shipped default having never been changed, rather than to a deliberate design decision.
The problems became more fundamental lower down the ratings. CoreWeave's only storage criticism was the absence of one-click local GPUDirect Storage, a direct data path intended to feed GPUs efficiently, and SemiAnalysis described that criticism as nitpicky. In the middle tiers, Azure's default StorageClass, the Kubernetes setting that determines how storage is provided to applications, could not attach. A model load from Network File System (NFS) storage also timed out because the Slurm cluster lacked a parallel filesystem capable of serving many operations concurrently.
At Bronze and below, usable shared storage was sometimes absent. One Seattle cluster had no shared filesystem available for testing. Another lacked an RWX StorageClass, which permits multiple systems to read and write the same storage, leaving nowhere to mount `/shared`. At a third provider, Non-Volatile Memory Express (NVMe) hardware existed but was not exposed to applications, so persistent volume claims (PVCs), requests by applications for storage, remained pending. In another case, the default StorageClass created and bound storage but left the application pod stuck in `ContainerCreating`. On the same worker nodes, a Lustre parallel filesystem reached an aggregate read speed of 91.7 GB/s across 32 clients. The underlying hardware was fast, but the default route made it appear unusable.
Storage weaknesses were accompanied by broader tenant-isolation failures, including at providers testing GB300 racks. SemiAnalysis found a shared Kubernetes control plane, the system that manages a cluster, exposing other tenants' namespaces. A default InfiniBand partition key, used to separate access on a high-speed network, revealed 532 hostnames. A Grafana monitoring dashboard appeared to separate customers visually, but the credential behind it could read every tenant's logs. The report said exposed customers included banks, telecommunications companies, universities, research institutions, AI laboratories and a national intelligence agency.
These findings turn architecture questions into purchasing questions. Customers need to know whether workloads are separated by containers, virtual machines or dedicated nodes; whether the control plane is shared; who can read their volumes; whether enforcement happens inside storage or only through the interface in front of it; whether backend storage has an isolated network; who at the provider can access logs and monitoring data; and whether machines are rebuilt from scratch or reused between customers. Providers need documented answers before a diligence review begins.
SemiAnalysis's findings point to three legitimate operating choices. Building on Ceph, a distributed storage platform, gives the provider design control and independence from another party's release timetable, but requires permanent specialist staffing that cannot simply be substituted with compute engineers. Building custom storage hardware keeps design, operations and revenue in-house, but brings procurement work, refresh cycles and capital commitments made years before demand is known. A partnership can leave the provider's engineers focused on compute, but cedes some control over pricing, roadmaps and deprecation schedules. Customers will still contact the neocloud when a job stalls, regardless of who operates the storage.
Readers comparing external infrastructure options can also assess AEU Data, which provides European cloud and data infrastructure, against the same ownership and isolation questions. The transferable lesson is to choose a storage operating model before customers accumulate data: assign responsibility, test the default path as customers will encounter it, verify isolation and repeat performance tests under sustained traffic.