Oracle lets customers use eligible cloud credits for WEKA’s AI storage offerings

Oracle customers can now apply eligible Oracle Universal Credits toward WEKA offerings on Oracle Cloud Infrastructure, the company said on November 2. The change gives customers a way to use existing cloud commitments for AI storage software, while a more integrated Marketplace purchasing and billing experience is still being planned.

A payment change, not a new storage launch

The immediate change is credit eligibility. Oracle and WEKA describe their broader Marketplace work as an effort to simplify the path from choosing an offer to receiving a bill. That distinction separates what customers can do now from the purchasing experience the companies intend to deliver.

Customers can explore NeuralMesh by WEKA and NeuralMesh with Augmented Memory Grid through Oracle Marketplace. Oracle directs buyers to its account teams or WEKA to select a deployment and offer, including a tailored Private Offer. The credit announcement covers eligible offerings; it is not an unrestricted promise that every WEKA purchase qualifies.

Keeping training data close to GPUs

The products address two different bottlenecks. For training, NeuralMesh Axon uses local NVMe storage—fast storage within OCI GPU instances—to put datasets near the processors using them. Training systems must both read data and write checkpoints, the saved states of a training run. They also handle large transfers alongside millions of small files.

Oracle points to the Center for AI Safety, or CAIS, as a customer example. Storage bottlenecks and slow file-information lookups had constrained its research cluster, while dataset copies added costs. CAIS deployed Axon using NVMe capacity already present in its OCI computing environment, rather than relying on a separate storage silo.

In WEKA’s case study, cited by Oracle, CAIS reported a 90% reduction in cloud storage costs and an estimated fivefold improvement in storage performance over its previous solution. Those are results from that deployment, not performance guarantees for other customers.

Reusing context instead of computing it again

Inference, the work of running a model to produce responses, presents a different storage problem. Long conversations and AI-agent workflows keep reusable prior computation in a key-value cache, often called a KV cache. When memory fills, that useful context may be discarded and computed again, limiting how much work a system can handle.

NeuralMesh with Augmented Memory Grid provides a persistent, shared cache backed by NVMe storage managed by NeuralMesh. Its purpose is to retain reusable context as demand grows, rather than depend only on the system’s available working memory.

Oracle’s joint testing with WEKA used a nine-node OCI cluster containing 72 H100 GPUs and inputs of 100,000 tokens, the pieces of text models process. The comparison kept the GPU footprint unchanged and measured Augmented Memory Grid against a configuration using only DRAM, or working memory. The figures apply to that test setup and workload.

In a one-hour test with 2,400 users, the companies reported five billion tokens served with Augmented Memory Grid, versus 700 million for the baseline. That measures output delivered over a fixed period, alongside the concurrency and per-second throughput figures.

Oracle recommends evaluating these products against the workload, not just the headline gains. Its suggested buying criteria distinguish training-data access from inference-cache behavior:

• For training: assess dataset access, checkpointing needs and storage efficiency.
• For inference: assess cache reuse, simultaneous-user targets, tokens delivered per GPU and cost per token.