Oracle explained its GPU repair feature on 26 October, the day Databricks announced new AI Search sources. Oracle protects hardware it cannot build fast enough. Databricks removes a data copy stored on someone else's cloud.
Oracle says it delivered more than 300,000 GPUs to AI cloud customers since the end of its fiscal fourth quarter. That is almost triple the capacity it delivered in the fourth quarter. A GPU earns money only while it stays in service.
Oracle's tool for keeping broken GPU nodes in service went live on 2 October, and Oracle explained it on 26 October. Databricks announced the same day that its search product can now read live tables directly.
The two products address different costs. Oracle is protecting capacity it says it cannot build fast enough. Databricks is removing a data copy that customers pay for on someone else's cloud. They compete in the layer above those products, which supplies context to AI agents.
How GPU repair works
Oracle's release notes date OKE GPU host repair to 2 October. OKE is Oracle's managed Kubernetes service. Repair tied to maintenance events followed on 21 October.
The documentation lists three repair types. The first reboots a node when GPU health monitoring reports a critical signal lasting at least 10 minutes. A 60 minute cooldown stops repeated reboots.
The second type starts when OKE links a GPU or network fault to an Oracle Compute maintenance event. OKE then either schedules a maintenance event that keeps the instance, or terminates it, depending on the event's action.
The third type starts when an operator flags a node as unhealthy. OKE drains the node, terminates the instance and restores the node pool's configured capacity. Oracle Support must enable it for a tenancy before live use.
Replacing a node requires available capacity. The documentation says a replacement node "depends on managed-node-pool reconciliation and available GPU capacity". A customer who needs a fast swap needs spare GPUs in the pool.
Oracle also warns that a "Completed" repair "does not confirm that the underlying GPU fault was corrected". The customer has to check the node's health afterwards.
The documentation tells customers to run dry run mode first, which proposes a repair without touching the node. Oracle's example policy repairs one node at a time and sets the unhealthy node threshold at 10%. A node must match exactly one repair configuration, or the request is rejected.
The feature has limited coverage. It applies to bare metal GPU nodes in enhanced clusters on Kubernetes 1.32 or later. Virtual nodes and non GPU nodes are excluded.
The supported shapes run from NVIDIA's A10 and A100 to the GB200 and GB300. They also include AMD's MI300X and MI355X, so the newest racks are covered.
Capacity growth and the cost of building
Uptime matters now because Oracle is adding GPUs faster than it adds anything else. Its Q1 results show cloud infrastructure revenue of $7.4bn, up 121%. Oracle added 850MW of datacentre capacity.
Remaining performance obligations, contracted revenue not yet booked, stand at $664bn. Oracle says demand for AI training and inference "continues to grow faster than supply".
With supply constrained, a broken node is capacity Oracle has built but cannot bill for. Repair turns a hardware fault into a managed event instead of a ticket.
Oracle guides cloud revenue up 65% to 71% for the quarter to 30 November. That is $13.2bn to $13.6bn, against $7.98bn a year earlier. It is $1.6bn to $2.0bn more than the $11.6bn reported for the August quarter.
Oracle also carries the build risk. It spent $28.5bn on capital in the quarter, helped by new equity of $19.9bn and customer prepayments of $11.4bn. Notes outstanding reached $125bn, and free cash flow was negative $5.4bn.
What Databricks removes
Databricks addresses customers' data costs rather than hardware. Its AI Search, formerly Vector Search, now indexes Streaming Tables, Materialized Views and managed Iceberg v3 or later tables, alongside Delta tables. Streaming Tables hold events as they arrive. Materialized Views hold maintained aggregates.
Before this, a customer built a separate Delta copy and pipelines to keep it in sync. The blog says AI Search now removes "the intermediate Delta table and custom synchronization pipeline". Fraud checks, personalisation and agents then search fresher data with no copy to pay for.
The blog names fraud detection, security monitoring and real time IoT analytics for streams. It names session aware search and agent assist for views. Managed Iceberg tables need no conversion to another format. That matters to customers who keep data open so more than one engine can read it.
Databricks runs on Azure, AWS and Google Cloud. Its savings on copies go to the customer, and its compute bill goes to those providers.
The company reported a revenue run rate above $7bn, growth above 80% and positive adjusted free cash flow over 12 months. Its data warehousing line runs above $1.5bn, Lakebase above $100m, and more than 1,000 customers pay above $1m a year. It raised $5bn at a $190bn valuation, at most 27 times that run rate.
Oracle's cloud infrastructure revenue of 7.4 billion dollars in the quarter to 31 August 2026 annualises to 29.6 billion dollars, against a Databricks total revenue run rate of more than 7 billion dollars, of which data warehousing is 1.5 billion and Lakebase 0.1 billion.
View data
| Revenue line | Annualised revenue, billions of dollars |
|---|---|
| Oracle cloud infrastructure (7.4 x 4) | 29.6 |
| Databricks total run rate | 7 |
| Databricks data warehousing run rate | 1.5 |
| Databricks Lakebase run rate | 0.1 |
The two businesses differ in size. Oracle's cloud infrastructure line, $7.4bn in the quarter, annualises to $29.6bn. That is roughly four times Databricks' whole run rate. Oracle's figure includes work that is not AI, and Databricks' figure is a floor.
Competition for the data layer
The companies compete over the data layer above the hardware. Oracle's 10 September results release says its new AI Data Platform "fully automates the creation of Enterprise Ontologies". Oracle names Palantir's hand built versions as the model.
That is the territory Databricks sells with its Genie assistant. A customer who buys GPUs and the data layer from Oracle gives it both the compute and the context. Databricks sells only the context, and on rivals' hardware.
Oracle's Investor Day at Oracle AI World on 28 October, live from 10:55 a.m. Central, is the first chance to tie repair to revenue. Suppose Oracle repeats the 300,000 GPU figure without an uptime number for bare metal nodes. Then the guide for $13.2bn to $13.6bn rests on new racks arriving. The repair tool then remains an operations feature, not a revenue argument.