In this briefing
- NVIDIA Vera Rubin NVL72 and the standalone NVIDIA Vera CPU service are generally available on Oracle Cloud Infrastructure. Oracle Cloud Infrastructure allows customers to scale CPU and GPU resources separately, but Oracle plans to make the NVIDIA Vera CPU service available only in selected regions this quarter.
- NVIDIA and Oracle place a 72-GPU rack system, a standalone CPU service and a planned 8-GPU system on one path. The three forms are respectively generally available, regionally supplied and planned, so an enterprise cannot treat the three states as the same procurement condition.
- AMD says that initial AMD Helios systems are undergoing tests at Meta, while AMD Instinct GPUs already support Meta's production workloads. VentureBeat's self-reported survey also indicates that many enterprises still do not fully use their existing GPUs, so a new-platform purchase needs to explain why existing capacity is idle.
Independent CPU and GPU scaling becomes a new signal
On 27 October, NVIDIA and Oracle announced that NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPU were generally available on Oracle Cloud Infrastructure. Each NVIDIA Vera Rubin NVL72 rack connects 72 GPUs through NVIDIA NVLink, and the associated superclusters are planned to expand to multi-gigawatt scale. The two companies classify model training and inference as GPU-intensive work, while they classify tool execution, code testing, data retrieval and Agent orchestration as CPU-intensive work.
Oracle's technical note says that Oracle Cloud Infrastructure initially plans to provide 150,000 NVIDIA Vera CPUs. Each NVIDIA Vera CPU contains 88 cores and 176 threads, and each NVIDIA Vera CPU provides up to 1.2 TB/s of memory bandwidth. A dual-processor instance contains 176 CPU cores, 352 threads, 1 TB LPDDR5X memory, 7.68 TB of local NVMe storage and 200 Gbps networking. Oracle provides the service through bare-metal instances and flexible virtual-machine instances, so customers can scale CPU services separately from GPU infrastructure.
For enterprise adoption, tool execution, code execution, retrieval and scheduling between model calls can create separate bottlenecks. An enterprise can determine whether additional CPUs or additional GPUs can shorten task time only when the enterprise separately records CPU time, GPU time, network waits, data access and external-service waits across the complete task. Separate hardware scaling provides configuration choices, but separate hardware scaling cannot by itself prove that the complete task will cost less.
General availability, regional supply and future products are three different states
NVIDIA and Oracle describe NVIDIA Vera Rubin NVL72 and NVIDIA Vera CPU as generally available, but Oracle's NVIDIA Vera CPU technical note restricts practical supply to selected regions during this quarter. Oracle does not provide instance prices, regional capacity or customer-workload results that would directly calculate an enterprise's expected cost in the two public materials. Oracle cites an up to 80% improvement in Agent CPU performance from NVIDIA platform benchmarks. Oracle Cloud Infrastructure was not the platform for the cited tests; actual results will still vary by workload and configuration.
Oracle also plans to introduce HGX Rubin NVL8. The planned system contains 8 Rubin GPUs, two 64-core Intel Xeon 6 CPUs, 4 TB DDR5 memory and 6.4 Tbps of remote direct memory access networking. The system provides a smaller configuration than a 72-GPU rack, but the system remains a future plan. The multi-gigawatt expansion similarly describes the provider's infrastructure direction, rather than a minimum purchase size for one enterprise.
For enterprise adoption, a procurement team needs to record product-release status, purchasable regions, deliverable dates, prices and test results as separate acceptance conditions. A generally available product can still face regional and capacity limits, while a planned smaller system cannot serve as an available solution before its delivery date and price are published. An enterprise can compare configurations only through tests that use the same model, context, batch size, concurrency and business process.
AMD and Meta extend hardware selection into co-design
On 27 October, AMD announced that Meta had installed initial AMD Helios rack-scale systems and was testing the systems in preparation for a broader future deployment. AMD also says that AMD Instinct GPUs already run in production for Meta's training, ranking and recommendation engines, while AMD EPYC CPUs support the dedicated computing environment behind the Meta Muse AI Agent. AMD also says that Meta has deployed millions of AMD EPYC CPUs across multiple product generations.
The AMD and Meta collaboration also covers AMD Pensando networking, AMD ROCm software, PyTorch and Triton. AMD says that the co-design will inform future AMD EPYC CPUs, a custom AMD Instinct GPU, the AMD Instinct MI500 series and next-generation rack-scale systems. AMD has provided day-one support for Meta Muse Spark Open on AMD Instinct GPUs and day-one support for Meta Muse Glimmer on the AMD Ryzen AI PC platform and AMD Radeon GPUs. The public materials about the AMD Helios tests do not provide power, memory, price, throughput or supply results.
For enterprise adoption, the co-design of chips, networking, runtime software and models can reduce compatibility delays, but the co-design also binds performance results to a particular software stack and customer scale. Meta's production workloads can prove that some AMD components already run in practice. AMD Helios testing by Meta cannot prove that the rack-scale system has completed a broader deployment. An enterprise needs to include portability, software versions, supply cycles, support responsibilities and rollback paths in a system-level comparison.
A low-utilisation survey gives expansion another test
VentureBeat Intelligence combined monthly surveys for month 7 (July), month 8 (August) and month 9 (September). Among 218 respondents who reported GPU utilisation, 79% said that their GPU utilisation was 50% or lower; 70% of that group still planned to add, replace or supplement a computing platform. The survey covers organisations with at least 100 employees, and consultants can answer for clients. The survey results come from respondents' reports rather than an audit of hardware telemetry.
VentureBeat's analysis distinguishes three different causes. A GPU might be waiting for a CPU to complete code execution, a GPU might be idle because no task is queued, or the bottleneck might sit in a remote service, database access or application design. A faster CPU might improve the first situation, but a faster CPU will not automatically create work and will not automatically remove waits in external systems. The available survey does not prove which cause dominates.
For enterprise adoption, one utilisation number cannot determine expansion or contraction on its own. An enterprise needs to place the total cost of an acceptable completed task, response time, success rate, peak capacity and recovery capability in the same test, and the enterprise needs to explain whether idle existing hardware reflects waiting, reservation or insufficient demand. Hardware specification improvements can become verifiable operating value only when a new configuration improves complete-task outcomes.
What to watch next
- Can Oracle publish NVIDIA Vera CPU instance prices, available capacity and Oracle Cloud Infrastructure instance-level workload results across different regions?
- Can NVIDIA, Oracle, AMD and Meta publish CPU waits, GPU utilisation, complete-task latency and total cost under the same model, toolchain, retrieval data and concurrency conditions?
- Can enterprise-infrastructure providers place delivered products, customer tests, future plans and vendor benchmarks in a comparable structure of states and evidence?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing