AMD Ships 256-Core EPYC SP7 CPUs While SambaNova Commits 10,000 Chips to Supermicro

AMD makes 256-core EPYC SP7 processors available

AMD released its 6th Gen EPYC SP7 processors on November 10, supporting configurations up to 256 cores. The chips, built on Zen 6 architecture, are shipping from major server vendors now. Microsoft Azure and Oracle Cloud Infrastructure plan to offer instances based on the new processors, though neither has published availability dates.

The launch matters because it gives enterprises a CPU platform they can buy today rather than waiting for future roadmap promises. Organizations refreshing virtualization clusters, database hosts, or AI-support infrastructure can evaluate EPYC SP7 against current Intel Xeon generations and incumbent AMD platforms. Azure and OCI distribution reduces procurement friction for buyers who prefer consuming infrastructure as cloud instances instead of purchasing physical servers.

AMD is also shipping Helion AI racks designed for agentic-AI workloads alongside the CPU launch. This positions AMD as a supplier of complete AI systems rather than a CPU alternative.

The gap: AMD has not published launch pricing or independent performance-per-dollar benchmarks. Buyers should request vendor-specific total cost of ownership figures, power consumption data, and workload-specific performance results before committing budget. The 256-core maximum is a data point, but without normalized cost and power metrics, it does not prove procurement advantage.

Google adds TPU 8tx and TPU 8i Max to accelerator lineup

Google Cloud announced two TPU additions: TPU 8tx, a more secure version of its TPU 8t training accelerator, and TPU 8i Max, a higher-capacity version of its inference-focused TPU 8i. The company did not publish pricing, capacity figures, benchmark results, or customer adoption numbers.

The products compete with NVIDIA GPUs, AMD accelerators, and alternative cloud-specific chips such as Amazon's Trainium and Inferentia. The higher-capacity inference option suggests Google is targeting production serving workloads rather than only model training.

For Google Cloud customers, the TPUs create another route to reduce dependence on scarce or expensive GPU capacity. However, the absence of published price-performance data, supported-model lists, and software-framework compatibility details makes immediate platform migration difficult to justify. Buyers should demand throughput-per-dollar comparisons, contractual capacity guarantees, and model portability assessments before moving inference workloads from GPUs to TPUs.

SambaNova and Supermicro commit to 10,000-chip inference systems

Supermicro and SambaNova announced a collaboration covering 10,000 SN50 Reconfigurable Data Units. Supermicro will build and sell disaggregated inference systems combining SambaNova RDUs with NVIDIA GPUs. The systems initially target neoclouds and sovereign-cloud providers, followed by enterprises operating AI in their own data centers.

SambaNova claims the SN50 uses a three-tier memory design supporting models with up to 10 trillion parameters and thousands of simultaneous AI sessions. Those figures are vendor claims, not independently reported benchmarks.

The design challenges all-GPU inference deployments by separating workloads across SambaNova inference hardware and NVIDIA GPUs. It also gives Supermicro an alternative architecture to server systems built exclusively around NVIDIA accelerators.

The 10,000-chip figure represents a supply commitment, not evidence of enterprise adoption. The announcement does not include system pricing, delivery schedules, power-per-token results, or production customer names. Buyers should treat this as a product roadmap signal rather than a procurement option until Supermicro publishes concrete commercial terms.

For organizations with data sovereignty requirements or on-premises AI deployment mandates, the collaboration could create another procurement path for high-volume inference. However, without disclosed pricing or delivery timelines, budget planning remains speculative.

What to watch

The most actionable development is AMD's EPYC SP7 availability. It combines a shippable product, a stated 256-core maximum, named cloud distribution through Azure and OCI, and server-vendor availability. Buyers refreshing infrastructure in the next two quarters should request EPYC SP7 quotes alongside incumbent platforms and demand workload-specific performance data before deciding.

The SambaNova–Supermicro deal is strategically significant because of its 10,000-chip scale, but its commercial importance remains unproven. Watch for pricing announcements, delivery commitments, independent benchmarks, and named enterprise deployments. Until those emerge, the collaboration is a supply agreement, not a validated procurement path.

Google's TPU additions expand cloud inference options but lack the published data needed to justify migration from GPU-based workloads. Demand price-performance comparisons and model portability assessments if considering a shift.