Microsoft Maia 300 vs Nvidia Rubin: Specs, Not Benchmarks

Microsoft unveiled Maia 300 on Oct. 12, 2026, announcing plans to expand its Azure Maia infrastructure to multi-gigawatt deployments beginning in 2027. The chip is designed to become Microsoft’s primary AI inference engine, challenging Nvidia’s dominance in data center computing.

Nvidia’s Vera Rubin platform serves a broader market, supporting AI training, inference, and scientific computing through partner-built systems. Both companies promise major efficiency improvements, but their published performance figures use different baselines and cannot establish which platform is faster.

Microsoft Maia 300 vs Nvidia Rubin at a glance

Nvidia’s published Rubin specifications are preliminary and subject to change.

How we compared

This comparison uses official Microsoft and Nvidia specifications, architecture documentation, product announcements, and Nvidia’s regulatory disclosures. Memory, bandwidth, rack configurations, and deployment plans were checked individually.

Manufacturer performance claims are reported with their original comparison baselines. No independent, workload-matched Maia 300 vs Rubin benchmark is available, so the comparison does not assign a performance winner.

Neither company provides a directly comparable public system price. Maia 300 is not offered as a separately rentable accelerator, while Rubin system pricing depends on commercial configurations and suppliers.

Microsoft Maia 300 specifications and rollout

Each Maia 300 accelerator includes eight HBM4 stacks providing 288 GB of memory and 18 TB/s of peak bandwidth within a 950-watt power envelope, according to Microsoft’s architecture breakdown.

The multi-chiplet design uses TSMC’s N2P process for compute, with separate memory-controller and I/O functions. Compared with Maia 200, Microsoft reports 2.5 times the memory bandwidth, 33% more memory capacity, and approximately twice the HBM bandwidth per watt.

Maia 300 also introduces a locality-aware, non-uniform memory architecture that helps software place data closer to the compute resources using it. Lower-precision formats and hardware-accelerated conversion reduce data movement during long-context and reasoning-intensive inference.

Microsoft’s two-tier Ethernet-based network can connect up to 12,288 accelerators in a superpod. Each rack contains 96 accelerators across 24 trays, with a total power envelope of 167 kW.

Liquid cooling removes more than 80% of the rack’s thermal load. Dual power feeds, automatic transfer switching, and expanded telemetry are designed to improve reliability.

Those specifications address data center power and cooling, where electrical capacity and heat removal can limit how densely operators deploy AI hardware.

Maia 300 will begin production deployments in 2027. Microsoft previously introduced Maia 100 in 2023, and Maia 200 is now serving production workloads, including Microsoft Copilot.

GitHub Copilot and Microsoft Foundry are expected to receive Maia 200 infrastructure support by the end of 2026. That timeline does not mean Maia 300 will be available before its scheduled 2027 deployment.

What Nvidia Rubin offers

Nvidia’s Rubin NVL72 combines 72 Rubin GPUs with 36 Vera CPUs, NVLink 6 networking, and additional components designed to support AI workloads at rack scale.

According to Nvidia’s official specifications, each Rubin GPU includes 288 GB of HBM4 memory and up to 22 TB/s of bandwidth.

That gives Rubin greater published bandwidth per accelerator than Maia 300, although the figures alone cannot predict application performance. They do not account for differences in model execution, memory management, networking, or software optimization.

An NVL72 rack provides 20.7 TB of aggregate HBM4 memory and approximately 1,580 TB/s of aggregate memory bandwidth.

Nvidia also positions Rubin NVL72 for agentic AI workloads requiring substantial throughput and long context windows.

Unlike Microsoft’s inference-focused chip, Rubin supports large-scale model training alongside inference. Nvidia’s CUDA software ecosystem and relationships with system manufacturers give enterprise customers additional deployment options.

Nvidia has begun production shipments of Vera Rubin hardware, according to its financial disclosures. Availability of specific configurations still depends on system manufacturers and cloud providers.

Maia 300 vs Rubin AI performance claims

Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than unnamed leading GPU-based systems, according to its Oct. 12 announcement.

That figure does not establish a performance advantage over Rubin because Microsoft does not identify the competing GPUs or provide a shared testing methodology.

Microsoft also claims up to seven times the rack-scale token throughput and more than 2.5 times the performance per watt of Maia 200. Both comparisons use Microsoft’s previous generation as the baseline.

Nvidia has published a separate efficiency projection for Rubin NVL72: up to 10 times more tokens per megawatt than GB200 NVL72, using the Kimi K2 Thinking model with 32,000 input tokens and 8,000 output tokens.

Nvidia also projects inference costs at one-tenth those of GB200 NVL72 under its stated workload assumptions. These are Nvidia’s estimates, not independent Maia comparisons.

The companies’ figures cannot be combined to calculate a winner. A reliable comparison would require the same model, precision, input and output lengths, latency targets, workload concurrency, and system-level power measurements.

Even their rack configurations differ. Microsoft’s 96-accelerator rack contains more HBM4 memory overall, while Nvidia publishes higher peak bandwidth per Rubin GPU. Neither specification establishes which system processes more useful tokens for a given amount of electricity.

Which platform should enterprises consider?

For most enterprise customers, Maia 300 is not a conventional hardware purchasing option.

Microsoft is developing it for its own infrastructure, with planned applications across Copilot, Teams, GitHub, Microsoft Foundry, and Dragon Copilot. Customers could eventually benefit through improved service capacity or responsiveness, but Microsoft has not quantified customer-level improvements or announced a separate Maia 300 rental service.

Nvidia’s Rubin systems, by contrast, target organizations that want accelerated computing infrastructure for their own AI training and inference workloads, whether through hardware partners or supported cloud environments.

Enterprises evaluating Nvidia systems can request workload-specific benchmark data, deployment configurations, pricing, and infrastructure requirements. Those comparisons should include power consumption and operational costs rather than GPU performance alone.

For Microsoft customers, the practical measures will be Azure service latency, availability, throughput, and pricing once Maia 300 enters production.

Microsoft has not disclosed how the transition will affect individual Azure service prices or whether customers will be able to select Maia 300-backed infrastructure.

Until production measurements become available, Microsoft and Nvidia’s efficiency claims remain separate projections rather than evidence of a direct performance winner.

Also read: AMD’s new Helion rack design offers another approach to combining CPU and GPU infrastructure for agentic AI workloads.