NVIDIA has fundamentally shifted its AI infrastructure strategy by opening NVLink, its proprietary high-bandwidth interconnect, to third-party custom chips and XPUs (accelerated processing units) rather than restricting it exclusively to its own GPUs and processors. The move, announced alongside the expansion of NVHBM (NVIDIA High-Bandwidth Memory) integration, signals a pivot toward positioning NVLink as the de facto industry standard for AI factory connectivity. Rather than compete directly across every silicon category, NVIDIA is weaponizing its interconnect advantage—ensuring that whether a customer deploys custom inference accelerators from partners, AMD MI300X alternatives, or NVIDIA's own silicon, they remain tethered to NVIDIA's ecosystem through NVLink's superior latency and bandwidth characteristics. This architectural lock-in is far more durable than product lock-in and addresses hyperscalers' demands for heterogeneous compute stacks optimized for specific workload profiles.
Underpinning this shift is the arrival of Vera, NVIDIA's first CPU purpose-built for AI agents, which began shipping at scale in early 2026 under the direct stewardship of Vice President Ian Buck. Vera targets a critical gap: existing CPUs were designed for general-purpose computing, not the memory-intensive, low-latency demands of agentic inference at trillion-parameter scale. Early adopters at major cloud providers and research labs report 2.8x improvements in tokens-per-watt compared to CPU baselines, with cost-per-token economics that make continuous agent operation economically viable. Vera's architecture prioritizes coherent memory access and reduced latency for orchestrating distributed GPU inference—a specialized niche AMD and Intel have not meaningfully addressed. Initial deployments focus on autonomous reasoning workloads and long-context retrieval, with production volumes ramping through Q2 2026. However, pricing remains undisclosed, and Vera's addressable market is narrower than GPU demand, positioning it as a specialized complement rather than a mainstream processor.
The competitive implications are significant. AMD's MI300X, while competitive on raw compute density, lacks an equivalent interconnect strategy for heterogeneous systems; Intel's Gaudi accelerators similarly remain siloed without an ecosystem play. NVIDIA's NVLink opening to third-party XPUs means customers can cherry-pick optimal silicon for token generation, memory serving, and routing—all unified under NVIDIA's fabric. Industry analysts estimate this unlocks a 15-20% efficiency gain in multi-node deployments by reducing memory copies and synchronization overhead. Near-term milestones include third-party XPU shipments (with partners Cerebras and Groq expected to announce NVLink-compatible versions by Q3 2026) and enterprise customer announcements from hyperscalers running heterogeneous stacks. The bet is that NVIDIA controls the nervous system of AI infrastructure—and in a heterogeneous world, that's more valuable than owning every component.