Share with your CTO
NVIDIA is no longer just selling chips. The Sands Capital Technology Innovators Fund Q2 2026 letter frames NVIDIA’s platform shift as a systems-level repositioning, with Q1 FY2027 revenue hitting $82 billion, up 85% year over year, and EPS growing 131%. The growth vector is no longer just GPU training clusters. NVIDIA is generating meaningful standalone CPU revenue and building across inference, agentic workflows, and enterprise deployments, a profile that looks less like a component supplier and more like the infrastructure layer enterprises can’t route around.
What this means for your business
If your AI infrastructure strategy still treats NVIDIA as a GPU vendor you buy from and Cisco or Dell as the systems integrators who stitch everything together, that mental model is now working against you. NVIDIA is collapsing the stack, moving from discrete accelerators into networking, CPUs, and software orchestration. Organizations that are mid-cycle on data center planning need to reckon with a vendor whose ambitions now overlap with nearly every other infrastructure relationship they hold.
The shift from training to inference and agentic workloads is the part worth pressure-testing. Training clusters are large, concentrated, and largely the domain of hyperscalers. Inference is distributed, latency-sensitive, and lives closer to enterprise applications. As AI moves from experimental to operational, the purchasing decision moves from centralized CapEx toward workload-level architecture choices that CTOs and platform engineers own directly. NVIDIA’s move into CPUs and systems software is a bid to be present at exactly that decision point, where enterprises are choosing how to serve models, not just how to train them.
Sands Capital, writing to investors in a fund positioned to benefit from NVIDIA’s continued ascent, unsurprisingly shades toward an optimistic read on the CPU revenue signal, treating early visibility as structural confirmation rather than early experiment. That tilt is worth noting, but it doesn’t make the thesis wrong. The falsification condition here is specific: if enterprise inference workloads get absorbed by hyperscaler managed services rather than on-premise or colocation deployments, NVIDIA’s systems-level expansion stalls at the data center door and the GPU moat holds while the platform story deflates.
Concept deep-dive: Inference infrastructure
Training is how an AI model learns from data, a one-time (or periodic) compute-intensive process. Inference is everything that happens after, running the trained model in production to answer a question, generate an output, or take an action. Inference is where enterprise AI actually lives, and it runs continuously at scale. Because inference traffic is distributed across applications and users rather than batched in a cluster, the infrastructure choices are different, and the vendor who owns that layer owns the ongoing operational relationship.
Based on reporting from NVIDIA (NVDA) Is Evolving Beyond GPUs Into a Full AI Infrastructure Platform, originally published 2026-08-07 11:26:00.
