Nvidia used the Hot Chips 2026 conference to lay out the technical details of Vera, the custom CPU that anchors its next-generation Vera Rubin systems.
According to Tom's Hardware, Nvidia broke down an 88-core design aimed squarely at what the company calls "agentic" data centers — the AI workloads where software agents run long, multi-step tasks rather than answering a single prompt. The presentation benchmarked spatial multithreading, an approach to squeezing more useful work out of the chip's cores, and detailed a memory system built on LPDDR5X with SOCAMM2 modules delivering 1.2 TB/s of bandwidth. Tom's Hardware reports Nvidia emphasized the power benefits of that memory choice, a notable framing at a moment when data center electricity consumption is a live constraint on AI expansion.
The commercial stakes are already visible. Yahoo Finance reports that an India-based AI data center firm has ordered 9,000 Nvidia Vera Rubin systems — an order placed for hardware whose architecture Nvidia is only now detailing publicly.
That enthusiasm arrives alongside harder questions. Yahoo Finance also reports that Nvidia faces a growth test as the Rubin debut coincides with increased scrutiny of how AI infrastructure is being financed — the loans, leases, and circular vendor arrangements underwriting the buildout.
What's worth understanding here is the shift in what Nvidia sells. For years the story was GPUs; Vera signals that Nvidia intends to supply the CPU, the memory subsystem, and the interconnect as one package, capturing more of each data center dollar and controlling more of the design.
It matters because the chips underneath AI agents are quietly being redesigned around them — and the size of orders like the 9,000-system deal will determine whether that bet, and the debt financing it, pays off.