Nvidia narrative shifts

Nvidia is selling Wall Street on a new artificial intelligence math: the biggest money may come after training, when models answer billions of user prompts. (forbes.com) At Nvidia’s GTC conference in San Jose on March 16, Chief Executive Jensen Huang said he sees $1 trillion in orders for Blackwell and Vera Rubin systems through 2027. CNBC reported that the figure doubled Nvidia’s prior view of a $500 billion opportunity tied to Blackwell and Rubin. (cnbc.com) The emphasis inside that forecast has shifted. Forbes reported that Huang framed inference — the step where a trained model generates an answer — as the new center point, while Yahoo Finance said analysts now see inference, not just training, driving current product roadmaps. (forbes.com, finance.yahoo.com) That change is visible in Nvidia’s own product pitches. Nvidia said in February that Blackwell deployments were cutting cost per token by as much as 10x for inference providers, and the company later highlighted lower-latency, lower-cost “token economics” for Blackwell Ultra in agentic artificial intelligence workloads. (blogs.nvidia.com, blogs.nvidia.com) Training is the expensive classroom phase for an artificial intelligence model; inference is the test day, when the model must respond quickly and cheaply for every user. Nvidia’s recent messaging has focused on that second phase because chatbots, coding tools and software agents create demand every time someone asks for an answer. (finance.yahoo.com, blogs.nvidia.com) The roadmap reflects that workload shift. Nvidia said its GB300 NVL72 rack system set new MLPerf reasoning-inference records, and company materials for Vera Rubin promise another step up in performance for large-scale inference. (blogs.nvidia.com, blogs.nvidia.com) The new story also leaves more room for custom chips. Nvidia’s NVLink Fusion program, announced in May 2025, lets partners build semi-custom artificial intelligence systems around Nvidia’s interconnect, and Nvidia said on April 1 that Marvell would join that ecosystem with custom XPUs and networking gear. (nvidianews.nvidia.com, nvidianews.nvidia.com) That is a notable adjustment for a company long defined by the graphics processing unit, or GPU. Instead of insisting every artificial intelligence job run on a standard Nvidia chip, the company is opening more of the system — networking, rack design, software and interconnect — to customers that want specialized silicon for predictable inference loads. (nvidianews.nvidia.com, trendforce.com) Nvidia still has the financial cushion to make that pivot from strength. The company reported fiscal 2026 revenue of $215.9 billion on Nov. 19, 2025, with fourth-quarter revenue of $68.1 billion, giving it room to invest in Blackwell, Vera Rubin and the broader systems business at the same time. (investor.nvidia.com) The pitch to investors is no longer just that bigger models need more training clusters. It is that the next wave of spending could come from running those models, at lower cost and higher volume, across entire data centers built around Nvidia’s hardware and links. (forbes.com, cnbc.com)