NVIDIA (NASDAQ:NVDA) executives used a question-and-answer session at the company’s event to expand on themes from the prior day’s presentation, with CEO Jensen Huang emphasizing what he described as a third major inflection point in AI: agentic systems. Huang argued that agentic AI—systems that can be given goals and can autonomously perform tasks—changes how customers should think about computing, moving from computers as tools to computers as “manufacturing equipment” that produces tokens.
Agentic systems and “token manufacturing”
Huang said agentic systems are increasingly used for tasks such as software development and suggested token consumption is becoming a standard input to knowledge work. He described a shift where engineers are no longer only given laptops, but also “tokens,” framing token budgets as a practical requirement for modern development teams. In his view, token production efficiency—measured in tokens per second and tokens per second per watt—becomes central to customer economics because customers are effectively buying systems to produce tokens that are monetized.
Updated demand visibility: “$1 trillion+” through 2027 for Blackwell and Rubin
Huang addressed prior comments about demand visibility by updating an earlier figure he said he provided a year ago. He said that while the prior visibility had been “$500 billion through 2026” for Blackwell and Rubin, the company now has “strong confidence and visibility of $1 trillion+” in demand, forecasts, and purchase orders for Blackwell plus Rubin through the end of 2027.
He repeatedly stressed what the figure does and does not include:
- It includes only Blackwell and Rubin.
- It does not include Rubin Ultra, Feynman, Vera standalone CPUs, “Groq,” or other products referenced elsewhere in the event.
- He said the figure is expected to grow “by definition,” as NVIDIA expects to close, book, and ship additional business between now and the end of 2027.
Huang also argued NVIDIA can book and ship incremental business quickly because it builds inventory and has a supply pipeline designed to support customers who suddenly need additional compute.
Inference economics, customer mix, and the 60/40 framing
Huang reiterated his “tokenomics” framing, saying the price of a system and the cost of tokens are only marginally related. He argued that higher-priced systems can still produce the lowest-cost tokens if performance and efficiency gains are large enough, which he tied to NVIDIA’s ability to “secure” gross margins by delivering higher token output per second and per watt each generation.
He said 2025 was NVIDIA’s “year of inference,” describing a progression from training to post-training to inference, and said the company expanded model coverage on its platform, citing additions such as Anthropic and Meta AI as “net new” to NVIDIA’s platform. He also said open-source models became a major share of inference usage, describing open models in aggregate as the second-most popular category by tokens generated, behind OpenAI.
On customer mix, Huang discussed a chart he characterized as roughly 60% hyperscalers and 40% non-hyperscaler demand, saying the right-hand portion includes regional clouds, industrial deployments, and enterprise on-premises. He said that without NVIDIA’s “full stack” and end-to-end platform approach, the 40% portion is “impossible” to serve because those customers “don’t buy chips” and instead “buy platforms.” He also emphasized that NVIDIA’s relationship with hyperscalers includes bringing customers to cloud service providers because CUDA developers tend to deploy on major CSP platforms.
Physical AI could shift the mix over time
In response to a question about how the mix could evolve, Huang said both sides of the 60/40 split may grow at similar rates “until the physical AI inflection happens” in a few years. He suggested physical AI workloads in industrial settings will require on-premises and edge deployments, and said he hopes the non-hyperscale portion could eventually become larger because industries tied to physical AI are significantly larger than purely digital AI markets.
Product timing: Rubin and “Groq,” plus networking transitions
NVIDIA executives also addressed product timing. Colette Kress said “Groq” is expected to ship in Q3 of the current year and that Vera Rubin is expected to ship before Groq, adding that Rubin is already in production. Huang described “Groq” as an extreme low-latency architecture compared with high-throughput systems, saying NVIDIA plans to fuse Groq with Vera Rubin and use it for the final stage of autoregressive language models, which he characterized as bandwidth-intensive.
When asked about incremental opportunity beyond the $1 trillion+ figure, Huang said adding Groq to 25% of workloads could increase compute spend by roughly 25% and described this as not included in the $1 trillion+ Blackwell-plus-Rubin figure. He also pointed to storage as a significant spending category and suggested CPUs for “tool use” would be a smaller add-on, offering a rough example that CPUs could represent another ~5%.
On NVLink scaling and connectivity, Huang said NVIDIA should scale with copper “as far as we can” but acknowledged distance limits. He said the next-generation “Ultra” platform will have two options—copper or copper plus CPO—while an NVL1152 configuration would be all CPO. He added that copper use would still remain substantial for Ethernet scale-out within racks and storage connections, even as scale-up transitions to CPO.
Separately, Kress said NVIDIA’s near-term cash priorities include funding growth and supply chain needs, including potential prepays and supplier capacity support, along with ecosystem investments. She added that after meeting existing commitments in the first half of the year, the company expects opportunities for stock repurchases and shareholder returns, and indicated NVIDIA would start with stock repurchases and dividends together at about 50% of free cash flow, with potential to do more depending on results.
About NVIDIA (NASDAQ:NVDA)
NVIDIA Corporation, founded in 1993 and headquartered in Santa Clara, California, is a global technology company that designs and develops graphics processing units (GPUs) and system-on-chip (SoC) technologies. Co-founded by Jensen Huang, who serves as president and chief executive officer, along with Chris Malachowsky and Curtis Priem, NVIDIA has grown from a graphics-focused chipmaker into a broad provider of accelerated computing hardware and software for multiple industries.
The company’s product portfolio spans discrete GPUs for gaming and professional visualization (marketed under the GeForce and NVIDIA RTX lines), high-performance data center accelerators used for AI training and inference (including widely adopted platforms such as the A100 and H100 series), and Tegra SoCs for automotive and edge applications.
