JPMorgan maintained its Overweight rating on Nvidia after the company’s Sept. 2 investor meeting and kept its $320 price target, based on about 20x CY2026 expected earnings per share of $15.87. The discussion featured Toshiya Hari, Nvidia’s vice president of investor relations. In a note released after the meeting, JPMorgan said Nvidia is comfortable with a framework that calls for 70% year-over-year growth in FY28, and said that view reflects supply constraints rather than weak demand.
Growth outlook tied to broad demand, not one driver
According to the meeting note, Nvidia’s confidence in FY28 growth comes from expansion across the market rather than from any single customer group. Hari said the 70% year-over-year framework is built on broad improvement in demand from hyperscalers, neo-cloud providers, AI labs, sovereign AI projects and enterprise on-premise deployments.
He also said one reason Nvidia chose to provide annual guidance was the significant gap between internal expectations and market consensus. Leaving that gap in place could make planning harder for supply-chain partners, he said.
Management repeatedly described 70% as a number it is “comfortable” with. Under the current framework, Nvidia still sees the business as supply-constrained, not demand-constrained. Hari said growth could have been more than double without supply limits.
Inference now larger than training in data center mix
The split between training and inference revenue came up several times during the meeting. Hari said Nvidia’s platform is highly interchangeable: customers can use Grace Blackwell products for training and then redirect the same assets to inference workloads, making a precise revenue split difficult.
Even so, he gave a directional update. About 18 months ago, the mix between inference and training was roughly 50/50. He now said he is confident that inference has become the larger part of Nvidia’s data center business and is still gaining share. Over time, inference should account for a bigger portion of total revenue, he said.
Advanced wafers and memory flagged as key bottlenecks
When asked about meeting next year’s demand and the supply limits that remain in place, Hari singled out memory and advanced wafers as two critical parts of Nvidia’s bill of materials.
Nvidia is working closely with Taiwan Semiconductor Manufacturing Co. and the three major memory suppliers — Micron, SK Hynix and Samsung — in discussions aimed at improving supply availability, he said.
Hari did not identify CoWoS advanced packaging as the main bottleneck at this point. The note said that is worth watching. It may suggest that the tightest pressure point has shifted upstream toward wafers and memory, though it does not mean CoWoS is no longer constrained.
Customer mix broadens as neo-cloud providers pass 50%
The note said end-demand concentration among AI model developers is declining. Hari said OpenAI and Anthropic currently account for about 20% of Nvidia’s end demand, and that share could approach 25% in FY28.
That end demand does not map directly onto Nvidia’s reported customer base. Nvidia sells compute capacity to hyperscalers and neo-cloud providers, which then resell capacity to model developers.
Hari said neo-cloud providers are already a major part of Nvidia’s business and now represent more than 50% of AI compute infrastructure. In JPMorgan’s reading, that points to growth driven by a broader customer ecosystem rather than only by the largest hyperscalers.
Open and closed models both remain in use
On the debate over open-source and closed-source large language models, Hari echoed Jensen Huang’s earlier view that AI progress requires both rather than a binary choice.
He said Nvidia internally uses closed models such as OpenAI and Claude, while combining open and closed approaches for critical tasks including chip design. As long as model developers keep improving the economics of their products, demand should flow through to chip suppliers, he said.
Hari added that model developers’ gross margins are improving, with Nvidia’s continuing decline in cost per token across generations serving as an important support.
Financing structures described as demand support, not circular funding
Hari also detailed several financing arrangements, including revenue-sharing agreements with some neo-cloud providers, the PORTS-Pike data center campus plan, and a $500 billion private capital financing platform involving Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.
Under the revenue-sharing model, Nvidia helps guarantee a base price and shares upside when compute is rented out above market rates. The note said that gives the company an option on recurring revenue on top of core hardware sales.
Management said these financing structures are limited and capped, and are supported by strong end demand, ecosystem returns and the credit quality of the ultimate buyers of compute capacity. JPMorgan’s summary said the arrangements are aimed at real demand rather than circular financial activity.
JPMorgan leaves target unchanged
Based on the meeting, JPMorgan kept its Overweight rating on Nvidia and reiterated its $320 price target, based on about 20x CY2026 expected EPS of $15.87. The note’s central view is that Nvidia’s FY28 70% growth framework represents a conservative floor under supply constraints, while inference has become the largest and still expanding part of the data center business and the customer base continues to broaden.
This article is a summary and interpretation of JPMorgan’s Sept. 2, 2026 research note. Any ratings, price targets, earnings forecasts and related judgments cited here are the views of the broker’s analysts and represent the institution’s position, not investment advice.
