NVIDIA (NASDAQ: NVDA) appears to be harboring far more aggressive growth expectations for the coming years than the market currently anticipates.

In a recent investor meeting, NVIDIA's management explicitly stated that the FY28 70% year-over-year growth framework does not represent an upper limit from the demand side. If not constrained by supply, the company's growth rate could have exceeded 100%. This signals a pivotal shift: the primary constraint on NVIDIA's expansion is no longer about sustaining demand but rather about scaling production capacity to match it.

Adding to the narrative, AI demand itself is expanding rapidly, and its composition is evolving. Roughly 18 months ago, NVIDIA's revenue was evenly split between training and inference workloads. Today, inference revenue has overtaken training and is projected to continue its ascent. Concurrently, emerging cloud service providers now contribute over 50% of AI compute infrastructure revenue, diversifying the growth engine away from simply the traditional hyperscale cloud vendors and toward a broader AI compute ecosystem.

The supply side emerges as the critical determinant for NVIDIA's ability to realize these higher growth figures. Management highlighted advanced wafers and memory as the two most significant bottlenecks, and they are in continuous coordination with TSMC, Micron, SK Hynix, and Samsung to increase capacity. With HBM still in tight supply, any improvement in these key component supplies could release previously suppressed order demand for NVIDIA.

Shifts in the customer base are also mitigating concerns about an over-reliance on a small number of major clients. OpenAI and Anthropic currently represent approximately 20% of NVIDIA's business on an end-consumer basis, a figure that could rise to around 25% by FY28. Meanwhile, the contribution from emerging cloud providers to AI compute infrastructure revenue has already surpassed the 50% mark. With demand expansion, a rising inference share, and a more diversified customer set, NVIDIA's growth story is transitioning from a simple "training compute cycle" to a broader "AI infrastructure cycle".

Why the 70% FY28 Growth Target Isn't the Ceiling: Supply is the Ultimate Constraint

The report clarifies that NVIDIA's VP of Investor Relations and Strategic Finance, Toshiya Hari, stated the FY28 70% year-over-year growth framework is not driven by a single customer or business line. Instead, it's fueled by a confluence of demand from hyperscalers, emerging cloud providers, AI labs, sovereign AI initiatives, and enterprise on-premise deployments.

The early disclosure of this multi-year growth framework stems from a notable discrepancy between market consensus expectations and the company's internal projections. If this gap persists, it could impact the capacity planning of supply chain partners. More striking, however, is Hari's direct admission that without supply constraints, NVIDIA's growth rate could have surpassed 100%.

In essence, the 70% target seems to be a growth floor the company is willing to publicly confirm under current supply conditions, rather than the true limit of demand. As production capacity expands, NVIDIA's actual growth potential could be significantly higher than this figure.

Inference Revenue Surpasses Training: AI Demand Moves from "Buying Chips" to "Continuous Operation"

The revenue mix between training and inference was a key topic at the meeting. Due to the workload-switchable nature of NVIDIA's GPUs, such as Grace Blackwell which can rapidly pivot between model training and inference, precise separation of these two revenue streams is challenging.

However, management provided a significant benchmark: approximately 18 months ago, training and inference revenues were roughly equal. Now, inference revenue has overtaken training, and this gap is expected to widen further. This indicates a structural shift in NVIDIA's demand. While training remains a critical driver of AI infrastructure expansion, inference is becoming a more sustained source of compute demand as model sizes grow and AI applications become more prevalent, potentially enhancing the durability of NVIDIA's revenue.

Advanced Wafers and HBM Remain the Two Major Supply Bottlenecks

Despite robust demand, NVIDIA's biggest challenge remains its supply chain. Hari specifically pointed to advanced wafers, which rely heavily on TSMC, and memory, including HBM from Micron, SK Hynix, and Samsung, as the key constraints for meeting next year's demand.

Continuous discussions are underway with TSMC and the three major memory suppliers to boost production of these critical components. This underscores that HBM capacity expansion will be a vital prerequisite for NVIDIA to unlock its full growth potential. For HBM suppliers like Micron and SK Hynix, the demand visibility from NVIDIA remains exceptionally high.

Customer Base Shifts: Emerging Cloud Providers Now Account for Over Half of Revenue

The customer landscape is also in flux. Hari revealed that OpenAI and Anthropic represent about 20% of NVIDIA's business in end-consumer terms, a share that might increase to roughly 25% by FY28. However, since these two companies primarily source compute power through cloud and emerging cloud service providers, this figure does not equate to NVIDIA's direct customer revenue concentration.

What is truly noteworthy is that emerging cloud service providers now contribute over 50% of NVIDIA's AI compute infrastructure revenue (ACIE). This indicates that NVIDIA's growth drivers are broadening from a few hyperscalers to include emerging cloud providers, model companies, and enterprise clients. The continuous dispersal of customers and compute demand also mitigates the risks associated with high single-customer concentration.

Open Source vs. Closed Source is Not a Zero-Sum Game; Lower Model Costs Stimulate Compute Demand

On the topic of open-source versus closed-source models, NVIDIA's management reaffirmed its stance: both models will coexist long-term, and the advancement of AI technology does not herald the replacement of one by the other. NVIDIA itself utilizes closed-source models like OpenAI's and Anthropic's, and employs a combination of open and closed-source models in critical workloads such as chip design.

Management also pointed out that the gross margins of model builders are improving. As NVIDIA's GPUs iterate and the compute cost per token declines, the economics for model companies improve. This could establish a new demand cycle: lower compute costs lead to better profitability for model firms, which accelerates AI application deployment, thereby boosting inference demand, and in turn, driving further compute procurement.

Addressing Financing Model Concerns: NVIDIA Responds to "Circular Financing" Skeptics

Regarding market concerns over financing, management outlined three primary arrangements: revenue-sharing agreements with certain emerging cloud providers, the PORTS-Pike data center campus plan, and a $500 billion private capital financing platform involving institutions like Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR.

The core mechanism of the revenue-sharing agreements sees NVIDIA help secure a baseline compute price, allowing the company to share in the upside when rental prices exceed agreed levels. This provides NVIDIA with an avenue for recurring revenue from compute infrastructure operations, beyond just direct hardware sales.

In response to concerns about "circular financing," management stated that the scale and caps of these financing arrangements are controlled, grounded in strong end-user demand, ecosystem returns, and the creditworthiness of the ultimate compute purchasers.

Based on the information from this meeting, the real challenge for NVIDIA is no longer "is there demand," but rather "can they convert demand into supply quickly." If the wafer and HBM bottlenecks continue to ease, the previously stated 70% FY28 growth framework may indeed have significant room for upward revision.