At a recent investor conference, NVIDIA management stated that the 70% growth target for fiscal year 2028 is not a ceiling; if supply were unconstrained, the growth rate could exceed 100%. The core constraint has shifted from demand to supply, with advanced wafers and HBM (High Bandwidth Memory) being the primary bottlenecks. Addressing market concerns regarding "circular financing," management clarified that the scale and limits of such arrangements are strictly controlled. The foundation remains actual end-user demand and the creditworthiness of purchasers, rather than unrestricted capital circulation.
$NVIDIA (NVDA.US)$ Growth expectations for the coming years may be far more aggressive than the market imagines.
In its latest report, JPMorgan noted that during recent investor meetings, NVIDIA’s management clearly stated that the framework of 70% year-on-year growth for fiscal year 2028 does not represent a ceiling on demand; absent supply constraints, business growth could have exceeded 100%. This indicates that the core constraint on NVIDIA’s growth has shifted from “whether demand can be sustained” to “whether production capacity can keep pace.”
Of greater note, AI demand itself continues to expand rapidly, and the structure of this demand is evolving. Approximately 18 months ago, NVIDIA's revenue from training and inference was roughly split evenly; today, inference revenue has surpassed training revenue and is expected to continue increasing its share. Meanwhile, revenue from AI computing infrastructure contributed by emerging cloud service providers has exceeded 50%, indicating that growth momentum is diffusing from traditional hyperscale cloud vendors to a broader AI computing ecosystem.
On the supply side, it has become the key determinant of whether NVIDIA can deliver higher growth. Management identified advanced wafers and memory as the two most critical bottlenecks and continues to coordinate with Taiwan Semiconductor, Micron, SK Hynix, and Samsung to expand supply. Particularly against the backdrop of continued tightness in High Bandwidth Memory (HBM), an improvement in the supply of key components is expected to further release order demand previously constrained by production capacity.
Changes in the customer structure are also alleviating market concerns about NVIDIA's over-reliance on a few major clients for growth. OpenAI and Anthropic currently account for approximately 20% of NVIDIA's business on an end-consumer basis, a figure that may rise to around 25% in FY28; however, at the same time, the share of emerging cloud service providers in AI computing infrastructure revenue has already exceeded 50%.
Amid continuous demand expansion, a rising share of inference workloads, and further customer diversification, NVIDIA's growth narrative is shifting from a pure "training compute cycle" to a broader AI infrastructure cycle.
70% growth in FY28 is not the ceiling; supply is the primary constraint.
The report notes that Toshiya Hari, Vice President of Investor Relations and Strategic Finance at NVIDIA, stated that the company's framework of 70% year-on-year growth in FY28 is not driven by a single client or business segment, but rather by combined demand from hyperscale cloud vendors, emerging cloud service providers, AI labs, sovereign AI initiatives, and enterprise and on-premises deployments.
A key reason for management providing a multi-year growth framework in advance is the significant gap between market consensus expectations and the company's internal assessments. If this expectation gap persists, it could impact the capacity planning of supply chain partners.
More notably, Hari explicitly stated that if there were no supply constraints, NVIDIA's business growth rate could have exceeded 100%.
In other words, the 70% growth target is more akin to the minimum growth level NVIDIA is willing to publicly affirm under current supply conditions, rather than the true ceiling on the demand side. As production capacity is further unleashed, NVIDIA's actual growth potential may significantly exceed this figure.
Inference revenue has surpassed training revenue, signaling a shift in AI demand from hardware acquisition to sustained operational usage.
The revenue structure of training versus inference was also a key topic of discussion at the conference.
Given that NVIDIA GPUs possess strong workload switching capabilities—for instance, Grace Blackwell can be used for model training and quickly transitioned to inference—it is difficult for the company to precisely segment these two business lines.
However, management highlighted a significant milestone: approximately 18 months ago, training and inference revenues each accounted for roughly 50%; currently, inference revenue has exceeded training revenue, and this gap is expected to widen further.
This indicates a structural shift in NVIDIA's demand profile. While training remains a major driver of AI infrastructure expansion, inference is becoming a more sustained source of computing power demand as model sizes grow and AI applications proliferate, which is also poised to enhance the sustainability of NVIDIA's revenue.
Advanced wafers and HBM remain the two primary supply-side bottlenecks.
Amid robust demand, NVIDIA's greatest challenge continues to stem from its supply chain.
Hari specifically pointed out that the key constraints on meeting next year's demand are concentrated primarily in advanced wafers and memory. Advanced wafers rely mainly on Taiwan Semiconductor, while high-bandwidth memory (HBM) involves Micron, SK Hynix, and Samsung.
NVIDIA is maintaining ongoing communications with Taiwan Semiconductor and the three major memory suppliers, with a focus on enhancing the supply capacity of critical components.
This also implies that as demand for AI continues to grow, the expansion of HBM production capacity will remain a crucial prerequisite for NVIDIA to unlock its growth potential. For HBM suppliers such as Micron and SK Hynix, visibility into demand from NVIDIA remains high.
Customers are no longer heavily reliant on hyperscale cloud providers, with contributions from emerging cloud providers exceeding half of the total.
NVIDIA's customer structure is also undergoing changes.
Hari revealed that OpenAI and Anthropic currently account for approximately 20% of NVIDIA's business on an end-consumer basis, a proportion that could rise to around 25% by FY28. However, since these two companies primarily access computing power through cloud providers and emerging cloud service providers, this figure does not equate to their share of NVIDIA's direct customer revenue.
What truly warrants attention is that emerging cloud service providers currently contribute more than 50% of AI Computing Infrastructure Equipment (ACIE) revenue.
This indicates that NVIDIA's sources of growth are expanding beyond a few hyperscale cloud providers to include emerging cloud service providers, model companies, and enterprise customers. The continued diversification of customers and computing demand has also helped mitigate risks associated with concentration in single customers to some extent.
Open-source and closed-source models are not mutually exclusive; declining model costs are instead stimulating demand for computing power.
Regarding the development of open-source and closed-source models, NVIDIA's management maintains its previous assessment: both models will coexist in the long term, and advancements in AI technology do not imply that one will replace the other.
NVIDIA itself utilizes closed-source models such as OpenAI and Claude, while adopting a combination of open-source and closed-source models for critical workloads like chip design.
Meanwhile, management noted that gross margins for model builders are improving. As NVIDIA GPUs continue to iterate, the computational cost per token is steadily decreasing, thereby enhancing the economic viability for model companies.
This could create a new demand cycle: declining computing power costs → improved profitability for model vendors → accelerated deployment of AI applications → growing inference demand → further driving computing power procurement.
Is the financing model self-consistent? NVIDIA responds directly to concerns about 'circular financing'.
Addressing market concerns regarding financing, management outlined three primary arrangements: revenue-sharing agreements with select emerging cloud service providers, the PORTS-Pike data center campus plan, and a $500 billion private capital financing platform involving institutions such as Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR.
The core mechanism of the revenue-sharing agreements is as follows: NVIDIA helps lock in base computing power prices, and when rental prices exceed agreed-upon levels, the company shares in part of the upside. Thus, beyond direct hardware sales, NVIDIA also has the opportunity to generate recurring income from the operation of computing power infrastructure.
In response to market concerns about 'circular financing,' management stated that the scale and caps of related financing arrangements are controlled, underpinned by strong end-user demand, ecosystem returns, and the creditworthiness of final computing power purchasers.
Based on information released during this meeting, the key challenge for NVIDIA is no longer whether there is demand, but rather how quickly it can convert demand into supply. If bottlenecks such as wafer and HBM availability continue to ease, the previously projected 70% growth framework for FY28 may indeed have significant room for upward revision.
Editor/lambor