Nvidia 70% Growth Forecast: Jensen Huang Says Supply, Not Demand, Sets the Limit

Jensen Huang says the Nvidia 70% growth forecast reflects a supply ceiling, not the limit of customer demand. Nvidia expects revenue to increase about 70% in fiscal 2028, an unusual year-ahead projection from a company navigating volatile AI spending.

The forecast signals confidence that Nvidia can keep growing despite custom chips from Amazon, Google, Microsoft, OpenAI, and Anthropic. It also raises a harder question. How much demand exists independently, and how much depends on Nvidia helping customers finance infrastructure built around its hardware?

Huang’s answer is that Nvidia can see demand across the entire AI supply chain. The company works with chip suppliers, cloud operators, model developers, enterprises, governments, and data center builders. Critics see many of those connections differently. They worry that Nvidia is increasingly financing or supporting the customers that buy its systems.

That conflict matters more than the headline percentage. Nvidia is no longer selling accelerators into an infrastructure market built entirely by other companies. It is helping secure manufacturing capacity, data center sites, financing, and cloud revenue commitments.

The result is a business with unusual visibility into future construction. It is also a business assuming new risks beyond semiconductor design.

The Nvidia 70% Growth Forecast Is a Supply-Constrained Promise

Nvidia’s forecast says the company expects demand to exceed what its manufacturing and infrastructure network can deliver.

Nvidia first presented the projection during its fiscal 2027 second-quarter earnings call on August 26. Chief Financial Officer Colette Kress said the company expected approximately 70% revenue growth in fiscal 2028.

The number was notable because Nvidia had not typically guided investors across an entire future fiscal year. Its guidance usually focused on the next quarter, where supply schedules and customer orders were easier to assess.

Huang later repeated the forecast during Goldman Sachs’ Communacopia and Technology Conference. His conference remarks framed the projection as a product of visibility across Nvidia’s market.

Nvidia reported approximately $96 billion in quarterly revenue for the period ending July 26, 2026. Data center revenue reached $89 billion, rising 18% from the preceding quarter.

The company also projected about $108 billion in revenue for its fiscal third quarter. These figures show that the 70% forecast starts from a much larger base than Nvidia faced earlier in the generative AI expansion.

Kress said customer forecasts pointed toward demand that could double during the following year. Nvidia still limited its revenue forecast to 70% because available supply could not support all those requests.

Huang reinforced that distinction during the earnings call. He said demand was substantially greater than the amount Nvidia expected to deliver. The company’s secured supply gave management confidence in the lower figure.

This is not simply a claim that more companies will order individual chips. Nvidia now describes its product at the scale of an entire computing system.

A modern Nvidia installation can include CPUs, GPUs, networking equipment, switches, cooling components, and extensive software. These parts operate as one data center-scale computer.

Huang used Nvidia’s GB200 NVL72 system to illustrate that shift. It combines 36 Grace CPUs and 72 Blackwell GPUs through Nvidia’s NVLink interconnect.

He said orders for the system were growing 27% month over month when he spoke in September. That figure represented current order momentum, not a guarantee that the pace would continue indefinitely.

Nvidia calls these installations AI factories because they turn electricity and data into model training and inference output. Inference is the process of using a trained model to generate answers, predictions, or other results.

The larger unit of sale changes how investors should interpret the forecast. Nvidia is not counting only chips shipped into existing data centers. It is planning around complete facilities that require power, cooling, land, financing, and construction.

This makes the forecast partly an industrial deployment schedule. Every revenue target depends on suppliers and customers completing a long chain of work on time.

That chain creates greater visibility because major projects require multiyear planning. It also creates more opportunities for delays.

AI Agents Are Driving a Much Larger Compute Requirement

Huang’s central demand argument is that AI agents perform longer tasks, making each user request far more computationally expensive.

A basic chatbot often responds to one prompt with one answer. An AI agent can plan a task, call software tools, inspect results, correct errors, and repeat that cycle.

Each step consumes inference capacity. The workload can grow quickly when an agent handles research, software development, customer service, scientific analysis, or business operations.

During Nvidia’s earnings transcript, Huang estimated that an agent can require 15 to 100 times more compute than direct human use. He said the amount varies with the problem being solved.

That estimate is a company claim, not an independent benchmark covering every application. However, the underlying mechanism is clear. Longer reasoning sequences and repeated tool calls require more generated tokens and more accelerator time.

Model development is also expanding beyond a small group of frontier laboratories. Nvidia says demand now comes from hyperscalers, neoclouds, AI startups, enterprises, industrial companies, research institutions, and sovereign customers.

Hyperscalers are the largest cloud providers, including Amazon Web Services, Microsoft Azure, and Google Cloud. Neoclouds are newer infrastructure companies built primarily around GPU computing.

In Nvidia’s reported second quarter, hyperscale revenue reached $49 billion. Another customer category, covering neoclouds, industries, and enterprises, generated $40 billion.

That second category grew 25% from the preceding quarter and 138% from the prior year, according to Nvidia. Its scale supports Huang’s argument that the market is broader than one AI laboratory or one cloud customer.

Kress said neocloud partners were expected to finish 2026 with eight gigawatts of installed capacity. They had approximately three gigawatts at the end of 2025.

A gigawatt measures power, not computing performance. Still, it offers a useful indicator of how large these projects have become. A new data center campus can require energy comparable to a major industrial development.

Amazon provides another demand signal. Nvidia said AWS would begin deploying two million additional Nvidia GPUs during the reported quarter. The deployment schedule runs through Nvidia’s fiscal second quarter of 2029.

AWS also plans to use Nvidia technology across models, cloud services, and warehouse robotics. Yet Amazon continues developing its own Trainium accelerators, creating both partnership and competition.

Google follows a similar strategy with its Tensor Processing Units. Microsoft has Maia accelerators, while OpenAI and Anthropic are pursuing custom hardware programs.

These projects pressure Nvidia because its largest customers want greater control over costs and supply. They do not automatically eliminate demand for Nvidia systems.

Custom accelerators work best when a company can optimize hardware, software, and workloads together. Nvidia’s advantage is broader compatibility across customers, models, clouds, and development tools.

Huang says Nvidia runs models from major closed laboratories and open-model developers. That reach lets a customer deploy similar software across several environments without rebuilding everything for one specialized chip.

CUDA, Nvidia’s programming platform for GPU computing, remains central to that position. Developers have spent years creating libraries, tools, and applications around it.

The Nvidia 70% growth forecast assumes this full platform remains preferable for a substantial share of new capacity. It does not require custom accelerators to fail.

Instead, the forecast assumes total demand grows fast enough for both approaches. Nvidia can lose some workloads to internally designed chips while expanding sales across a larger computing market.

The uncertainty lies in timing. Customers can overestimate near-term agent adoption, while useful applications can take longer to generate revenue than infrastructure takes to build.

If enterprises struggle to convert agent experiments into dependable workflows, cloud operators could face lower utilization. That would weaken the economics behind another round of equipment orders.

For knowledge workers, the practical signal is usage depth rather than model announcements. Systems that become embedded in software development, research, sales, and operations will consume compute repeatedly.

A searchable AI knowledge base, for example, can combine retrieval with repeated model calls. At scale, those ordinary workflows matter more than a single demonstration.

Nvidia’s Full-Stack Reach Explains Huang’s Confidence

Nvidia believes it can forecast further ahead because it now monitors both semiconductor supply and the physical infrastructure awaiting its systems.

Chip companies traditionally estimate demand from customer orders, manufacturing schedules, inventory, and channel activity. Nvidia now tracks another layer of information.

Huang said the company follows data center land, electrical power, and building shells around the world. A shell is the physical structure prepared before servers and networking equipment arrive.

That knowledge matters because constructing AI capacity takes years. A customer cannot install an advanced system without securing utility connections, cooling equipment, contractors, permits, and capital.

Nvidia has positioned itself on both sides of the deployment schedule. Upstream, it works with foundries, memory suppliers, packaging providers, and other component manufacturers.

Downstream, it coordinates with cloud operators, data center developers, energy companies, enterprises, and financial institutions. That position gives Nvidia an early view of projects before hardware is delivered.

Its spending commitments reflect the scale of that coordination. Nvidia’s quarterly filing said supply and capacity commitments increased from $119 billion to $279 billion by July 26.

Those commitments help secure production for current and future architectures. They can also become a liability if customer deployments slow or technology transitions move differently than expected.

Nvidia’s sales value per gigawatt has increased as its systems have become more integrated. Huang said Hopper represented about $18 billion of Nvidia equipment per gigawatt.

He placed Grace Blackwell at approximately $25 billion per gigawatt and Vera Rubin at about $40 billion. These are Nvidia’s estimates for system value, not universal construction costs.

The progression shows why revenue can grow faster than the number of powered data centers. Each generation can place more valuable computing and networking equipment behind the same electrical connection.

New platforms also pull Nvidia into categories once supplied by several vendors. The company sells accelerators, CPUs, networking components, interconnects, software, and complete rack-scale systems.

That expansion gives Nvidia more revenue from each deployment. It also means that Nvidia’s forecast depends on customers accepting more of its architecture.

Competitors can attack different layers of that stack. AMD sells accelerators, Broadcom helps companies develop custom chips, and cloud providers offer their own silicon.

Cerebras and other specialized developers propose alternative ways to build AI computing systems. Some target inference economics, while others emphasize memory capacity or simpler scaling.

Nvidia’s defense is fungibility, meaning a system can serve many models and customers rather than one narrowly defined workload. That flexibility is particularly important for cloud operators.

A neocloud wants to rent available capacity to whichever customer offers the best return. Hardware optimized for only one model can become difficult to repurpose when demand changes.

Nvidia argues that broad software compatibility protects the resale and rental value of its equipment. Financing institutions also care about that claim because they must evaluate long-term asset risk.

However, computing hardware depreciates faster than conventional infrastructure. A six-year cloud commitment can span several Nvidia product generations.

Customers might still prefer older systems for less demanding inference. Yet that secondary demand must remain strong enough to support the expected residual value.

The Nvidia 70% growth forecast therefore rests on two forms of visibility. The first comes from booked supply and physical projects. The second comes from Nvidia’s belief that its platform will remain economically useful after newer systems arrive.

The first can be measured through commitments and construction. The second will only become clear through utilization, rental rates, and customer renewals.

The Circular Financing Question Does Not Disappear

Nvidia’s deals can unlock genuine demand while still transferring more customer and infrastructure risk onto its own balance sheet.

The circular financing criticism is straightforward. Nvidia invests in AI companies or supports their infrastructure, and those companies use capital to purchase access to Nvidia systems.

Critics worry that this arrangement can make demand appear more independent than it is. If customers cannot finance expansion without vendor support, current sales might pull future demand forward.

Nvidia rejects that interpretation. Kress told analysts that the company views its support as infrastructure development backed by customer usage and independently supplied capital.

For neocloud projects, Nvidia can provide a take-or-pay commitment covering part of a facility’s capacity. Take-or-pay means Nvidia agrees to pay for a minimum amount, even when actual use falls short.

That floor can help a cloud operator secure project financing. Nvidia then receives a share of revenue generated above the guaranteed level.

Kress said the company gets paid through the hardware sale and can later receive rental revenue. Nvidia reported $36 billion of these cloud service commitments as of July 26.

The filing says cloud providers can stop supplying capacity to Nvidia and sell it to third parties at better rates. Nvidia’s commitment decreases as outside customers or Nvidia itself consume the capacity.

This structure differs from simply lending a customer money for a hardware purchase. Nevertheless, Nvidia bears costs if external utilization fails to develop.

Its exposure extends to physical projects. Nvidia disclosed up to $3.5 billion in guarantees related to land, power, and building obligations for selected AI cloud partners.

The company also entered guarantees connected to SB Energy’s PORTS Technology Campus in Ohio. Nvidia capped that support at an aggregate $105 billion, subject to conditions.

The planned campus covers approximately 4.25 gigawatts of IT load. Its nine data centers are expected to begin entering service during Nvidia’s fiscal 2029.

OpenAI is expected to lease the campus for 20 years, with limited exceptions. Nvidia’s exposure declines as OpenAI makes lease payments and can end following specified events.

That arrangement creates a clear connection among Nvidia, an AI laboratory, a data center developer, and future hardware demand. It is reasonable for investors to examine the entire chain.

Nvidia says the guarantees address a financing mismatch. Frontier laboratories can have fast-growing demand without the credit history required for multidecade infrastructure contracts.

The company reported investing nearly $50 billion in frontier AI laboratories. Nvidia characterizes that amount as a small portion of expected free cash flow over the same period.

In August, Nvidia also announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The financing announcement targeted more than $500 billion in third-party capital over time.

These were memorandums of understanding, not completed deployments of the full amount. Nvidia’s filing warned that preliminary arrangements might not produce definitive agreements.

The proposed platforms would have outside institutions underwrite individual projects. Nvidia can still provide limited residual-value support at its discretion.

Huang’s simplest defense came during his September conference appearance. He argued that Nvidia contributes a small amount while much more outside capital returns to the infrastructure market.

The response captures Nvidia’s thesis, but it does not settle the accounting or economic question. A financial loop does not need equal contributions from every participant to concentrate risk.

The more important question is whether end users generate enough cash to support cloud rental charges. That revenue must eventually cover electricity, facilities, financing, hardware, and operator margins.

Positive utilization can make Nvidia’s intervention look like sensible market formation. Weak utilization can expose guarantees, reduce rental prices, and leave customers with rapidly aging equipment.

Investors have already asked for greater disclosure. Before Nvidia’s August earnings, Morgan Stanley analysts requested extensive transparency about investments and financing arrangements.

Visible Alpha research leader Melissa Otto called the 70% forecast a significant increase over market expectations near 45%. Her assessment, reported alongside wider investor scrutiny, highlights why the debate matters.

A forecast so far above expectations depends on more than superior chip performance. It relies on a financing and construction system scaling without creating excessive unused capacity.

Nvidia’s own filings acknowledge that shortage and oversupply risks coexist. Insufficient power, land, capital, or components can delay revenue.

Customers can also postpone new architectures or adopt them more slowly than expected. Changing market conditions can make Nvidia’s commitments less valuable and hurt financial results.

The balanced conclusion is not that every Nvidia-supported project produces artificial demand. Nor is it that independent underwriting removes Nvidia’s risk.

The company is using its market position to solve real infrastructure bottlenecks. In doing so, it is linking its balance sheet more closely to customer utilization and credit quality.

That tradeoff is the central test behind Huang’s confidence.

Three Signals Will Test Jensen Huang’s Forecast

Investors should watch supply conversion, outside utilization, and customer concentration rather than treating the 70% figure as self-validating.

The first signal is Nvidia’s conversion of its $279 billion in supply and capacity commitments into shipped, paid-for systems. Those commitments show preparation, but revenue requires completed products and ready data centers.

Memory represents one immediate constraint. Nvidia said heavy AI demand had tightened supply and pressured costs.

The company reported a 75% gross margin for the quarter. Management expected that measure to decline as low as 71% before recovering.

If memory availability improves and new architectures enter volume production on schedule, Nvidia can deliver more of the demand it sees. Delays would weaken the supply-based explanation for the forecast.

Investors should therefore compare quarterly revenue guidance with inventory, purchase commitments, margins, and platform transition commentary. Rising commitments without proportional shipments would deserve closer attention.

The second signal is third-party utilization at neoclouds and supported data centers. This is the clearest test of the circular financing concern.

Nvidia says its cloud commitments decline when outside customers consume capacity. Strong third-party usage would move risk away from Nvidia and validate the projects’ independent economics.

Utilization must also produce sustainable rental rates. A data center can appear busy while generating insufficient returns because providers discount GPU access to attract customers.

Cloud operators should demonstrate that revenue grows alongside installed capacity. They must also show that customer demand extends beyond a few frontier laboratories and hyperscalers.

Nvidia’s reported growth across enterprises, sovereign projects, and regional clouds supports that case. Future disclosures need to show that these customers become durable users rather than announced pipelines.

If Nvidia’s take-or-pay exposure falls while partner revenue expands, the company’s market-building argument becomes stronger. Rising guarantees paired with weak outside usage would strengthen the critics’ case.

The third signal is the balance between Nvidia systems and customer-designed chips. Amazon, Google, Microsoft, OpenAI, and Anthropic have strong reasons to diversify.

Custom silicon can lower costs for stable workloads and reduce dependence on one supplier. It can also give model developers tighter control over memory, networking, and software design.

Nvidia does not need to prevent those deployments. It needs its platform to retain enough workloads while the overall market expands.

Watch whether cloud providers continue ordering Nvidia’s complete systems while deploying their own accelerators. A mixed strategy would support Huang’s claim that Nvidia serves a broad, growing market.

A meaningful shift toward proprietary hardware across both training and inference would weaken that conclusion. It could also reduce Nvidia’s pricing leverage and the value captured per gigawatt.

The fiscal 2028 forecast will be tested progressively, not on one earnings date. Construction schedules, supply purchases, cloud usage, and competitive chip deployments will reveal whether its assumptions remain aligned.

Developers and enterprise buyers should also follow availability and pricing. Greater supply can make advanced inference easier to access, even if organizations never purchase Nvidia hardware directly.

Knowledge workers will experience the forecast through software. If agents become reliable enough to manage longer workflows, compute demand can rise without users seeing the infrastructure beneath them.

The Nvidia 70% growth forecast is ultimately a claim about behavior, not just manufacturing. It assumes businesses will use enough AI to justify a historic expansion in computing capacity.

Huang has explained why Nvidia sees that outcome. The company touches models, clouds, enterprises, suppliers, power projects, and capital providers.

That reach gives Nvidia information few competitors possess. It also means Nvidia is helping create the future it says it can already see.

The next question is whether customers generate enough independent value to sustain it. Watch delivered systems, third-party cloud utilization, and custom-chip adoption over the coming quarters.

Those indicators will show whether 70% represents durable demand or a supply chain moving ahead of its users.