A New Story in the Earnings Report: Hardware Giant Becomes a "Compute Landlord"

While the market was still debating whether the AI bubble was too large and when capital expenditure would peak, NVIDIA delivered a surprising answer with its latest earnings report. According to the Q2 fiscal report released in early August 2026, NVIDIA not only continued its strong data center business performance but also surprised the market with its "compute leasing" segment—classified as "GPU-as-a-Service" revenue in the report—which exceeded $8.5 billion for the quarter, a 210% year-over-year increase. This figure's share of total revenue jumped from 6% in the same period last year to 18%, making it the second-largest revenue pillar after direct data center GPU sales.

For investors who have been following US stocks for a long time, the signal from this earnings report is far more profound than the revenue numbers themselves. Over the past two years, the market's main concerns about NVIDIA have centered on two points: the cyclical risk of AI capital expenditure and the competitive threat from AMD, Broadcom, and tech giants' in-house chips. However, the explosion of the GPU leasing business precisely hedges against these two risks to some extent—it transforms NVIDIA from a pure hardware vendor into a compute platform operator with "recurring revenue" characteristics.

How "Compute-as-a-Service" Reshapes Valuation Logic

Traditionally, NVIDIA's valuation framework fits the typical cyclical growth stock model. Hardware sales are highly dependent on downstream customers' capital expenditure cycles: when cloud providers like Microsoft, Google, and Amazon build AI infrastructure on a massive scale, NVIDIA's revenue and profits soar; once capital expenditure enters a digestion period, performance could face a cliff-like decline. This "boom-and-bust" expectation has long suppressed NVIDIA's P/E ratio. Even during the AI explosion of 2024-2025, its forward P/E ratio often hovered between 25-35x.

But the rise of the GPU leasing business is changing this valuation logic from three dimensions:

• Enhanced Revenue Predictability. Unlike one-time hardware sales, GPU leasing adopts a subscription or pay-as-you-go model, with customer service contracts typically lasting 1-3 years. This means NVIDIA can lock in revenue for several future quarters, significantly reducing earnings volatility. The earnings report showed that as of the end of July 2026, the unbilled contract balance for GPU leasing (i.e., signed orders not yet recognized as revenue) had exceeded $30 billion, providing high visibility for future income.
• Diversified Customer Base and Increased Stickiness. Previously, NVIDIA's GPU customers were highly concentrated among a few hyperscale cloud providers, with the top five customers accounting for over 60% of revenue at one point. The customer base for GPU leasing is much more fragmented, covering small and medium-sized AI startups, AI departments of large enterprises, research institutions, and even national laboratories. These customers often lack the capability or willingness to build their own data centers, making them more dependent on NVIDIA's compute platform and facing higher switching costs.
• Optimized Margin Structure. The market once worried that as competition intensifies, GPU hardware gross margins would fall from their peak. However, the gross margin for GPU leasing is actually higher than hardware sales—the earnings report showed this segment's gross margin reached 78%, about 6 percentage points higher than data center GPUs. The reason is that the leasing model maximizes the entire lifecycle value of the chip: new cards are prioritized for high-end training, previous-generation cards shift to inference services, and older cards are used for edge computing or low-priority tasks. This "compute tiering" strategy allows each GPU to generate more profit than a simple sale would.

Industry Sector Analysis: The "Second Half" of the Compute Market

To understand the strategic value of NVIDIA's GPU leasing business, it must be viewed against the backdrop of the entire AI infrastructure industry. In 2026, the AI compute market is undergoing a critical inflection point: shifting from "training-driven" to "inference-driven."

In the previous phase of AI development (2023-2025), compute demand primarily came from training large models. This is an "explosive, pulse-like" demand pattern—training a large model requires a massive, concentrated investment of compute power over several months, but demand drops sharply once training is complete. This model naturally suits hardware sales, as customers need to purchase a large number of GPUs in a short period.

Entering 2026, as large models are deployed across various industries, inference demand is becoming the dominant force in the compute market. The characteristic of inference demand is "continuous, stable, and distributed"—thousands of enterprises need to run AI services continuously, but each company's usage is relatively small and requires flexible, elastic scaling capabilities. This demand pattern naturally suits leasing services rather than hardware procurement. NVIDIA's GPU leasing business precisely captures this structural shift.

Another noteworthy industry dynamic is that the US Federal Trade Commission (FTC) has recently expressed concern about the competitive landscape of the cloud service market, hinting at a possible review of "exclusivity agreements" by large cloud providers. If this policy direction materializes, it will benefit independent third-party compute platforms like NVIDIA, as it provides small and medium customers with an alternative choice beyond AWS and Azure.

Capital Flow Signals: What Are Institutions Buying?

Looking at US stock capital flow data from July 2026, institutional investors' attitude towards NVIDIA is undergoing a subtle but important change. In the past, hedge funds and mutual funds held NVIDIA primarily for "trading" purposes—increasing positions before earnings reports and reducing them after results were announced, showing clear swing trading characteristics. However, in the weeks surrounding the Q2 earnings release, several large pension funds and sovereign wealth funds showed signs of "strategic position building."

According to Morgan Stanley's latest institutional holdings analysis report, long-term investors with over $100 billion in assets under management (such as pension funds and endowments) made record net purchases of NVIDIA in Q2 2026. This type of capital typically does not pursue short-term price differences but focuses on a company's 5-10 year long-term growth prospects. The reason they are increasing allocation at this point is largely because they see the "business model upgrade" brought by the GPU leasing business—NVIDIA is transforming from a cyclical hardware company into a technology infrastructure company with platform attributes.

Meanwhile, some aggressive growth funds, represented by ARK, slightly reduced their NVIDIA holdings in Q2. This changing of hands, with "long-term capital entering and trading-oriented capital exiting," is actually a positive signal for NVIDIA's stock price stability, as it means shares are shifting to more stable holders.

Valuation Re-rating: What P/E Multiple Is Reasonable?

For deep value investors focused on fundamental valuation, NVIDIA's current pricing is an unavoidable question. As of early August 2026, NVIDIA's 12-month forward P/E ratio was approximately 42x. If viewed as a pure hardware company, this valuation is indeed high—Broadcom is around 28x, and AMD is around 25x. But if viewed as a platform company with both hardware sales and subscription services, referencing the valuation framework of SaaS companies, 42x might not be unreasonable.

A simple sum-of-the-parts valuation can be done here:

• Data Center GPU Sales (~60% of revenue): Assign a 25-30x P/E, reflecting its hardware cyclicality.
• GPU Leasing Services (~16% of revenue): Assign a 50-60x P/E, reflecting the high visibility and high margin of its subscription-based revenue.
• Other Businesses (Gaming, Automotive, Robotics, etc., ~24%): Assign a 20-25x P/E.

Based on this weighted calculation, NVIDIA's overall reasonable P/E ratio is roughly in the 35-40x range. The current 34x is at the lower end of this reasonable range, which may explain why long-term capital chose to increase positions at this time.

Of course, risks cannot be ignored. The biggest uncertainty comes from geopolitics: 2026 is a US midterm election year, and chip export control policies towards China could escalate at any time. About 15% of NVIDIA's GPU leasing revenue comes from international customers in the Asia-Pacific region. Secondly, the pace of tech giants developing their own chips has not slowed—Google's TPU, Amazon's Trainium, and Microsoft's collaboration with AMD are all advancing steadily. If these alternatives achieve breakthroughs in cost-effectiveness, NVIDIA's compute monopoly could face challenges.

Implications for Investors

The insight this NVIDIA earnings report brings to US stock investors goes far beyond one company's performance. It reveals a profound structural change occurring in the AI industry: compute power is transforming from a "capital good" into a "utility."

Under this trend, the traditional framework of "valuing a hardware company based on its P/E ratio" may no longer apply. Investors need to consider: when a company possesses both the high explosiveness of hardware sales and the high persistence of subscription services, what valuation premium should it receive? When compute power is supplied on demand like electricity, which business models will be the biggest beneficiaries?

For investors building US stock portfolios, NVIDIA's case also provides an important asset allocation perspective: in the second half of AI investment, the line between pure hardware suppliers and pure software service providers is blurring. Companies that can close the loop on "hardware + services" may possess a wider moat and a longer growth runway.