NVIDIA: Who Pays For the Next Wave?

Analysis

The company powering the AI buildout is becoming one of its financial backers. What that means for an owner of its shares.

Evidence: business · financials. [S011] [S012]

From drawing pixels to running AI.

NVIDIA designs the processors, networking and software that make many of the world’s AI systems work. Its customers buy the machinery for a new form of computing. Shareholders own a claim on the profits—and the risks required to keep that machinery selling.

A graphics processing unit, or GPU, performs many calculations at once. NVIDIA originally used that ability to make games look better. It became useful for scientific computing and then for training AI models: adjusting a model through repeated calculations over enormous amounts of data. Running a trained model to answer a question, generate software or carry out a task is called inference. Both can require GPUs, but their economics and alternatives differ.

NVIDIA earns most of its revenue selling chips and complete computing systems, including the networking that lets thousands of processors cooperate. It also sells software and services. Customers include large cloud companies, specialist GPU operators, equipment manufacturers, enterprises, governments and gamers. Manufacturing is outsourced; NVIDIA owns designs, software, relationships and inventory rather than wafer factories. Follow a sale from order to cash. (Evidence: business)

1993–2006 · The foundation

Founded by Jensen Huang, Chris Malachowsky and Curtis Priem in 1993. CUDA, introduced in 2006, made GPUs programmable for work beyond graphics.

2012–2020 · A broader platform

AlexNet demonstrated GPUs’ usefulness in modern AI. The 2020 Mellanox acquisition added networking critical to large computing clusters.

2023–2026 · The spending surge

Revenue rose as generative AI moved into deployment. Blackwell systems, networking and the Rubin transition expanded NVIDIA’s possible content per installation.

2026 · A larger financial role

Longer customer credit, supply reservations, investments and lease guarantees expanded. A proposed $11.9bn Hugging Face acquisition would deepen developer reach; it has not closed.

Financial history: accounts (Evidence: financials). Chronology and pending acquisition: NVIDIA history [S003], Mellanox [S004], September 3 filing [S005].

4Trim

NVDA · $222.27 Regular close · 18 September 2026

Trim direct NVIDIA exposure at the stated reference price. Today's price assumes customers can keep earning enough from AI to fund new systems, and that NVIDIA can sustain growth without much heavier investment. The public evidence does not yet establish either. Exceptional profits, funding and platform strengths support a smaller diversified holding.

Investment stance 4 out of 10 — Trim. The balance favors reducing direct exposure, while continued ownership of a smaller position remains reasonable.

Four favors reducing exposure; it is not an exit judgment, a probability or a fixed portfolio weight.

Why now Competing chips have real deployments. NVIDIA can retain system value, but the price leaves little room for weaker margins or costly customer support.

What could prove us wrong Profitable adoption and richer systems could justify adding even at our 11% requirement. Funding progress and repeat GPU use give this case substance.

What changes the view Representative customer cash returns, better collections and resilient system profit could remove the case to trim. Persistent credit deterioration, unrecovered support or sustained workload losses would favor a stronger reduction.

Our measurement window 90 calendar days from actual publication; research revisited roughly quarterly. Long-term values are conditional, not 90-day targets.

Real growth. Uneven economics.

AI demand has progressed well beyond announcements. Microsoft reports strong Azure growth and millions of paid Copilot seats. Oracle reports GPU deliveries and expanding cloud business. Specialist operators earn revenue from services. These observations are stronger than a budget or a construction promise; they still do not establish that every installation earns an adequate return after equipment, electricity, financing and replacement.

1 · Funding Cash, customer deposits, equity or borrowing pay for construction and orders.

2 · Equipment NVIDIA and system partners supply computing and networking.

3 · Deployment Power, cooling and software make delivered equipment usable.

4 · Customer value Paid use or productivity gains must ultimately cover the recurring cost.

Revenue at one stage does not establish the economic success of the next. The same demand passes through several stages; the amounts cannot be added together.

Microsoft generated $63.9bn in fiscal 2026 after cash equipment spending and finance-lease principal. Amazon presents a different picture: over the year to June, operating cash fell $9.5bn short of net cash equipment spending and the disclosed lease and other financing principal payments. Its substantial cash, investments and borrowing capacity still support expansion. Both companies have large future commitments. The distinction is funding resilience: a cash-spending gap can be financed, but sustained expansion then depends more on capital markets or future cash growth. Customer cash and obligations. (Evidence: customer funding)

CoreWeave’s Q2 revenue was $2.575bn, with $640m of net interest expense and a $626m net loss. Those company-wide figures obscure an important favorable observation: in February, management reported contribution margins in the mid-20s for its fully running contracts. In its illustrated method, this is profit after direct operating costs and equipment depreciation, but before corporate overhead, interest and tax. September disclosures add higher prices, customer prepayments and better expected contract margins. The evidence supports profitable mature operations and makes expansion costs a plausible contributor to company-wide losses; it still lacks a representative record of cash recovered after all costs. Contract results and their limits. (Evidence: coreweave economics)

OpenAI’s March update reported $122bn committed capital, broader investors and an approximately $4.7bn undrawn credit facility. This strengthens payment capacity without proving cash receipt or profitability. Oracle’s customer prepayments likewise fund construction while committing it to future service. Financing can improve NVIDIA’s collections before productive customer returns are established. What is funded, committed and conditional. (Evidence: funding update)

Can a customer fund the next equipment cycle from productive use of the last one?

NVIDIA expects approximately 70% revenue growth in fiscal 2028. Management has repeatedly exceeded its revenue guidance, which lends this outlook weight; the forecast still depends on customers accepting and paying for the systems. Public data do not establish fleet-wide utilization or returns after equipment, power, engineering and finance costs. Our research weighs disclosed contract terms alongside a clearly assumed break-even example. Customer evidence and its limits. (Evidence: ecosystem)

Evidence: financials. [S011] [S012]

The sales grow faster than cash arrives.

For the six months ended July 26, NVIDIA reported $118.0bn net income and $74.4bn operating cash flow. Investment gains increased profit without equivalent operating cash. Customers owed more money, inventory grew, and other operating assets absorbed funds.

NVIDIA reported $118.0 billion net income, including $23.7 billion equity gains, $74.4 billion operating cash flow and $69.9 billion free cash flow. These bars compare measures; they are not an additive bridge.

H1 FY27, USD billions. The equity-gain bar explains part of the gap; do not subtract it from cash again. Free cash flow deducts equipment/intangible purchases and related principal payments, excluding equity investments and several other uses. Complete reconciliation. (Evidence: financials)

Some investment-grade customers can pay over 90 days to a year for large builds. Receivables reached $63.1bn. Longer terms can accommodate larger orders from creditworthy customers; they also mean NVIDIA supplies financing while waiting to collect.

Raw materials and work in process rose while finished goods fell. That is consistent with preparation for production; it does not prove profitable future sales. Q2 also had heavier tax-payment timing than Q1. These facts weaken an inference of collapsing demand, while leaving collections and inventory recovery to be tested.

We keep employee stock compensation in the economics: stock awards cost owners even when they do not immediately use cash. GAAP means standard U.S. accounting rules; non-GAAP measures make company-specified adjustments. NVIDIA now retains stock compensation in its adjusted presentation, making older adjusted figures unsuitable for casual comparison. Investment gains are not recurring chip earnings. A headline earnings multiple can look more reassuring than operating cash supports.

The balance sheet reaches into the customer’s building.

NVIDIA’s supply and capacity commitments reached $279bn in July. They secure components for future deliveries and expose the company if products become obsolete or purchases uneconomic. Normal purchases become product costs; deducting all commitments again would count the same cost twice.

NVIDIA also holds financial investments, purchases cloud services and supports partners’ infrastructure leases. An August guarantee covers specified obligations for an initial 4.25-gigawatt Ohio campus used by OpenAI. It is conditional, phased and capped at $105bn, with the first phase expected in fiscal 2029. It is not a bill due today.

The favorable explanation

Scarce components and power must be secured years ahead. Supporting creditworthy deployments can protect NVIDIA’s position, create service revenue and earn returns.

The adverse explanation

If customer economics disappoint, sales can slow as receivables take longer to collect, investments weaken and support costs rise.

In a combined stress, the first two years meet our middle path, then sales run 20% lower and operating margins ten percentage points lower. Delayed cash recovery, weaker investments and an extra $10bn a year of unrecovered support take conditional value to about $124. These assumptions illustrate a mechanism; they are not an expected loss estimate. Cost mapping, timing and the combined stress. (Evidence: support model)

We found disclosed dependencies, without establishing circular revenue or misconduct. Power and construction also affect timing: a delivered processor is not an energized data center. Ohio’s initial capacity is planned for 2028, subject to infrastructure and approvals. Cooling technology does not remove those constraints. Manufacturing and deployment evidence. (Evidence: environment)

Customers can diversify before NVIDIA loses its franchise.

NVIDIA combines processors, software, networking and systems engineers know how to deploy. Migration costs, cluster reliability and useful work per dollar matter as much as a chip benchmark. The advantage is strongest where workloads keep changing or enterprises lack the staff to engineer an alternative.

Some sophisticated customers already use meaningful alternatives. Anthropic reports more than one million Amazon Trainium2 chips training and serving Claude. Cerebras reported $56.8m quarterly OpenAI revenue after warrant amortization, with OpenAI separately offering a limited service powered by its hardware. That revenue includes data-center costs passed through to the customer and need not represent fresh quarterly cash; service billings can settle earlier customer funding. Broadcom’s $16.7bn quarterly AI semiconductor revenue spans custom accelerators and networking. These different measures cannot be summed as displaced NVIDIA sales. Dated evidence by workload and customer. (Evidence: ecosystem)

Repeated inference on well-understood models is particularly open to specialization. Frontier labs also have the scale to fund alternative training systems. Enterprises buying general computing place greater value on flexibility; hosted-model users often never choose a chip. More AI usage need not mean the same NVIDIA share or price.

The defense is tangible. Microsoft’s offered AMD MI300X systems use NVIDIA network connections, showing that a rival processor can still bring NVIDIA business. NVLink Fusion extends that possibility to other architectures. Licensing Groq technology helps NVIDIA adapt to inference, at a substantial cost. Nscale’s prospective Anthropic Rubin contracts and selected CoreWeave renewals support continued GPU demand. Older equipment can remain useful alongside new generations, although longer lives can delay replacement purchases. Equipment life and mixed-system evidence. (Evidence: replacement evidence)

Our judgment is that NVIDIA can remain central while losing bargaining power. The middle case allows gross margins to fade and engineering/support costs to take a larger sales share. A stronger case has richer systems and workload growth more than offset substitution. Lower computing cost per task can stimulate demand; the investment benefit depends on who supplies that work and keeps the savings. How these mechanisms reach sales and profit. (Evidence: operating bridges)

A strong company, with a close allocation decision.

The $222.27 share price is roughly 45 times trailing-year free cash flow after an illustrative stock-compensation charge. Growth can improve that picture, provided it produces cash after financing customers and keeping the platform competitive.

Our middle path reaches $590bn sales in the first rolling year, linked to quarterly guidance and an assumed fiscal 2028 ramp. Sales then rise to $1,270bn over nine further years, an 8.89% compound annual average. Growth starts near 24% and slows toward 2%. A 50% operating margin then means $50 profit per $100 sales after product and operating costs, before financing and tax. Gross margin deducts product costs only. Retaining such high profit is demanding even after growth slows.

After taxes and reinvestment, this path gives about $190 per share at an 11% required return. The rate expresses the compensation we require for risk, not a promised stock return. Roughly half the modeled business value comes from cash after year ten, when uncertainty is greatest.

Conditional present values · ten annual periods · 11% discount requirement

Conditional present values, not 90-day targets. Annual periods end September 2027–2036, then continuing value. All cases retain stock compensation, finance growth and use 24.3bn shares. No probabilities or loss floor. Assumptions, annual calculations and sensitivities. (Evidence: valuation)

A stronger-growth scenario also uses an 11% required return. It assumes the same first two years, followed by sales 20% higher and operating margins three percentage points higher from year three. That gives about $235 today. Paid adoption, retained networking value, better-funded customers and continued GPU orders support its possibility. Its roughly 6% excess over the quote retains the original capital-efficiency assumptions. Requiring more capital to sustain growth after year ten lowers this case to about $224 at a 20% return on new capital, or $218 at 15%. Those returns are sensitivity assumptions. The broader favorable scenario is $331; neither case has an established probability, and 6% is not a limit on possible upside.

The reverse calculation also matters: 17% more sales throughout the middle projection would match the quote. At unchanged operations, a 9% requirement gives about $255 and 13% about $152. The 11% choice is an analyst judgment, about six percentage points above the 5.01% ten-year Treasury yield. Neither the rate nor the middle case is a market consensus.

Why Trim? Today's price requires exceptional profits to last for years. Meanwhile, NVIDIA is extending more credit and financial support to customers whose ability to earn back their AI investment is still uncertain. Rival chips are also winning real deployments. Those risks justify selling some shares, rather than simply avoiding new purchases. We would keep a smaller holding because NVIDIA already earns substantial profits, customers continue to secure funding and renew equipment, and its networking can earn revenue even when customers choose rival processors. Evidence: rating rationale.

A broad-market fund is a comparison, although it owns NVIDIA. A semiconductor fund concentrates the same company and cycle. Hypothetical cash at 4.14% annual simple accrual earns about 1.02% over a full 90 days before tax and cost; scoring uses the actual interval between equity marks. None establishes future outperformance.

Let the next evidence challenge the argument.

The next earnings call is scheduled for November 17. Cash catching up with sales would challenge our caution. These five predictions give favorable and unfavorable evidence the same treatment.

1. Sales reach $108–116 billion. For Q3 FY27, we expect revenue within this inclusive range, versus $96.221 billion in Q2. Management guides to $108 billion ±2%; our higher upper end allows another shipment beat. Any result outside our range is a miss, including a larger positive surprise.
2. Gross margin remains 73–75%. Q3 GAAP gross margin stays in this inclusive band, versus 75.0% in Q2. This allows some memory and transition pressure around management’s 74% ±0.5 percentage-point outlook. A result outside the band is a miss.
3. Cash conversion recovers. Q3 operating cash flow reaches at least 40% of quarterly revenue, from 25.0% in Q2. Our judgment expects some recovery after Q2 tax timing and a large receivable build. Below 40% is a miss and would challenge the case for retaining exposure.
4. Customer credit stops stretching rapidly. Quarter-end receivables divided by Q3 revenue, multiplied by the actual number of days in the quarter, stay at or below 65 days. Q2 was about 60. Above 65 is a miss; the measure tests cash tied up in sales, not whether a customer has defaulted.
5. Growth contains a customer’s interest burden. CoreWeave’s calendar Q3 net interest expense is no more than 25% of revenue, versus 24.85% in Q2 ($640 million / $2.575 billion). This predicts that growth keeps the financing burden from worsening materially. Above 25% is a miss and strengthens the funding concern. Passing challenges that concern, but does not establish ability to repay debt or independently fund new equipment.

Revenue and margin forecasts build on management guidance. Cash conversion, receivable days and CoreWeave’s interest burden are our judgments. NVIDIA fiscal Q3 and CoreWeave calendar Q3 are different periods. Evidence must be public by the fixed 90-day endpoint; missing data remain unobservable. Definitions and outcome conventions. (Evidence: forecasts)

Five durable tests also track competing deployments, retained system economics, collections and financing, replacement demand, and net support costs. Each identifies the model assumption new evidence would change. Baselines, sources and decision rules. (Evidence: thesis tests)

Examine the Evidence & Assumptions →