NVIDIA Corporation

01 — Executive summaryThe operating business is extraordinary. The earnings quality is not.

NVIDIA Corporation reported revenue of $96.2 billion for the quarter ended July 26, 2026. That is a 106% increase over the same quarter a year earlier and an 18% increase over the immediately preceding quarter. Data Center revenue was $89.0 billion, up 117%. GAAP gross margin was 75.0%. GAAP operating income was $63.7 billion, a 66% operating margin. The company guided the following quarter to $108.0 billion. Management has described approximately $1 trillion of committed orders through 2027. At $210.94 per share, NVIDIA carries a market capitalization of roughly $5.09 trillion, the largest of any company in the world.

None of that is in dispute, and none of it is the subject of this report. The question we set out to answer is narrower and less comfortable: at the current price, what does the market already believe — and is that belief reasonable?

Our answer, compressed into a sentence: NVIDIA's operating business is one of the most remarkable in corporate history, and the equity is no longer priced for perfection. It is priced for a specific and, we argue, somewhat excessive degree of cyclical pessimism. The offsetting problem — the reason this is not an unqualified buy — is that the quality of reported earnings has measurably deteriorated, and the company has begun borrowing money to finance the demand it depends on.

Six findings follow from the primary record. Each is developed in full later in this report; each is stated here with the number that supports it.

1. The business is real, and the financial statements prove it

This is not a narrative stock. A 75.0% gross margin, a 66.2% GAAP operating margin, $74.4 billion of operating cash flow in six months, and a guided $108.0 billion quarter are accounting facts. The capital intensity that usually accompanies a business at this scale is absent: research and development consumed 7.3% of revenue in the quarter, and purchases of property and equipment were $2.7 billion against $96.2 billion of revenue. NVIDIA is a design and software company that happens to sell hardware. Anyone arguing that the revenue is fictional has to explain why the cash keeps arriving.

2. But it has become a single-segment company

Data Center was 89.7% of FY2026 revenue and 92.5% of revenue in the most recent quarter, up from roughly 78% two fiscal years earlier. Gaming, professional visualization, and automotive — the three businesses that made NVIDIA a household name among gamers and designers — are now collectively smaller than the revenue line item NVIDIA reports for networking alone. This matters less as a diversification argument than as an analytical one: it means the company's entire future is now a function of a single variable, namely the capital spending plans of a small number of buyers.

3. The customer base is dangerously concentrated — and strategically conflicted

NVIDIA discloses direct customers that individually exceed 10% of revenue. In FY2026 there were four of them, and together they accounted for 61% of total revenue. The largest single customer was 22%. A year earlier, only three customers cleared the 10% threshold, at roughly 12% each, for a combined 36%. In other words, customer concentration nearly doubled in twelve months.

The standard defense — "big company, big customers" — fails here for a specific reason. Those same four buyers are each running in-house accelerator programs at scale. Google's TPU runs the majority of Gemini inference. Amazon's Trainium handles more than half of Bedrock tokens. Meta has deployed hundreds of thousands of MTIA parts. Microsoft's Maia program continues. The customers are simultaneously the competitors, and they are the ones with the strongest economic incentive to reduce their dependence.

4. Earnings quality has thinned in three measurable ways

This is the core of our concern, and it is the part of the story that receives the least attention.

Working capital. Days sales outstanding moved from 51 days at FY2026 year-end to 45 days in Q1 FY2027 and then to 60 days in Q2 FY2027. Accounts receivable grew 64% over the two quarters while revenue grew 41%. Operating cash flow was 40% of net income in Q2 FY2027, against 95% in the comparable prior-year half.

Non-operating income. In the first half of FY2027 NVIDIA booked $23.7 billion of pre-tax gains on equity securities — roughly 17% of pre-tax income — from marking up stakes in public and private companies. Some of those companies are customers or suppliers. Non-GAAP results strip these gains out, which is why GAAP net income of $59.7 billion in Q2 FY2027 sits alongside non-GAAP net income of $54.0 billion. The gap is not an accounting technicality; it is the difference between operating performance and portfolio appreciation.

Financing the ecosystem. In a single quarter NVIDIA issued $24.9 billion of debt, lifting long-term borrowings from $7.5 billion to $32.4 billion, while holding $56.6 billion in cash and marketable debt securities. Over the first half it deployed $42.4 billion into equity securities. A company does not borrow when it is sitting on $56 billion of liquidity unless the borrowing is doing something other than funding operations. What it is doing, on the evidence, is financing the customers and partners who buy its products — a structure the company describes as "strategic compute financing" and which critics describe more bluntly as vendor financing.

5. China is now zero, and guidance assumes it stays there

Data Center revenue attributable to China was $4.6 billion in Q1 FY2026. In Q1 and Q2 FY2027 it was zero, and NVIDIA's Q3 guidance explicitly assumes no Data Center compute revenue from China. Two doors closed in sequence: the United States blocked the compliant B30A part in November 2025 and now licenses older H200 shipments subject to a 25% revenue share and quantity caps, while China's own certification regime for domestic AI processors structurally excludes foreign silicon. Sell-side models as recently as early as 2026 still carried $12–22 billion of China revenue. That number is gone, and its absence is already inside the guidance the company has given.

6. Competition is real, but the moat is not yet broken

NVIDIA holds roughly 80% of AI accelerator revenue. AMD's MI350X matches the Blackwell B200 on FP8 throughput and exceeds it on memory capacity, at an estimated 30–50% lower price. Broadcom's custom ASIC business is running at an $8.4 billion quarterly revenue rate with a $73 billion backlog. On specifications and on price, the gap has narrowed materially. What has not narrowed is the software and systems layer: CUDA's developer base, the NVLink interconnect, and the rack-scale integration that makes a thousand accelerators behave as one machine. Those are the reasons NVIDIA still earns 75% gross margins while its nearest merchant competitor earns in the mid-50s.

What would change our mind

We would upgrade this to a high-conviction position on any of three developments: receivables days returning below 50 and staying there across two consecutive quarters; evidence that the equity investment program is generating strategic lock-in rather than revenue round-tripping; or a credible path to nonzero China revenue under a durable licensing framework. We would downgrade on any of three others: a single hyperscaler publicly reducing its NVIDIA-attached capital budget; gross margin guidance below 72%; or a disclosure that a material share of revenue growth came from customers in which NVIDIA itself holds an equity stake.

02 — Company & business modelA design house that learned to sell the whole building

NVIDIA was founded in 1993 to build graphics accelerators for personal computers. It invented the term "GPU" with the GeForce 256 in 1999 and spent the following decade selling discrete graphics cards to gamers. The decision that created the modern company came in 2006, when NVIDIA released CUDA — a programming model that let general-purpose software run on its graphics processors. For roughly a decade CUDA was a strategic liability, subsidized by the gaming business. It is now the single most valuable asset the company owns.

The business model that emerged from that history is unusual in semiconductors. NVIDIA is fabless: it designs chips and outsources manufacturing entirely to TSMC and packaging to TSMC and its partners. It owns almost no factories. Purchases of property and equipment were $2.7 billion in the most recent quarter against $96.2 billion of revenue — a capital intensity of 2.8%. For comparison, a memory manufacturer at similar scale would be spending multiples of that. The consequence is that incremental revenue converts to gross profit at an extraordinary rate, and that the company's balance sheet is asset-light in a way that its income statement is not.

Two reporting segments, and what the change conceals

Beginning with fiscal 2027, NVIDIA reorganized its reporting into two market platforms: Data Center and Edge Computing. The prior structure — Data Center, Gaming, Professional Visualization, Automotive, and OEM — is gone. Data Center is further described as comprising two sub-markets, "Hyperscale" and "ACIE" (AI Clouds, Industrial and Enterprise), for which no separate revenue figures are disclosed.

We flag this as a disclosure-quality issue rather than a scandal. Reorganizing segments around how management runs the business is legitimate, and the new structure does map to a real strategic shift: NVIDIA increasingly sells complete systems, and the boundary between a gaming GPU and a workstation GPU and an automotive inference module has genuinely blurred. But the practical effect is that an investor now receives less granularity at precisely the moment when granularity matters most. Edge Computing at $7.2 billion is a residual line that mixes consumer graphics, professional visualization, automotive design wins, and robotics — businesses with wildly different growth rates and margins. A reader can no longer see whether gaming is growing or shrinking.

How the revenue actually reaches the income statement

NVIDIA sells through three broad channels. The largest, and the one that drives the concentration statistics in this report, is direct sale to a handful of hyperscale operators and large cloud providers — the companies building AI factories at gigawatt scale. The second is through OEMs and original device manufacturers, who integrate NVIDIA parts into servers and systems. The third is distribution, which serves enterprise, academic, and consumer buyers.

What has changed structurally is that NVIDIA now sells more than chips. Rack-scale systems such as the GB300 NVL72 and the Vera Rubin NVL72 are sold as integrated units containing 72 accelerators, the NVLink switch fabric that connects them, the CPUs that orchestrate them, and the software stack that runs on them. The reported Data Center networking revenue line — $14.8 billion in Q1 FY2027, up 199% year over year — exists because NVIDIA captured the interconnect rather than ceding it to Ethernet vendors. That is a strategic choice with a large margin consequence, and it is the part of the business that custom-ASIC competition struggles most to replicate.

03 — The financial recordFive years, and one number that needs explaining

NVIDIA's revenue was essentially flat between fiscal 2022 and fiscal 2023, at $26.9 billion. Five years later it was $215.9 billion. No company of this size has ever grown this quickly, and it is worth stating plainly that the growth is not the product of accounting choices. It is the product of a new class of customer — the AI factory — arriving with a budget and a deadline.

The FY2026 gross margin line is the one that needs explaining. GAAP gross margin fell from 75.0% in FY2025 to 71.1% in FY2026 — a 390 basis point compression in a year when revenue grew 65%. That is counterintuitive: scale is supposed to expand margins. The cause is a $4.5 billion charge taken in Q1 FY2026 against H20 inventory and purchase commitments after US export restrictions halted the sale of that part into China. Strip the charge out and the underlying margin trend is flat to modestly positive, as the quarterly series shows.

This matters for two reasons. First, it means the FY2026 margin decline is a policy event, not a competitive one — a distinction that several commentaries have blurred. Second, it establishes that export controls can impose a nine-figure quarterly charge at will, and that the company has no control over when the next one arrives. NVIDIA took a $4.5 billion hit in one quarter and guided straight through it. That is a demonstration of earnings power. It is also a demonstration of exposure.

The quarterly record: acceleration, then a plateau in margin

The quarterly series is more informative than the annual one because it shows the inflection points as they happened.

Q3 FY2026 figures are derived from full-year totals less reported quarters. Q3 FY2027 revenue and gross margin are company guidance of $108.0 billion ±2% and 74.0% ±50bp. Q3 FY2027 Data Center and Edge Computing figures are Farstar estimates. Q2 FY2026 non-GAAP EPS is stated on the restated basis that includes stock-based compensation.

Three observations from this table deserve emphasis.

Growth is accelerating, not decelerating. Year-over-year growth rose from 56% in Q2 FY2026 to 106% in Q2 FY2027. Companies of NVIDIA's size do not normally accelerate. This is the strongest single argument against the "peak" thesis: a business at the top of its cycle does not typically see its growth rate double.

Gross margin has plateaued at 75%, not expanded. Across the last four quarters the margin has been 73.5%, 75.0%, 74.9%, and 75.0%, and guidance for the next quarter is 74.0% — a step down. A 75% gross margin is extraordinary, but the absence of further expansion at a moment when revenue is scaling rapidly suggests either that pricing power has reached its limit, that mix is diluting, or that input costs are rising faster than they are being offset. The guidance step-down to 74.0% is small but it is in the wrong direction.

The GAAP-to-non-GAAP gap is widening. In Q2 FY2026, GAAP diluted EPS of $1.08 and non-GAAP EPS of $1.01 were closely aligned. By Q2 FY2027, GAAP EPS of $2.46 exceeded non-GAAP EPS of $2.22 by 11%. The divergence is not caused by restructuring charges or amortization — it is caused by gains on equity securities, which NVIDIA excludes from non-GAAP. In other words, the company's GAAP earnings are increasingly flattered by the appreciation of its own investment portfolio.

The table below isolates the effect. Over the first half of FY2027, NVIDIA recorded $23.7 billion of pre-tax gains on equity securities against $141.4 billion of pre-tax income. That is 16.8% of pre-tax profit generated by marking up financial assets rather than by selling accelerators.

04 — Segment deep diveData Center: the whole company

Data Center revenue was $39.1 billion in Q1 FY2026 and $89.0 billion in Q2 FY2027. That is a 128% increase in six quarters, and it is the single most important revenue series in the technology industry today. Growth in the most recent quarter was 117% year over year and 18% sequentially.

Two components make up this line, and they have very different characteristics.

Compute — the accelerators themselves — was $60.4 billion in Q1 FY2027, up 77% year over year. This is the part of the business that competes directly with AMD's Instinct line, with hyperscaler in-house silicon, and with every startup building a matrix-multiply engine.

Networking — NVLink switch fabric, Spectrum-X Ethernet, and InfiniBand — was $14.8 billion in Q1 FY2027, up 199% year over year and 35% sequentially. Networking is growing roughly two and a half times as fast as compute, and it is the most strategically important number in the segment. When a customer buys a rack-scale system, they buy NVIDIA's interconnect whether or not they buy NVIDIA's accelerator — and increasingly they buy the accelerator because the interconnect makes the alternative unattractive. NVIDIA has spent a decade converting a commodity networking market into an extension of its own platform.

The demand signal management is pointing to

At GTC in March 2026, management described approximately $1 trillion of committed orders through 2027. That figure should be handled carefully. It is a management statement rather than an audited backlog disclosure, NVIDIA does not define what constitutes a "commitment," and the company has historically converted visibility into revenue at high rates without publishing the conversion ratio. What can be said is that the guidance record supports the general direction: NVIDIA has guided to a sequential increase in each of the last six quarters and has met or exceeded guidance in each.

Geographic mix

NVIDIA does not disclose Data Center revenue by geography in its quarterly releases, but the direction is unambiguous from the China disclosure and from management commentary: revenue is increasingly concentrated in the United States, with growing contributions from sovereign programs in the Gulf states, Japan, Korea, and Europe. The company announced a partnership with the Japanese government on what it described as the world's first national AI infrastructure, and expanded relationships with SK Telecom, NAVER, and SK hynix in Korea. Sovereign demand is structurally attractive because it is less correlated with commercial hyperscaler capital cycles — but it is also politically determined, and therefore lumpy.

05 — Segment deep diveEdge Computing: small, growing, and now opaque

Edge Computing revenue was $6.4 billion in Q1 FY2027 and $7.2 billion in Q2 FY2027, up 27% year over year. Within the old reporting structure this line corresponds roughly to Gaming plus Professional Visualization plus Automotive, and the growth rate is therefore a blend of businesses that have historically moved in different directions.

The composition matters. Gaming was, until 2023, NVIDIA's largest segment. It is now roughly one-thirteenth the size of Data Center. That is not because gaming shrank — gaming revenue grew 41% in FY2026 — but because Data Center grew so much faster. A reasonable question for any investor is whether a business contributing 3% of revenue and receiving a proportionate share of management attention is being managed for growth or harvested for cash.

The genuinely interesting part of this segment is physical AI — robotics and autonomous vehicles. NVIDIA has built a full stack here: the DRIVE Hyperion platform for autonomous driving, the Jetson Thor compute module, the Isaac GR00T reference humanoid robot design, the Cosmos world-model family for simulation, and Halos, a safety architecture for physical AI. Design wins have been announced with Foxconn, VinFast, Uber, Hyundai, Kia, BYD, Geely, Isuzu, and Nissan. The robotaxi market is projected by third parties to reach $400 billion by 2035.

We treat this as a call option rather than a valuation input. Automotive and Robotics revenue was $2.3 billion in FY2026 — approximately 1.1% of the company — and even aggressive growth from that base will not move the consolidated numbers within our forecast horizon. What it does do is extend the addressable market for NVIDIA's platform into a domain where CUDA compatibility, simulation tooling, and safety certification create switching costs that are, if anything, higher than in the data center. It is the most credible long-duration part of the story and the least relevant to the next eight quarters.

06 — Product roadmapAn annual cadence, and the depreciation problem it creates

NVIDIA has moved to a yearly architecture cadence, a decision with consequences that extend well beyond product marketing. The sequence is Blackwell (fiscal 2025), Blackwell Ultra, Vera Rubin, Rubin Ultra, and Feynman.

Vera Rubin entered what the company calls full production at CES in January 2026 — ahead of the original second-half 2026 target. The platform is a six-chip system, and its disclosed specifications represent a generational jump: roughly 336 billion transistors, HBM4 memory at 288GB per package with bandwidth in the region of 13 TB/s, NVLink 6.0 at 3.6 TB/s per GPU, and a rack-scale NVL72 configuration rated at 3.6 exaFLOPS of FP4 throughput. Deployment partners announced at launch include CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius.

The platform extends beyond the accelerator. NVIDIA has introduced the Vera CPU, positioned as the first processor purpose-built for agentic AI workloads; BlueField-4 STX for agentic storage infrastructure with in-silicon security; Spectrum-6 switch systems supporting both pluggable and co-packaged optics; and the DSX platform for designing, building, and operating AI factories. Management also disclosed that the Groq 3 LPX inference accelerator is in full production, following a $2.9 billion payment related to Groq recorded in the Q2 FY2027 cash flow statement.

Software as the retention mechanism

The roadmap's commercial purpose is not only to sell new silicon but to make the software layer indispensable. Dynamo 1.0, an open-source inference orchestration layer, is claimed to improve generative and agentic inference throughput on Blackwell by up to 7×. NemoClaw, OpenShell, and the Agent Toolkit extend the platform into autonomous agent deployment. Nemotron provides open model weights; BioNeMo targets life sciences. The pattern is consistent: NVIDIA gives away the software layer that makes its hardware easy to use, and captures the value in the hardware. It is the same strategy that made CUDA decisive, applied at the systems level.

07 — The moatSoftware, interconnect, and the rack as the unit of sale

The most common mistake in analyzing NVIDIA is to treat the moat as a manufacturing advantage. It is not. NVIDIA's chips are made by TSMC on the same process nodes available to AMD, Google, Amazon, and anyone else with a purchase order. The moat is software and integration, and it operates at three distinct levels.

Level one: CUDA and the switching cost of a decade of code

CUDA has been in market for roughly eighteen years and now supports a developer base measured in millions, with a library ecosystem — cuDNN for deep learning primitives, cuBLAS for linear algebra, TensorRT for inference optimization, NCCL for multi-GPU communication, CUTLASS for custom kernels — that has no equivalent on any competing platform. The relevant question is not whether AMD's ROCm can run a transformer; it can, and it does. The question is whether an engineering organization can port a decade of accumulated kernels, profiling tools, and operational runbooks without a measurable productivity loss and without the risk that some dependency has no ROCm equivalent.

The commercial expression of this is NVIDIA AI Enterprise, priced at roughly $4,500 per GPU per year. That is a software line item attached to hardware, and it exists because the software is worth paying for. It also creates a recurring revenue stream that scales with the installed base rather than with unit shipments — a distinction that matters enormously in a cyclical downturn.

Level two: NVLink, and the reason the rack is now the product

The performance of a large language model training run is determined less by the throughput of an individual accelerator than by how quickly thousands of them can exchange data. NVIDIA's NVLink provides roughly 1.8 TB/s of per-GPU interconnect bandwidth on Blackwell, rising to 3.6 TB/s on Vera Rubin. AMD's competing Infinity Fabric delivers approximately 128 GB/s of peer-to-peer bandwidth on the MI350X generation. That is not a marginal difference; it is an order of magnitude, and it determines how large a model can be trained efficiently on a single coherent domain.

This is why NVIDIA sells racks, not cards. A GB300 NVL72 or Vera Rubin NVL72 is a single failure domain with 72 accelerators connected by a non-blocking switch fabric, sold as one unit at a price in the low millions. Competitors must either match the fabric — which requires designing a switch chip, a topology, and a protocol stack — or accept that their systems top out at a smaller coherent domain.

Level three: the ecosystem gravity that follows

Because CUDA is where the researchers are, new model architectures appear on NVIDIA hardware first and run best there. Because they run best there, cloud providers offer the widest selection of NVIDIA instances. Because the selection is widest, enterprises build on NVIDIA. This is a classic two-sided network effect and it is the most durable part of the position — but it is also the part most exposed to a single technological development: compiler frameworks that abstract away the hardware.

08 — Competition IMerchant silicon: AMD has closed the spec gap, not the utilization gap

AMD's Instinct MI350X is a genuinely competitive part. On paper it matches or beats NVIDIA's Blackwell-generation B200 in several dimensions, and it does so at a substantially lower price.

Sources: vendor specifications, MLPerf submissions, and third-party benchmarks. MFU denotes model-FLOPs utilization and measures the share of theoretical peak throughput achieved in real workloads. Price and margin figures are third-party estimates, not vendor disclosures.

The specification table reads well for AMD. The economics do not, quite. Independent benchmarking has consistently found that AMD's hardware achieves a lower share of its theoretical peak than NVIDIA's in real training workloads — the gap is commonly cited at 45% versus 50–55% utilization — and that the gap widens at multi-node scale where interconnect quality dominates. A part that costs 35% less but delivers 15–20% less usable throughput is a much narrower bargain than the price list suggests.

Total cost of ownership modeling makes the same point. On a three-year, 32-GPU training cluster, AMD's hardware advantage of roughly 50% narrows to something in the range of 15–25% once power, cooling, networking, software licensing, and the engineering cost of kernel optimization are included. For inference workloads, where memory bandwidth matters more than interconnect and where the MI350X's 288GB capacity advantage is meaningful, AMD is genuinely at parity or better.

Where AMD has already won

AMD's commercial progress is real and should not be dismissed. OpenAI has committed to a multi-gigawatt deployment of MI450 parts beginning in the second half of 2026 — described by AMD as the largest win in the company's history. Meta has reportedly committed to a comparable gigawatt-scale deployment. These are not pilot programs. They are the largest AI infrastructure commitments in the industry, and a meaningful share of them is going to AMD.

That said, the structure of these wins is instructive. The commitments are for incremental capacity, not replacement capacity. OpenAI and Meta are not swapping out NVIDIA fleets; they are adding AMD alongside them, which is rational behavior for a buyer facing a single-source constraint. The commercial question for NVIDIA is not whether it loses the next gigawatt, but whether it loses the next gigawatt at the same price. So far, the gross margin line says it has not.

09 — Competition IICustom ASICs: the structurally more dangerous competitor

AMD is the visible competitor. The custom ASIC programs are the consequential one. Broadcom's AI ASIC business generated $8.4 billion of revenue in a single quarter — growth of 106% year over year — and the company has disclosed a $73 billion backlog extending through mid-2027. Marvell is building a comparable franchise at smaller scale.

The programs behind those numbers are the reason to pay attention:

• Google TPU. The seventh-generation TPU, Ironwood, runs the majority of Gemini inference. Google has committed a million Trillium chips to Anthropic by 2027. Google does not need to buy NVIDIA accelerators to serve its own models.
• AWS Trainium. Trainium3, on TSMC's 3nm process, handles more than half of Bedrock token throughput. Amazon's internal demand alone is large enough to justify a multi-generation silicon program.
• Meta MTIA. Hundreds of thousands of units deployed for ranking and recommendation inference, with the program expanding toward generative workloads.
• Microsoft Maia. Maia 200 continues to develop, though later than originally planned.

The economics driving these programs are straightforward. A hyperscaler running a stable, well- characterized workload at massive scale can design a chip that does that workload with less silicon area and less generality than a merchant GPU, and can amortize the design cost over hundreds of thousands of units. The result is a lower cost per token for that specific workload. For inference — which is now roughly two-thirds of all AI compute spending, and trending toward 70–80% by the end of the decade — that arithmetic is compelling.

There are three things custom silicon cannot easily do, and they define the limits of the threat. It cannot serve external customers, because no hyperscaler has built a merchant silicon business and doing so would mean competing with its own cloud customers. It cannot respond quickly to architectural change, because a fixed-function design is by definition optimized for the workloads known at tape-out. And it cannot match the generality that lets an enterprise run a newly released open model on day one. These constraints are why we expect NVIDIA's share to decline gradually rather than collapse.

Our working estimate — triangulated from unit-shipment and revenue-share datasets, and shown in the share chart above — is that NVIDIA's revenue share of AI accelerators falls from roughly 87% at its 2024 peak to approximately 75% in 2026 and toward the mid-60s by 2028, with AMD reaching low double digits and custom silicon taking the remainder. That is a share loss of more than twenty points over four years. It is also, importantly, compatible with NVIDIA's revenue growing substantially, because the total market is expanding faster than share is eroding. Both things are true at once, and most commentary insists on choosing one.

10 — DemandThe capex question, and why it decides this investment

Everything above concerns supply, competition, and execution. None of it determines the outcome. The outcome is determined by one variable: how much money NVIDIA's customers spend on AI infrastructure over the next three years, and whether they continue to spend it when the returns disappoint.

Current spending levels are extraordinary. Moody's Ratings raised its forecast for the six largest hyperscalers — Microsoft, Amazon, Alphabet, Meta, Oracle, and CoreWeave — to $785 billion for 2026 and to near $1 trillion for 2027. That 2026 figure was revised upward three times in under a year, from an original estimate near $600 billion. The upward revisions were driven by an earnings cycle in which Google Cloud revenue grew 63%, AWS posted its fastest growth in fifteen quarters, and Microsoft's AI revenue run-rate exceeded $37 billion with 123% growth.

The chart above contains the most important analytical finding in this report. NVIDIA's Data Center revenue, expressed as a share of combined hyperscaler capital expenditure, rose from roughly 21% in 2024 to approximately 39% in 2025 to an estimated 48% in 2026 — and then flattens at approximately 49% in 2027. In other words, NVIDIA's growth is now approximately coincident with the growth of hyperscaler capital budgets, rather than exceeding it. For as long as NVIDIA was capturing a rising share of a growing pool, it enjoyed two compounding drivers. Going forward it has one. That is a material change in the character of the equity, and it is not reflected in the consensus price target.

The depreciation question

The bear case on AI infrastructure is not that demand is fake. It is that the accounting used to justify the investment overstates its returns. Google, Oracle, and Microsoft have all assigned useful lives of up to six years to AI compute equipment. Critics argue the correct figure is two to three years, on the grounds that a part which is two generations behind the current architecture cannot be rented at a price that covers its cost of capital.

The argument has force, and NVIDIA's own annual cadence supplies the ammunition. But the bear framing has a weakness too: the useful life of an accelerator is not a physical constant, it is a function of what workloads it can still serve profitably. A Blackwell part that is no longer competitive for frontier training may remain perfectly economic for inference on smaller models, for fine-tuning, for synthetic data generation, or for the long tail of enterprise workloads. The evidence so far — rental rates for older generations holding up better than the depreciation skeptics predicted — suggests the truth is somewhere between the two positions.

Circular financing, examined without hysteria

NVIDIA has committed up to $100 billion of investment to OpenAI. It has established "strategic compute financing" platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, intended to mobilize over $500 billion of third-party capital. And, as documented earlier, it has itself deployed $42.4 billion into equity securities in six months while issuing $24.9 billion of debt.

The concern this raises is simple to state. If NVIDIA invests in companies that then buy NVIDIA products, revenue growth is partly a function of NVIDIA's own capital, and the reported growth rate overstates the organic demand for the product. The parallel drawn by critics is to Lucent and Nortel, which financed their telecom customers in the late 1990s and recognized revenue on equipment those customers could not ultimately pay for. The 1990s comparison is imperfect — those customers were startups with no cash flow, whereas NVIDIA's are among the most cash-generative companies in history — but the accounting mechanism is not different in kind.

Our assessment is that this is a real risk of indeterminate magnitude, and that NVIDIA's disclosure is not sufficient to size it. The company does not disclose what proportion of revenue comes from entities in which it holds an equity stake. Absent that disclosure, an investor cannot determine whether the $23.7 billion of equity gains in the first half are a mark-to-market of strategic positions or the leading indicator of a revenue round trip. We treat this as a governance issue rather than a fraud allegation, and we size it as a risk rather than a thesis. But it is the single disclosure we would most like to see.

11 — Supply chainNVIDIA's revenue is gated by things NVIDIA does not make

NVIDIA designs chips and outsources everything else. That makes it dependent on a set of suppliers who are, collectively, the real constraint on how fast revenue can grow.

Advanced packaging is the narrowest gate

Every high-end accelerator requires chip-on-wafer-on-substrate packaging, and TSMC's CoWoS capacity has been the binding constraint on accelerator shipments for three years. Capacity has expanded substantially and the constraint has eased, but the migration to larger interposers and co-packaged optics means packaging demand continues to grow faster than supply. NVIDIA's ability to hit its own guidance is, to a first approximation, a function of TSMC's packaging expansion schedule.

Memory is the second gate, and the more interesting one

HBM4 is required for the Vera Rubin generation, and the supply base is three companies: SK hynix, Micron, and Samsung. NVIDIA has signed a multiyear memory partnership with SK hynix and expanded its Korean relationships. The strategic implication is that memory makers, not NVIDIA, capture the economics of the memory constraint — and that a memory shortage is a cost problem for NVIDIA rather than a revenue problem. This is a plausible partial explanation for the guidance step-down in gross margin from 75.0% to 74.0%.

Power, land, and the physical layer

The most consequential constraint is not a chip. It is electricity and the buildings to put it in. Grid interconnection queues in the United States now run longer than construction schedules, and the permitting and transmission frameworks governing data center development were not designed for investment at the current pace. NVIDIA has begun addressing this directly — securing land, power, and shell capacity through an arrangement with SB Energy at a technology campus in Ohio, and announcing partnerships to expand capacity with Australian cloud partners.

12 — GeopoliticsChina: from $4.6 billion to zero, and the option value that went with it

For most of the past two decades, China was a material part of NVIDIA's business. In January 2026 that ended, and the sequence of events is worth reconstructing precisely because it explains why the revenue is unlikely to return quickly.

The result is a two-door squeeze. The US door is closed by veto and, where ajar, policed and taxed. The Chinese door is closed by procurement policy designed to route state demand to domestic vendors. Even a fully compliant US export must now clear a second gate that exists specifically to exclude it.

Data Center revenue attributable to China was $4.6 billion in Q1 FY2026. It has been zero in every quarter since, and NVIDIA's Q3 FY2027 guidance explicitly assumes no Data Center compute revenue from China. Meanwhile Huawei's Ascend 910C is reported to deliver roughly 60% of an H100's real-world inference performance, and Chinese domestic vendors took 41% of the China AI accelerator market by unit volume in 2025. The domestic alternative is not equivalent to NVIDIA's best product. It does not need to be.

The Taiwan concentration

Every leading-edge NVIDIA product is manufactured by TSMC in Taiwan and packaged by TSMC and its partners. This is the largest single unhedgeable risk in the equity, and it is not a risk that management can mitigate by operational excellence. It is a geopolitical tail risk with a low probability and an enormous consequence. We do not assign it a probability in our scenario model, because doing so would produce a number that implies false precision. We note it, and we note that it applies with equal force to AMD, Broadcom, Apple, and most of the semiconductor industry — it is a systemic exposure rather than a company-specific one.

Sovereign AI as the partial offset

The countervailing development is sovereign AI: national governments building domestic compute capacity for reasons of policy rather than profit. NVIDIA has announced a partnership with the Japanese government on national AI infrastructure, expanded relationships with SK Telecom, NAVER, and SK hynix in Korea, collaborations with Australian cloud partners, and thirty-five new AI and HPC supercomputers in development across Europe. Sovereign buyers are structurally attractive because their budgets are political rather than commercial and are therefore less correlated with hyperscaler capital cycles. They are also, for the same reason, less predictable.

13 — Quality of earningsThree numbers that do not appear in the press release

This section is the analytical core of the report. NVIDIA's headline results are not in question. What is in question is whether the headline results represent the operating business, or the operating business plus a set of financial transactions that would not exist if the operating business were smaller.

Finding one: receivables are growing much faster than revenue

Days sales outstanding, computed as period-end receivables divided by quarterly revenue, moved from 51.4 days at FY2026 year-end to 45.4 days in Q1 FY2027 and then to 59.6 days in Q2 FY2027. In absolute terms, accounts receivable rose from $38.5 billion to $63.1 billion over two quarters — a 64% increase against 41% revenue growth over the same span.

There are benign explanations. When a company's revenue mix shifts toward enormous, complex, multi-party rack-scale system deployments with acceptance milestones, billing and collection naturally lengthen. Revenue recognized on delivery with payment due on acceptance is a normal commercial structure for capital equipment of this size. And a 60-day DSO is not alarming in isolation — many enterprise hardware companies operate at 70 days or higher.

But the direction of travel matters, and it matters more here than it would elsewhere. NVIDIA's revenue is concentrated in four customers representing 61% of the total, and receivables are even more concentrated than revenue: at FY2026 year-end the largest single customer represented 25% of the receivable balance. When a small number of very large counterparties are simultaneously stretching payment terms and being extended financing by the vendor, the distinction between a sale and a financing arrangement becomes one an investor should be able to test. NVIDIA does not provide the disclosure that would allow it.

Finding two: cash conversion has fallen sharply

In the first half of FY2026, NVIDIA generated $42.8 billion of operating cash flow against $45.2 billion of net income — a conversion rate of 95%. In the first half of FY2027, it generated $74.4 billion of operating cash flow against $118.0 billion of net income — a conversion rate of 63%. In Q2 FY2027 alone the ratio was 40%.

Part of the gap is the equity gains discussed below, which are non-cash by construction. Removing them, adjusted conversion is closer to 80%. That is still a meaningful deterioration from 95%, and the residual is absorbed by the working-capital build: receivables up $24.6 billion and inventories up $10.2 billion in six months. Inventory days moved from 114 at FY2026 year-end to 119 in Q2 FY2027. Neither figure is alarming. Together they mean that a growing share of NVIDIA's reported profit is sitting on the balance sheet rather than in the bank.

Finding three: nearly a fifth of pre-tax profit came from marking up investments

In the first half of FY2027, NVIDIA recorded $23.7 billion of gains on equity securities. Against pre-tax income of $141.4 billion, that is 16.8%. In Q1 FY2027 alone the figure was 22.8%.

These gains arise from two buckets. The first is marketable equity securities — stakes in public companies, marked to market each quarter, which grew from $12.9 billion to $42.8 billion in six months. The second is non-marketable securities — private company stakes, carried at fair value on a periodic valuation basis, which grew from $22.3 billion to $51.2 billion. Over the same period NVIDIA purchased $42.4 billion of equity securities.

Two things follow. First, GAAP earnings are increasingly sensitive to movements in a portfolio that an operating analyst has no way to underwrite; a markdown in a single large private position could produce a multi-billion-dollar charge in a quarter. Second, and more importantly, the company is deploying an amount of capital equivalent to roughly a third of its operating cash flow into equity stakes in an ecosystem that purchases its products — while simultaneously borrowing $24.9 billion in the debt markets.

14 — Balance sheetFrom fortress to levered in one quarter

NVIDIA's balance sheet has historically been one of the strongest in technology: net cash, minimal debt, and a reputation for conservative financing. That description no longer fits cleanly.

The company still holds more cash and marketable debt securities — $56.6 billion — than total debt of $33.4 billion, so it remains net cash positive by roughly $23.2 billion on a narrow definition. But the trajectory is what matters. Long-term debt increased by $24.9 billion in a single quarter, a more than fourfold increase. Total liabilities grew 84% while shareholders' equity grew 46%.

It is worth stating the obvious: NVIDIA did not need to borrow. It held $49.7 billion of cash and marketable debt securities at the start of the period, generated $74.4 billion of operating cash flow during it, and spent $39.0 billion on share repurchases. A company in that position that chooses to raise $24.9 billion of debt is making a deliberate capital-structure decision, and the most plausible reading is that the proceeds fund the investment and financing platforms described above. That is a legitimate strategic choice. It is also a choice that increases the company's correlation with the creditworthiness of the AI ecosystem it is financing.

Capital allocation: buying back stock at 26 times earnings with borrowed money

NVIDIA returned approximately $26.0 billion to shareholders in Q2 FY2027 — $19.7 billion of repurchases and $6.0 billion of dividends. It has roughly $99.0 billion remaining under its repurchase authorization, which the board increased by $80.0 billion in May 2026 without expiration. The quarterly dividend was raised from $0.01 to $0.25 per share — a 25-fold increase — beginning in Q1 FY2027.

The dividend increase is a genuine signal of confidence in durable cash generation, and it is welcome. The buyback is more debatable. Basic shares outstanding fell from 24,366 million in Q2 FY2026 to 24,190 million in Q2 FY2027 — a reduction of just 0.7% over four quarters, despite $39.0 billion of repurchases in the first half alone. The reason is stock-based compensation of approximately $2.0 billion per quarter, which offsets most of the repurchase. NVIDIA is spending heavily to stand still on share count, while borrowing at corporate rates to do it.

15 — GovernanceFounder control, and a disclosure gap that matters

NVIDIA's governance is conventional in structure and founder-dominated in practice. Jensen Huang has led the company since its founding in 1993 and holds a stake of roughly 3.5%. The company has a single class of common stock, an independent board majority, and no controlling shareholder structure — all of which are points in its favor relative to many founder-led peers.

The relevant governance issue for an investor is not board mechanics but disclosure. Three specific gaps limit what an outside analyst can establish:

• Related-party revenue. NVIDIA does not disclose the share of revenue attributable to companies in which it holds an equity interest, despite having invested $42.4 billion in such securities in six months.
• Customer identity. The four customers exceeding 10% of revenue are disclosed only as anonymized accounts A through D. Their identity is inferable but not established, and the concentration of receivables is disclosed without corresponding identity.
• Segment granularity. The reorganization into two market platforms removed the visibility into gaming, professional visualization, and automotive that investors previously had, at a time when the mix within Edge Computing is changing rapidly.

None of these is unusual for a large-cap technology company, and none is evidence of misconduct. But taken together they mean that the single most important question in the investment case — how much of NVIDIA's demand is organic versus self-financed — cannot be answered from public filings. We treat that as a reason to require a higher margin of safety, not a reason to avoid the security.

16 — ValuationThe multiple has already compressed. That is the whole argument.

The conventional bear case on NVIDIA runs as follows: the stock trades at an absurd multiple, and when growth normalizes the multiple will collapse. That argument was correct in 2024. It is not correct now, and an investor who repeats it without checking the arithmetic is missing the most important fact about this security.

On annualized current-quarter earnings, NVIDIA trades at roughly 21 times. On our FY2028 base case it trades at 18 times. These are not the multiples of a bubble. They are, if anything, below the multiple of the broad US equity market. On enterprise value — approximately $5.08 trillion after netting cash — the company trades at 16.8 times trailing revenue and roughly 25 times trailing EBITDA.

This changes the shape of the debate entirely. If NVIDIA already trades at a market multiple, then the bear case cannot be multiple compression, because there is no large multiple left to compress. The bear case must be earnings compression: a decline in revenue, margins, or both severe enough to make the current multiple look expensive in hindsight. The question is not "how much will investors pay for earnings" but "will the earnings be there."

Reverse engineering the price

The cleanest way to frame this is to invert it. Rather than forecasting a value, ask what earnings the current price requires at various exit multiples.

At a 25× exit multiple, today's price implies earnings of roughly $8.44 per share — below what NVIDIA will likely earn in FY2027 and barely above FY2026's $4.90 plus one year of growth. At 20×, it implies $10.55, slightly above our FY2027 estimate. At 18×, it implies $11.72, which is essentially our FY2028 base case. In other words: the market is pricing NVIDIA as though a 20 times multiple is the correct multiple for this business. Everything else is a disagreement about what that multiple should be.

We think 20× is too low for a company with a 75% gross margin, a demonstrated ability to grow revenue at triple-digit rates, a software annuity attached to the installed base, and an interconnect position that competitors cannot currently match. But we also think 30× is too high for a company with 61% customer concentration, a self-financing ecosystem, and a capex cycle controlled by third parties. Our base case uses 24×.

Scenario valuation

The table below sets out our three cases in full. Every input is stated so that a reader can disagree with a specific assumption rather than with the conclusion.

Weighting the three cases at 25/50/25 produces a scenario value of $286 per share. We then apply a discount for time value and for the width of the outcome distribution to arrive at a 12-month target of $258, approximately 22% above the current price.

Against a street consensus 12-month target of approximately $344 across 79 covering analysts, our $258 sits meaningfully below. We are comfortable with that divergence. Consensus is anchored on FY2028 revenue estimates that assume continued acceleration; our base case assumes deceleration to 28% growth. If the hyperscaler capital cycle continues at the pace Moody's is now projecting, our bull case is the right one and the street is too conservative. If it moderates, our base case holds and the street is too optimistic. The asymmetry is the point.

17 — The bull caseSix arguments we take seriously

We present the bull case at full strength, not as a straw man. If we cannot state the strongest version of the opposing argument, our own position is not worth publishing.

18 — The bear caseSix arguments we cannot dismiss

19 — Risk matrixRanked by probability times impact

The matrix below scores each identified risk on a 1–5 scale for probability and impact. The score is the product. Risks scoring 12 or above are the ones we would expect to drive a material re-rating in either direction.

Cells read left to right as impact rises from 1 to 5, top to bottom as probability rises from 1 to 5. Superscripts identify risks from the register above. R12 is not shown; it falls below the plotted range.

20 — CatalystsWhat we are watching, and when

21 — ConclusionConstructive, at $258, with the bear case priced in

NVIDIA is not a bubble. The revenue is real, the margins are real, the cash is real, and the multiple — at roughly 21 times annualized current-quarter earnings — is approximately the multiple of the broad market. Anyone arguing that this security is priced for perfection has not done the arithmetic recently. The de-rating from the 2024 peak multiples has already occurred, and it occurred while estimates were rising.

NVIDIA is also not a clean story. Its revenue depends on the capital allocation decisions of four customers who are also its competitors. Its earnings are increasingly flattered by the appreciation of a portfolio it has assembled partly with borrowed money. Its cash conversion has fallen from 95% to 63% in a year. Its share of hyperscaler capital budgets has stopped rising, which means its growth is now capped by the growth of a pool it does not control. And its single largest market has gone to zero in a policy decision it could not influence.

The tension between those two paragraphs is the investment case. Our scenario-weighted value is $286 per share, and our 12-month target is $258, approximately 22% above the current price. We rate the shares Constructive.

22 — AppendixFinancial summary tables

A. Quarterly income statement detail

B. Key operating ratios

C. Valuation summary

23 — SourcesPrimary and secondary references

Company filings and releases

• NVIDIA Corporation, Financial Results for the Second Quarter Fiscal 2027, 26 August 2026.
• NVIDIA Corporation, Financial Results for the First Quarter Fiscal 2027, 20 May 2026.
• NVIDIA Corporation, Financial Results for the Fourth Quarter and Fiscal 2026, 25 February 2026.
• NVIDIA Corporation, Form 10-K for the fiscal year ended 25 January 2026, filed with the SEC.
• NVIDIA Corporation, Form 10-Q for the quarter ended 26 July 2026.
• NVIDIA Corporation, Quarterly CFO Commentary, Q1 and Q2 FY2027.
• NVIDIA Newsroom, press releases and technical announcements, January–September 2026, including the Vera Rubin platform launch at CES 2026.

Industry and third-party data

• Moody's Ratings, hyperscaler capital expenditure forecast revisions, 2026, as reported by Compute Forecast, 18 May 2026.
• Silicon Analysts, AMD vs NVIDIA AI GPU Market Share 2026, April 2026.
• Daloopa, NVIDIA Customer Concentration: A Big 4 Earnings Preview, based on NVIDIA Forms 10-K and 10-Q.
• MLPerf Training and Inference benchmark results, 2025–2026 submissions.
• US Bureau of Industry and Security export control framework analyses and licensing reports, 2025–2026.
• IDC and Counterpoint Research, China AI accelerator unit shipment data, 2025.
• CNBC and other financial media reporting on AI compute depreciation assumptions, November 2025 onward.

Market data

• Consensus price targets and analyst coverage counts as compiled by third-party market data aggregators, September 2026.
• Share price and market capitalization as of 15 September 2026.

24 — DisclosureConflicts, limitations, and revision policy