NVIDIA

IThe business

NVIDIA designs the chips and builds the systems that AI models are trained and served on, and it owns none of the factories. TSMC makes the chips in Taiwan. NVIDIA designs the silicon and the software, and sells the result as complete systems.³

That model explains the numbers. In the twelve months to July 2026, cost of goods was 25% of revenue, R&D 7.8% and capital spending 2.4%; operating margin was 65.2% C (calculated).³²

Revenue grew from $6.9B in FY2017 to $215.9B in FY2026, 46.6% a year. Data centers were 12% of revenue in FY2017 and 90% in FY2026.⁴¹³³

What it sells

• Compute: GPUs and CPUs, sold as boards, servers and racks. The NVL72 rack ties 72 Blackwell GPUs and 36 Grace CPUs into one machine. Data-center compute revenue was $162.4B in FY2026.³
• Networking: NVLink inside the rack, InfiniBand and Spectrum-X Ethernet between racks. Revenue was $31.4B in FY2026, up 142%.³
• Edge computing: gaming cards, workstations and car chips, reported together since FY2027. They brought in $7.2B in the quarter to July 2026, 7% of the total.²

Who pays

Since FY2027 NVIDIA splits Data Center revenue by customer type. In the quarter to 26 July 2026 hyperscalers paid $48.7B, up 102% on the year: 55% of Data Center revenue and 51% of all revenue C (calculated). AI clouds, industrial and enterprise customers paid $40.3B, up 138%.²⁵ The second group grows faster. It includes AI clouds such as CoreWeave, in which NVIDIA held a stake worth $4.7B at 30 June, and part of its purchases is supported by NVIDIA’s capacity commitments and stakes.¹⁴²

The filings measure concentration by direct customer, and NVIDIA’s direct customers include “ODMs, system integrators and distributors” that build systems for others.³ Two direct customers were 22% and 14% of FY2026 revenue; in the first half of FY2027 three were 16%, 15% and 13%.³² NVIDIA estimates that 76% of FY2026 Data Center revenue from Taiwan-headquartered customers was attributed to end customers in the United States and Europe.³ By end buyer, the hyperscaler share above is the better measure of concentration.

Now, compute is revenue.

Huang’s point: an AI company’s sales now scale with the compute it runs. That helps NVIDIA today and hurts it later. Buyers pay a premium for the fastest system available, and they also hunt for the lowest cost per token, where custom chips compete.

Cycle or utility

The case for durable demand is usage. Alphabet’s models processed more than 16 billion tokens a minute through its API in the quarter to March 2026, about 60% more than the quarter before, and 22 billion in the quarter to June.¹⁶¹⁷ Over the same two quarters NVIDIA’s quarterly revenue rose 41%, from $68.1B to $96.2B.⁴² Usage is growing faster than spending on the machines that serve it, which is what demand for a utility looks like.

The case against is Cisco. In the quarter to April 2000 Cisco’s shares closed as high as $80.06, about 222 times its diluted earnings of $0.36 a share for the fiscal year to July 2000 C (calculated).¹⁸ NVIDIA trades at about 17 times my estimate of its next twelve months’ earnings E (estimate). Its multiple is a fraction of Cisco’s. The resemblance is in its customers’ spending: Microsoft, Alphabet, Amazon and Meta spent 91% of their operating cash flow on property and equipment in the first half of 2026, up from 66% a year earlier C (calculated).¹⁹ In the third quarter alone they spent 115% of it, $201.9B against $175.5B C (calculated).^44454647

I lean toward utility for usage and toward cycle for spending. Tokens can keep compounding while capacity is bought ahead of them, as it is now. Once spending outruns the cash that funds it, spending pauses. That pause is the flat-to-down year in my base case.

Where

By customer headquarters, the United States was $149.6B of FY2026 revenue. China, including Hong Kong, was $19.7B, down from $25.0B a year earlier, and the outlook for the quarter to October 2026 assumes no data-center compute revenue from China.³⁵ Licensed H200 shipments to China were under 1% of Data Center revenue in the quarter to July 2026.² Only my bull case counts on China, at about $20B a year from year two E (estimate); my bear and base cases assume none.

NVIDIA’s fiscal year ends on the last Sunday of January, so FY2026 ended on 25 January 2026. FY2027 has 53 weeks. TTM on this page means the four quarters to 26 July 2026.²

IIThe moat

I think NVIDIA’s moat is an ecosystem: a developer network around CUDA, held in place by switching costs and paid for by scale in R&D. It is wide. It is widening at the level of the full system and narrowing at the level of the chip, among the handful of customers large enough to design their own.

CUDA dates from 2006. More than 7.5 million developers use it and NVIDIA’s tools, across hundreds of libraries and about 6,000 applications, and more than half of NVIDIA’s engineers work on software.³ Developers write for CUDA because the installed base runs it, and buyers pick NVIDIA because their software already runs on it. For the long tail, CUDA is the moat. For the few buyers who pay half the bills it binds less: Anthropic runs its models on Google TPUs, Amazon Trainium and NVIDIA GPUs at once.⁷ At that scale porting software is an engineering cost, and the advantages that bind are the system and the supply.

The last bad year shows the floor. In FY2023 revenue was flat, gross margin fell 8 points to 56.9% and return on operating capital fell from 64% to 23% C (calculated). It never went negative.⁴ Today’s figures show scarcity as much as moat: gross margin was 75.0% in the quarter to July 2026 while memory costs rose, and return on operating capital was 217% in the TTM, or 79% counting the $98.9B of equity stakes as capital C (calculated).²⁵⁴

Evidence it is widening

Networking. NVIDIA paid $7.1B for Mellanox in April 2020.¹¹ Mellanox brought about $1.3B of revenue in the year before the deal, on NVIDIA’s pro forma figures C (calculated), and NVIDIA’s networking revenue was $31.4B in FY2026, up 142%.¹¹³ Not all of that came from Mellanox: NVLink, the link inside the rack, is NVIDIA’s own. A rack of 72 GPUs only works as one computer if the links inside it keep up, so a rival now has to match the network, the CPU and the software as well as the chip.

Supply. Supply and capacity commitments rose from $95.2B in January 2026 to $279B in July, mostly for memory.³²⁵ In a market short of memory, NVIDIA has reserved capacity a rival would need. The commitments are $87B for FY2028 and $88B for FY2029, a sign that management plans FY2029 at about FY2028’s level.⁵

Cadence. NVIDIA ships a new generation every year, and Vera Rubin, the successor to Blackwell Ultra, began production shipments in the quarter that started in July 2026.³² The 2026 proxy statement claims Vera Rubin cuts the cost per token by up to ten times against Blackwell. That is a company claim, without an independent test.¹⁰

Evidence it is narrowing

The largest buyers now use more than one supplier. Anthropic agreed to use up to one million Google TPUs. OpenAI agreed to deploy 6 gigawatts of AMD GPUs, and Anthropic plans up to 2 gigawatts of AMD’s MI450.⁷²⁰⁹ Broadcom’s AI revenue grew 221% on the year to $16.7B and AMD’s data-center segment 107% to $6.7B, against 117% for NVIDIA’s Data Center revenue.⁸⁹²

Set Broadcom’s AI revenue against NVIDIA’s Data Center revenue. NVIDIA’s share of the pair fell from 89% a year earlier to 84% C (calculated), with quarters that end a week apart.²⁸ Broadcom’s figure includes the Ethernet switches it sells for AI clusters, so the series overstates custom chips. It is a small shift from a high base. I track it each quarter in the signposts table.

The moat is strongest where customers are small and weakest where they are large, and the large ones pay about half the bills.

IIIThe competitive advantage

NVIDIA’s advantage is scale in revenue and in design. In each company’s latest reported quarter, ending June to August 2026, it took in 3.8 times as much data-center revenue as Broadcom’s AI business and AMD’s data-center segment combined C (calculated).²⁸⁹

The margin gap shows the pricing power. NVIDIA’s Compute & Networking segment earned 71.0% of revenue as operating profit in the quarter to July 2026, and AMD’s Data Center segment earned 31.3% C (calculated).²⁹ AMD’s segment includes server CPUs, so the filings do not show the gap on accelerators alone.

The 10-K names the field: AMD, Huawei and Intel in chips, in-house designs at Alibaba, Alphabet, Amazon, Baidu, Huawei and Microsoft, and Arista, Broadcom and Cisco among others in networking.³

Where NVIDIA wins

NVIDIA sells a whole system and replaces it every year. Its 2026 proxy statement says Blackwell Ultra cut the cost per token 35 times against Hopper.¹⁰ The 35x is a company figure. The annual reset of the target is what matters.

Where it loses

Custom chips win on narrow, high-volume work. On the 26 August call (unofficial transcript), Huang described many of them as inference chips built for one cloud or one service.²¹ That is his defense, and it also names the risk: a cloud that serves its own model at scale can tune a chip to that model and skip the margin it would pay NVIDIA.

NVIDIA’s answer is to own more of the stack. It agreed to acquire Hugging Face for about $11.9B and pledged to keep it open, including to other chip vendors.²² The 10-K lists open models running on rivals’ hardware as a risk.³ The deal is a hedge: if models become open and portable, NVIDIA wants to own the place developers get them.

A custom chip is designed against this year’s NVIDIA system and ships against next year’s.

IVManagement

Jensen Huang co-founded NVIDIA in 1993 and has been its chief executive for all 33 years. He is 63, worked at LSI Logic and AMD before, and studied electrical engineering at Oregon State and Stanford.³ At the 23 March 2026 record date he beneficially owned 870.6 million shares, 3.58%, of which 58.7 million sat in a foundation in which he has no economic interest. His own 811.9 million shares, 3.34%, are worth about $166B at the Q4 2026 close C (calculated).¹⁰⁴⁸

Pay

Huang’s FY2026 pay was $36.3M, 129 times the median employee’s $282,050. Of his target pay, 96% was at risk, and he receives no stock that vests on time alone.¹⁰ His stock awards are earned on one year of non-GAAP operating income and vest over four years, plus a three-year award on shareholder return against the S&P 500. His cash bonus pays on revenue.¹⁰

The plan’s FY2026 stretch goals were $190.0B of revenue and $120.0B of non-GAAP operating income, cut automatically to $160.0B and $96.0B after the H20 export rules. Actual results were $215.9B and $137.3B, so the plan paid its maximum.¹⁰ In a boom, annual targets lag the business. The three-year return award is the part of his pay that tests him.

His stake matters more than his pay: each $1 move in the share price changes the value of his own shares by about $812M C (calculated). Pay “actually paid” under SEC rules, which marks his awards to market, was $344.2M in FY2025 and $162.2M in FY2026.¹⁰ Under a trading plan adopted in March 2025, he sold the 6.0 million shares it allowed between 20 June and 29 October 2025, 0.7% of his holding and about $1.1B at the prices of the time E (estimate). His Form 4 filings since then show gifts and tax withholding, and no sales.²⁵²⁶⁴⁸

The plan pays him to grow the year’s revenue and operating income, and none of its measures is cash. His own money is the counterweight: 3.34% is about $166B, and the Forms 4 show no sales since October 2025.¹⁰²⁶

The bench

Colette Kress has been chief financial officer since 2013. Debora Shoquist has run operations since 2007, and Timothy Teter has been general counsel since 2017.³ Two senior roles changed in 2026. Ajay Puri retired as head of field operations after 21 years, and Nicholas Parker joined from Microsoft to replace him, with $35.0M of restricted stock on hire.²³ Scott Gawel, formerly Intel’s chief accounting officer, took the same role at NVIDIA in May.²⁷

The board has a lead independent director, Stephen C. Neal, instead of an independent chair. It grew to 11 members when Suzanne Nora Johnson joined the audit committee in July 2026, after Persis Drell resigned in January.¹⁰²⁸²⁹ The 2026 proxy gives the full board oversight of management development and names no successor to Huang, and the 10-K lists failure to ensure effective succession planning among its risks.¹⁰³

The record of the last downturn is mixed. In May 2022 NVIDIA paid $5.5M to settle SEC charges that it had not disclosed that crypto mining drove a significant part of its FY2018 gaming growth.³⁰ In FY2023, as gaming revenue fell 27%, it took $2.17B of provisions for excess inventory and purchase obligations.³¹

Capital allocation

Huang spends to extend the platform first and returns what is left. The record holds one great deal and one failed one, and a decade of buybacks that mostly offset the shares issued to employees.⁴

The great deal was Mellanox, bought for $7.1B; networking revenue reached $31.4B in FY2026.¹¹³ The failure was Arm, abandoned in FY2023 at a cost of $1.35B.³² The decision that mattered most came in the slump: in FY2023, when operating income fell 58%, R&D rose 39% to $7.3B, and revenue more than doubled the following year.⁴

Buybacks were $96.6B from FY2017 to FY2026, yet diluted shares fell only 5.6%, from 25.96B to 24.51B C (calculated).⁴ Stock pay over those years was $22.8B, and NVIDIA paid $7.9B of employee taxes on vesting shares in FY2026 alone.⁴³ The average price paid rose from $46.41 in FY2024 to $143.22 in FY2026 and $196.28 in the first half of FY2027 C (calculated), about my base value; those 203 million shares cost $39.8B.⁴² In May 2026 the quarterly dividend went from $0.01 to $0.25, about $24B a year at today’s share count C (calculated).² Spending on the product has paid; spending on others is less proven. The first half of FY2027 put $42.4B into stakes, and the $17.0B for Groq bought no products or customers.²³

The newer habit is buying into the ecosystem. NVIDIA held $99B of equity investments at July 2026 and had committed $25B more.² In the first half of FY2027 it spent $42.4B on equity securities, 57% of its operating cash flow C (calculated).² Its 30 June filing of holdings showed Intel at $30.0B, from $5.0B paid at $23.28 a share in 2025, and SpaceX at $21.0B, a stake Techstrong.ai reports NVIDIA received for its xAI shares when SpaceX acquired xAI.¹⁴³³³⁴ It agreed to pay $17.0B for a non-exclusive license to Groq’s technology and the Groq engineers it hired, and had paid $15.9B by July 2026. The 10-K records $14.4B of goodwill and a $2.5B technology intangible, and says no products or customers were acquired.³²

Shareholder returns, stakes, employee stock taxes and Groq took $25.2B more than free cash flow in the first half, about what the June notes raised C (calculated). Free cash flow itself was held down by a $24.6B rise in receivables.² NVIDIA borrowed in the bond market while extending more credit to its customers, and kept $56.6B of cash and debt securities.²

In the first half of FY2027 NVIDIA put $42.4B into stakes in other companies and returned $45.3B to shareholders, against $69.9B of free cash flow as it defines it, and raised $24.9B of new debt.²⁵

VFinancials

Profit is growing faster than cash, and the cash flow statement shows a business that now lends to its customers.

Gross margin went from 56.9% in FY2023 to 75.0% in FY2025. It fell to 71.1% in FY2026, after a $4.5B charge for H20 chips caught by US export licensing for China, and was 74.7% in the TTM.⁴³² Operating margin went from 15.7% to 65.2% over the same span.⁴²

NVIDIA has had two bad years in the last seven. In FY2020 revenue fell 7% and operating income fell 25%. In FY2023 revenue was flat and operating income fell 58% C (calculated).⁴ Both hit a business less than a tenth of its current size C (calculated).

Earnings quality

TTM net income of $192.9B includes $30.6B of gains on equity investments, 13% of pretax income C (calculated).³² The 10-Q does not name the holdings behind the gains; the largest marks in the 30 June holdings filing were Intel and SpaceX.²¹⁴ Without the gains, at the current tax rate, earnings per share fall from $7.91 to $6.85, and the price-to-earnings ratio rises from 25.9 to 29.9 C (calculated).³²⁴⁸

Cash is arriving more slowly. In the first half of FY2027 operating cash flow was $74.4B against net income of $118.0B, 63% C (calculated), and net income included $23.7B of gains on equity securities. In the second quarter alone it was $24.1B against $59.7B.²⁵ Receivables rose 64% in six months while quarterly revenue rose 41%, and days sales outstanding went from 45 to 60 in a single quarter.²⁵ NVIDIA attributes it to payment terms of 90 days to one year on large multi-quarter agreements with investment-grade customers.²

The balance sheet

Cash and debt securities were $56.6B at July 2026 against $33.4B of debt, which includes $25.0B of notes issued in June 2026 at coupons of 4.25% to 5.625%.² Equity investments add $98.9B at carrying value C (calculated).² Off the balance sheet, NVIDIA has guaranteed up to $108.5B of other companies’ obligations and carries $366B of commitments.²⁵

VIValuation

How this goes badly wrong

Before any numbers I ask what would most likely wreck this, and work back. I see three ways.

• Customers run out of cash. The big four spent 115% of their operating cash flow on equipment in the third quarter C (calculated).^1944454647 My bear has revenue falling by a quarter over two years.
• NVIDIA props up its own demand. It has guaranteed up to $108.5B of others’ obligations and holds $98.9B of stakes in companies that also buy its products.² If they stumble, the loss comes back to NVIDIA. My model deducts nothing for the guarantees.
• The margin goes home. Gross margin was 56.9% in FY2023 and is 74.7% now C (calculated).⁴ My bear settles at 58%, my base at 63%.

None needs a surprise. Each is in the filings today.

What an owner can take out

Net income over the last twelve months was $192.9B, including $30.6B of gains on stakes. Free cash flow after stock pay and capital spending was $119.8B, 62% of net income C (calculated).³² Standing still takes little plant, only $7.4B. It takes credit: working capital rose $34.9B, and before that rise the figure is $154.6B, 80% of net income C (calculated).² So the cash an owner can count on sits between $120B and $155B, depending on whether customers keep paying on terms of up to a year.

What I can’t judge

I can’t tell you how long a GPU stays worth running. Baker says at least six years, and Huang, who sells them, says eight or longer.⁵⁵⁶⁰ Nor can I tell you whether 115% of cash flow is a peak or normal. So I show the range, $80 to $359, not one number, and the $195 is the middle of it, not a measurement.

What the price implies

NVIDIA fell 4.2% on 30 October with the chip sector: the Philadelphia Semiconductor Index lost 7.2% and the Nasdaq Composite 2.9%, the day after Intel guided fourth-quarter revenue to $15B to $16B.^48505152

Start with earnings. On the company’s outlook, NVIDIA trades at about 17 times my estimate of its next twelve months’ earnings and 13.1 times my estimate of FY2028 earnings E (estimate).⁴⁸⁵⁶ The FY2027 figure under that is mine: $177.8B in the first half, $108.0B guided for the third quarter and about $125B for a fourth quarter of 14 weeks, $411B in all. The preliminary outlook of about 70% growth takes FY2028 to about $698B E (estimate).²⁵⁶ Before the 26 August call, analysts had modelled FY2028 growth of about 44%, roughly $570B of revenue, as Implicator reported.³⁶

Then the return. At $204.64, my cash flows return 9.7% a year in the base case, 14.2% in the bull and 5.5% in the bear, a little above the 5.00% ten-year Treasury yield E (estimate).⁴⁸³⁷

A constant growth rate is the wrong shape for a company in the middle of a build-out, so the table reads the price two more ways. Take the company’s outlook for two years, $550B and $750B of revenue at my base-case margins, and the price needs free cash flow to grow 2.2% a year after that E (estimate). The second of those years includes two FY2029 quarters the company has not guided. Solve instead for one rate from TTM free cash flow after stock pay, before the $34.9B rise in working capital, and the price needs 12.5% a year for ten years, to about $504B in year ten C (calculated). At my base-case margin of 40.4% that is about $1.2 trillion of revenue E (estimate), 4.1 times the TTM.³²

At 13.1 times the earnings NVIDIA’s own outlook implies for FY2028, the price is a bet on FY2029 and after.

The discount rate

I use 10%, which is what the capital asset pricing model gives for a beta of about 1.2 C (calculated). The model with NVIDIA’s measured betas gives more: the 5.00% Treasury yield plus Damodaran’s 4.20% equity risk premium, times betas of 1.62 to 2.22, gives 11.8% to 14.3% C (calculated). Damodaran states his 4.20% against a 5.29% Treasury yield.R (reported) On that pair the same betas give 12.1% to 14.6%, and I keep 10%.C (calculated)³⁷³⁸³⁹ I would rather use a rate I would apply to any large business with net cash and put the cyclical risk into the cash flows, where it is visible. The table shows what other rates do.

Three cases

I build each case from the income statement down: revenue, gross margin, operating spending, a 17% tax rate, capital spending of 2.4% of revenue against depreciation of 1.2%, and working capital at 15% of revenue, against 14.6% of the run rate at July 2026 C (calculated).⁴²⁵ Stock pay is a cost inside those margins, as under GAAP, so free cash flow is after it.

The base case takes most of the company’s plan and then assumes a normal cycle.

• Revenue of $540B in year one, to July 2027, a little under the company’s path, and $700B in year two, the FY2028 level the 70% outlook implies, six months late E (estimate).
• Revenue 2% lower in year three, half of which is FY2029.
• Growth of 5% to 6% a year after that, as custom chips take share of new spending.
• Gross margin from 72.5% in year one to 63% by year nine, the middle of the 61.2% to 64.9% NVIDIA earned from FY2019 to FY2022, before the boom. Operating spending rises from 8% to 12% of revenue, against 12.6% in FY2025 C (calculated).⁴
• No China data-center compute revenue, as in the company’s current outlook.⁵

Why I override the FY2029 signals

Four facts point to growth in FY2029. Supply commitments are $88B for FY2029, flat with FY2028.⁵ The first phase of the SB Energy campus, where management estimates $150B to $200B of NVIDIA revenue for each generation of equipment, is expected in FY2029.⁵³⁵ In March Huang said he sees at least $1 trillion of Blackwell and Rubin through 2027.⁴⁰ On the August call (unofficial transcript), Kress put top-five hyperscaler capital spending at nearly $800B in 2026 and $1.3 trillion in 2027.²¹

My base case agrees on the level and differs on the trend. It has revenue of $700B in the year to July 2028 and 2% less in the year after. The mechanism I expect to bind is funding. The big four’s capital spending grew 84% in the first half of 2026 and their operating cash flow 33%, so spending reached 91% of cash flow C (calculated).¹⁹ The third quarter went further: spending of $201.9B against cash flow of $175.5B was 115%, from 96% in Q2 C (calculated).^44454647 Kress’s $1.3 trillion for five companies is twice the four’s first-half cash-flow run rate of about $645B a year C (calculated). Power is the second limit: $700B a year at management’s $25B to $40B of revenue per gigawatt is 17 to 28 gigawatts of new capacity a year C (calculated).²¹ Moving the down year from year three to year four changes my base value by $0.29 E (estimate).

The bear case has NVIDIA miss its FY2028 plan as customers run short of power and funding, a risk the 10-Q names.² Revenue falls 15% and then 12%, gross margin falls to 58%, close to FY2023’s 56.9%, and NVIDIA takes a $35B loss on excess memory and capacity commitments in year two, 20% of the $175B committed for FY2028 and FY2029 E (estimate).⁵⁴ In FY2023 a demand miss showed up the same way: $2.17B of provisions for excess inventory and purchase obligations, 8% of that year’s revenue, against 9% of year-two revenue in my bear case C (calculated).³¹⁴

The bull case has AI infrastructure spending reach $3.5 trillion a year, the middle of the $3T to $4T Kress expects by the end of the decade (unofficial transcript), and NVIDIA keep a third of it; its hyperscaler revenue was 30% of big-four capital spending in the June quarter, the latest NVIDIA has reported.⁴¹²¹⁹ Year-five revenue of $1.17T is 29 gigawatts a year at the $40B per gigawatt Huang gives for Vera Rubin E (estimate).²¹ The bull includes about $20B a year of China data-center compute from year two E (estimate). Weighted 25%, 50% and 25%, the three cases give $207 a share E (estimate), against a price of $204.64.⁴⁸

What closes the gap

Any one of three changes takes my base value past the price. With the plan on time, $550B and $750B in the first two years, it is $208. With year three up 5.5% instead of down 2%, it is $207. With gross margin settling at 70% instead of 63%, it is $214 E (estimate). The discount rate moves it more than any of them: at 9% it is $229, and at 9% with 4% growth after year ten, $243 E (estimate). Dating the cash flows from Q4 2026, the price date, instead of 26 July, the last balance sheet, adds $4 E (estimate).⁴⁸

The model is generous in one place. It treats NVIDIA’s stakes in its customers as free, and in the first half of FY2027 net purchases of equity securities were $35.2B, half of free cash flow.² If strategic investment ran at 20% of free cash flow for five years and proved worth half its cost, my base value would be $190. At 5% of revenue every year, written off, it would be $172 E (estimate). The model adds the $98.9B of stakes at carrying value, $4.07 a share, and deducts nothing for the $108.5B of guarantees; a 10% chance of paying out all of them would cost $0.45 a share E (estimate).²

The margin of safety

• Price$204.64Close, Q4 2026 [48]
• Base value$195My estimate E (estimate)
• Margin of safety below$106set from my bear case C (calculated)

I set the margin of safety from the bear case instead of as a discount to base value, because one cycle drives this business. Below $106, 46% under my base value, my bear value of $80 is at least three quarters of the price paid, and my base case would return about 16% a year E (estimate). That is about 9 times my estimate of next-twelve-month earnings. The bear case is built on FY2023, when gross margin fell 8 points and operating income 58% C (calculated).⁴ I do not cut the value again for the cycle or for customer concentration, because both are in the cash flows. The cushion between the bear value and the margin-of-safety price is for what the model does not hold: a Taiwan supply shock, a change in China policy, the founder, and my own error. The price sits 93% above the margin-of-safety price.

Where this could be in 5 and 10 years

For the base to stand at 5 years, the plan has to arrive about six months late and free cash flow has to reach $354B in year 5. For it to stand at 10, growth has to settle at 5% to 6% a year, gross margin has to return to 63%, and free cash flow after stock pay has to reach $407B, a 40% margin E (estimate). My base value goes from $195 today to $226 in 2031 and $267 in 2036, and the owner cash paid on the way is $146 a share.

The bear and the bull differ on revenue and margin. In the bear, revenue is $372B in year 5, gross margin settles at 58% and free cash flow is $129B; from $204.64 that loses 2.9% a year. In the bull, revenue is $1,170B in year 5 and free cash flow is $584B. Capital needed is the same in all three: capital spending of 2.4% of revenue and working capital of 15%. The discount rate is 10% in all three; at 9% my base value today is $229 E (estimate).

The most likely way the base is wrong is margin. I have gross margin falling from 72.5% to 63%. In FY2023 it fell 8 points in a year. Nearly all of the year-10 value is cash after year 10, so it is the part I know least.

The next best use of the same dollar is another business here. Broadcom’s base case returns 7.5% a year against 9.3% for NVIDIA, and the ten-year Treasury pays 5.00% E (estimate).³⁷ The business is excellent. The price is a separate question: it sits 5% above my base value and 93% above the margin-of-safety price.

The price needs only one of three things: the plan on time ($208), no down year ($207) or a 70% margin ($214). The bull needs all of year-five revenue of $1.17T, China sales and a gross margin above 70%, and at two chances in three each, all three come to about 30% E (estimate).

These are estimates of value, not forecasts of the share price, and the market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Old chips are renting for more. Silicon Data’s daily index priced an H100 at $2.00 an hour on 9 December 2025 and $2.90 on 30 October 2026, up 45% C (calculated).⁵³⁵⁴ Those chips are nearly four years old, Gavin Baker noted in February.⁵⁵ Newer ones ought to push the rent down. They have not, which suggests buyers still cannot get enough computing power. The index reads prices at smaller AI clouds and GPU marketplaces, so I trust the direction more than the cents.

The 70% outlook is also a limit on supply. On the August call (unofficial transcript), Colette Kress said customers’ forecasts point to growth doubling, and that NVIDIA expects about 70% because supply stays short at least through FY2028.²¹ Price increases already made take effect in Q1 of FY2028, she said, lifting gross margin to 72% to 73% after a Q4 low of 71% to 72%.²¹

The bull case

Baker makes the strongest version. On 25 February he wrote that the market was pricing NVIDIA near a peak in earnings, ahead of too much capacity being built. His answer: the world is short of power and chip wafers, that may take years to fix, and it may prevent an overbuild.⁵⁵ He added that six-year-old A100s were still fully used, per AWS, so a GPU may last at least six years, longer than most customers take to write it off.⁵⁵ Jensen Huang, the seller, said eight years or longer in October, Benzinga reported.⁶⁰ His firm, Atreides, held 1.25M shares, worth $250M, at 30 June, under 2% of its reported book C (calculated).⁵⁹

“I think Nvidia will be the first $10 trillion company,” Brad Gerstner said on CNBC in May. He said he could wait two to three years.⁵⁶⁶² In September, on the All-In podcast, he pointed to 14 times next year’s fully taxed earnings.⁵⁷⁶³ NVIDIA was Altimeter’s largest reported position at 30 June: 9.41M shares, worth $1.88B.⁵⁸ BNP Paribas raised its target to $345 in early October, GuruFocus reported, with NVIDIA keeping at least 75% of AI computing by dollar value.⁶¹ Bloomberg’s 14 October headline has Cathie Wood favoring NVIDIA and AMD over memory makers.⁶⁴ ARK reported 1.38M shares at 30 June, up from 1.04M in March.⁶⁵

Against reality

The price asks for less than these investors do. If NVIDIA ships $550B in the year to July 2027 and $750B in the next, my base value is $208. If it skips the down year instead, $207. Both, $221. With gross margin settling at 70%, not my 63%, $214 E (estimate). So if they hit the plan on time, or keep growing in year three, or keep the margin, the price is possible. Gerstner’s $10 trillion is about $412 a share C (calculated), above my $359 bull, which already needs $1.17 trillion of revenue in year five.

The last four quarters brought in $303.0B. Q3 is guided at $108.0B, and I have about $125B for Q4.⁵ That leaves the two quarters to July 2027 averaging about $159B, 65% above Q2’s $96.2B C (calculated).² Gross margin is 75.0%, guided to 71% to 72% in Q4, so 70% is in reach for now.²¹ The test comes once supply catches up and custom chips take share. Funding is furthest away. The big four spent 115% of their operating cash flow in Q3 C (calculated),^44454647 and Kress’s $1.3 trillion for five companies in 2027 is about twice the four’s first-half cash flow, annualized.²¹¹⁹ A shortage explains why they cannot overbuild. It does not say who pays.

Where I come down

The rents and the supply-limited outlook moved me: the down year in my base looks less certain. They did not move my $80, $195 and $359, because the funding test is getting worse. FY2029 supply commitments rising, gross margin at the top of its guide, 73%, in FY2028, and big-four spending back under 90% of cash flow would raise my values.

VIIIRisks, and what would change my view

1. The cycle. Microsoft, Alphabet, Amazon and Meta spent $295B on property and equipment in the first half of 2026, up 84% on a year earlier and 91% of their operating cash flow C (calculated).¹⁹ In the third quarter they spent $201.9B, 115% of cash flow C (calculated).^44454647 NVIDIA sells the new capacity, so when they slow I expect its revenue to fall first and furthest.
2. Customer financing. NVIDIA has signed residual value guarantees, capped at $105B, on OpenAI’s 20-year leases at SB Energy’s Ohio campus. If OpenAI defaults, NVIDIA covers the gap between a guaranteed lease value and what re-letting or a sale recovers, and OpenAI must reimburse it. Each guarantee starts when its lease begins, with the first sites expected to be ready in 2028, and all of them end if OpenAI earns a satisfactory credit rating.³⁵² The cap is about a third of a year’s revenue C (calculated). NVIDIA has also committed $36B to purchase capacity from AI clouds that procure its systems, capacity they may instead resell to others; it gives some customers up to a year to pay, and holds $99B of stakes in companies that also purchase its products.² The disclosure is unusually complete. It also shows that part of today’s demand rests on NVIDIA’s own balance sheet.
3. Custom chips and AMD. Broadcom’s AI revenue and AMD’s data-center revenue together equal 26% of NVIDIA’s Data Center revenue C (calculated). Broadcom’s figure includes AI networking and AMD’s includes server CPUs. OpenAI and Anthropic have both signed multi-gigawatt deals with AMD.⁸⁹²²⁰
4. Memory. The CFO described memory pricing as extreme.⁶ On the 26 August call she guided gross margin to a low of 71% to 72% in the quarter to January 2027, a figure I have from an unofficial transcript and a secondary report.²¹³⁶ Memory is also where my bear case takes its write-down: most of the $279B of supply commitments is memory bought at today’s prices.⁵
5. China and regulators. The FY2026 H20 charge was $4.5B. China’s market regulator made a preliminary finding in September 2025 that NVIDIA breached the terms of its Mellanox approval, and regulators in the EU, the US, the UK and South Korea have asked for information.³²
6. Concentration of supply. TSMC makes the chips in Taiwan.³
7. One founder. Huang has set the strategy for 33 years, and the company has no history without him.³

NVIDIA short report: Culper Research, “The China Problem”

Who and what. Culper Research is a short seller. It says “We are short Nvidia”, so it gains if the price falls.⁷⁸⁷⁹ On 13 May 2026 it published “NVIDIA (NVDA): The China Problem”.⁶⁶⁷⁸⁷⁹ I could not open it: the site blocks automated access and I found no usable archive copy. This is as described by Newsquawk, Benzinga and Fortune, so it is secondhand.⁷⁸⁶⁷⁷¹

What Culper says. More than 20% of NVIDIA’s FY2026 compute revenue was driven by Chinese demand, through diversion and Southeast Asian intermediaries.⁶⁷ At 20% of $162.4B of compute revenue, that is about $32.5B, 15% of the year’s $215.9B C (calculated).³ Culper says Alibaba finances Megaspeed, a Singapore cloud firm, through shell companies.⁶⁷ It calls the route in the March 2026 Super Micro indictment “one of many”.⁷¹

NVIDIA’s response. I found no reply to the report by name; Benzinga said it had asked NVIDIA for comment and printed none.⁶⁷ In October 2025 NVIDIA said its own inquiry into Megaspeed found no “reason to believe products have been diverted”.⁷⁵ On 29 October 2026 a spokesperson said the diversion allegations against Megaspeed were “thoroughly investigated and debunked last year”.⁷⁰ The 10-Q is general risk language, not a reply: it says reports of diversion can hurt NVIDIA “even when unsubstantiated and untrue”, and that it is “effectively foreclosed” from China’s data-center compute market.²

Established and not. A Southern District of New York indictment, unsealed on 19 March 2026, charges three people tied to Super Micro in an alleged scheme to send about $2.5B of servers to China. Neither Super Micro nor NVIDIA is charged.⁷²⁷³ These are charges, not findings, and I could not find the docket number. Singapore charged individuals and four companies from 1 July 2026 in a separate server-fraud case; whether the servers held export-controlled chips is unconfirmed.⁷⁴ Singapore police said in October 2025 that they were investigating Megaspeed, and the New York Times reported a US Commerce inquiry too, citing unnamed officials.⁷⁵ I found no charge against Megaspeed or NVIDIA, and no BIS or Commerce action against NVIDIA.

On 8 November 2026 NVIDIA’s Board of Directors issued a press release announcing an independent investigation after reports that Huang “has been the target of a foreign intelligence effort”. A committee of independent directors, assisted by outside counsel, will oversee it. The board says it has no reason to believe national security or NVIDIA secrets were compromised.⁶⁸ The release names only NVIDIA and Huang, not Megaspeed or Culper, and EDGAR shows no Form 8-K since 3 September.⁶⁸⁶⁹ Press coverage links the reports to the CEO’s contacts with a Megaspeed executive; the New York Times cited unnamed officials, per The Hill, and nothing filed or ruled confirms it.⁷⁰⁷¹ Culper’s May report also called the CEO “uncomfortably close” to Megaspeed’s leadership, per Benzinga.⁶⁷ I found no primary source tying the inquiry to Culper’s report.

Where it touches my model. My base and bear cases have no China data-center compute; my bull has about $20B a year from year two. NVIDIA reports China by customer headquarters: $19.7B in FY2026, 9.1% C (calculated). Culper’s claim is about where demand ends up, so the figures do not conflict. If Culper is right, part of the bull’s $20B may already sit in base revenue.

I re-ran my published paths in scratch; the page’s model is unchanged. Each 1% of revenue is about $1.90 a share in the base case E (estimate). If NVIDIA is right, nothing moves. If Culper is right as a flat $32.5B a year, the base is $188 and the weighted value $200. Taking Culper’s “more than 20%” as exactly 20%, that is 15% of revenue every year: the base is $167 and the weighted value $177, against $207 E (estimate). With gross margin held and spending fixed, 15% less TTM revenue cuts operating income 17% and operating margin from 65.2% to 63.5% C (calculated). A closed channel could bring a one-off charge; H20’s was 2.1 points of FY2026 gross margin C (calculated). The CFO said supply stays short through FY2028 (unofficial transcript), so lost sales might be resold.²¹

So what

The question for a reader valuing NVIDIA is how much revenue ends up with Chinese users: NVIDIA’s outlook assumes none in data-center compute, and Culper says over 20% of compute revenue, about 15% of the total. At Culper’s 20% the base case is about $167 a share against $195, before any charge or fine. Culper is right if chips sold elsewhere reached China at that scale and were not resold; NVIDIA is right if its checks would have caught diversion at that scale and found none. Culper gains if the price falls, and management and insiders are paid in the stock; I could not see Culper’s method or the results of NVIDIA’s checks, and no court or regulator has found NVIDIA’s sales diverted or cleared them. Watch 17 November, when NVIDIA reports the quarter to 25 October and gives its next outlook; the last one assumed no China data-center compute.⁷⁶ Three House chairs asked Huang for a closed-door briefing on 18 November, and the board inquiry has no date.⁷⁷

NVIDIA reports the quarter to 25 October on 17 November, after this page’s figures were set. It guided $108.0B of revenue and a 74.0% gross margin for it [5][43]. The base, bear and bull values above rest on the July balance sheet and do not change until it is out. I will grade that report against the thresholds in the table, as written.

What would raise my base value: gross margin at 73% or more in FY2028, the top of the company’s 72% to 73% guide, AI clouds and enterprises outgrowing the hyperscalers without new NVIDIA financing, and FY2029 supply commitments rising. What would lower it: days sales outstanding above 75, any new guarantee, big-four capital spending above their operating cash flow for two quarters, or rivals’ combined AI revenue passing 35% of NVIDIA’s Data Center revenue.²⁵¹⁹