Semiconductors · #1 in the Retail 100

NVIDIA

NVDA

Paying for the AI build-out↓The AI spending bet↑The robotaxi race↑China's cheap AI models↓

Our rating: buyhigh conviction
12-month target: $293.14 +41.8%
Expected vs. S&P 500: +23.6 ptsSee the math
Next earnings: Nov 17In 12 daysEarnings calls page ↗

NVIDIA now guarantees up to $105B of OpenAI's data center leases, so it carries the credit risk of one of its biggest buyers.

Summary

Forget whether AI demand is real: NVIDIA still can't make chips fast enough, and it ships a faster generation every year. What matters for the stock is the 75% gross margin, meaning NVIDIA keeps 75 cents of each sales dollar after the cost of making it. Two things threaten that margin. The cloud giants that buy most of its data center chips are designing their own, and memory is getting more expensive. NVIDIA wins as long as its systems stay the best use of scarce data center power.

$206.69

+$24.05 (+13.17%)Past year

What NVIDIA does

NVIDIA designs the chips, networking and software most AI runs on, with about 86% of AI chip sales in 2025. Its CUDA software makes switching costly, so NVIDIA keeps 75% of sales after production costs.^1,2,33,34,44

What it sells

Systems for cloud giants51% of revenue, Q2 FY2027
Whole AI racks for Amazon, Microsoft, Google, Meta and Oracle. Sales doubled from a year earlier, to $48.7B last quarter.
AI clouds and enterprises42% of revenue, Q2 FY2027
The same racks for AI labs, rental clouds, companies and governments. The fastest-growing part, up 138%, and management expects it to reach half.
PCs, cars and robots7% of revenue, Q2 FY2027
GeForce gaming cards, where NVIDIA has led for decades, plus workstation, car and robot chips. Growing 27%, far slower than data centers.

Who buys it

The five biggest cloud and internet companies buy about half, and one unnamed direct customer was 16% of revenue last quarter. A few buyers' spending plans set NVIDIA's sales, which is why their capex forecasts move the stock.

Where you've seen it

GeForce graphics cards in gaming PCs carry NVIDIA's name, while its AI systems sit unseen in cloud data centers.

How it got this way

Huang bet early that graphics chips could do any math done in parallel. NVIDIA released CUDA in 2006 so programmers could use them that way, and in 2012 AlexNet, trained on NVIDIA chips, kicked off modern AI. It bought Mellanox in 2020 for about $6.9B, which gave it the networking that ties thousands of chips together. Since then it has shipped a new data center design every year.^34,43

What a customer buys

NVIDIA sells a $5 million to $8 million computer the size of a refrigerator, not a chip

A Vera Rubin rack holds 72 of NVIDIA's AI chips and 36 of its own processors, wired together so they work as one giant computer. NVIDIA also supplies the switches that link racks into a data center and the software that runs on top. A buyer gets one system from one vendor.

That is the business model in one sentence. A rival with a good chip still has to match the links, the networking and the software, which is why NVIDIA can charge for the whole package.

The cost of that package is shifting. Wells Fargo estimates memory is about 31% of what it costs to build a Rubin rack, and memory makers plan to raise prices more than 50% next year. NVIDIA designs the rack, but it buys the memory.^{38,39,40,16,3,2}

Blackwell rack price: About $5M
Rubin rack estimate: About $7.8M
Revenue per gigawatt: $18B → $40B

Part	Per rack	What it does
Rubin GPU	72	Does the AI math. Each carries 288GB of HBM4, the fast memory stacked beside the chip.
Vera CPU	36	NVIDIA's own processor, feeding data to the GPUs. CPU sales topped $5B over the past year.
NVLink switches	One set	Lets all 72 GPUs share memory and act as a single chip.
InfiniBand or Spectrum-X	Per data center	Networking that ties thousands of racks together, from the Mellanox purchase.
CUDA and AI Enterprise	Every GPU	The software developers build on. AI Enterprise lists at $4,500 per GPU a year.

Where we landSelling the whole rack is the advantage that matters most after CUDA. It is also why memory is the near-term risk: NVIDIA now has to pass a third of the rack's cost through to buyers, and the January quarter will show whether it can.

Who NVIDIA depends on

NVIDIA owns no factories, and it has promised suppliers $279B to make sure it gets parts

Every NVIDIA chip is made by TSMC in Taiwan, which also stacks the chip and its memory into one package, a step called CoWoS. The memory comes from SK hynix, Micron and Samsung. Foxconn and Wistron assemble the racks. NVIDIA designs everything and makes nothing.

In a shortage, that only works if you get first call on the scarce parts. NVIDIA reportedly booked more than half of TSMC's 2026 packaging capacity. Its supply commitments, mostly for memory, rose to $279B from about $46B a year earlier.

Those commitments cut both ways. While demand outruns supply, they lock rivals out of parts. If demand ever drops, NVIDIA still owes the money, and $279B is nearly three quarters of revenue at today's pace.^{2,35,33,34,41}

Supply commitments: $279B
Inventory: $31.6B
Where chips are made: Taiwan, Korea

Where we landToday the commitments are an edge: they are why a rival with a good design can still be short of parts. We'd worry if NVIDIA kept raising them while the cloud giants cut capex plans, because then NVIDIA carries its customers' risk.

The backstory

AI companies want more computing than exists, and the real limit is electricity. That sets up the threat: NVIDIA's biggest customers, about 55% of data center sales, are designing their own chips to cut their NVIDIA bill. NVIDIA's answer is a faster system every year that does more work with each unit of power. While power stays scarce, the most efficient system is the one buyers pay up for.

How the business makes money

NVIDIA charges a premium because it sells the scarcest thing in AI: work per watt^1,2,3,13

Step 1
AI companies need far more computing than exists. The five biggest cloud companies are expected to spend nearly $800B on data centers in 2026.
Step 2
Electricity is the real limit, so buyers pay most for the system that does the most AI work per watt. NVIDIA argues that's its own.
Step 3
NVIDIA designs the chips and pays others like TSMC to build them, so it keeps about 75¢ of every $1 of sales after the cost of making them.
Step 4
That cash pays for a new chip design every year, which keeps customers from switching to chips they build themselves.
What shareholders get
Spare cash goes back to shareholders: $26B last quarter.

Advantages

CUDA: 20 years of lock-in

Major advantage

NVIDIA's biggest advantage. NVIDIA launched CUDA, the software developers use to program its chips, in 2006, and it runs only on NVIDIA chips. Most AI software is built on it, so moving to a rival means rewriting code. Only the largest labs can afford that, and they are exactly the customers now signing deals with rivals.^3,9

Rivals have to copy the whole system

Strong advantage

Hard to copy. NVIDIA designs the chips, the links between them and the networking as one system, so a rival has to match all three at once. Networking is where it's still catching up: in AI Ethernet switches it ranks second, behind Broadcom-based gear from Celestica.^3,42

First claim on scarce parts

Strong advantage

Strong. NVIDIA gets first call on high-bandwidth memory and TSMC's advanced packaging, the scarcest parts in AI chips, and has committed $279B to lock in supply, mostly memory. In a shortage, a rival can have a good design and still not get the parts to build it.^2,16

Its deals help, but anyone with cash can do them

Supporting

Useful, but not a lasting edge on its own. NVIDIA uses its cash to buy technology, including a Groq license reported at $20B. It also invested in Intel and agreed a $12.9B deal for Hugging Face.^13,14,17,31

What's changing

Rubin is shipping, and memory just got expensive

Blackwell still brings in most sales, but its successor, Vera Rubin, is now in full production for the big cloud companies. Rubin runs on a new memory, HBM4, whose price rose faster than NVIDIA's finance chief expected. That is why NVIDIA expects margin to dip this quarter and next before it recovers. The January quarter will show whether it can pass the cost on.^1,2,3,15,16

Competitors

Who else an investor would consider, and how they compare.

Company	Market value	What it does	Compared with NVIDIA
NVIDIAthis stock	$5.08T	NVIDIA designs the chips, networking and software most AI runs on, with about 86% of AI chip sales in 2025. Its CUDA software makes switching costly, so NVIDIA keeps 75% of sales after production costs.
Advanced Micro DevicesAMD	$867.61B	Makes AI GPUs and server CPUs; the only other GPU maker at scale.	AMD is the second-source GPU; NVIDIA sells $89B of data center chips to AMD's $6.7B.^1,9,32
BroadcomAVGO	$1.57T	Designs custom AI chips for one big customer at a time, like Google's TPU.	Broadcom builds chips each giant designs for itself; NVIDIA sells one chip to everyone.^1,8
Alphabet (Google TPU)GOOGL	$3.88T	Designs its own TPU AI chips and rents them through Google Cloud, including to Anthropic.	TPUs mostly run inside Google Cloud; NVIDIA chips run on every cloud and on-site.^10,11
Amazon (Trainium)AMZN	$2.69T	Makes its own Trainium AI chip for AWS customers, including Anthropic.	Amazon builds chips to cut its NVIDIA bill, yet adds about 2 million NVIDIA GPUs.^3,22,23
IntelINTC	$478.55B	Long the PC chip leader, now testing a cheaper chip for running AI models.	Intel has no AI chip in volume yet; NVIDIA owns $5B of it and partners.^17,21

Leadership

Who runs the company, and how the stock has done under the CEO.

NVDA 5,532xS&P 500 +895.6%Total return with dividends since listing in Jan 1999 · log scale

2000200420082012201620202024

Jensen HuangFounder, President and CEO33 yrs, since 1993

Colette KressEVP and CFO13 yrs

Nick ParkerEVP, Worldwide Field Operations3 mo

Debora ShoquistEVP, Operations17 yrs

Wall Street on Huang: PositiveMorgan Stanley and BofA both rate the stock a Buy. Morgan Stanley's main worry is getting enough power and parts; BofA stayed positive after NVIDIA agreed to pay up to $105B of OpenAI's data center rent if OpenAI can't.^{L9,L11,L12,L13}

Leadership sources (13) · researched Oct 2026

L1Executive bios · NVIDIA Newsroom, 2026-10-08
L2Jensen Huang, Founder and CEO · NVIDIA Newsroom, 2026-10-08
L3Nvidia · Wikipedia, 2026-10-08
L4Colette Kress, EVP and Chief Financial Officer · NVIDIA Newsroom, 2026-10-08
L5Nick Parker, EVP, Worldwide Field Operations · NVIDIA Newsroom, 2026-10-08
L6NVIDIA Form 8-K: Nicholas Parker appointed EVP Worldwide Field Operations; Ajay K. Puri to retire · Stock Titan (SEC filing), 2026-07-02
L7Debora Shoquist, EVP, Operations · NVIDIA Newsroom, 2026-10-08
L8NVIDIA Q2 fiscal 2027 earnings call transcript · NVIDIA Investor Relations, 2026-08-26
L9Morgan Stanley Has Strong Message For Nvidia Stock Investors · GuruFocus, 2026-10-02
L10NVIDIA at Goldman Sachs conference: Huang sees AI buildout still early · Investing.com, 2026-09-10
L11Nvidia offers investors 'compelling valuation,' trading at significant discount to FCF: BofA · Seeking Alpha, 2026-08-18
L12BofA sees Nvidia stock climbing 55% on cheap valuation and buyback potential · PrimeXBT, 2026-08-18
L13OpenAI data center deal with Nvidia comes in $145 billion lower than reported, signaling concerns of artificial demand for chips · Fortune, 2026-08-18

Bull vs. bear: two bets on the same stock

1
A share is a small slice of the company. Its value comes from the profits the company will make in the future.
2
Nobody knows those profits yet. Buyers (bulls) pay today's price betting they'll beat what most people expect. Sellers (bears) bet they'll fall short.
3
When results come in, the price moves toward whichever side was right, and that side makes money.

The question they disagree on

Can NVIDIA keep its 75% margin while its biggest customers build their own chips and memory costs rise?^{1,3,4,8,9,10,12,20,32}

Bull case: why buyers buy

Demand still outruns supply and each chip generation earns more than the last, so a few quarters of thinner margins won't slow the growth that drives the stock.

Demand isn't the problem. NVIDIA says supply is what limits its forecast, and CEO Jensen Huang sees $1 trillion in orders through 2027.
Each generation earns more from the same power. Huang estimates Vera Rubin will bring in about $40B per gigawatt of data center power, more than double Hopper.
The buyers' budgets keep growing. NVIDIA says the five biggest cloud companies are expected to spend $1.3 trillion on infrastructure in 2027.
Sales are still more than doubling. Data center sales rose 117% from a year earlier last quarter.

Bear case: why sellers sell

Every dollar big buyers move to their own chips or AMD gives them more bargaining power over NVIDIA's prices, while pricier memory squeezes margin at the same time.

Custom chips are growing faster than NVIDIA. Broadcom, which designs custom AI chips for companies like Google, more than tripled its AI revenue from a year earlier.
The biggest AI labs are hedging their bets. Anthropic signed for several gigawatts of Google and Amazon chips, and OpenAI and Meta each signed large deals with AMD.
The margin is about to fall. NVIDIA expects pricier memory to cut its gross margin to 71–72% in the quarter ending in January.
China is gone for now. NVIDIA expects no data center sales there.

Where Delphi landsbuyWe think the bull case is more likely than today's price assumes.See why

NVIDIA as its customers' banker

NVIDIA is now funding and guaranteeing its own customers, and the sales numbers don't show how much

NVIDIA holds $99B of stakes in other companies. Its private stakes alone rose to $47.9B from $3.8B a year earlier. The biggest is $30B in OpenAI. It has also committed up to $10B to Anthropic, and $2B each to CoreWeave and Nebius, AI clouds that rent out NVIDIA chips. Most of these companies buy NVIDIA systems.

Critics call it circular: NVIDIA invests, the customer buys chips, NVIDIA books the sale. Short sellers Michael Burry and Jim Chanos have both made the case. NVIDIA's answer is that young AI labs need financing their balance sheets can't yet support, and that it doesn't rely on vendor financing, meaning lending customers the money to buy from you.

This year the support went past equity. NVIDIA guarantees up to $105B of OpenAI's 20-year leases on an Ohio campus that will run only NVIDIA chips. It has promised $36B to rent back capacity from AI clouds it sells to. And customers now take 60 days to pay, up from 45.

Profit carries some of it too. Gains on these stakes added $7.8B to last quarter's $59.7B of net income, mostly on paper. NVIDIA doesn't say how much revenue its investees bring in.^{2,33,46,47,48,49,50,51,52,53,54,55,60,61}

Stakes in other companies: $99B
Guarantees: Up to $108.5B
Capacity it rents back: $36B
Days to get paid: 45 → 60

Where we landIt matters, but for 2028 more than next year. Anthropic's run rate passed $65B, so the labs are growing. OpenAI is the risk: it forecasts about $278B of cash burn through 2030. If it can't raise money, NVIDIA loses a top buyer and owes rent.

What to watch

The numbers that will settle the debate.

Data center sales▲ improving

$89.0B last quarter, up 117% from a year ago^1,2

The simplest test of whether AI buyers still choose NVIDIA.

Gross margin▼ worsening

75% last quarter; 74% expected this quarter and a 71–72% low next quarter^1,3

The number that matters most. Rival chips or memory costs show up here first.

Broadcom AI sales■ unclear

$16.7B last quarter, more than triple a year ago; $21.7B expected next quarter⁸

Broadcom makes Google's AI chips, so its growth shows how fast big customers are moving to their own.

Next earnings report

Nov 17

In 12 days

Earnings calls page ↗

Expected earnings per share: $2.47
Expected revenue: $108.80B

What to listen for: Gross margin, above everything else. A result near 74%, with 71–72% confirmed as the low, would show Rubin can cover pricier memory.^1,3,18,19

What would prove this wrong

✕The largest cloud providers cut 2027 capital spending below 2026 levels.
✕Gross margin stays at 71–72% after the January quarter: NVIDIA can't pass memory costs on.
✕A top AI lab moves most new capacity to TPU, Trainium or AMD.
✕Vera Rubin is delayed by HBM4 or packaging, and MI450 or custom chips fill in.
✕OpenAI can't raise money, leaving NVIDIA owing its Ohio lease guarantees.

Our ratingBuyHigh convictionas of Nov 3 close, $206.69 · what ratings mean

Earnings are expected to grow about 39% next year, yet the stock trades at 17 times expected earnings, half its five-year norm. The market is pricing a sharp slowdown.

12-month target: $293.14 (+41.8%)
S&P 500, same method, with dividends: +18.5%
Expected vs the market: +23.6 points

The math behind the recommendation

Three 12-month price scenarios, weighted by probability, compared with the S&P 500 under the same method.

Now $206.69

Target $293.14

bull case30% likely

$359.50 +74%

Vera Rubin, NVIDIA's next-generation chip, ships on time, margin returns to about 75%, and custom chips gain share slowly, so earnings hit the top of forecasts.

$20.72 earnings per share × 17.4 = $359.50

base case50% likely

$298.97 +45%

Earnings match analyst forecasts as margin recovers to about 73% and custom chips win some share.

$18.95 earnings per share × 15.8 = $298.97

bear case20% likely

$179.00 -13%

Cloud giants shift more work to their own chips and AMD, margin stays near 71%, and earnings fall well short of forecasts.

$15.35 earnings per share × 11.7 = $179.00

Why these odds: Analysts raised fiscal 2028 estimates 24% in 90 days, to $15.80, with no cuts last month, and NVIDIA has beaten estimates four quarters in a row.

How the prices are built: Wall Street expects earnings per share of $18.95 for the year starting a year from now (analysts range from $11.75 to $22.48). Today the stock trades at 15.2 times next year's earnings per share; its five-year middle is 35.1. Because earnings per share is forecast to grow fast, valuations are cut 9%, since fast growth rarely keeps its price.^{1,4,25,26,28,29}

What tests the call: Fiscal Q3 2027 earnings: whether gross margin holds 71 to 72% as the low (Nov 18)

Buy means expected to beat the S&P 500 by 10+ points over 12 months; Sell means expected to trail it by 10+. Inputs gathered Oct 5; the recommendation updates with the price every day. Model-generated, not personalized investment advice. How ratings work · Track record

What the price assumes

At $206.69, investors pay less for each dollar of NVIDIA's earnings than for the average stock's

NVIDIA trades at 15 times the earnings analysts expect over the next 12 months. That ratio is the multiple: what investors pay today for each dollar of a year's earnings. That is cheaper than it has been 80% of the time over the past five years, when the middle was 35.

Analysts expect earnings to grow about 39% the year after, against about 16% for the S&P 500. The S&P 500 trades at 19 times. At the market's multiple, today's price would only make sense if NVIDIA earned about 19% less than forecast over the next year.

NVIDIA multiple: 15×
S&P 500 multiple: 19×
Base case: +45%
Bear case: −13%

The skewOur bear case loses about 30¢ for every $1 our base case makes. That ratio, how much you stand to lose against how much you stand to make, is the skew.

What margin is worth

The margin drop everyone is watching costs less than one quarter of NVIDIA's growth

Gross margin is the share of each sales dollar left after paying to make the product. NVIDIA's was 75% last quarter. It expects pricier memory to push it down to 71–72% in the quarter ending in January.

On a $108B quarter, each point of margin is worth about $1.1B of gross profit. A drop from 75% to 71.5% costs about $3.8B a quarter.

Now set that against growth. NVIDIA forecasts revenue rising from $96.2B to $108B in a single quarter. At a 72% margin, those extra sales bring in about $8.5B of gross profit, more than twice what the margin drop takes away.^2,1,3

One margin point: About $1.1B
The expected drop: About $3.8B
One quarter's growth: About $8.5B

Where we landIgnore the margin dip while sales grow this fast. Margin matters if it stays near 71% and growth slows at the same time, because that is what would happen if customers move to their own chips. Watch the two together.

Where the market is wrong

The price treats NVIDIA's boom as nearly over, and we think the market has the timing wrong

A low multiple usually means slow growth. NVIDIA's earnings are expected to grow far faster than the market's, yet investors pay less for each dollar of them. That is how they price chip companies at the top of a cycle, when they expect profits to fall.

The worry is fair. Chip demand has always come in cycles, and NVIDIA's own history includes sharp drops. Where we differ is timing. NVIDIA says supply, not demand, limits its forecast, and Huang sees $1 trillion of orders through 2027.^12,25,29

Where we landThis is why we'd own it: growth well above the market's, priced as if it were about to stop. The call is wrong if the cloud giants cut their 2027 capex plans.

Last earnings

No reported quarters available.