Nvidia Stock Crashed — Understand It Better Than 99% of Investors
Get smarter on investing, business, and personal finance in 5 minutes.
If you were forwarded this email, click here to subscribe for freeStock Breakdown
This newsletter is adapted from a recent video Drew Cohen did on Nvidia.
NVIDIA has a long history of looking years ahead of where its business actually was.
Back in 2016, the company was already talking about how its GPUs would uniquely benefit from emerging end markets such as AI, autonomous vehicles, and robotics.
A decade later, that bet on AI looks correct.
Revenue at the time of that 2016 prediction was about $5 billion, with less than $1 billion of profit.
A decade later, revenue has topped $300 billion and profit has reached $200 billion, and the stock has risen roughly 100xover that same stretch.
Going back to NVIDIA's 1991 IPO, the stock stands as one of the strongest performers of all time.
A $1,000 investment at IPO would be worth about $10 million today.
What has become genuinely open-ended is what comes next, especially after this week's announcement that CEO and founder Jensen Huang will resign following two separate personal scandals in the same week.
Jensen Huang led NVIDIA for more than three decades and had a track record of seeing around the corner of nearly every risk facing the business.
This departure arrives as several serious risks are piling up.
Some of NVIDIA's biggest customers, including Microsoft, Amazon, Meta, and Google, spend hundreds of billions of dollars together on NVIDIA chips while also actively trying to design their own chips and cut NVIDIA out of the data center.
CUDA, the platform once considered an unbreachable moat, has started showing real cracks, with genuine competition gaining real adoption over the past couple of years.
NVIDIA has also become something of a credit backstop behind more than $100 billion in financing guarantees tied to circular arrangements across the AI ecosystem.
We will cover all of this and more in this breakdown of Nvidia!
Business Overview.
Over the last 12 months, NVIDIA generated $302 billion in revenue, including $96 billion in the most recent quarter alone, up 106% year over year, with growth actually accelerating over the past several quarters.
Jensen Huang has said NVIDIA expects at least 70% growth in its own chip capacity next year, while actual demand growth could reach as much as 100%, a gap between what the company believes it can supply and what customers want.
Gross margins now stand at 75%, up 12 percentage points over the past 5 years.
Operating margins have expanded even more, reaching 65% over the last 12 months, up 35 percentage points over that same 5-year stretch from a prior 30% range.
That improvement came fairly evenly from leverage on sales and marketing spending and on research and development spending: sales and marketing costs fell from around 10% of revenue to about 2%, and research and development fell from about 20% of revenue to about 8%, even as the dollar amount spent on research and development stayed large given the overall size of the business.
That combination of revenue growth and expense leverage has produced roughly $200 billion in trailing annual profit
History.
NVIDIA was founded in 1993 by Jensen Huang and two co-founders.
In 1999, the company released its first GPU, the GeForce, a gaming-focused graphics card; NVIDIA did not invent the GPU, but it popularized it.
A GPU differs from a traditional CPU in how it processes information: a CPU executes instructions largely in sequence, while a GPU runs many independent calculations in parallel, such as rendering separate pixels on a screen that do not depend on one another.
That parallel structure made GPUs excellent at graphics rendering from the start.
Around 2006, Jensen Huang began to see that this same style of parallel computing could be useful well beyond graphics, leading to the creation of CUDA, a programming platform letting developers tap a GPU's parallel processing power for general-purpose tasks.
In 2012, a neural network called AlexNet dramatically improved image recognition after training on NVIDIA GPUs, among the earliest signs that GPUs were unusually well suited to AI.
By NVIDIA's 2016 Investor Day, Jensen Huang was already describing a future built around AI, autonomous vehicles, and robotics as major new sources of demand, years before most of the market took that seriously.
At that point, NVIDIA's business was still almost entirely gaming graphics cards, a cyclical business tied to consumer demand for computers.
Revenue peaked at $4.1 billion in 2008 and did not surpass that level again for 5 years.
A crypto-driven boom beginning around 2017 pushed revenue to $11 billion by 2019 with 32% margins, a run that continued through most of COVID, before the aftermath hit hard: by 2022, revenue had gone flat and NVIDIA was discounting excess GPU inventory to clear it, with operating margins falling from 37% to 16% and revenue at $26 billion.
4 years after that trough, revenue has reached $300 billion with 65% margins, a transition driven by the AI boom that took off around 2023 following the arrival of ChatGPT, after a decade of positioning for exactly this kind of demand.
Business Segments.
Under NVIDIA's older disclosure format, the business broke into 4 market segments: data center, gaming, professional visualization, and automotive.
Data center already represented about 90% of revenue in the last breakdown given, with gaming at about 7%, professional visualization at about 1.5%, and automotive at about 1%, and its share has only grown since, which is likely why NVIDIA stopped breaking out these categories in detail.
Within data center, NVIDIA has disclosed a further split between compute, mostly GPUs along with some CPUs, at 83%, and networking at 17%, a share becoming increasingly central to NVIDIA's competitive position.
NVIDIA's newer disclosures split revenue differently: hyperscale customers generated $169 billion over the last 12 months, growing 106%, against $109 billion from AI clouds, industrial, and enterprise customers combined, growing 68%.
The company has also simplified its segments into 2 categories, data center and edge compute, where edge compute is a catch-all covering gaming, professional visualization, and automotive together.
Edge compute totals about $25 billion in revenue, growing 35%, while data center totals $277 billion, growing 89% and representing 92% of total revenue.
Products.
NVIDIA's GPU architecture has moved through several generations, from Hopper, the H100 and H200, to the current Blackwell generation, the B100 and B200, with Rubin next, alongside CPUs called Grace and Vera.
HGX combines multiple GPUs through NVIDIA's own networking, such as NVLink, letting NVIDIA sell an entire server as one product, and DGX packages 8 B200 GPUs with 2 Intel Xeon CPUs, 2 NV switch chips, and NVLink into one integrated AI system, with Intel still a partner for certain CPU needs.
The next step up is rack-scale AI supercomputers, NVIDIA's biggest current product category.
The flagship version, the Vera Rubin NVL72, packages 72 Rubin-generation GPUs with 36 Vera CPUs, NVLink, and a large amount of high-bandwidth memory into a single rack ready to drop into a data center.
Over time, NVIDIA has moved from selling an individual chip to selling an entire AI system, often described by Jensen Huang as an AI factory, and that shift has itself become a performance advantage: even when a rival's individual chip tests faster in isolation, performance across NVIDIA's fully integrated system tends to hold up, since the chip, the rack-level networking, and the software all get optimized together.
NVLink, NVSwitch, InfiniBand, and Spectrum-X Ethernet, technology NVIDIA acquired roughly 6 years ago, together connect chips, servers, and racks across the data center into that unified platform.
Gaming remains sizable, built around the GeForce GPU line, though now far less consequential relative to data center.
Automotive centers on NVIDIA Drive, an autonomous vehicle platform, while Jetson, a compact AI computer built for robots, points toward a potential future wave of demand once AI moves into physical robotics, demand not yet material and not included in any of NVIDIA's current numbers.
SubscribeSix Key Considerations.
Beyond the historical overview, 6 ongoing considerations help frame the debate around NVIDIA right now: some are genuine risks, others are simply dynamics worth understanding, where it remains unclear whether they ultimately help, hurt, or do little to the business.
Consideration#1: Hyperscalers vs. Neo Clouds.
The first consideration involves the relationship between hyperscalers and neoclouds.
Hyperscalers are broadly in the business of selling compute, which gives them an incentive to design their own chips: because they buy enormous volumes and operate at massive scale, they can often get better performance per dollar from a custom chip than from paying NVIDIA's margin, and that scale generally only exists at the largest companies.
Google has built its own TPU chip, Microsoft has built a chip called Maia, Amazon has built a GPU alternative called Trainium, and Meta has built its own chip called MTIA.
Oracle remains one of the few major hyperscalers without a custom chip of its own, and Meta, while not technically a hyperscaler, now spends heavily enough on capital expenditures that it increasingly behaves like one.
NVIDIA's response has been to fund a wide range of smaller cloud competitors, often called neo clouds, including CoreWeave, Nebius, nScale, Lambda, and Crusoe, supplying them with GPUs and taking equity stakes instead of selling exclusively to the largest hyperscalers.
The logic is straightforward: a large hyperscaler with enough scale can eventually justify designing its own chip and cutting NVIDIA out, but a smaller neo cloud generally cannot, which keeps those customers dependent on NVIDIA's architecture and gives NVIDIA a broader, more diversified base that is far less likely to ever fully walk away.
NVIDIA is unlikely to be cut out of the largest hyperscalers entirely, since its chips still tend to offer the best performance per dollar and the most flexibility, but certain workloads are gradually migrating to those hyperscalers' own custom chips, with Google's TPU standing out as a particularly credible, relatively general-purpose alternative.
A related dynamic is sovereign AI, a theme Jensen Huang has discussed for several years.
A country seeking sovereignty over its own data centers usually finds it easiest to build that capability by buying NVIDIA's pre-built AI factory racks directly, since a country such as Saudi Arabia has no realistic path into designing its own chips.
Funding neo clouds and courting sovereign AI buyers serve the same strategic goal: keeping NVIDIA's customer base diversified, not concentrated among a handful of powerful buyers capable of designing their way out of paying its margin.
Consideration #2: Individual Chip vs. AI Factory.
On a pure per-chip basis, NVIDIA's advantage has narrowed, with competitors including Google's TPU and, in some cases, AMD's GPU now matching or surpassing NVIDIA on certain chip-level benchmarks, though that comparison usually involves NVIDIA's prior-generation chip and measures the individual chip alone, not a complete system.
An analysis from SemiAnalysis comparing Google's TPU to NVIDIA's B300 chip found the TPU 19% cheaper on a per-token basis at the chip level, but once the comparison shifted to NVIDIA's full integrated system, the GB300 NVL72 rack, NVIDIA's performance per dollar came out 30% better than the TPU.
Competitors have closed much of the gap at the chip level, often landing roughly a year behind NVIDIA's latest release, but have not closed the systems-level gap, largely because NVIDIA ties the chip, the rack-level networking, and the software together as one integrated product, a much harder thing to replicate than a single chip.
Competing claims, such as Amazon's statement that its Trainium chip delivers 30% better performance per cost than a general GPU, often compare individual chips instead of full systems, and these comparisons can be shaded depending on how a test is structured.
Consideration #3: The CUDA Moat.
The third consideration is CUDA, technically a programming platform, not a programming language.
It takes code written in a language such as PyTorch or JAX and compiles it into low-level instructions telling a GPU's millions of transistors how to move electricity to execute that program.
A long-running feedback loop between chip design and CUDA, where improvements in one inform improvements in the other, has historically been central to NVIDIA's ability to extract the best possible performance from its chips.
Large customers including Google, Amazon, AMD, and OpenAI have all built alternatives to avoid depending entirely on one proprietary platform: Google built its own compiler called XLA, Amazon built Neuron, AMD built an open-source platform called ROCm, and OpenAI built an open-source compiler called Triton that can compile PyTorch code into instructions for either CUDA or ROCm.
Being able to compile for either platform does not automatically mean a program runs equally well on both, since the chip architectures differ enough that code often still needs platform-specific tuning.
What has changed more recently is that AI itself is becoming genuinely good at translation and optimization, the two capabilities most relevant to chip programming, able to grasp the intent behind code, translate it to a different architecture, and then run that translated code repeatedly to learn the best way to execute it.
That could meaningfully lower the switching cost for a developer built around CUDA, though translation remains imperfect.
The direction of travel does look like a shrinking CUDA moat relative to where it stood historically, even if it remains far from gone, and NVIDIA combined with CUDA still tends to deliver the best results today, helped by a head start built on real-world AI workloads it can use to keep optimizing both software and hardware.
Consideration #4: Training vs. Inference.
NVIDIA was assumed to hold a stronger position in training, since training requires many chips communicating with extremely low latency, an area where NVIDIA's strong networking gives it a real advantage.
Inference was assumed to be an easier target for specialized chips such as Amazon's Trainium, which would be a negative for NVIDIA as more AI workloads shift from training toward inference over time.
More recently, inference itself has grown more complex, particularly with agentic workflows coordinating many steps or agents at once, often waiting on one piece of work to finish before another proceeds, which starts to resemble training's own coordination demands.
In those cases, strong interconnects such as NVLink matter for inference much the way they matter for training, narrowing the advantage simpler inference chips were expected to gain.
This risk, while worth surfacing, does not currently look like a serious threat to NVIDIA's position.
Consideration #5: Open-Source vs. Closed-Source Models.
An open-source model can be downloaded without restriction and hosted anywhere, while a closed-source model, such as those from OpenAI or Anthropic, is hosted and controlled entirely by the company that built it, giving that company power over where workloads run, including the ability to direct them toward its own custom chip.
This is not hypothetical: OpenAI is already designing its own chip, possibly to be hosted with Microsoft, and Anthropic is reportedly doing the same.
A world dominated by a small number of extremely powerful closed models would raise NVIDIA's exposure to this risk, since those companies would have outsized leverage to pressure NVIDIA on pricing or shift workloads elsewhere.
That does not look like the market's direction, though, since the broader AI landscape looks more likely to include a wide range of different models than to consolidate around a couple of closed ones, and NVIDIA's own strategy of funding diversified neo cloud customers works against this kind of concentration risk.
On balance, this does not look like the single biggest risk facing NVIDIA.
Consideration #6: Chip Obsolescence.
NVIDIA's GPUs, being relatively general-purpose and programmable across workloads such as training and inference, were long assumed to have longer useful lives than narrower, specialized chips built for one task, since a data center operator wants equipment that lasts, not something cheap today that becomes obsolete quickly.
That advantage still appears to hold, making buyers more comfortable committing to NVIDIA chips, but competitors including Google's TPU are closing the gap, and other specialized chips are likely to follow, which may erode what is currently only a moderate advantage.
In the current environment of severe compute constraints, where essentially every chip produced gets deployed regardless, this risk matters somewhat less right now than it otherwise would.
NVIDIA's Moats.
Stepping back from those 6 considerations, NVIDIA's competitive moats can be summarized directly.
The CUDA platform remains formidable, even as it appears to be eroding gradually relative to its historical strength, and competitors still generally have to close a real performance gap to catch up.
The clearest advantage right now is system-level performance, where NVIDIA still delivers the best performance per cost at the full AI-factory level, helped by a large base of real-world AI workloads it uses to keep refining both chips and software.
Owning NVLink, NVSwitch, InfiniBand, and Spectrum-X lets NVIDIA compete as a full AI system instead of merely a chip supplier, an edge hard to replicate by outsourcing networking to a third party such as Broadcom or Arista, since an outsourced, less integrated setup tends not to perform as well.
NVIDIA has also sped up its own innovation cycle, moving from roughly 2 years between new chip architectures to about 1 year now, and a broad developer ecosystem already familiar with CUDA adds a final layer of stickiness.
The single most important of these advantages is simply having the best performance per cost on a per-token basis.
As long as that remains true, NVIDIA's position stays secure, but sustaining it requires continuous outperformance of the competition to justify its price premium.
A rival such as Google could, in principle, undercut NVIDIA on price with a somewhat less performant chip and still match NVIDIA on performance per dollar, meaning NVIDIA has to keep improving faster than everyone else simply to preserve its current margins, a meaningful risk in its own right.
Other Key Risks.
Circular Financing.
One widely discussed risk involves NVIDIA's circular financing arrangements.
NVIDIA has made roughly $100 billion in equity investments, including sizable stakes estimated around $30 billion to $40 billion each in Anthropic and OpenAI, with a pattern where some of that invested money gets spent by those companies buying NVIDIA chips: NVIDIA invests in a neo cloud such as CoreWeave to help it grow, that company then spends a large share of that money on NVIDIA chips, and NVIDIA ends up with both an equity stake and a margin on the chips sold, an arrangement that entrenches it deeper into the AI ecosystem.
Beyond equity investments, NVIDIA has roughly $25 billion in additional future investment commitments, $36 billion in neo cloud compute purchase commitments, and an estimated $100 billion in credit guarantees supporting financing arrangements across the industry.
Added together, these commitments run to roughly $100 billion to $200 billion, comparable to about a year of NVIDIA's own trailing earnings, a pattern drawing some comparison to the excess of the early-2000s technology bubble, though nowhere near that scale relative to NVIDIA's own size and cash generation.
The current amounts do not look alarming given how cash-generative the business is, but if these arrangements keep growing substantially from here, they could become a genuine source of concern.
Leadership Transition.
The biggest risk right now looks less like a financial or competitive dynamic and more like straightforward execution risk during a leadership transition.
Jensen Huang had roughly 60 people reporting to him directly, an unusually flat structure that worked only because he was the company's founder.
Colette Kress, NVIDIA's CFO since 2012, has stepped in as interim CEO, but inserting management layers between those 60 direct reports and the top of the company is likely to create friction, with some executives frustrated at reporting to someone other than the founder and some simply choosing to leave.
Large, unconventional strategic moves, such as the roughly $100 billion in equity investments and $100 billion in credit guarantees already discussed, are considerably harder for an interim leader to repeat with the same conviction a founder could bring, making further acquisitions less likely in the near term.
Jensen Huang had a demonstrated pattern of spotting risks and opportunities years in advance, from funding neo clouds early, to pivoting toward full AI systems, to pursuing sovereign AI deals, to recent moves that positioned NVIDIA in inference chips and the open-source model community.
Losing that kind of forward-looking leadership, just as NVIDIA has compressed its own product cycle to about a year with little room for error, raises real risk at exactly the moment competitors are most eager to close the gap.
A useful parallel is Intel's old development cadence, which alternated between manufacturing and architecture improvements roughly every 2 years and worked reliably for about a decade before Intel's execution could no longer keep pace, a reminder that falling behind even once in a race this demanding can be very hard to reverse, since losing the performance lead would likely also mean losing the pricing power behind NVIDIA's current 65% operating margin.
AI Spending and Capacity Risk.
A broader risk worth noting is the general possibility of an AI spending bubble.
Regulatory pushback at the state or federal level, including specific concern in Texas, where a large share of current data center construction is concentrated, could slow new buildouts.
Separately, hundreds of billions of dollars in capital spending cannot continue indefinitely without a clear return, and if returns disappoint, hyperscalers would likely pull back on that spending.
There is real evidence of return on investment already showing up at large hyperscalers such as Meta, Amazon, and Google, but a meaningful share of current AI spending remains experimental, and it stays genuinely unclear how much sustainable economic activity AI is generating relative to the trillions of dollars being committed to it.
Historically, every major capital spending cycle has eventually overshot demand to some degree, and there is no strong reason to assume this one will be the exception, though it is also possible this simply turns out to be an unusually long cycle.
Valuation.
Turning to valuation, at a $167 stock price, NVIDIA carries a $4 trillion market capitalization against $192 billion in trailing earnings, a multiple of 21x trailing earnings, about 14x forward 12-month estimates, and about 10x 2027 earnings estimates.
Before settling on any particular multiple, the more important question is whether NVIDIA should still be treated as a cyclical chip business or as a secular grower, since that distinction changes everything about how to value it.
Cyclical businesses are notoriously difficult to value with a standard discounted cash flow model, since the analyst has to guess which future years bring growth and which bring contraction, with no reliable way to know where the business sits within its own cycle at a given moment.
That uncertainty is why chip stocks have historically traded at unusually low multiples on peak-cycle earnings.
Micron, for instance, traded at a low point of around 3x earnings, arguably even lower than it should have, and at correspondingly higher, sometimes double-digit, multiples when earnings sit near a cyclical trough.
Whether a given multiple turned out right is really only knowable after the fact.
As one concrete framework, a bottom-up revenue model can be built around data center gigawatt capacity through 2031, NVIDIA's assumed share of that capacity, and NVIDIA's revenue per gigawatt, a figure Jensen Huang has referenced directly on earnings calls, rising from about $18 billion to $25 billion with Blackwell and now toward $40 billion with Vera, as NVIDIA's products increasingly include networking and other data center equipment alongside chips.
The model separately tracks new-build gigawatts and replacement gigawatts, since chips are generally assumed to last roughly 4 to 6 years before needing replacement, a dynamic that, if it holds, would make a meaningfully larger share of future revenue recurring instead of tied purely to one-time buildouts, not unlike smartphone users replacing their phones every couple of years.
The model also assumes an eventual maintenance phase, once the initial buildout matures, during which growth slows sharply because revenue mostly reflects replacing existing capacity instead of adding new capacity, though even that slower revenue base could still land higher than today's revenue if the installed base is large enough by then.
In a base case using 55% NVIDIA market share and 40 billion gigawatts of total data center capacity, a figure that may itself run high, the model produces roughly $1 trillion in data center revenue plus about $50 billion in edge revenue, translating to roughly $460 billion in NOPAT.
Applying a range of multiples to that base-case profit figure produces a wide spread of outcomes: a 15x multiple implies a roughly $7 trillion valuation, a 20x multiple implies about $9.2 trillion, and a 25x multiple implies about $11.5 trillion.
Over the stretch from today through 2031, NVIDIA is projected to generate about $1.7 trillion in cash, and if that cash were returned to investors as generated, those same 3 multiples translate to annualized returns of roughly 18%, 24%, and 29% from today's price.
Every one of these base-case assumptions remains fundamentally bullish, requiring a long AI cycle, continued NVIDIA dominance, and maintained margins throughout.
A bear case looks meaningfully different.
If earnings were instead cut roughly in half to around $118 billion, with NVIDIA by then viewed as a proven cyclical business instead of a secular grower, the multiple the market assigns would likely compress severely too, landing somewhere in the high 30s times earnings, cutting the stock considerably from where it trades today.
These figures are inherently uncertain, which helps explain why NVIDIA trades at the multiple it does: real disagreement exists about how long this AI cycle can run.
Forcing a precise discounted cash flow valuation on a business this unpredictable tends to produce a false sense of precision instead of a genuinely useful answer.
A more honest approach is to look out to a year an investor feels reasonably comfortable forecasting, such as 2027, see what multiple the stock trades at on that year's estimates, about 10 times, and compare that to where cyclical chip multiples have typically traded at a trough versus still growing.
There is nothing wrong with declining to take a firm position on something this uncertain.
For a business this dependent on continued flawless execution, having real confidence in the people running it matters a great deal, and that confidence is precisely what is missing right now with no permanent CEO yet named, even though the stock has already fallen about 30% and still trades at multiples that may or may not have fully priced in this leadership uncertainty.
That gap may buy the company time to sort out its leadership internally, but it also makes NVIDIA a considerably less comfortable investment without Jensen Huang at the helm, and that discomfort may be exactly why the current opportunity, if there is one, exists at all.
For more on Nvidia, check out this video below.
YouTube FAQ
Nothing in this newsletter is investment advice nor should be construed as such. Contributors to the newsletter may own securities discussed. Furthermore, accounts contributors advise on may also have positions in companies discussed. Please see our full disclaimers here.