Nvidia Is the Best Business Alive. That's Not the Same as a Good Price.

Nvidia trades at $206.84. Market cap: $5.01 trillion, as of July 25, 2026. That number is so large it stops meaning anything, so try this instead: it is bigger than the entire stock market of every country on earth except the United States and, some days, China.

Here is the harder number. If Nvidia's earnings simply stop growing for three years and the market decides to pay 15 times earnings for it instead of today's 32 times, the stock is worth $98.80. That is not a stress-test fantasy. It is one of three scenarios built directly off Nvidia's own trailing numbers, and it implies 52% downside from here.

The moat is real, and nobody is arguing about that part

Nvidia's gross margin (revenue minus the direct cost of making the product, divided by revenue: a measure of pricing power before overhead) has stayed above 70% through an entire demand supercycle, 74.9% in the most recent quarter, per Nvidia's Q1 fiscal 2027 press release. A commodity hardware vendor riding a temporary shortage does not hold a number like that for four straight years. A company selling something genuinely hard to substitute does.

Return on invested capital, which measures how much profit a company generates for every dollar tied up in the business, has been triple digit for three straight fiscal years: 191.2% in FY2025, 145.8% in FY2026, 117.4% on a trailing basis today. Debt to equity, a simple ratio of borrowed money to shareholder money, sits at 0.06. Cash plus short-term investments ($80.6 billion) is more than 6.5 times total debt. This is about as far from a fragile balance sheet as a large company gets.

Because Nvidia doesn't run its own factories (it designs chips and pays TSMC to manufacture them, the "fabless" model), capital spending eats only about 6% of operating cash flow. The rest falls through to free cash flow at roughly a 45% margin on revenue. Almost half of every dollar that comes in the door leaves as cash the company can return to shareholders or reinvest, per Nvidia's cash flow statement.

The mechanism behind all of it is CUDA, the software layer Nvidia built starting in 2006 that over 4 million developers and more than 3,000 optimized applications now run on. Switching off it isn't like changing a supplier. A customer that wants to leave has to retrain engineers, rewrite low-level code, and often re-architect the physical layout of a data center that was sized around Nvidia's rack designs. That is what a switching-cost moat looks like when it's the strongest one observable in large-cap technology. A quieter second version of the same playbook is playing out in networking gear: Nvidia's Spectrum-X and NVLink product lines, built substantially on its 2020 purchase of Mellanox, generated over $31 billion in fiscal 2026 and made Nvidia the number one revenue vendor in data-center Ethernet switching by early 2026, per the research firm IDC, after Meta and Oracle standardized on it.

None of this is a secret. The current price already assumes it.

What changed in 2026 is that the risks stopped being hypothetical

For years, the bear case on Nvidia was a list of things that could go wrong. This year, several of them went from "could" to "did."

The demand backdrop is why anyone is still buying at all. Hyperscaler capital spending, the best leading indicator for this whole market, is on track to hit somewhere between $630 billion and $900 billion in 2026, up 62% to 77% from 2025, spread across Amazon, Google, Meta, and Microsoft. Nvidia's most recent quarter shows data-center revenue growing 92% year over year, faster than even that extraordinary capex growth. The market Nvidia sells into is not shrinking, and by its own most recent numbers is not even saturated yet.

The clearest threat inside that growth: Nvidia's biggest customers are building their own chips to avoid paying its margin. Google's TPU, Amazon's Trainium, and Microsoft's Maia are collectively growing about 44.6% a year against 16.1% for the merchant GPU market they are chipping away at, nearly three times faster, and they are aimed squarely at inference (running an already-trained AI model to answer queries), which is now somewhere between half and two-thirds of total AI compute demand, depending on whose estimate you use. Google and Anthropic have already disclosed, per reporting from SemiAnalysis and Tom's Hardware, that Anthropic has deployed over a million of Google's Ironwood TPU chips for Claude, the first custom AI chip to reach a seven-figure single-customer deployment.

The result shows up directly in Nvidia's numbers. Its share of the data-center GPU market has slipped from roughly 90% in 2024 to somewhere between 80% and 86% today, depending on the source. That is still dominant. It is also a real, measured decline, not a forecast.

China is the sharper version of the same story, and it's already happened, not pending. Nvidia's China chip market share was roughly 95% before US export controls took hold. Chinese domestic chipmakers Huawei and Cambricon are now projected to hold 56% of China's AI server market in 2026, up from 46% in 2025, while Nvidia and AMD's combined foreign share falls to roughly 21% from 34% in a single year, per reporting from the South China Morning Post. Jensen Huang has said Nvidia's China chip share is now effectively zero and that the company has "largely conceded" the market to Huawei, striking words for a CEO to use about his own company's largest addressable growth market.

Market	Then	Now
Data-center GPU, global (2024 to 2026)	~90% share	80-86% share
China AI server market (pre-controls to 2026)	~95% share	Effectively zero, per Huang
Custom silicon vs. merchant GPU growth (2026)	n/a	44.6% CAGR vs. 16.1%

Four separate regulatory fights, all still open

This is the part of the story that gets flattened into "regulatory risk" as a single line item. It shouldn't be. There are four distinct, live threads, and none of them has resolved.

First, the 15% revenue-share arrangement that lets Nvidia sell certain chips into China in exchange for handing 15% of that China revenue to the US government. The White House has publicly said the legality and mechanics of this are "still being ironed out." Legal analysis published by Lawfare has separately argued it may function as an unconstitutional export tax.

Second, the AI OVERWATCH Act, a bill that would treat the most advanced AI chips as strategically sensitive, similar to weapons, and give Congress a veto over export licenses. It cleared a House committee 42-2 and was secured in the Senate's defense authorization bill in mid-July 2026. It is advancing, not stalling.

Third, China's own restriction, separate from anything Washington does: Beijing has banned Nvidia's top-tier GPU products domestically to support its homegrown suppliers. Nvidia needs permission from both governments to sell into China, and each has its own reason to say no.

Fourth, the Department of Justice has escalated its antitrust inquiry into Nvidia to legally binding subpoenas, examining whether Nvidia makes switching costly, ties scarce product access to broader purchase commitments, and whether its CUDA lock-in constitutes anticompetitive conduct. No formal complaint has been filed. A subpoena is not a lawsuit. But it is a step closer to one, and it targets the exact mechanism, CUDA switching costs, that underpins the moat described above.

Four unresolved fronts. Not one of them requires a bear thesis to imagine. They're already filed, drafted, or subpoenaed.

The July selloff was the dress rehearsal

In early July 2026, semiconductor stocks sold off hard on a specific worry: that AI infrastructure spending has outrun the revenue it's supposed to generate. Intel fell as much as 21%. Micron dropped about 13% in a single session, erasing roughly $138 billion of market value. Cited catalysts included reports that SK Hynix was slowing its memory expansion plans and a more hawkish tone from the Federal Reserve.

That is not a Nvidia-specific event. It's a sector-wide repricing of a specific question: does AI capex convert into durable revenue, or is the industry building ahead of proven return on investment? Nobody in this research can answer that question with certainty, and neither can the sell side.

What is observable is that the market treated it, for a few weeks in July, as a live risk rather than a hypothetical one. Nvidia, as the largest single beneficiary of the capex boom, is also the company with the most to lose if that spending decelerates: two customers alone, unnamed in Nvidia's disclosures but almost certainly among Microsoft, Amazon, Google, and Meta given the order size, account for roughly 39% of its revenue.

What the price already assumes

Here is the valuation math, run three ways over a three-year horizon, using Nvidia's trailing earnings per share of $6.59 as the starting point.

Scenario	Assumption	Target price	Change from $206.84
Bull	25% annual earnings growth, 35x multiple	$450.50	+117.8%
Base	15% annual earnings growth, 25x multiple	$250.60	+21.1%
Bear	0% earnings growth, 15x multiple	$98.80	-52.2%

A word on where those multiples come from: they are not arbitrary. The bear-case multiple, 15x, roughly mirrors TSMC's current trading multiple, a reasonable stand-in for what the market pays for semiconductor cash flow once growth expectations disappear. The base and bull multiples, 25x and 35x, hold Nvidia at a premium to that floor for continued (if decelerating) growth, with the bull case deliberately built off a 25% growth assumption, below Nvidia's recent trajectory, to leave room for sell-side consensus to be wrong on the high side and still land inside this range.

A word on the base case specifically: 15% annual earnings growth is well below what Nvidia has actually delivered over the past three years (revenue alone grew 126%, then 114%, then 66% year over year), and it's below current sell-side consensus for the coming year (roughly 52% net income growth is the consensus estimate). Building a "base case" this conservative and still landing at +21% upside tells you the market is not pricing in a blow-out scenario. It also is not pricing in any real probability of a slowdown. The bear case assumes earnings merely go flat, not decline, and still costs the stock half its value.

Trailing price-to-earnings, a measure of how many years of current earnings you're paying for in the stock price, sits at 31.68x today, down from 119.77x three years ago as earnings caught up to the price. That makes Nvidia cheaper than its own recent history, and cheaper than AMD (174x trailing) or Broadcom (63.6x) on the same measure. It does not make Nvidia cheap in absolute terms. A 31.68x multiple on a business whose growth rate is mechanically decelerating (126% to 114% to 66% revenue growth, in three consecutive years) prices in continuation, not surprise.

Management: real discipline, one real flag

Jensen Huang has run Nvidia for its entire 33-year history, and the CUDA platform that defines the company today predates the AI boom by 15 years: it wasn't built in response to the cycle, it was already there when the cycle arrived.

His capital allocation record has one clear failure and one clear act of discipline. The failure: a $40 billion attempt to acquire chip designer Arm, blocked on antitrust grounds by regulators in the US, UK, and EU, terminated in 2022 with Nvidia paying a $1.25 billion breakup fee. The discipline: Nvidia's proposed investment in OpenAI, originally a letter of intent for up to $100 billion, was scaled back to roughly $30 billion, with Huang telling investors in March 2026 that the larger figure is "not in the cards." That is a company choosing restraint over the largest possible headline, at a moment when other chipmakers were still stacking so-called circular-financing deals: arrangements where a chip supplier invests in an AI lab, the lab spends that money (and more) on cloud capacity, and the cloud provider uses part of it to buy chips back from the original supplier, a loop that inflates everyone's reported revenue without proving any new demand exists outside the loop itself.

Set against that: across 15 insiders over 18 months, there has been zero recorded open-market buying, against more than $450 million in insider sales in a single four-month stretch (December 2025 through March 2026). Sales made under pre-scheduled 10b5-1 trading plans, a mechanism that lets executives set up stock sales in advance to avoid any appearance of trading on inside information, are ordinary and not by themselves a red flag. A complete absence of buying, across the entire leadership bench, at a $5 trillion valuation, is harder to wave away. Read charitably, it is routine diversification by people whose net worth is already concentrated in one stock. Read skeptically, it is the people with the best seat in the house for judging whether today's AI-capex demand is durable choosing not to bet new money on it. The data (zero buys, broad-based sales) is fact; which reading is correct is not verifiable from outside the company.

What this means for the price you'd pay

None of the above is a secret. That's the point. The moat, the margins, the balance sheet, all of it is fully visible and fully priced. What's newer is that the offsetting risks, custom silicon, Chinese substitution, a four-front regulatory fight, and capex-durability skepticism, are no longer things an analyst has to imagine. They're line items with numbers attached to them now.

That changes what "own Nvidia" should mean in a reader's head. It is not a question of whether this is a good business; every metric in this piece answers that question the same way. It's a question of what you're willing to pay for a business whose quality is already fully known, when the price leaves almost no room for the risks that are now sitting in plain view to actually matter. A pullback into the $150-175 range, without a matching deterioration in the fundamentals above, would be a genuinely different entry point than today's. At $206.84, the honest read is: own it if you already do, for the long run, but this is not the price at which the case for starting a new position writes itself.

Sources

This article is LLM-assisted, disclosed per site policy, and does not constitute financial advice or a recommendation to buy or sell any security.

Key facts