NVIDIA designs the chips that train and run AI models, and sells the networking gear and software around them. More than 90 cents of every revenue dollar now comes from data centers. It is the best business in the market's biggest theme, and the price knows it. One wrinkle: about a quarter of next year's business comes from customers NVIDIA itself helped finance, so part of the demand is NVIDIA-funded.
Scale is 0–50%/yr. The shape of the verdict: the base case clears the 12% growth bar with room to spare. The 30% bar needs the full bull case.
Almost every serious AI model in the world trains on NVIDIA hardware, and the CUDA software platform is why they stay. CUDA has been built up over fifteen years, and moving a large AI operation off it would mean rewriting years of engineering work. So NVIDIA doesn't just sell chips, it rents out the road the whole industry drives on, and it collects at both ends: the hardware sale and the software ecosystem that keeps customers buying the next generation.
The 12% bar math says NVIDIA compounds about 16% a year from here, reaching roughly $468 by 2031. That assumes earnings keep growing fast enough to carry the stock while the multiple drifts down from around 25x toward 20x, which is what usually happens when a growth company gets this large. It's a reasonable path, not a heroic one. It just doesn't get you to 30%.
Here's the uncomfortable part. NVIDIA has been investing its own money into its customers: up to $100 billion committed to OpenAI (later finalized around $30 billion), up to $10 billion into Anthropic, $2 billion into CoreWeave. Those companies turn around and spend it on NVIDIA systems. On the last earnings call, NVIDIA's CFO said companies supported by NVIDIA's balance sheet will be about a quarter of next year's business. Management calls it accelerating the ecosystem. Skeptics call it circular financing. The honest read: the end demand for AI compute is real, but about a quarter of it is NVIDIA-financed, which means the growth rate flatters itself a little.
Vera Rubin, the chip platform after Blackwell, entered full production in March 2026, and management calls it the fastest product ramp in company history. The CFO expects it to be around 20% of data center revenue this quarter. That matters because the bear case on NVIDIA has always been 'what happens when the current cycle ends,' and the answer keeps being: there's already a next cycle shipping.
China used to be more than a fifth of NVIDIA's revenue. Export restrictions have shrunk that to a rounding error, and the latest guidance assumes literally zero data center revenue from China. That's a headwind baked into the numbers, but it's also the upside nobody models: if restrictions ever loosen, it's pure optionality on top of the base case.
Quietly, NVIDIA's networking business is turning into something real. Spectrum-X Ethernet revenue grew 2.6x year over year. As AI clusters get bigger, the networking between the chips matters almost as much as the chips themselves, and NVIDIA now sells both. It's still small next to the GPU business, but it's the kind of second leg that makes the base case less fragile.
At around $222, NVIDIA trades near 28 times trailing earnings, which sounds almost reasonable until you remember the company is worth over $5 trillion. The multiple has already compressed a lot: earnings grew faster than the stock. You're paying for a company that has to keep growing at historic rates just to hold the multiple, let alone expand it. That's the selective call in one sentence: great business, full price.
Four hyperscalers, Amazon, Microsoft, Google, and Meta, plus a wave of sovereign AI programs, are the demand base. Their capital spending plans underwrite the next few years, and right now those plans keep going up. But concentration is concentration: if two of them pause spending in the same year, the quarter gets ugly fast. The visibility is excellent until it isn't, and there's no diversifying away from it.
High conviction would mean the base case clears 30% a year. It doesn't, it clears about 16%. The 30% bar needs the full bull case: everything going right, multiple holding up, no capex pause, no export escalation. That's a lot of things going right at once. Selective is the honest label: a great company at a full price, worth owning, not worth chasing.
Two things would change this. Up: if the base case math ever clears 30% on its own, without leaning on financed demand, this becomes high conviction. That would take either a much lower entry price or earnings growing even faster than the current blistering pace. Down: if the financed share of revenue keeps growing faster than real demand, if the investment stakes start getting written down, or if hyperscaler capex rolls over, the verdict drops. Watch the financing disclosures each quarter. They're the canary.
NVIDIA at 25 to 28 times earnings doesn't need a disaster to fall 30%. It just needs growth to slow from spectacular to merely excellent. When the market stops believing the next quarter will be a blowout, the multiple comes down first and asks questions later.
Four hyperscalers plus a few sovereign buyers are most of the story. If Microsoft or Amazon pauses GPU orders for one quarter to digest what they've bought, NVIDIA feels it immediately. Concentration is the price of selling to the biggest spenders on earth.
No capital spending supercycle lasts forever. Right now every hyperscaler is racing to build, because nobody wants to be short on compute. But if AI returns disappoint, or if the economy turns, capex gets cut, and NVIDIA's order book is the first place it shows up.
If the AI labs NVIDIA has funded run into trouble, NVIDIA gets hit twice: the investment stakes lose value and the revenue from those customers dries up in the same downturn. The two risks are correlated, not diversified. A quarter of next year's business tied to balance-sheet-supported customers is the number to watch.
China is already at zero in the guidance, so the risk isn't losing China, it's the controls spreading. Tighter rules on what's shippable, or new restrictions on other countries, would shrink the addressable market further. And on the supply side, every advanced NVIDIA chip is made by TSMC in Taiwan, which is a geopolitical risk no diversification plan can fully remove.