Every time you type a prompt into ChatGPT, or ask an AI tool to summarise a report, or watch a company announce its big new AI product, there’s a good chance a chip from one company is quietly doing the heavy lifting behind it.
That company is Nvidia. And it’s now worth more than $5 trillion, which makes it the single most valuable company on the planet. Not Apple. Not Microsoft. Not Saudi Aramco. Nvidia.
That’s a strange thing for a chipmaker most people had never heard of a decade ago. So let’s actually unpack how it got here, what it does, and why practically everyone in tech seems to have a stake in whether Nvidia keeps winning.
What Nvidia Actually Makes
Nvidia builds a kind of chip called a GPU. Originally, GPUs existed to do one job: render graphics for video games fast enough that they didn’t look choppy.
Here’s the detail that ended up mattering more than anyone expected. A regular computer chip is good at doing one complicated task at a time, quickly, one after another. A GPU is built differently. It’s good at doing thousands of simple tasks all at once, in parallel.
For years, that was just useful for making games look good. Then researchers building AI realised something. Training an AI model is really just millions of small, repetitive calculations happening at the same time. Which is exactly what a GPU is built for.
So a chip designed for rendering explosions in video games turned out to be the perfect engine for building ChatGPT. Nobody planned it that way. It just happened to fit.
Why Nvidia, Specifically
Plenty of companies make chips. What set Nvidia apart is something less obvious than the hardware itself, a piece of software called CUDA.
CUDA is what lets developers actually programme Nvidia’s chips to do useful work. Over nearly two decades, Nvidia made CUDA the default toolkit that AI researchers learned on. Once a company builds its entire AI system around CUDA, ripping it out and switching to a competitor’s chips becomes a genuinely painful, expensive exercise.
That’s the real moat. Not just faster chips, but an entire ecosystem that’s expensive to leave. It’s a big part of why Nvidia can charge premium prices and still sell every chip it makes.
The Business, in Simple Terms
Nvidia earns money across four areas, though one of them now dwarfs everything else.
Data Center is the business that matters. It’s the sale of GPUs, and increasingly entire computing systems, to the companies and governments racing to build AI infrastructure. It now makes up close to 90% of everything Nvidia earns.
Gaming is where Nvidia started, selling graphics cards to PC gamers. Still profitable, just small next to data center now.
Professional Visualisation covers GPUs for designers and engineers doing 3D work.
Automotive and Robotics is the newest, smallest piece, Nvidia’s bet that AI eventually moves out of chatbots and into physical machines like self-driving cars and warehouse robots.
Three years ago, data center was one business among several. Today it basically is the company.
What the Last Quarter Actually Showed
Nvidia’s most recent results, for the three months ending in July 2026, give a sense of the scale involved.
Revenue came in at $96.2 billion for the quarter. More than double what it made in the same quarter a year earlier. That’s more money in three months than most large public companies make in an entire year.
Profit moved in step. Nvidia earned $59.7 billion in net income for the quarter, roughly double the year before. And it kept about 75 cents of every dollar it earned after paying for the cost of making its chips, a margin most hardware companies would consider almost unreasonable.
For the current quarter, Nvidia expects around $108 billion in revenue, its first-ever $100 billion quarter. And looking further out, it’s guiding to 70% revenue growth for the year ahead. A company this size, expecting to grow 70% in a year, is not normal.
Two Very Different Kinds of Customers
For a while, the standard worry about Nvidia went like this: its entire business rides on four giant customers, Amazon, Google, Microsoft, and Meta, and all four are quietly building their own chips to need Nvidia less.
That worry isn’t baseless, but it misses something Nvidia only recently made visible. The company has started splitting its AI revenue into two buckets. One bucket is those four hyperscalers. The other, which Nvidia calls its AI Clouds, Industrial and Enterprise business, is essentially everyone else, smaller cloud providers, governments running national AI projects, hospitals, banks, drug companies, trading firms.
Last quarter, the hyperscaler bucket brought in $48.7 billion, up 13% from the quarter before. The everyone-else bucket brought in $40.3 billion, up 25% from the quarter before, and up 138% from a year earlier. That second bucket is now nearly half of Nvidia’s AI business, and it’s growing roughly twice as fast as the first one.
Here’s why that matters. Most customers in that second bucket will never build their own chip. A drug company doing AI-assisted research isn’t going to spend three years and a billion dollars designing a custom processor. It wants a machine that shows up on a truck and works. So even as the four giants slowly build more of their own silicon, an entirely new, faster-growing set of buyers has shown up behind them, with no interest in ever leaving Nvidia.
The Metric Nvidia Actually Cares About Now
Chip market share sounds like the obvious thing to track, and some analysts do expect Nvidia’s share of the AI chip market to slip from around 90% today to somewhere near 70% over the next couple of years, as hyperscalers deploy more in-house silicon.
But Nvidia itself has started talking about a different number entirely: revenue per gigawatt of power. During its Hopper generation of chips, the ones that powered the original ChatGPT boom, Nvidia earned roughly $18 billion for every gigawatt of computing power its chips helped deploy. With the current Blackwell generation, that’s climbed to about $25 billion. With the newest generation, Vera Rubin, it’s around $40 billion.
Why frame it this way? Because power, not chip supply, is the actual bottleneck in AI right now. You can pour concrete for a data center reasonably fast. Building a new electricity substation, getting the transmission lines approved, that can take years. So the real ceiling on how much AI computing exists in the world is set by how many gigawatts of power are available, not by how many chips get manufactured. And every generation, Nvidia is selling more around each chip, CPUs, networking, its Groq-based inference chips, so even a smaller share of raw chip volume can still mean more revenue per gigawatt of power deployed.
There’s a second number worth knowing here too: tokens per watt, basically how much useful AI output you get for each unit of electricity. Once a data center is drawing all the power it’s allowed to draw, you can’t add more chips to get more output. The only way to grow is to make each watt smarter. Nvidia claims its newest systems deliver up to 30 times more output per watt than the previous generation in ideal conditions, and real-world testing by at least one cloud provider found something closer to 10 times. Either way, that’s a big enough jump that customers have mostly shrugged off Nvidia raising prices on its newest systems by around 15%, a move driven by rising memory costs.
A Chain That Runs Well Beyond Nvidia
Nvidia’s fortunes don’t stay contained to Nvidia. They ripple through an entire chain of companies, which is worth understanding because it explains why Nvidia’s earnings move markets far beyond its own stock.
The big cloud companies are expected to spend close to $800 billion this year building data centers, climbing to $1.3 trillion next year.
Much of that flows to Nvidia directly, or through smaller “neocloud” companies that rent out AI computing power to others.
Those neoclouds need somewhere to physically build their data centers and someone to assemble the servers, which is where a company like Foxconn comes in, the same company that assembles iPhones now makes more money building AI servers than it does building phones.
The chips get manufactured by TSMC in Taiwan, and the specialised memory that feeds them comes almost entirely from Samsung and SK Hynix.
So when Nvidia does well, it’s really a signal running through cloud giants, neoclouds, server assemblers, chip foundries and memory makers all at once. Which also means if AI spending ever slows meaningfully, a lot of companies feel it, not just Nvidia.
The Part That’s Genuinely Complicated: Nvidia Is Now Financing Its Own Customers
Here’s where the story gets less straightforward, and it’s the part most worth paying attention to.
Nvidia has started investing directly in, guaranteeing loans for, and extending long payment terms to, some of the very companies that buy its chips.
Critics call this circular financing, and the comparison people reach for is the dot-com telecom crash, when equipment makers lent money to startups so they could buy gear, and got wiped out when those startups folded.
Look at Nvidia’s own filings and the picture gets more textured. Net income for the first half of the fiscal year came in at $118 billion, an enormous number. But buried in the cash flow statement are two quieter drains: $23.71 billion in unrealised gains sitting in equity investments, and $24.89 billion parked in money customers owe Nvidia but haven’t paid yet. On top of that, Nvidia spent $42.4 billion in direct cash purchases of stakes in other companies.
Total assets jumped from $206 billion to $320 billion in just six months. Nearly half of that, $93.94 billion, now sits in equity investments, and another $63 billion sits in unpaid customer bills. Combined, those two categories make up 49% of Nvidia’s total assets, up from 35% just six months earlier. Nvidia has also started letting select customers pay over 90 days, sometimes a full year, instead of upfront. In effect, it’s helping finance its customers’ data center builds using its own products as the collateral.
The market actually noticed. Nvidia’s stock dipped in after-hours trading the moment these numbers came out, before recovering once management gave its forward guidance on the earnings call.
Three specific arrangements are worth knowing by name. Nvidia has guaranteed up to $105 billion in credit support for a massive data center and power buildout in Ohio, built for an OpenAI affiliate, securing 4.25 gigawatts of capacity, on the condition that the site runs exclusively on Nvidia hardware through 2030.
It’s also committed roughly $36 billion in long-term capacity deals with smaller neocloud providers, letting those companies use Nvidia’s backing to secure their own bank loans, since banks want long-term certainty that AI startups often can’t offer on their own. And it’s signed deals with some of the world’s largest asset managers, including BlackRock, Blackstone and KKR, to mobilise more than $500 billion in outside capital for AI infrastructure. Nvidia’s filings say any additional backstop here is technically optional, but in practice, institutional investors rarely sign off on deals this size without one.
There’s a defence to all this worth hearing too. The total scale of Nvidia’s financing arrangements is around $165 billion, which is roughly one and a half times its trailing twelve-month cash flow, a very different risk profile from the dot-com telecom lenders, who were financing companies whose own finances couldn’t support the debt. And if a customer does fail, what’s left behind is Nvidia hardware still capable of running someone else’s AI workload, not years of unused fibre-optic cable sitting dark underground.
Why Nvidia Wants Hugging Face
All of this financing points toward one strategic goal, and Nvidia’s talks to acquire a company called Hugging Face for roughly $12.9 billion make that goal clearer.
Hugging Face is one of the biggest hubs online where developers find, share and build AI models, sitting a level above Nvidia in the AI stack. Buying it would bring Nvidia much closer to the people actually deciding what gets built on top of its chips, not just the chips themselves. Owning it also gets Nvidia the engineering team behind a tool that lets AI models run on non-Nvidia hardware too, giving it a foothold even outside its own CUDA ecosystem. Think of it as similar in spirit to Microsoft buying GitHub, a move less about the platform’s existing revenue and more about owning a piece of infrastructure that developers build their habits around.
Nvidia has also been locking down its supply chain aggressively, with outstanding commitments for memory and manufacturing capacity reaching $279 billion recently, up sharply from the quarter before. That secures Nvidia’s own production, but it also makes it harder for competitors and custom-chip programmes to get their hands on the same scarce components.
What Could Actually Go Wrong
A few things are worth watching honestly, rather than glossing over.
The biggest tech companies buying Nvidia’s chips are also quietly building their own, Google’s TPUs, Amazon’s Trainium, Microsoft’s Maia. That’s not an immediate threat given how sticky CUDA is, but over several years it could mean Nvidia keeps a smaller slice of a market that’s still growing fast overall.
Then there’s China. The US restricts which Nvidia chips can legally be sold there, for national security reasons. That’s already cost Nvidia billions in lost sales, and the situation still isn’t fully resolved. A real chunk of this business depends on political decisions made far from Santa Clara.
And there’s the financing question covered above, unresolved, genuinely worth watching, and not something to dismiss just because Nvidia’s counterarguments sound reasonable.
Should You Actually Care, as an Investor
Nobody, not even the analysts covering this stock full-time, knows exactly where Nvidia goes from here. What’s clear is that the company sits at the centre of one of the largest infrastructure buildouts in business history, and its numbers are about as close to a real-time read as you’ll get on whether the world is genuinely investing in AI, or just talking about it.
A lot needs to keep going right for the current trajectory to hold. Demand needs to stay strong, competition from custom chips needs to stay manageable, China needs to remain a minor irritant rather than a real drag, and the financing arrangements need to hold up rather than unwind.
Indian investors can buy Nvidia shares directly through platforms like Vested, using the RBI’s Liberalised Remittance Scheme, which lets individuals invest a set amount abroad each year. As with any single stock, however dominant it looks today, it’s worth treating it as one piece of a diversified portfolio, not a bet made on its own.
This article is for informational purposes only and should not be construed as investment advice. Please consult a financial advisor before making investment decisions.
