Nvidia

The company the AI era is built on

Nvidia is the most valuable company in the world - about $5.06 trillion in late July 2026 - and it earned that title by becoming the indispensable supplier of the single most sought-after resource of the decade: AI compute. Roughly 80 to 85% of the money spent on data-center AI accelerators in 2026 flows to Nvidia, and when a frontier lab, a hyperscaler, or a sovereign wants to train or serve a large model, Nvidia's systems are still the default answer. In a period when trillions of dollars of capital are being redirected toward artificial intelligence, Nvidia sits at the toll booth every one of those dollars has to pass through.

What makes the company remarkable is not just that it sells the chips, but that it defined the entire category. Jensen Huang, who co-founded Nvidia in 1993 and still runs it, spent two decades turning a graphics-card maker into the substrate of modern machine learning, and then rode the generative-AI explosion from a roughly $1 trillion valuation in 2023 to the first-ever $4 trillion and $5 trillion closes in 2025. Few companies have ever been so completely identified with a technological revolution, and none at this scale.

Source

From a Denny's booth to the GPU

Nvidia was planned, according to company lore, in a Denny's in East San Jose, where Jensen Huang, Chris Malachowsky, and Curtis Priem sketched out a bet that the personal computer would become a platform for real-time 3D graphics. Founded in April 1993 with about $20 million in venture capital, the company survived several near-death moments in the brutal 1990s graphics-chip market before shipping the GeForce 256 in 1999, the first product it marketed as a 'GPU' - a single chip that offloaded the heavy geometry and lighting math from the CPU.

For most of its first fifteen years Nvidia was a gaming-hardware company. The pivot that made it a trillion-dollar business came in 2006 with CUDA, a software platform that let developers run general-purpose parallel code on the GPU. It was an expensive, unproven gamble at the time, but it turned the GPU into a general computing engine. When AlexNet won the ImageNet competition on Nvidia GPUs in 2012, deep learning took off with CUDA as its foundation, and every wave since - from the first large language models on the Hopper H100 to today's agentic systems - has been trained and served largely on Nvidia silicon.

Source

The real moat is software and the full stack

It is tempting to explain Nvidia's dominance by raw chip performance, but the deeper and more durable advantage is software. CUDA has accumulated roughly 6 million developers and two decades of libraries, frameworks, and tooling that assume Nvidia hardware underneath. A competitor can match a transistor count; it cannot easily replicate the ecosystem of code, expertise, and muscle memory that makes Nvidia the path of least resistance for anyone building AI. Switching costs, not clock speeds, are what keep customers locked in.

Nvidia has widened that moat by selling not chips but entire AI data centers. Its rack-scale systems bundle GPUs with its own high-speed interconnects - NVLink for GPU-to-GPU links and Spectrum-X and InfiniBand networking - so that dozens of accelerators behave like one giant machine. Networking alone reached $14.8 billion in a single quarter in early 2026. The company has also committed to an annual cadence of new platforms, a punishing pace that forces rivals to chase a target that moves every twelve months. Owning the compute, the interconnect, and the software stack together is what makes Nvidia so hard to route around.

Source

Blackwell to Vera Rubin: the product engine

The current workhorse is the Blackwell platform, whose GB300 NVL72 systems link 72 GPUs into a single accelerator and have driven record data-center revenue since shipping in September 2025. But the story that matters most is the transition already underway to Vera Rubin, the next-generation platform Huang declared in full production at GTC in March 2026 - ahead of the early-2027 timeline analysts had expected. Vera Rubin is a seven-chip system spanning the Vera CPU, the Rubin GPU, NVLink 6, and, notably, a Groq 3 low-latency inference processor integrated after a December 2025 licensing deal.

Two things about Vera Rubin are strategically significant. First, its efficiency claims - training mixture-of-experts models with a quarter of the GPUs Blackwell needs, and up to 10x the inference throughput per watt - are aimed squarely at the economics of serving AI at scale. Second, the Vera CPU marks Nvidia's move into the data-center processor market that AMD and Intel have long owned, giving cloud builders a reason to buy Nvidia CPUs alongside its GPUs. Partner systems and cloud instances are arriving in the second half of 2026, with OpenAI among the first to deploy at scale, and Huang has projected roughly $1 trillion in cumulative Blackwell-plus-Rubin orders through 2027.

Source

Financials at a scale rarely seen

The numbers read like a typo. Fiscal 2026, which ended in January 2026, brought in $215.9 billion of revenue, up 65%, with GAAP net income of $120.1 billion - meaning Nvidia earned more profit in a year than most of the world's largest companies book in total sales. The momentum then accelerated: the quarter ended April 2026 set a record $81.6 billion in revenue, up 85% year over year, of which $75.2 billion was data center, and management guided the following quarter to $91 billion.

Two features stand out beyond the growth rate. The first is profitability: non-GAAP gross margins around 75% are extraordinary for a hardware company and reflect how little effective competition Nvidia faces at the leading edge. The second is concentration - data center is now roughly 92% of revenue, so the company's fortunes are tied almost entirely to the AI buildout. Confident enough in that demand to return cash, Nvidia paired its results with tens of billions in buybacks and a twenty-five-fold dividend increase, even as supply, not demand, remained the binding constraint on how much it could sell.

Source

Kingmaker of the buildout, and the circular-financing question

Nvidia is no longer just a supplier to the AI economy; it is one of its principal financiers. It has committed up to $100 billion to OpenAI to deploy at least 10 gigawatts of Nvidia systems, up to $10 billion to Anthropic, a central role in the $500 billion Stargate infrastructure project, and GPU commitments to Meta and a widening roster of AI clouds. These deals help guarantee demand for its chips and shape the direction of the entire industry, positioning Nvidia as the kingmaker deciding which labs get the compute to compete.

That same web of deals is the sharpest criticism leveled at the company. Skeptics call it circular financing: Nvidia invests in customers like OpenAI, CoreWeave, and Anthropic, who then turn around and spend that money on Nvidia hardware, a loop that critics argue can inflate perceived demand and tie the fate of every AI company to the same few balance sheets. Defenders, including Anthropic's Dario Amodei, frame it as the natural way to finance a capital-hungry technology shift where the chipmaker has money and the labs have conviction but not yet the cash. Whether these arrangements represent healthy ecosystem investment or a self-reinforcing bubble is the central debate hanging over Nvidia's valuation.

Source

What to watch next

China is the clearest risk in the numbers. US export controls forced a $4.5 billion charge in 2025, and although a December 2025 arrangement cleared H200 sales with a 25% cut paid to the US government, a Commerce Department official told Congress in July 2026 that actual shipments remained 'trivial.' Nvidia's own filings describe the company as effectively foreclosed from China's data-center market, and its guidance assumes zero China data-center revenue - so the upside if that market ever reopens is real, but the base case now excludes it entirely.

The competitive threat has also shifted. AMD is finally landing meaningful wins with its MI450 GPUs and Helios racks, including multi-gigawatt deals with OpenAI, Meta, and Anthropic, but the more structural challenge comes from Nvidia's own customers building custom silicon - Google's TPUs, Amazon's Trainium, and Broadcom-designed ASICs - to reduce their dependence. Nvidia's counter is partly to lock up scarce supply, having reportedly booked a large share of TSMC's advanced CoWoS packaging capacity that every rival accelerator also needs.

The largest question of all is the durability of AI capital spending. Nvidia's revenue is now a direct read on how much compute the world is willing to build, and any sustained pullback in hyperscaler and lab spending would hit it squarely. For now, an order book stretching toward $1 trillion, sold-out supply, and a clean transition from Blackwell to Vera Rubin suggest the demand is real. The things to watch are whether the Rubin ramp stays on schedule, whether margins hold as production scales, and whether the circular-financing structure that is fueling growth proves to be a foundation or a fault line.

Source