Opinion | How Jensen Huang Is Using Wall Street's Half A Trillion Dollars To Defend Nvidia's AI Empire

In the board game Go, a strong player commits a stone while already reading ten moves ahead. Nvidia chief executive Jensen Huang runs his company the same way. Nvidia is living through the most profitable stretch in its history: quarterly revenue above $80 billion, gross margins holding near 75 percent, GPU demand still climbing, and a market capitalization that sits at the very top of the global corporate league table. Yet even as Nvidia's empire looks unassailable, Huang is already several moves further down the board, asking how long the company's scarcity driven windfall can last once AMD closes the gap, cloud giants speed up their own chip programs, and China builds an entirely separate AI computing stack.

On August 10, Huang made his latest move. Nvidia announced a partnership with six financial heavyweights, Apollo Global Management, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, to mobilize more than $500 billion in third party capital for AI infrastructure, capital meant to make it easier for AI labs, cloud operators and enterprises to get their hands on GPUs and computing power.

It is a strange thing for a company sitting on enormous cash reserves and staggering profits to go looking for outside financiers to fund its own customers. The answer sits inside nearly two decades of strategy Huang has been quietly building: the CUDA software ecosystem, the 2019 acquisition of networking company Mellanox, an enormous portfolio of strategic equity stakes, and now a $500 billion financing platform. The common thread running through all of it is Huang's steady conversion of Nvidia from a GPU maker into something closer to a systems company that controls AI computing, networking, the surrounding ecosystem and even the flow of capital that pays for it all.

When Nvidia paid $6.9 billion for the Israeli networking company Mellanox Technologies in 2019, much of the market still saw Nvidia purely as a chip company and struggled to understand why Huang would pay so much for a maker of network interface cards, switches and high speed InfiniBand interconnect technology. The logic became clear only years later. As AI computing scaled up, the bottleneck stopped being how fast a single GPU could run and became how quickly thousands or tens of thousands of chips could exchange data with each other; a GPU cluster is only as fast as its slowest link, and if networking, memory bandwidth and data transfer cannot keep pace, the whole cluster idles while waiting. Mellanox closed exactly that gap, carrying Nvidia from GPU supplier into data center networking and giving it control over both computing and the communication layer that ties it together.

Nvidia's investment in CUDA runs even deeper. Years of building software tools, libraries and a developer ecosystem meant engineers were not just buying Nvidia GPUs, they were building inside Nvidia's software environment, and by the time generative AI took off, CUDA had already created enormous switching costs. From CUDA to Mellanox, and from GPUs onward into CPUs, switches, full server racks and data center design, Huang's logic has become unmistakable: control the chip first, then control what connects the chips, and finally define the entire data center itself as one giant computer.

If Mellanox pushed Nvidia from chips into systems, its equity stakes push Huang from technical systems into industrial ones. According to Nvidia's second quarter 13F filing for 2026, its eight disclosed equity positions were worth roughly $63.4 billion combined as of the end of June. Intel accounted for about $30 billion of that and SpaceX around $21 billion, together making up roughly 80 percent of the disclosed portfolio, with the remainder spread across CoreWeave, Coherent, Nokia, Synopsys and Nebius.

That $63.4 billion is a mark to market figure rather than money Nvidia has actually spent, and it leaves out private holdings and other economic interests entirely, but the list of names still lays out Huang's strategic map with unusual clarity. Intel connects Nvidia to CPUs, the x86 architecture, domestic manufacturing and advanced packaging. Coherent secures a position in optical communications. Nokia extends AI into 5G, 6G and telecom networks. Synopsys locks in chip design and simulation tools. CoreWeave and Nebius supply AI cloud computing capacity. SpaceX ties the portfolio to enormous AI demand, satellite communications and future computing infrastructure. Taken together, it reads less like an investment portfolio and more like a map of AI infrastructure, with Huang using equity stakes to stitch different corners of the industry into a network built around Nvidia's platform.

Huang's strategy now breaks down into three distinct layers. The first is technical: CUDA controls the software and the developers who build on it, GPUs supply raw computing power, and Mellanox, NVLink and BlueField handle the data transfer that ties chips, CPUs and full racks together into what Nvidia calls AI factories. The second is industrial: through acquisitions, equity stakes and supply chain partnerships, Nvidia has woven together CPU makers, foundries, optical communications firms, chip design software and cloud providers along with its largest AI customers. The third layer, only now taking shape, is financial. As AI servers and data centers grow ever more expensive, tech companies' cash flow has started falling behind the pace of the buildout, prompting Huang to pull Wall Street in to mobilize more than $500 billion in outside capital. Financial institutions can now use project financing, leasing, private credit or special purpose vehicles to buy GPUs and AI equipment and lease them out to computing providers, collecting interest and returns while AI companies pay rent instead of a lump sum upfront. In the process, GPUs are turning from a technology product into something that can be pledged as collateral, financed, leased and potentially even securitized. Where the $63.4 billion in equity stakes represents Nvidia cultivating its ecosystem with its own balance sheet, the $500 billion financing platform mobilizes Wall Street's balance sheet to do the same thing at far greater scale.

The more complete this strategy becomes, the more it exposes a deeper problem. Huang has repeatedly argued that AI data centers are becoming a new layer of infrastructure and that GPUs are essentially machines for manufacturing intelligence, something every country and every company will eventually need the way they need electricity, telecommunications and cloud services today. The trouble is that once a technology becomes genuine infrastructure, the cost per unit of use tends to keep falling as supply expands, specifications standardize and new competitors show up, squeezing the outsized margins equipment makers once enjoyed.

In the near term, Huang can still hold that line through rapid product cycles: a new generation of AI systems sells at a higher price, but if performance improves even faster, the cost per million tokens keeps falling for customers even as Nvidia sells increasingly expensive racks. Whether that balancing act can continue depends on two things, whether global AI demand keeps growing fast enough and whether Nvidia's customers can find real alternatives, and both of those conditions are already starting to shift.

The first challenge comes from AMD. Major AI developers including OpenAI, Meta and Anthropic have gradually deepened their use of AMD chips, a sign that AMD is moving from being a fallback option during GPU shortages to becoming a genuine second source of supply for large AI companies. AMD does not need to beat Nvidia outright; capturing even a modest share of the training and inference market gives customers real leverage to diversify their purchasing, which chips away at Nvidia's scarcity premium and pricing power on its own.

The second challenge comes from Nvidia's own customers. Google has its TPU chips, Amazon is expanding its Trainium line, Meta continues developing its MTIA chips, and OpenAI has partnered with Broadcom on custom silicon. These same companies keep buying enormous volumes of Nvidia GPUs while simultaneously building a second path meant to reduce their dependence on Nvidia. As the industry shifts from training models to running them at massive scale, the threat from purpose built chips may grow further still, since inference workloads are easier to optimize for a specific model, and the payoff from shaving down the cost of each inference only grows as usage scales up.

The third challenge comes from China. Chinese chipmakers remain behind Nvidia in advanced manufacturing, high bandwidth memory and per chip efficiency, but Huawei is trying to close that gap by combining high speed interconnects, networking equipment and sheer volume of chips. Huawei, Cambricon, Alibaba and a cluster of Chinese model developers are building a parallel ecosystem designed to steadily reduce its reliance on CUDA. Chinese suppliers do not need to match Nvidia's performance, only get close enough at a lower price while supporting domestic model training and inference, a combination that could win them enormous state owned enterprise, cloud and government contracts. American export controls limit China's access to advanced GPUs in the near term, but over a longer horizon they have also created a protected domestic market for China's chip industry, forcing Chinese firms toward homegrown chips, turning orders into research funding, and letting real world deployment build up software and systems expertise. Global AI computing may eventually split into two separate ecosystems, and even if Nvidia keeps its technical lead, CUDA's once unrivaled global dominance is likely to face mounting pressure.

Seen from this angle, the strategic purpose of the $500 billion financing platform comes into sharper focus. GPUs are expensive and depreciate quickly, and plenty of AI companies want computing power without having the cash on hand to pay for it upfront. With Wall Street now in the picture, that enormous capital expenditure can be converted into financing, leasing and collateral arrangements paid off over time. Financial leverage lowers the barrier for customers to buy into Nvidia's ecosystem and effectively recreates GPU orders that might otherwise have evaporated for lack of available cash.

Even more important is what this locks in early. Once a data center commits to Nvidia's GPUs, CUDA, NVLink, networking equipment and full server racks, the software, engineering talent, models and expansion plans that follow tend to grow up around that same architecture, and the more a customer invests, the higher the cost of ever switching away. It resembles the early days of railroads before track gauges were standardized: whoever lays down the most track first stands the best chance of becoming the eventual standard. Huang is effectively borrowing Wall Street's money to lay down Nvidia's AI tracks faster, racing to build an installed base large enough before AMD's software matures, cloud giants deploy their own chips at scale, and China's parallel ecosystem comes fully online. What $500 billion actually buys, in other words, is not just GPUs, but something far more valuable: time.

The most powerful part of this strategy may also be its biggest source of future risk. Nvidia invests in AI infrastructure companies, which in turn buy Nvidia GPUs; customers secure more financing and expand their data centers, lifting Nvidia's revenue; that revenue growth pushes up valuations across Nvidia and the companies in its orbit, generating still more capital to plow back into AI. In good times, this is an extraordinarily powerful capital flywheel.

But as GPU sales lean more heavily on financing, the way anyone reads AI demand has to change too. Imagine an AI company with $1 billion of its own money, enough on its own to buy $1 billion worth of GPUs; if Wall Street lends it another $2 billion, that company suddenly commands $3 billion in purchasing power. From Nvidia's earnings alone, GPU demand looks sharply higher, but a meaningful share of that new demand is simply purchasing power manufactured by financial credit. Watching this industry going forward means asking a further question every time: is this GPU demand being paid for with money AI companies earned, or money they borrowed? If AI investment returns fall short of expectations, the same flywheel can spin the other way, as customers cut capital spending and Nvidia's order book shrinks, valuations at AI infrastructure companies fall and shrink the value of Nvidia's own equity stakes, and falling GPU rental rates, utilization and resale values push up credit risk across the financing platform itself. Financial leverage can amplify an AI boom just as easily as it can amplify the correction that follows one.

Huang clearly understands that GPU performance leadership alone cannot sustain today's extraordinary margins forever, which is why he is building three deeper moats at once: CUDA and its software create technical lock in, equity stakes and partnerships create industrial lock in, and the $500 billion financing platform creates capital lock in. Even once GPUs themselves become a standardized commodity, as long as a large share of the world's AI systems keep running on CUDA, Nvidia's networking and Nvidia's system architecture, the company has a real shot at evolving from a chipmaker into something closer to a toll booth for the entire AI economy.

That also explains why Huang is moving now. This is the moment when Nvidia's margins are at their highest, its cash position at its strongest, its technical lead at its clearest, and Wall Street at its most willing to bankroll the AI story. In another three to five years, once AMD wins more orders, cloud giants deploy custom chips at scale and China's computing ecosystem matures, Nvidia's pricing power today could erode even if its technical lead survives intact. Looked at in sequence, Mellanox, the $63.4 billion equity portfolio and the $500 billion financing platform turn out to be three stages of a single strategy: Mellanox gave Nvidia control over the flow of AI data, the equity stakes embedded Nvidia throughout the AI supply chain, and the financing platform now channels Wall Street's own capital into that same ecosystem. What Huang is ultimately fighting for is the right to define the AI system, to provide the capital that builds it, and to keep collecting a share of revenue every time the world produces another unit of artificial intelligence. What he is really trying to buy is more time for Nvidia to keep ruling the AI world.

Original Article in Chinese

（Related： Jensen Huang Saw It Coming: How a $6.9 Billion Gamble Turned Nvidia Into AI's Infrastructure King ｜ Latest ）