In July this year, NVIDIA learned that OpenRouter, a startup operating an AI model marketplace, was about to be acquired. The three-year-old company had already become a popular platform for trading AI models, hosting a wide range of cost-effective open-source models. Databricks and Stripe were already in negotiations to acquire the startup, leaving NVIDIA late to the game. Upon learning the news, NVIDIA’s CEO, Jensen Huang, wanted to join the acquisition.
A person familiar with the negotiations said that NVIDIA executives expressed interest in acquiring OpenRouter and were prepared to make a generous offer, but needed more time to evaluate the deal before making a formal bid. However, OpenRouter’s founder was unwilling to wait. Ultimately, NVIDIA did not issue a formal acquisition offer—something never previously reported by any media outlet—and Stripe secured OpenRouter with an $8 billion bid.
NVIDIA's deal team quickly shifted focus to other acquisition targets. Over the next two months, NVIDIA finalized a variety of transactions totaling over $140 billion, including a $105 billion credit facility. By the end of July, NVIDIA held nearly $100 billion in equity investments and had an additional $25 billion in future investment commitments, with more transactions potentially on the horizon.
Despite investor concerns that NVIDIA and other tech giants have overextended financially in their AI-related investments, multiple investment banks, law firms, and investors collaborating with NVIDIA say the chip designer is expected to continue this wave of large-scale transactions over the coming months. NVIDIA is actively seeking startups to invest in or acquire, particularly those focused on robotics, autonomous driving technologies, and AI models capable of running directly on local devices such as smartphones and home computers.
A source revealed that, in previously undisclosed news, NVIDIA has discussed making an additional $1 billion investment in the humanoid robotics company Figure. Figure’s valuation prior to this round was approximately $38 billion, and the company is currently raising new funds. (NVIDIA is already an existing investor in Figure; the company completed over $1 billion in funding at this valuation a year ago.)
According to insiders, NVIDIA may also acquire startups to support its self-developed Nemotron open-source large model and expand into other AI application sectors. One reason NVIDIA agreed to invest $6 billion to acquire Poolside’s software licenses and integrate its team is to advance the Nemotron project. This year, the Poolside team launched the Laguna open-source AI model.
This series of investment moves is led by a manager who has always maintained a strong sense of urgency regarding NVIDIA’s industry position. Huang has previously stated that he has long been concerned that NVIDIA’s lead could erode—due to a cooling in AI demand or the emergence of competitors whose chips offer sufficient performance to replace NVIDIA’s.
With $99 billion in cash and marketable securities, along with a steady stream of cash flow, Jensen Huang aims to leverage NVIDIA’s strong financial position to build an industry ecosystem featuring thousands of high-quality AI models, rather than allowing just a few model providers to dominate. This approach helps mitigate the risk of excessive revenue dependence on a small number of major clients like OpenAI. NVIDIA disclosed that, in the six-month period ending in July, its three largest customers accounted for 44% of total sales.
An investor familiar with NVIDIA’s trading strategy and investments in AI infrastructure and applications said: “If I were in Jensen Huang’s war room, I would do everything in my power to create a scenario where thousands of models serve millions of different use cases.”
This also explains why NVIDIA's trading engine is running at full speed.
Reports indicate that NVIDIA has recently been in talks to invest approximately $2.5 billion in Thinking Machines Lab, an AI laboratory founded by Mira Murati, former Chief Technology Officer of OpenAI. If finalized, this investment would be the latest in NVIDIA’s series of investments in model companies, following prior investments in Anthropic, Elon Musk’s xAI, and open-source model company Reflection AI.
NVIDIA is also a major financial backer of large-scale data center projects: investing $3 billion in SB Energy, which is building a massive data center for OpenAI; meanwhile, NVIDIA is providing a $105 billion guarantee for OpenAI’s lease agreement for the project.
Jensen Huang wrote in a blog post that the data center campus developed by SB Energy in collaboration with OpenAI can support approximately $600 billion in NVIDIA computing power. NVIDIA chose to support OpenAI’s project because “leading AI labs have enormous demands for training and inference computing, but many companies are expanding faster than their balance sheets and long-term creditworthiness can sustain.”
Although he missed out on OpenRouter, when competitors emerge for other targets, Huang Renxun acts quickly.
Using the recent acquisition of Hugging Face as an example, investment banks and individuals close to the company say that this decade-old startup has become a well-known platform for hosting open-source AI models and has consistently received acquisition offers. Sources indicate that NVIDIA’s corporate development team has long been interested in investing in Hugging Face.
A person familiar with the matter revealed that after OpenAI’s AI agents launched an attack on Hugging Face in early summer this year, OpenAI initiated preliminary discussions with Hugging Face regarding a potential $100 million investment. Around July, Hugging Face’s existing investors, including competitors like Salesforce, also expressed interest in acquiring the company.
Hugging Face co-founder Clément Delangue approached Jensen Huang with these potential acquisition offers. According to people familiar with the matter, Huang moved swiftly to advance the deal, assuring Delangue that NVIDIA would be the only trustworthy partner capable of ensuring the continued operation of Hugging Face’s open-source model community. NVIDIA offered $12.9 billion for the acquisition—an offer too compelling to refuse, equivalent to more than 80 times the startup’s annualized revenue of $150 million.
Delaunay said at the press conference: "Throughout its journey, Hugging Face has received numerous investment and acquisition offers, all of which we declined in the past. But this summer, everything fell into place, and the timing was just right."
Local computing赛道
Several individuals who have worked with NVIDIA say that over the coming months, NVIDIA plans to pursue more deals to expand GPU usage beyond data centers and into home devices, enabling users to run AI locally on personal computers and small devices.
This year, the widespread adoption of AI agents—software that can execute multi-step tasks such as booking flights or organizing email—has led to a surge in demand for local AI. These agents have driven users to purchase large quantities of the Apple Mac mini, a device well-suited for running such tasks.
NVIDIA has launched new products such as the DGX Spark, aiming to enable local execution of AI agents. NVIDIA's recent acquisitions also reflect its intent to expand its product portfolio in this space.
According to insiders, the startup Perplexity initially began as an AI search engine and later developed the AI agent Perplexity Computer. In June this year, Perplexity engineers demonstrated to NVIDIA’s team that their software could run on two DGX Spark systems.
After hearing the demonstration, Huang Renxun became highly interested, and the two companies spent the entire summer negotiating a potential deal. According to knowledgeable sources, Aravind Srinivas, co-founder and CEO of Perplexity, proposed to Huang Renxun that NVIDIA could acquire Perplexity directly.
Subsequently, both parties shifted to discussing a technology licensing and talent acquisition deal; according to insiders, NVIDIA is expected to invest at least $20 billion, and possibly more, to acquire Perplexity’s related technologies.
Both parties officially announced their partnership at the end of August. Under the agreement, Perplexity has launched its new application, Portable Computer, which has been custom-optimized to run more efficiently on DGX Spark devices. According to a person familiar with the matter, NVIDIA plans to invest approximately $3 billion in Perplexity in this round of equity financing, valuing the company at $35 billion prior to the investment.
Financial risk
Like most tech companies, NVIDIA has a corporate development team responsible for executing deals. The team is led by Vishal Bhagwati, formerly of Hewlett Packard and Oracle. Nevertheless, Jensen Huang frequently engages deeply in deal details, personally finalizing transaction prices and leading high-value acquisitions such as Hugging Face.
Jensen Huang also frequently meets with startup founders, investors, and private equity executives to understand how they use NVIDIA’s products and how NVIDIA can better support them. Executives like Satya Nadella, CEO of Microsoft, also use this informal research approach to stay attuned to technological trends.
Over the past few months, a new reality has confronted Jensen Huang, potentially limiting NVIDIA’s acquisition ambitions: NVIDIA may have reached the upper limit of funds that can be safely allocated for investment, acquisitions, and large-scale project guarantees.
SB Energy, a subsidiary of SoftBank, is developing a massive data center project on federal land in Ohio, and its financing plan reflects market concerns. Earlier this summer, SoftBank initially explored with NVIDIA the possibility of the chip giant providing up to $250 billion in credit support to OpenAI, which plans to lease the data center and use NVIDIA chips to train and run its models. (NVIDIA also participated in OpenAI’s recent fundraising, investing $30 billion, with the final $10 billion settled on October 1.)
In August, NVIDIA's credit default swap spread widened, reflecting concerns among some investors that the company was taking on excessive risk. Huang noticed this metric and frequently asked his colleagues about changes in the spread.
NVIDIA ultimately provided only $105 billion in credit guarantees for Phase One of the project—a still massive amount, but less than half of the initially discussed figure. The project will be built in phases, giving NVIDIA several years to decide whether to provide funding support for Phase Two.
Meanwhile, NVIDIA continues to advance another investment, planning to invest $3 billion in SB Energy prior to and during its IPO.
Huang emphasized that the continued advancement of AI requires greater financial investment from companies in chips, data centers, and power projects. In early August, he brought together six Wall Street institutions—including Blackstone, Apollo, and Goldman Sachs—to collectively raise $500 billion for purchasing NVIDIA hardware. NVIDIA stated that, in certain related transactions, it may provide a guarantee of up to 25% of the total financing amount.
But Huang Renxun downplayed the risks associated with such massive investments, believing that capital-rich companies like NVIDIA are not overinvesting.
In September, Jensen Huang told the audience at Goldman Sachs’ annual technology conference in San Francisco: “People are beginning to realize that when I invest in something, it usually works out, because I have an information advantage. I don’t take blind risks. We’re not as brilliant as people think—we just want projects with a sure win.”
Editor-in-Chief: Chen Yujia