From $2.2B to $99B in Two Years: Inside Nvidia's AI Empire-Building Balance Sheet

Two years ago, Nvidia’s equity investment portfolio was worth $2.2 billion — a rounding error for a company that now flirts with a $5 trillion market capitalization. One year ago it was roughly $7 billion. As of July 26, 2026, according to CNBC’s analysis of the chipmaker’s latest filing, that number is $99 billion — a 10x expansion in twelve months and a 45x expansion in two, making Nvidia not just the dominant supplier of AI compute but arguably the single most important investor in the AI economy.

The scale is staggering enough. The structure is what should worry — or excite — you.

What’s in the $99 billion

Nearly half the portfolio, close to $50 billion, sits in AI labs themselves. The centerpiece is a planned $30 billion stake tied to OpenAI’s $110 billion round announced in February — a round that valued the ChatGPT maker at a $730 billion pre-money valuation and also drew $30 billion from SoftBank and $50 billion from other investors. Another roughly $10 billion went into Anthropic, which Nvidia CEO Jensen Huang has since signaled may be its last such bet as both labs march toward IPOs that would force Nvidia to unwind or restructure its positions.

Then there are the neoclouds. Nvidia has written $2 billion checks into CoreWeave and Nebius, the GPU rental firms whose fortunes rise and fall with demand for Nvidia’s own silicon. These are the companies that buy Nvidia chips, wrap them in data centers, and rent them by the hour to the same AI labs Nvidia also funds — a loop that critics have spent all of 2026 calling “circular financing.”

Beyond the labs and clouds, roughly $6.5 billion committed since March has flowed into photonics and optical networking — Lumentum, Coherent, and Marvell — the companies building the interconnects that let hundred-thousand-GPU clusters behave like a single machine. And the portfolio is increasingly not just equity: Nvidia’s $12.93 billion acquisition of Hugging Face (filed via 8-K on September 2, closing expected in H1 2027) converts the open-source AI hub from a strategic stake into outright ownership, alongside earlier asset deals like the $20 billion Groq buy.

The empire logic

Strip away the numbers and a coherent strategy emerges. Nvidia is no longer content to sell picks and shovels in the gold rush — it is buying stakes in the gold mines, the land, the railroad, and the assayer’s office simultaneously.

The vertical integration runs in every direction:

• Downstream into customers: OpenAI, Anthropic, CoreWeave, Nebius — the entities that account for a huge share of ultimate GPU demand
• Sideways into the supply chain: Lumentum, Coherent, Marvell — the photonics layer that determines how large GPU clusters can physically scale
• Horizontally into platforms: Hugging Face, the distribution and community layer for the open-source models that run on Nvidia hardware
• Financially into the money itself: the $500 billion financing facility organized with Wall Street asset managers in August, built on Huang’s argument that Nvidia chips are an “investable asset” that can be financed like infrastructure

That last piece deserves its own scrutiny. The WSJ described the arrangement as an “exotic money pipeline” — chip purchases become securitizable assets, Wall Street provides the leverage, and Nvidia’s ecosystem gets its funding without Nvidia itself carrying the whole balance sheet. Standardize the financing of GPUs the way mortgages were standardized, and the size of the market stops being limited by anyone’s cash on hand.

The circular financing debate

Every dollar of this expansion has intensified the central critique of the AI buildout: Nvidia invests in companies that buy Nvidia chips, generating revenue that Nvidia reports to investors who then value Nvidia higher, enabling more investment. When Nvidia-backed CoreWeave takes delivery of GPUs, books rental revenue from OpenAI — also Nvidia-backed — the question of where real, external, end-customer demand begins and vendor financing ends gets genuinely hard to answer.

Huang’s defense, delivered to CNBC on August 26, is that the risk is low because the positions are equity, not receivables, and that this is a “once in a generation” opportunity to secure the ecosystem that his chips anchor. He has a point about structure: equity stakes don’t come due like loans, and Nvidia’s stake in OpenAI was restructured precisely to avoid the regulatory tangles of the originally discussed $100 billion deal, which Reuters reported the two sides abandoned in February amid doubts about AI-sector financial health.

But even Huang has begun drawing lines. In March he told Reuters that Nvidia “will not be able to invest $100 billion in OpenAI” due to the lab’s IPO path — a notable walk-back from the headline number that framed the original partnership. And Sam Altman, OpenAI’s CEO, separately warned this week of “the first signs of what feels to me like unsustainable silliness” among neocloud providers pledging capacity with no matching buyers — a warning that, read carefully, indicts parts of Nvidia’s own portfolio.

Why $99 billion is the number that matters

There is a version of this story where the portfolio is simply brilliant: Nvidia converts a once-dominant market position into permanent structural advantage, earning chips-sales margins and venture-style returns on the entire AI stack. If AI demand compounds the way bulls expect, the $99 billion will look like the cheapest empire ever assembled.

There is another version, and it rhymes with Cisco 2000. A supplier inflates demand for its own product by financing its customers, revenue quality degrades as end-user demand fails to keep pace with capacity commitments, and the unwinding — when it comes — hits the financier and the supplier simultaneously because they are the same entity. Nvidia today is both the largest beneficiary of AI capex and, increasingly, its largest private underwriter. Those are the two positions that suffer most in a downturn.

Which version plays out depends on a variable nobody controls: whether the applications being built on all this rented compute start generating revenue at anything close to the pace of the infrastructure bills. Broadcom’s guidance of $21.7 billion in AI chip revenue for next quarter, Korea’s record $46.65 billion in chip exports, and the PwC projection of $31.6 trillion in data center capital mobilization through 2040 all say the demand side is still rising. The $99 billion says that if it ever stops rising, Nvidia will feel it from every direction at once — as chipmaker, as investor, and as the AI economy’s de facto central bank.

For now, the machine is running. Two years, forty-five times, one balance sheet. The most consequential question in AI may no longer be who has the best model — it’s whether anyone can afford for Nvidia to be wrong.