Nvidia Just Crossed $40 Billion in AI Equity Bets. The Customer-Investor Loop Is the Real Moat.

CNBC ran the tally on Friday: Nvidia's 2026 equity commitments to AI companies have crossed $40 billion. That number is anchored by a $30B stake in OpenAI from late February, and it grew by another $5.3 billion this week alone (up to $3.2B into Corning and up to $2.1B into the data center operator IREN). Add roughly two dozen private startup rounds and seven multi-billion deals in publicly traded companies, and a chip vendor is now running one of the largest active venture programs on the planet.

Critics call it circular. They are not wrong about the mechanics. They are arguably wrong about what the strategy is for.

The Numbers, In One Place

Nvidia's investment portfolio is fragmented across a $30B anchor, a string of multi-billion public-equity stakes, and the private rounds the company files quietly. Here is the disclosed 2026 picture as of this morning.

The line item that matters most for understanding the program is the rightmost column. Every single one of these checks pairs with a commercial commitment in the other direction. Nvidia is not running a passive index fund.

What Each Deal Actually Trades

The IREN deal is the cleanest read. Up to $2.1 billion of equity, paired with a partnership to deploy up to 5 gigawatts of Nvidia DSX-branded infrastructure across IREN's global facilities. DSX is the reference design Nvidia is pushing as the standard rack template for AI workloads, and it is mostly Nvidia silicon by spec. Five gigawatts at current rack densities is on the order of 2 to 3 million accelerators' worth of deployment runway. IREN takes the capital, builds the facilities, and the system orders flow back into Nvidia's top line on the same paper.

The Corning deal trades capital for fab capacity. Three new US plants dedicated to optical technologies, which Corning will retool to Nvidia's spec. The shift here is from copper interconnect to fiber-optic cabling for rack-scale systems, a switch that pencils out at the GB200 NVL72 form factor and is a hard requirement for the Rubin generation that follows. Nvidia did not have to acquire Corning to get this. They had to make sure Corning built the fabs in time, and equity is the lever that gets a public company to commit the capex on someone else's timeline.

The OpenAI stake is the strangest of the three on paper, because OpenAI is not a hardware buyer in the conventional sense. It buys cloud capacity from Microsoft, Oracle, AWS, and now (per the Anthropic playbook everyone is copying) directly contracted megawatts. The $30B equity check is closer to a ten-year option on OpenAI's silicon roadmap. As long as OpenAI's training and inference workloads stay on Nvidia, every dollar of that relationship throws off margin Nvidia keeps.

That option matters more now than it did six months ago. Anthropic just locked in $200B of Google TPU capacity, and we walked through the math on that here. If Anthropic is the proof of concept that a frontier lab can run on non-Nvidia silicon at scale, OpenAI is the next domino Nvidia is paying to keep upright.

The Circular Investment Critique

Wedbush's Matthew Bryson put it on tape this week: the deals fall "squarely into the circular investment theme." The shape is familiar. Nvidia hands a customer money, the customer hands it back as system orders, the revenue books as growth, the growth supports the multiple, the multiple funds the next equity check.

The 1999 comparison everyone reaches for is Cisco's vendor financing program. Cisco extended credit to its customers (mostly competitive local exchange carriers and dot-com builders) so they could buy Cisco routers. When the customer base went bust in 2001 and 2002, Cisco wrote down billions of receivables and the stock lost three quarters of its value inside two years. The cautionary tale is real.

But the analogy is incomplete in two ways that matter.

First, the asset side. Cisco was extending unsecured credit. Nvidia is buying equity. If IREN's 5 GW deployment underperforms, Nvidia's upside on the equity goes down, but the equity does not become a write-down on receivables sitting against shipped product. The accounting failure mode is genuinely different.

Second, the customer base. Cisco's 1999 buyers were funded by junior debt and IPO proceeds. Nvidia's 2026 buyers are funded by Microsoft, Google, Amazon, SoftBank, and (in IREN's case) hyperscaler take-or-pay contracts that survive a recession. The counterparty quality is a different category.

Where the comparison does hold: capital that flows in a circle is capital that masks the underlying demand signal. If 30% of Nvidia's 2026 booking growth is funded by checks Nvidia itself wrote, the actual end-customer demand is smaller than the headline. That is a real thing to watch in the next two earnings cycles.

What the Loop Actually Locks In

The cleaner read is that Nvidia is buying defense, not growth. The competitive landscape in 2026 is the most credible threat the company has ever faced.

Google's TPU economics are 40 to 50% lower than equivalent Nvidia capacity at the very top of the buyer list, by the math we ran on the Anthropic deal. AWS Trainium and Inferentia are running production inference for some of Anthropic's and Bedrock's workloads. Microsoft's Maia silicon is shipping in Azure regions. AMD's MI400 family is real and credible enough that Meta and Oracle have both signed multi-billion-dollar commitments for it. The buy-side has options it did not have eighteen months ago.

When you look at the equity portfolio through that lens, the strategy reads differently. The OpenAI stake locks in roughly $20B per year of frontier-lab compute spend on Nvidia. The IREN stake locks in 5 GW of capacity that defaults to Nvidia silicon. The Corning stake locks in the optical components without which the next two Nvidia generations cannot physically scale. None of these are growth bets in the standard sense. They are perimeter fences.

The Three Risks Worth Watching

The strategy is coherent, but it carries real failure modes. The three I am tracking through the next two quarters:

Concentration in a falling market. If AI capex decelerates in late 2026 or 2027, the equity book carries the loss twice: once on the equity mark and once on the system orders that no longer materialize. Cisco lived through that symmetry, and Nvidia is more concentrated in this cycle than Cisco was in 1999.

Antitrust on the customer-investor pair. The FTC and the EU Commission have both telegraphed interest in vendor-financed AI deals. A consent decree that limits Nvidia's ability to bundle equity with capacity commitments would defang the strategy without anyone needing to break up the company.

A non-Nvidia frontier model that wins on benchmarks. The single fastest way to break the loop is for an Anthropic, a Google, or a DeepSeek successor to ship a state-of-the-art model trained entirely on non-Nvidia silicon and run inference at half the per-token cost. Anthropic is one TPU generation away from that experiment. We are tracking the silicon mix on every frontier release on the models tracker.

Our Take

Nvidia is not the first chip company to try to lock in its customers with capital. Intel tried it in the late 1990s with the Communications and Computing initiative. ARM has flirted with equity-for-design-win deals for two decades. What is new is the scale, the velocity, and the fact that the customer base is a small enough number of frontier labs that a $40B program can reach most of them in a year.

The customer-investor loop is real, the circular-investment critique is real, and they are both rounding errors next to the question that actually matters: how many years of frontier AI training run on Nvidia silicon by default. If the answer is three or more, the equity book is the cheapest moat money can buy. If a non-Nvidia frontier lab ships in 2027, the $40B looks very different on the next 10-K.

We are adding the Nvidia portfolio to our funding tracker this week, with each commitment tagged by counterparty silicon dependency. The most useful single number for tracking whether the loop is working is going to be the share of frontier training compute (in FLOPs) that runs on Nvidia in any given quarter. We will start publishing it.