Nvidia's SEC Filing Doesn't Name Its $70 Billion in Bets

Nvidia's balance sheet held $47.9 billion in equity stakes in privately held companies as of July 26, 2026, according to the quarterly report it filed with the Securities and Exchange Commission. The number that matters is not that total. It is what the total is buying: Bank of America estimates that $70 billion of Nvidia's cumulative equity commitments has gone directly into companies that are also among its largest chip customers, including as much as $30 billion into OpenAI alone. Nvidia's own filing does not name a single one of them.

The 10-Q, certified on Aug. 26, 2026 by Executive Vice President and Chief Financial Officer Colette Kress, reports $31.0 billion in new non-marketable equity investments added during the first half of fiscal 2027, alongside $7.5 billion in unrealized gains on those holdings during the same period. Cumulative gross unrealized gains across the whole non-marketable portfolio stood at $9.1 billion, against $250 million in cumulative losses and impairments, a ratio that shows the paper gains on Nvidia's stakes running more than thirty-six times its recorded losses to date. Nvidia separately reported $3.3 billion in equity-method investments in what it calls infrastructure financiers, carrying a maximum loss exposure, including future committed amounts, of $4.7 billion. Beyond all of that, Nvidia has committed to $25 billion in additional equity investments in AI model makers, infrastructure financiers and other private companies, spanning fiscal 2027 through 2032 and beyond, subject to contingencies the filing does not spell out.

The $70 billion breakdown Nvidia's own filing won't give you

None of those SEC figures come with company names attached. For that, the only public breakdown traces to equity research, not to Nvidia. Bank of America analyst Vivek Arya put Nvidia's committed direct equity investments at $70 billion in an Aug. 12 note: $30 billion into OpenAI, up to $10 billion into Anthropic, $5 billion into Safe Superintelligence, and an unspecified remainder spread across other ecosystem partners. Arya called the exposure "easily manageable," pointing to projected free cash flow of roughly $470 billion over 2026 and 2027 combined, which would make the $70 billion figure about 15 percent of that two-year total.

Recipient	Committed equity	Source
OpenAI	$30 billion	BofA analyst estimate
Anthropic	Up to $10 billion	BofA analyst estimate
Safe Superintelligence	$5 billion	BofA analyst estimate
Unspecified remainder	~$25 billion	BofA analyst estimate

1. OpenAI30 $ billion
2. Anthropic (up to)10 $ billion
3. Safe Superintelligence5 $ billion
4. Unspecified remainder25 $ billion

Source: Bank of America analyst Vivek Arya, via Yahoo Finance, Aug. 12, 2026, accessed 2026-09-12

Why the same dollar showing up twice is the actual question

The mechanism worth tracing is not whether Nvidia can afford these stakes; at 15 percent of a projected two-year free-cash-flow figure, Arya's math says it plainly can, even while Nvidia keeps its separate commitment, announced in May 2026, to return half of free cash flow to shareholders. The mechanism worth tracing is what happens to a dollar of Nvidia equity once it lands inside a lab like OpenAI or Anthropic and a portion of it comes back out as a purchase order for Nvidia accelerators. Reported industry demand for AI chips does not distinguish a dollar spent by a customer's independent revenue from a dollar that started on Nvidia's own balance sheet three steps earlier. Nvidia's 10-Q discloses the stakes as assets; it does not, and is not required to, disclose which portion of any single customer's chip orders trace back to Nvidia's own capital. The $3.3 billion Nvidia carries specifically as equity-method stakes in what it calls infrastructure financiers is the clearest version of this loop: those are companies whose entire business is arranging capital for someone else to buy and operate Nvidia hardware, and Nvidia's maximum loss exposure on that category alone, including committed but unfunded amounts, already runs to $4.7 billion.

Oracle's GPU delivery backlog, one of the clearest public proxies for how much AI infrastructure spending is booked rather than delivered, sits at roughly seven times Oracle's own full-year capital spending guidance. None of Oracle's backlog figures disclose whether any portion of the demand behind them traces to a chip supplier's own equity, either, which is exactly the disclosure gap this filing shares with the rest of the industry's AI capital chain: Nvidia's move into Hugging Face is the same pattern at outright acquisition scale rather than minority-stake scale.

What would change this read is a company-by-company breakdown from Nvidia itself, rather than an analyst's reconstruction, or evidence that a meaningful share of these stakes' recipients are buying compute from Nvidia's competitors instead. Absent either, the next filing to watch is Nvidia's third-quarter fiscal 2027 10-Q, which will show whether the equity book kept growing at the same pace as the $31.0 billion added in the first half of the fiscal year alone, or whether Kress's own certification eventually comes with a company-by-company table attached rather than a single aggregate line.