# Everyone Trades NVIDIA. Nobody Watches What NVIDIA Buys.

> We pulled NVIDIA's latest 13F: $18.4B filed, worth ~$35B today, mapped across four AI bottlenecks. Intel +364%, Nokia +130%, Nebius +948% on cost.

- Author: Barebone Research, Barebone AI
- Published: 2026-06-10
- Canonical: https://barebone.ai/resources/nvidia-13f-portfolio-ai-bottleneck-map
- Publisher: Barebone AI (https://barebone.ai)

---

## The Most Watched Company on Earth Has a Portfolio Almost Nobody Reads

NVIDIA is the most analyzed stock in the world. Every earnings call is live-blogged, every Jensen Huang keynote is parsed word by word, every supply chain rumor moves billions. And yet four times a year, NVIDIA publishes a document that most of that audience never opens: its 13F — the SEC filing where any institution managing over $100 million must disclose the stocks it owns.

That's right: NVIDIA owns stocks. Not as a hedge fund, and not to trade. NVIDIA buys companies it cannot afford to lose — its suppliers, and the customers it needs to keep buying chips.

We used Barebone to pull NVIDIA's filings going back two years, every investment press release, and a year of price history for every name it holds. The latest filing, covering March 31, 2026, shows **$18.4 billion across just seven stocks**. Marked to today's prices, those same seven positions are worth roughly **$35 billion** — the portfolio nearly doubled in ten weeks, and four of the seven holdings are beating the S&P 500 this year by 85 points or more.

But the returns aren't the story. The story is what the portfolio *is*: a map of the exact bottlenecks the AI buildout has to squeeze through, drawn by the one company that can see the entire supply chain from the inside. Read it correctly, and the most important man in AI is showing you where the money goes next.

## The Map

Here is the complete portfolio — all seven holdings, what NVIDIA paid, and what they're worth now:

| Holding | Role in the map | NVIDIA's cost | Value at filing (Mar 31) | Value today (Jun 9) | Return on cost |
|---------|----------------|---------------|--------------------------|---------------------|----------------|
| **Intel** (INTC) | The factory | $5.0B at $23.28 | $9.5B | **$23.2B** | **+364%** |
| **CoreWeave** (CRWV) | The customer | ~$2.4B blended | $3.7B | $4.6B | +146% vs IPO price |
| **Coherent** (COHR) | The wires | $2.0B at ~$257 | $1.9B | $2.8B | +39% |
| **Nokia** (NOK) | The wires | $1.0B at $6.01 | $1.3B | $2.3B | **+130%** |
| **Synopsys** (SNPS) | The blueprints | $2.0B at $414.79 | $1.9B | $2.2B | +12% |
| **Nebius** (NBIS) | The customer | $25M at $21.00 | $124M | $262M | **+948%** |
| **Generate Biomedicines** (GENB) | The wildcard | undisclosed | $10M | $11M | — |

<Chart name="NvidiaPortfolioCompositionChart" />

Seven names, but really four ideas. NVIDIA's own framing for its products is "AI factories" — and its stock portfolio buys the factory's choke points: the **customers** who buy chips by the hundred thousand (CoreWeave, Nebius), the **blueprints** every chip is designed with (Synopsys), the **factory** that could one day build them outside Taiwan (Intel), and the **wires** that connect them at the speed of light (Coherent, Nokia).

Each one is a gate the entire AI buildout has to pass through. Everyone needs them. Only a few companies make them. Let's take the layers one at a time.

## Layer 1: The Customers

NVIDIA's deepest relationship is with CoreWeave — a "neocloud," a company that buys NVIDIA GPUs by the hundred thousand and rents them to everyone racing to build AI. NVIDIA invested around $100 million in 2023, anchored the March 2025 IPO, then added **$2 billion more at $87.20 in January 2026**. It now owns about 11% of the company.

But the equity is the smaller half of the relationship. In September 2025, CoreWeave disclosed that NVIDIA had signed an order committing to purchase CoreWeave's **unsold cloud capacity — up to $6.3 billion of it — through April 2032.** Think about what that does: NVIDIA sells CoreWeave the chips, then guarantees the revenue those chips generate if no one else shows up. The money leaves NVIDIA, and comes right back. CoreWeave, for its part, grew revenue 112% last quarter and carries a $99.4 billion backlog.

Nebius is the same idea with a better entry. NVIDIA put roughly **$25 million** into the Amsterdam-born neocloud at $21 a share in December 2024. The stock closed at $220.12 on June 9 — the stake is up **roughly 10x**, the best trade in the book on a percentage basis. Microsoft then signed a capacity deal with Nebius worth up to $19.4 billion, Meta followed with $27 billion, and in March 2026 NVIDIA announced another $2 billion investment that will show up in the next filing.

Notice the pattern: NVIDIA doesn't just pick the customers. It *manufactures their demand* — with its own capacity purchases, and by being the supplier whose chips make the customer fundable in the first place.

## Layer 2: The Blueprints

On September 9, 2025, Synopsys — the company whose software is used to design almost every advanced chip on earth, NVIDIA's included — had the worst day in its 33 years as a public company. A weak quarter and a guidance cut sent the stock down **36% in one session**.

Less than three months later, NVIDIA bought **$2 billion of it** at $414.79, alongside a partnership to accelerate Synopsys's chip-design and simulation software on NVIDIA hardware.

This is the most quietly strategic position in the book. Electronic design automation — EDA, the software that turns a chip architecture into something a fab can actually manufacture — is dominated by two companies, and Synopsys is one of them. Control the tools that design chips and you have leverage over everyone trying to build one, including every custom-silicon project at every hyperscaler that might one day compete with NVIDIA. The position is up a modest 12% — the only holding not beating the market this year. NVIDIA wasn't buying momentum. It was buying the dip in infrastructure it cannot operate without.

## Layer 3: The Factory

The biggest position started as the boldest call. In September 2025, NVIDIA announced a **$5 billion investment in Intel** at $23.28 per share — back when Intel was the industry's punchline, and the deal came packaged with joint development: Intel building custom x86 CPUs for NVIDIA's AI platforms, NVIDIA GPU chiplets inside Intel PC processors.

The logic was geographic. Nearly every advanced AI chip on earth is manufactured in one place — Taiwan. TSMC has no real substitute, and that makes the entire AI economy hostage to a single strait. Intel — backed by a ~10% US government stake and $2 billion from SoftBank announced the month before — is the only credible path to advanced chipmaking at scale on American soil. NVIDIA's $5 billion was a hedge against the single largest tail risk in its business: the day Taiwan goes dark.

Then the hedge became the jackpot. Intel's 18A manufacturing process started yielding, reports emerged of Google and Apple foundry interest, and the stock went vertical: **+192% this year, +364% from NVIDIA's entry.** The position is now worth $23.2 billion — about **65% of the entire portfolio** — and generated the bulk of the $15.9 billion in equity gains that showed up in NVIDIA's own income statement last quarter. A geopolitical insurance policy is currently the single best-performing big-cap trade of 2026.

## Layer 4: The Wires

Once you have the chips, the next bottleneck is connecting them. A modern AI cluster is hundreds of thousands of GPUs that must behave like one computer, and the data moving between them travels on optics — lasers, transceivers, fiber. This is exactly where NVIDIA's two newest large positions sit.

**Coherent** makes the optical transceivers and photonics that carry data between racks. In March 2026, NVIDIA invested **$2 billion**, paired — same pattern as CoreWeave — with a *multibillion-dollar purchase commitment* and capacity rights for next-generation optical technology. Coherent says its AI transceiver demand is booked through 2028. The stock is up **93% this year** — it nearly doubled, touching an all-time high of $440 on June 3 before giving some back in this week's selloff.

**Nokia** — yes, that Nokia — got a **$1 billion** NVIDIA investment at $6.01 in October 2025, tied to building AI-RAN: NVIDIA compute embedded in telecom radio networks. The stock jumped 22% the day of the announcement and is up **130% since**, as its AI and cloud networking sales grew 49% last quarter. A twenty-year smartphone casualty is now an AI infrastructure play, and NVIDIA called the turn.

If you want the cleanest confirmation that the wires are the bottleneck, it's in NVIDIA's own results: last quarter, data center *networking* revenue grew **+199% year over year** — the fastest-growing line in the entire company, more than double the growth of the data center business overall.

(And the seventh position? **Generate Biomedicines** — a $10 million stake in an AI drug-discovery company that IPO'd in February. It's 0.03% of the book, the only position that was a genuine surprise in the filing, and best read as a sign of where NVIDIA thinks AI compute goes after the infrastructure is built: biology.)

## The Kingmaker Effect

Here's where we need to correct a popular narrative — including one we've leaned on ourselves. The story goes: *NVIDIA files its trades quietly, almost nobody reads 13Fs, and when the market finds out, the stocks explode.*

That was true once. When NVIDIA's first detailed 13F dropped in February 2024, SoundHound jumped **+66% in a day**. Serve Robotics popped **+187%** when NVIDIA's stake surfaced in July 2024. WeRide ripped **+84%** on the February 2025 filing. The filings were genuine surprises, and the market repriced them violently.

But the mechanism has changed — NVIDIA got too big to be quiet. Every major 2025–26 position was announced with a press release long before any filing: Intel jumped **+23%** on announcement day, Nokia **+22%**, Nebius **+16%**, Coherent **+15%**. By the time the May 15 filing formally confirmed them, there was nothing left to reveal — no holding popped on filing day at all.

<Chart name="NvidiaKingmakerChart" />

The kingmaker effect is real — arguably more powerful than ever, since the moves now come with billion-dollar commercial contracts attached. What's gone is the lag. You no longer get paid for reading the filing nobody else reads; you get paid for understanding, faster than the market, what each announcement means for the *next* bottleneck. The 13F's remaining edge is the details: position sizes, what got trimmed, and the occasional unannounced wildcard like Generate.

And the scoreboard says the announcements are worth understanding:

<Chart name="NvidiaPortfolioReturnsChart" />

Every holding except Synopsys is beating NVIDIA's own stock this year, and four of the seven are beating the S&P 500 by 85 points or more. The companies NVIDIA points at outperform the company doing the pointing — the same pattern we found when we tracked Jensen's GTC keynote themes, where the infrastructure *around* NVIDIA (Vertiv +205%, TSMC +95%) crushed NVIDIA itself.

## The Circular Question

Now the uncomfortable part, because an honest read of this portfolio requires it.

The bear case has a name and a famous track record. Michael Burry spent late 2025 arguing that the AI boom is held up by exactly the structures this article describes — chip vendors investing in customers who use the money to buy chips. His verdict, verbatim: **"And once again there is a Cisco at the center of it all, with the picks and shovels for all and the expansive vision to go with it. Its name is Nvidia."** The CoreWeave backstop is Exhibit A: if NVIDIA guarantees its customer's unsold capacity, how much of that demand is real?

NVIDIA took the criticism seriously enough to send a rebuttal memo to Wall Street analysts — itself a remarkable event — stating that it "does not use special-purpose vehicles, does not rely on vendor financing," and that its strategic investments totaled $4.7 billion year-to-date against revenue that ran $57 billion *per quarter*. The scale point is fair: the entire $18 billion public portfolio is about three weeks of NVIDIA's revenue. The circularity exists, but it is not load-bearing for NVIDIA's income statement.

What the portfolio *is* increasingly load-bearing for is NVIDIA's reported profits: last quarter, **$15.9 billion of NVIDIA's $58.3 billion in GAAP net income — about 27% — was mark-to-market gains on these very positions**, mostly Intel. That's not fraud; it's disclosed plainly. But it means NVIDIA's headline earnings now move with its stock picks, and a bad quarter for Intel becomes a bad optics quarter for NVIDIA.

Two more caveats before you treat the 13F as a treasure map. It's stale by design — filed 45 days after quarter-end, so positions may have changed. And it only shows *public* equities: NVIDIA's private book is now **larger than its public one** — $43 billion of non-marketable investments including $30 billion into OpenAI's latest round and up to $10 billion in Anthropic. The 13F is the visible tip of a much bigger allocation.

## What This Means

Strip it down and NVIDIA's portfolio answers one question with real money: *where are the choke points in the AI buildout?* Compute capacity to rent (CoreWeave, Nebius). Design software (Synopsys). Non-Taiwan manufacturing (Intel). Optical interconnect (Coherent, Nokia). NVIDIA can see the entire order book of the AI economy, and this is where it concluded the squeezes are.

How to actually use the map:

**Watch the announcements, not the filings.** The repricing now happens the morning a partnership press release drops. The question to ask each time: *which bottleneck did NVIDIA just point at?* The next 13F (mid-August) matters mainly for sizing and surprises.

**Watch the wires.** NVIDIA's networking revenue growing twice as fast as compute, plus two fresh multi-billion-dollar optics positions, is the same signal said two ways: connectivity is the current constraint. That's a theme with an entire supply chain behind it, not just two tickers.

**Don't confuse the map for a buy list.** NVIDIA buys for strategic control, not returns — position size reflects how scared NVIDIA is of losing something, not how much upside it sees. Synopsys is flat this year. CoreWeave is still 46% below its 2025 high and runs on leverage. These stocks rip *and* crater — CoreWeave fell 65% from peak to trough in 2025 before recovering, and Synopsys lost 36% in a single session before NVIDIA touched it.

The market spends every day debating what NVIDIA is worth. The better question, most days, is the one almost nobody asks: what does NVIDIA — with more information about AI demand than any entity on earth — think everything *else* is worth? As of the latest filing, you have its answer, priced to the share: the factory, the customers, the blueprints, and above all, the wires.

---

*Data: Barebone | Sources: NVIDIA 13F-HR filings (SEC EDGAR), NVIDIA Q1 FY2027 results, NVIDIA, Nokia, Coherent, and CoreWeave press releases and SEC filings | Prices as of June 9, 2026 close*