Nvidia's 13F: $30B in Intel, $21B in SpaceX

Nvidia filed its quarterly 13F on August 14, and the two biggest lines are worth an operator's attention: roughly $30 billion in Intel and $21 billion in SpaceX, both as of June 30. That is a chip company holding a nine-figure-times-three equity position in a manufacturing partner and in a customer that has committed to buying its hardware. The AI supply chain is not a market with arm's-length buyers and sellers anymore. It is a cap table.

What actually happened

Per Fortune, Nvidia disclosed 214.8 million Intel shares worth about $30 billion and 122.8 million SpaceX shares worth just under $21 billion. The Intel position came from a $5 billion investment made in 2025 alongside an agreement to co-develop chips for data centers and PCs; it was valued near $9.5 billion only three months earlier. The SpaceX line is the first public mark on Nvidia's xAI exposure — Nvidia had committed as much as $2 billion to xAI in 2025, and xAI was folded into SpaceX before its listing earlier this year. Smaller positions include Coherent, Nebius Group, Nokia, Synopsys and Generate Biomedicines.

Two facts sit on either side of that. SpaceX's xAI has committed to Nvidia hardware for its AI data centers. Intel is a foundry and CPU rival that Nvidia is now both invested in and building products with. Nvidia's two largest disclosed holdings are, in effect, a customer and a supplier.

Why Nvidia's vendor concentration matters for your business

You are three or four layers downstream of this, and it still reaches you.

Demand signals get harder to read. When a chip vendor funds the entity buying its chips, "record demand" stops being a clean market signal. We wrote about the same shape in Nvidia's $250B OpenAI backstop and the $500B Wall Street financing. Don't build a three-year budget on the assumption that token prices fall because the supply chain looks healthy. Price what you use today and re-check quarterly.

Concentration is now a single-name risk. Over $50 billion of one company's balance sheet sitting in two names means an Intel or SpaceX repricing moves the vendor everyone's inference runs through. That is not a prediction of trouble. It is a reason to keep the switch cheap.

The practical control is portability, not prediction. Keep your model layer behind an interface you own, keep contracts short, and know what it actually costs to move a workload to a different provider. If you can't answer "what happens to this feature if our provider doubles the price in ninety days" with a number and a week, you have a dependency, not a vendor.

Sources: Fortune, Seeking Alpha.