Vector Core Compute commits $3.5bn to SambaNova chips for edge inference

Vector Core Compute has committed to spend about $3.5 billion over three years on SambaNova hardware, a bet that agentic AI inference will be served close to end customers. The company is backed by Vista Equity Partners and Cambium Capital and uses SambaNova's Reconfigurable Data Unit chips in a platform that also includes Intel hardware. It gave its first interview to Data Center Dynamics, which published the feature in its Agentic AI supplement on October 22.

Both sponsors were already SambaNova investors. Cambium co-founder and managing partner Landon Downs said the firm, which invests mostly in hardware and has backed Groq, Rivos and SambaNova, decided to move down the chain: “We think these investments are great for data centers, and given the way the world is moving, we decided we should actually start building data centers ourselves.” Vista, historically primarily a software investor, watched portfolio companies spend growing sums on agentic AI or develop it in-house; Downs framed the answer as verticalizing enterprise use cases “all the way down through the hardware stack.”

Downs also argued against the mega-project reflex. A property in Canada, he said, taught him that large-scale builds do not necessarily deliver economies of scale, and that smaller projects can in some cases produce significantly better economics.

The $3.5 billion is a supply commitment as much as a capital plan, fixing a large slice of VC2’s chip cost for three years and giving SambaNova a multi-year demand anchor. Both sponsors hold SambaNova stakes, and some of the customers VC2 intends to serve are companies the sponsors have also invested in, so chip supply, sponsor equity and a prospect list sit inside one relationship. A weak reception for the RDU line or for edge inference would hit the same parties on both sides.

Nebius, part of the first GPU-backed neocloud wave with CoreWeave and Iren, has financed capacity against 12-year tenancy and raised $5.75 billion of debt. VC2 sells a purpose-built stack to enterprises rather than long-dated leases to an anchor tenant, so its credit story rests on enterprise contracts and on whether keeping agentic inference close to the customer commands a premium over centralised capacity. Downs’ claim about smaller projects producing better unit economics is the number to test first.