CoreWeave had plenty to announce in San Francisco. Nvidia's Vera Rubin NVL72 is live on its cloud, and the company says it was among the first to get there.
New rack engineering, a developer platform called Forge and a partner network filled out the bill. Taken together, the day was less about faster chips and more about what kind of company CoreWeave wants to become.
The headline act
Being first to new Nvidia silicon is CoreWeave's oldest trick. It brought up Vera Rubin early, pointed to Cognition's benchmark numbers and promised quick access for customers.
Speed to the newest hardware is a genuine edge in a market where every lab wants the latest GPUs yesterday.
Being first in the queue is worth less the moment rivals get their own allocations, and it does little for margins when the underlying business is still renting out depreciating hardware.
Why the software matters more
That is what makes Forge the more interesting announcement. CoreWeave is pitching it as a single environment for training, inference, evaluation and agent development, sold in free, Pro and Enterprise tiers.
Software of this kind carries far higher margins than compute, and it is stickier: once a team builds its workflow on Forge, moving becomes a chore. Management has been signalling the direction for a while.
On a recent earnings call it noted that 2026 guidance excludes any potential benefit from licensing CoreWeave's proprietary cloud stack beyond its own data centres. Forge, the partner network and the shared search layer for agents all point the same way, towards selling tools and lock-in rather than only capacity.
Timing is not an accident
CoreWeave is one of the most heavily geared names in AI infrastructure, carrying about $35bn of debt at the end of its second quarter, while remaining lossmaking as higher interest expense widens the gap.
Revenue more than doubled year on year and the contracted backlog has passed $100bn.
But the model depends on raising ever more debt to buy ever more GPUs, and that is precisely the funding tap tightening across the sector as banks and bond investors turn picky on neoclouds.
CoreWeave already tried the obvious fix, agreeing to buy data-centre operator Core Scientific to own its sites and borrow more cheaply.
Core Scientific's shareholders voted the deal down last October, and it was terminated. With that route to vertical integration closed, moving up the software stack is the other way to better economics.
Crowded ground
None of this is a free win. Selling a development platform drops CoreWeave into a fight with the hyperscalers it has so far avoided competing with head on, as well as with the model labs' own tooling and, increasingly, Nvidia's own software ambitions. The marquee pieces are also early.
The Vera CPU has no production customers yet, Forge is days old, and the named adopters are a short list. Concentration is another overhang, with OpenAI and Meta dominating the revenue base. Investors noticed the gap between ambition and proof. The shares rose only about 1% on the day, and the stock still sits well below its 2025 high.
What to watch
The test is not whether CoreWeave can deploy Vera Rubin first. It is whether Forge and the partner network turn into recurring software revenue and higher utilisation, rather than a thin wrapper around the same rented GPUs. If licensing and platform fees start showing up in the numbers, the neocloud label stops fitting and the financing worries ease. If they do not, CoreWeave remains what it has always been: a fast, well-run and very indebted landlord for other people's AI.
