Amazon Reveals AWS AI Revenue Over $15B

On April 9, Meta signed a fresh $21 billion agreement with CoreWeave for additional cloud computing capacity, extending through December 2032 and layering on top of the companies’ earlier $14.2 billion deal signed in September. Reuters frames it as part of Meta’s rush to catch up after an underwhelming model release last year. That’s accurate, but it’s incomplete. The bigger story is that the AI race has become a procurement game, and Meta just wrote a very large check to stay in the match.

The headline feature is hardware access. Reuters reports the arrangement gives Meta early deployments of Nvidia’s next-generation Vera Rubin chips – chips Reuters describes as twice as fast as Blackwell, the current platform. In older cloud eras, you paid for capacity. In this one, you also pay for priority. The queue is now an asset class.

Meta’s willingness to do that is consistent with the scale of its stated ambition. Reuters says the company plans to spend up to $135 billion on its AI buildout this year as Silicon Valley pursues artificial general intelligence. Whether you buy the AGI framing or not, the cost posture is clear: Meta is treating compute as the irreducible input.

CoreWeave’s side of the ledger shows why this market is consolidating around a small set of “neocloud” specialists. Reuters notes Meta is now among CoreWeave’s largest customers and that Microsoft accounted for about 67% of CoreWeave’s revenue last year. That concentration is risky – until it becomes a moat. If your business is built to serve a handful of customers who can sign 10-figure commitments, you don’t need a broad customer base. You need credibility with the buyers who are rationing the world’s high-end accelerators.

And then there’s the financing. In the same Reuters report, CoreWeave disclosed plans to sell $1.25 billion of bonds and $3 billion of convertible bonds. The company’s own disclosures add the missing mechanics: CoreWeave’s April 9 SEC 8‑K describes a $1.25 billion senior notes offering due 2031 and a $3.0 billion convertible senior notes offering due 2032, with an option for an additional $450 million, and it details the use of capped call transactions associated with the convertibles. This isn’t background noise. It’s the business model: sign long-duration capacity contracts, then finance the hardware buildout with debt and equity-linked instruments built around those contracts.

The structure also clarifies what customers like Meta are paying for. They’re not just renting GPUs from CoreWeave’s existing fleet. They’re funding the vendor’s ability to order, deploy, power, and operate the next wave of systems – at pace.

There’s also a subtle competitive read-through. Bloomberg reports CoreWeave said it now holds $35 billion in contracts with Meta. That number, if you take it at face value, suggests Meta is effectively creating a parallel compute supply chain beyond its own data centers – and it’s willing to commit multi-year demand to do it.

The open question is whether this becomes a durable advantage for Meta or simply an expensive bridge. Deals like this can smooth the path to shipping AI features at scale – especially across Meta’s massive surfaces – without waiting for internal capacity to come online. But they also deepen dependency on a vendor whose fate is tied to hardware availability, energy constraints, and its own ability to refinance and roll capital forward.

In 2026, “model strategy” is increasingly downstream of this: if you can’t guarantee inference capacity, you can’t guarantee product momentum. And if you can’t guarantee product momentum, the frontier model is just a lab artifact.

Sources: Reuters · Bloomberg · SEC