Amazon and Nvidia open their wallets to lock in OpenAI's business while SoftBank keeps the lights on

ChatGPT maker announces $110B in new investment amid flurry of self-serving deals

The headlines say OpenAI on Friday announced $110 billion in new investment from Amazon, Nvidia, and SoftBank at a $730 billion pre-money valuation, though terms and conditions apply.

Both Amazon's $50 billion and Nvidia's $30 billion investments are tied to massive customer commitments by OpenAI and its partners. 

Of Amazon's investment, $35 billion will only be paid out "when certain conditions are met." From what we gather, those conditions include renting out two gigawatts worth of Amazon's Trainium AI accelerators and the deployment of OpenAI models and services in AWS. The cloud provider will also be the "exclusive third-party cloud distribution provider for OpenAI Frontier."

Announced earlier this month, Frontier is OpenAI's new agent builder aimed at enterprise customers. Curiously, OpenAI was quick to emphasize that "nothing about today's announcements in any way changes the terms of the Microsoft and OpenAI relationship," and that Azure remains the exclusive cloud provider of its stateless OpenAI APIs and first party products, like Frontier.

Nvidia's $30 billion investment appears to have similar provisions. In a blog post published on Friday, the AI flag bearer announced an expanded partnership with Nvidia that would see the deployment of three gigawatts of inference and two gigawatts of training capacity built on the GPU slinger's Vera Rubin systems. Announced at CES in January, the racks are expected to begin shipping in the second half of 2026.

The economics of datacenters can be tricky to pin down as it varies by region, but at a power usage effectiveness of 1.1, a gigawatt is enough for about 3,600 Vera Rubin NVL72 systems, assuming 250 kW per rack. 

At an estimated cost of $8.4 million each, that's roughly $30 billion per gigawatt. However, compute only accounts for about half the cost of standing up a modern AI datacenter. 

The land, shell, power, and plumbing costs make up the remainder. This puts the cost of building and deploying 5 gigawatts worth of Vera Rubin accelerators at more than $300 billion. OpenAI is unlikely to take this on itself. Instead, we expect CEO Sam Altman to broker purchase agreements through its hyperscaler and neocloud partners, as it has done with Oracle and Crusoe for its Stargate facility in Abilene, Texas.

Amazon is one of those partners. While OpenAI may be playing up Trainium in today's disclosures, its existing $38 billion tie-up with Amazon, announced back in November, was specific to Nvidia GB200 and GB300 systems. That contract has now been extended to $100 billion over the next eight years.

Both Amazon's and Nvidia's investments are structured in such a way as to guarantee a return on every dollar invested in OpenAI. The funding is essentially a discount on compute infrastructure that doesn't dilute their revenues while driving up OpenAI's valuation.

This kind of circular dealing has become commonplace amid the AI boom, and all the big boys are taking part. Back in October, Nvidia rival AMD issued OpenAI a warrant for roughly 10 percent of its stock. To claim it, all Altman needs to do is deploy six gigawatts of the chip designer's Instinct accelerators. This week, AMD copied and pasted that deal, extending the same offer to social media and would-be AI magnate Meta.

While Nvidia and Amazon are clearly engaged in financial engineering, the same doesn't appear to be the case for SoftBank. Infamous spendthrift Masayoshi Son plans to invest another $30 billion into the company to keep the lights on while Altman and crew continue their search for artificial general intelligence. SoftBank's contribution will be paid out in three tranches of $10 billion beginning in April and wrapping up in October.

Even with annualized recurring revenue reportedly exceeding $20 billion and more than 50 million paying subscribers on the books, OpenAI is going to be reliant on outside funding for a while. The company isn't expected to achieve profitability until at least 2029. ®