TheChek News Flow

The Takeaway

• OpenAI projects to spend three-quarters of its 2030 revenue on computing and technical talent, more than any company of its scale.

• OpenAI plans to spend $350 billion on server rentals this year through 2030.

• Projected R&D costs are far higher as a percentage of revenue than those of other large technology firms.

The most popular consumer internet services usually figure out how to make money even after a decade or more of losses. Amazon, Uber, Netflix and Snap all started generating cash by raising prices or leveraging their large audiences to sell ads or other services.

OpenAI’s growing audience would seem to put it in a similar position. Its three-year-old ChatGPT business is closing in on 1 billion users and $10 billion in subscription revenue from millions of customers just this year. But as investors buy OpenAI shares at an implied $500 billion valuation, up 17 times from three years ago, they are taking a risk that is as unprecedented as OpenAI’s costs.

The Information recently reported that OpenAI plans to burn through $115 billion in cash through 2029. According to OpenAI’sprojections, its computing and talent costs will remain high as a percentage of its revenue for at least the next five years.

While OpenAI projects it will start generating free cash flow from its business by 2030, its high expenses raise questions about how profitable OpenAI can be long term relative to its publicly traded tech peers.

Consider OpenAI’s research and development costs. Based on how tech giants such as Microsoft consider R&D, these include the expense of paying for servers to produce new artificial intelligence, known as research compute or training costs, as well as paying its cutting-edge researchers, including stock-based compensation. According to The Information’s analysis, in 2030, when OpenAI projects an astounding $200 billion in revenue, its R&D costs would amount to about 45% of its revenue, or roughly $90 billion. (Compensation costs that year will be slightly lower than computing and software costs, including paying for data and services that improve the quality of its AI models.)

That’s a far higher percentage than at Amazon, Microsoft, Oracle and Alphabet, whose R&D costs are currently between 10% and 20% of revenue. Meta’s is around 25%.

OpenAI’s persistently high R&D costs likely reflect the cutthroat competition for talent and the need to stay ahead of rivals such as Google, Meta and Anthropic, some of whom have highly profitable core businesses that generate cash to pay for AI costs.

Then there’s the computing cost to power ChatGPT and other applications for consumers and businesses (including buyers of its AI models, such as Cursor and Salesforce). Those costs, known as inference computing, also remain high in the latest projection—about 25% of revenue in 2030, or nearly $50 billion that year.

Only Snap spends such a high percentage of revenue on cloud computing costs to run its services, according to The Information’s Cloud Database. And as Snap CEO Evan Spiegel acknowledged in a staff note on Monday, that is why Snap’s gross margin is “meaningfully below many of our competitors”—a gap he now aims to close.

It isn’t clear whether OpenAI’s latest projections factored in potential savings from launching GPT-5 in August. That system aims to route simple ChatGPT queries to a model that generates an answer using less computing power than reasoning models that are better suited to answer harder questions. (Update: an OpenAI employee has pointed out that the GPT-5 launch in ChatGPT also allowed non-paying users to experience reasoning models for the first time, and those models are more expensive to run. So the overall costs of serving ChatGPT users may have risen following the launch.)

In OpenAI's prior projections, research compute costs peaked in 2028 and actually fell slightly over the next two years; that wasn't the case in its latest projections.

Slim Margin of Error

All told, OpenAI projects to spend half its revenue—$100 billion—on renting servers to both train and run its models in 2030 alone. OpenAI projects to spend another $250 billion renting cloud servers this year through 2029. That’s $350 billion over six years.

It’s fantastic news for OpenAI’s cloud vendors: Oracle (which reportedly will get $300 billion in revenue just from OpenAI between 2027 and 2032), Microsoft, CoreWeave and Google Cloud.

No other major company spends three-quarters of its revenue on computing and technical talent the way OpenAI plans to do. And that’s not including another $100 billion in losses OpenAI projects between 2026 and 2030 from other costs related to data center servers that the company implied it would control, as opposed to servers it currently rents.

OpenAI presumably will save money over time if it can control enough servers efficiently and avoid paying more fees to Oracle and Microsoft, which surely will seek to generate a decent profit margin on their server rentals to OpenAI. It makes sense OpenAI is going this route, but it’s a long-term project.

OpenAI may do well to build for the long term, as long as investors allow it. “We’re not a public company. We’re owning and building the frontier era of AI,” said an OpenAI spokesperson. “Continued investment will deliver outsized results. Our motivation isn’t to create incremental gains.”

You might wonder if OpenAI was being conservative in its most recent financial projections and plans to beat them, including by lowering its costs. That doesn’t seem to be the case: The long-term costs it projected as of this summer were far higher than the cost projections it made in the first quarter. In those prior projections, research compute costs peaked in 2028 and actually fell slightly over the next two years; that wasn’t the case in its latest projections. (It isn’t clear how often the company updates its internal projections.)

These massive costs leave OpenAI with little room for error as it races to match the current revenue of firms such as Nvidia, Meta, Chevron and General Motors by 2030.