At the end of a short video clip posted in early May, Mark Cuban says something that should probably be printed on a banner inside every AI boardroom in San Francisco. "You just spent a trillion dollars," he says, "to be an app."
The audience laughs. Cuban laughs. The host laughs. It is a very good line. It is also, depending on how the next five years unfold, either a devastating summary of the greatest capital misallocation in history or a joke that ages very badly.
Here is the situation Cuban is responding to: OpenAI has committed to spending somewhere between $600 billion and $1.4 trillion on computing infrastructure by the end of the decade. The company recently raised $110 billion in a funding round, which by most measures was roughly three times larger than any technology company has ever raised before going public. It has signed deals with Oracle, Microsoft, and chip manufacturers that lock in spending for years. And it is currently losing more money, faster, than almost any company in history - not as a side effect of the spending, but as a direct consequence of it.
The question Cuban is actually asking is not whether artificial intelligence is real. He is asking something more specific and more uncomfortable: does the business case for the companies currently spending the most money actually work? And his answer, delivered with the calm of someone who has been on both sides of a market bubble, is a flat no.
The background
To understand why anyone would spend this much money, you need to understand what these companies are actually building and why it costs so much.
Foundational AI models - the large systems that power products like ChatGPT, Google's Gemini, Anthropic's Claude, and others - require enormous quantities of specialised computing hardware to train and run. Training a model means feeding it vast amounts of text, code, and other data so that it learns statistical patterns in language. Running it - generating a response every time a user asks a question - is called inference, and it has its own costs. Every single query to ChatGPT costs money in electricity and hardware.
The hardware at the centre of this is primarily made by Nvidia, the American chip company whose market value has swung violently based on demand from AI companies. The facilities that house the hardware are data centres - essentially enormous, power-hungry warehouses filled with specialised computers. They require water for cooling, land, construction, and ongoing electricity. They cannot be easily repurposed if the AI market slows down.
OpenAI's compute capacity - a measure of how much processing power it operates - grew from 0.2 gigawatts in 2023 to 1.9 gigawatts in 2025, according to the company's own financial disclosures. A gigawatt of electricity is roughly what it takes to power one million homes. OpenAI now runs nearly two million homes' worth of electricity through its systems continuously, just to answer questions and generate code.
The logic behind all this spending is straightforward: more compute means better models, better models means more users, more users means more revenue, more revenue means you can afford even more compute. It is a flywheel. The problem is that the flywheel currently runs at a loss, and the losses are getting larger faster than the revenue is.
What is actually happening
OpenAI's revenue has grown at a historically unprecedented rate. The company went from $2 billion in annual revenue in 2023 to $6 billion in 2024 to more than $20 billion by the end of 2025, according to CNBC. That is roughly a tenfold increase in two years - faster than any software company has ever grown at that scale.
The problem is that the costs are growing faster still.
According to internal projections reported by The Decoder, OpenAI now expects to burn through $25 billion in cash in 2026 and $57 billion in 2027 - against projected revenue of $30 billion and $62 billion respectively. The company does not expect to reach positive cash flow (the point at which money coming in exceeds money going out) until 2030. Total projected spending through the end of the decade has now been revised to approximately $665 billion, up from earlier estimates, driven by training and inference costs - the cost of running the models every time a user sends a message - that have repeatedly come in higher than expected.
To put that cash burn in context: the entire Apollo space program, which put humans on the moon across 13 years, cost roughly $288 billion in today's money.
Cuban's point, as he makes it in the clip, is not that the revenue is fake or that the technology does not work. It is that the math of return on investment - the amount of profit a business eventually generates relative to the money put in - may simply never work for most of the companies currently spending this way. "They'll never get it," he says, when asked about the returns needed to justify the infrastructure commitments. "They're throwing away that money at scale."
His second point is more structural. The AI industry does not yet know what kind of market it will become. Cuban lays out two historical comparisons: streaming - where one or two companies dominate and a handful of others survive - and search, where effectively one company (Google) won almost everything. If AI foundational models end up looking like search, then perhaps one company, or at most two, will capture most of the value. Everyone else will have spent hundreds of billions of dollars to end up as, in his phrase, "just an app."
Research from Andreessen Horowitz found that fewer than 10 percent of ChatGPT weekly users visited any other major AI provider during most of 2025, and only 9 percent of paying users subscribed to more than one AI service. The consumer market, at least, already looks more like "winner takes most" than a diversified ecosystem.
The money trail
Follow the money carefully, because it moves in a loop that is hard to exit.
OpenAI raises capital from investors - SoftBank, Microsoft, sovereign wealth funds - on the promise of future dominance. It uses that capital to sign long-term deals with hardware and cloud providers, locking in spending commitments years in advance. Those commitments require it to keep raising capital, because stopping mid-build would leave it with partial infrastructure and no competitive advantage. So it raises again, at a higher valuation each time, which requires projecting even larger future revenues to justify the price.
As Yahoo Finance reported, OpenAI projects losses of up to three-quarters of revenue by 2028 - its worst year before the projected turnaround - driven by compute costs that scale with usage. The company currently spends approximately $1.69 for every dollar of revenue it generates, a ratio that only works if you believe the revenue eventually catches up and that the company survives long enough for it to happen.
There is a second problem embedded in the cost structure that Cuban gestures at but does not fully unpack: inference costs are not fixed. The more people use ChatGPT, the more it costs to run. Unlike traditional software, where a product can scale to millions of users with minimal extra cost, AI models require fresh compute for every query. The more popular the product becomes, the faster the cash burns. Internal data leaked to The Register and reported in early 2026showed that OpenAI's inference costs were at certain points exceeding its subscription revenue - meaning the company was losing money specifically because people were using its product.
Who benefits from this arrangement? The hardware and cloud companies do, in ways that are largely locked in. Nvidia's chips are the primary input. Microsoft, which has a revenue-sharing arrangement with OpenAI, benefits from the cloud computing fees. Oracle signed a deal reportedly valued at $300 billion to provide data centre capacity. These companies get paid regardless of whether OpenAI ever becomes profitable.
The investors betting on a single winner - or a small group of winners - are making a different kind of bet. They need the company to survive long enough to reach the point where revenue exceeds costs, competition thins out, and the remaining players can start charging prices that actually cover their infrastructure. That is the streaming model: lose money for years to build an audience, then raise prices once the audience has no real alternative.
The risk is that the AI market never reaches that consolidation point. Competition from open-source models, from Chinese AI companies building at a fraction of the cost, and from cloud providers who could choose to prioritise their own models at any time, keeps the pricing pressure high. A Chinese lab released a competitive AI model in January 2025 for a reported training cost of $5.9 million, compared to the $100 million or more that OpenAI reportedly spent on comparable training runs. That single data point, more than anything else, illustrates the fragility underneath the trillion-dollar commitments.
What people are doing about it
The responses to this dynamic fall into roughly three categories: doubling down, hedging, and quietly waiting.
OpenAI is in the first category, aggressively. It recently revised its infrastructure commitments downward from $1.4 trillion to roughly $600 billion by 2030, according to CNBC - a reduction that was framed as discipline but also reflects the pressure of investors questioning whether the original number was ever realistic. The company launched an advertising pilot in early 2026 to generate revenue from its free users, a clear signal that subscription revenue alone is not sufficient.
Apple, as Cuban points out, has taken a deliberately different approach. Rather than building its own foundational models from scratch and racing to spend the most on compute, it has focused on integrating existing models into its devices and operating system - spending comparatively little while retaining control of the customer relationship. The economic logic is clear: if foundational AI becomes a commodity, the company that owns the interface to the customer captures the margin.
Google occupies a structurally unique position. Its Gemini model is embedded directly into Google Search, meaning every AI-powered search result is an opportunity to sell advertising. The company has an existing revenue engine - search advertising - that can fund AI development without requiring the same kind of pure capital commitment that OpenAI needs.
Enterprises are hedging by distributing their AI spending across multiple providers. Only 9 percent of paying consumers subscribe to more than one AI service, but in business markets the picture is different - companies are increasingly building on top of multiple models to avoid dependence on a single provider and to play competing vendors against each other on price.
Hospitals, universities, and research institutions - the holders of the proprietary data and intellectual property that could give any model a genuine edge in specialist domains - have started to reassess publication and patent strategies, according to Cuban. Traditionally, academic institutions published research to build reputation and influence. Publishing now means the research becomes training data for every AI model simultaneously, eliminating any competitive advantage. Some institutions are being advised to sell access to their data rather than publish it freely.
The bottom line
OpenAI is spending money at a rate that requires it to become not just profitable, but one of the most profitable companies ever built, within roughly four years. The revenue growth is real and historically unprecedented. The losses are also real and historically unprecedented. What is not yet clear is which of those two facts will define the outcome - and the answer depends almost entirely on whether the AI market consolidates quickly enough around a small number of winners for those winners to charge prices that cover what it actually costs to run the technology. If it does not, the most expensive infrastructure build in peacetime history produces a product that competes primarily on price, in a market where the cheapest competitor is building the same thing for a fraction of the cost.
Timeline
- 2015 - OpenAI founded as a nonprofit AI research laboratory.
- 2022 - ChatGPT launches, reaching 100 million weekly active users faster than any product in history.
- 2023 - OpenAI reports $2 billion in annualised revenue. Compute capacity: 0.2 gigawatts.
- January 2025 - Chinese AI lab DeepSeek releases a model reportedly trained for $5.9 million, compared to OpenAI's $100 million-plus training runs, triggering a $600 billion single-day loss in Nvidia's market value.
- Mid-2025 - OpenAI reaches $1 billion in monthly revenue for the first time.
- November 2025 - CNBC reports OpenAI is on track to exceed $20 billion in annualised revenue; company has committed to $1.4 trillion in infrastructure over eight years.
- January 2026 - OpenAI CFO Sarah Friar confirms $20 billion in annualised 2025 revenue; company begins testing advertising in ChatGPT's free tier.
- February 2026 - CNBC reports OpenAI revises compute spending target down from $1.4 trillion to approximately $600 billion by 2030; cumulative cash burn through 2030 revised upward to $665 billion.
- May 4, 2026 - Mark Cuban tells the Big Technology Podcast that OpenAI and rivals spending at this scale "will never get the return" and are "throwing away that money at scale."
Summary
Who: OpenAI, its investors, and the broader field of companies building large-scale AI foundational models - including Google (Gemini), Anthropic (Claude), Meta, and xAI (Grok). Mark Cuban, entrepreneur and investor, is the critic.
What: OpenAI has committed to spending between $600 billion and $1.4 trillion on computing infrastructure by 2030 while projecting losses through at least 2028. The company currently spends more money than it earns on every dollar of revenue, and its cash burn is increasing faster than its revenue.
When: The infrastructure commitments were announced through 2025 and into early 2026. Cuban's comments aired on May 4, 2026.
Where: The spending is concentrated in data centre construction across the United States, with deals involving Microsoft, Oracle, and chip suppliers. Cuban's assessment was made on the Big Technology Podcast.
Why: The underlying bet is that AI foundational models will consolidate into a winner-takes-most market, similar to search or streaming, in which the company with the most infrastructure and the largest user base captures enough pricing power to eventually cover its costs. The counterargument - the one Cuban is making - is that the market may remain competitive, that infrastructure costs may fall faster than expected as the technology matures, and that the return on investment may never arrive.