Yuval Avidani
Author
"xAI lost $6.4 billion on revenue of $3.2 billion." That number, pulled straight out of SpaceX's IPO filing, sums up the biggest riddle in the AI industry right now: everyone's growing like crazy, everyone's raising insane amounts of money, and almost everyone's losing cash. Let's understand why, and why that's somehow not scaring anyone off.
Let's break it down slowly, because this is where the economic logic of the entire industry hides.
The case that sums it all up: SpaceX, xAI and Grok
Let's start with the freshest story. In February 2026, SpaceX acquired xAI, the company behind the Grok chatbot, in a deal that valued the combined entity at around $1.25 trillion. Then, when SpaceX filed for its IPO, xAI's numbers got exposed, and they tell the whole story.
Look at that gap. xAI brought in about $3.2 billion in 2025, but lost $6.4 billion, meaning it burned roughly a billion dollars a month. Revenue was half the loss. How is that even possible? The answer in one word: infrastructure. xAI's data center, called Colossus, consumes about a gigawatt of electricity, and in Q1 2026 alone the company spent $7.7 billion on equipment. In the AI world, to bring in a dollar you first have to burn a bunch of dollars on chips, electricity and data centers.
It's not just xAI: everyone's in the same boat
Now let's zoom out, because this pattern repeats across the board. Let's look at revenue versus losses for the three biggest players:
Here's the picture. OpenAI is hitting an annualized revenue run-rate of about $20 billion, but is projected to lose around $14 billion in 2026 alone, and burn roughly $115 billion cumulatively through 2029. Anthropic is growing to a run-rate of $30 billion plus, with compute expenses of around $19 billion, and isn't projected to be profitable until around 2028. Even when revenue is huge and growing fast, infrastructure costs are running even faster.
The riddle: why a $20 subscription loses money
And here's the part that surprised me the most, and the part that actually touches us as users. A lot of these companies are losing money even on paid subscriptions, not just on free users. Sam Altman, OpenAI's CEO, said himself that the company is losing money on the $200/month Pro subscription, because people are using it way more than they expected.
Let's understand the mechanism, because this is the heart of it. Every request we send to a model consumes seconds of an expensive GPU plus electricity, and that's the direct cost (what's called COGS, cost of goods sold). With a flat subscription, we pay one fixed amount, but we consume a variable amount. A heavy user sending thousands of long requests a day costs the company way more than $20, or even $200.
Think of it like an all-you-can-eat restaurant with a fixed price. Most diners eat a reasonable amount, but it only takes a few heavy eaters for the restaurant to lose money on the deal. The difference is that at a restaurant, the marginal cost of one more plate is low, whereas in the AI world every "plate" of inference costs real electricity and real compute time. So the more we use these tools, the more we sometimes actually cost the company.
So why are investors still pouring in billions
The obvious question: if everyone's losing money, why do Google, Microsoft and venture funds keep pumping in more and more? Anthropic has raised around $125 billion, OpenAI is valued at about $852 billion, and investors aren't stupid. There are a few bets happening here.
The first: gross profitability improves with scale. As models and hardware get more efficient, the cost per request drops, and at some point revenue per user crosses over the cost of serving them. The second: it's a land grab. Whoever controls the infrastructure and the users once the market matures will profit for years to come. The third is a bet on usage growth itself: as AI gets cheaper and better, we use it more, and that grows the entire market.
On the flip side, the skeptics say something worth remembering: maybe every efficiency gain just gets swallowed up by growing demand, and AI stays "software with thin margins," because unlike regular software, here every single use costs real money in electricity and compute. Who's right, we'll only know in a few years.
Bottom line, and my take
So let's sum up. AI companies are losing money because infrastructure costs, chips, electricity, data centers, are running faster than revenue, and sometimes even paid subscriptions lose money because heavy users consume more than they pay for. Investors keep going because they're betting on costs dropping at scale, on capturing the market, and on usage growth.
To me, this is what makes this moment so strange and so fascinating: we're enjoying services worth a lot more than what we're paying for them, because someone else, the investors, is subsidizing us in the meantime. The big question is how long that lasts. Important note: all the numbers here come from reports and filings and keep changing all the time, and this is not investment advice or financial guidance, just an attempt to understand the mechanism.
So here's the question I'll leave you with: on the day investors stop subsidizing this, how much will we actually be willing to pay for the AI we've gotten used to?