16-05-2025
Will OpenAI ever make real money?
BEING SAM ALTMAN is a glamorous gig. Since the launch of ChatGPT in November 2022 the boss of its creator, OpenAI, has turned into a global business superstar. He is the darling of both the starch-collared Davos set and Silicon Valley's dishevelled techno-Utopians. He hangs out with everyone from Katy Perry to Donald Trump, whom he accompanied on a visit to Saudi Arabia this week. It would shock no one if by its next funding round his startup, currently worth $300bn, overtook SpaceX and ByteDance to become the world's most valuable unlisted firm. The AI wunderkind recently told the Financial Times that he has the 'coolest, most important job maybe in history'. No kidding.
Being Sarah Friar is not nearly as fun. As OpenAI's chief financial officer, the Irishwoman has two main tasks. The first is to make sure that the numbers add up. The second is to persuade investors to part with the billions of dollars the firm needs in order to train and run ever cleverer artificial-intelligence (AI) models.
Happily for Ms Friar, moneymen swept up in the AI mania need little persuading. They are falling over themselves to fund OpenAI. On May 13th SoftBank, a Japanese tech piggy-bank, said that its $30bn investment in the firm was unaffected by Openai's recent decision not to ditch its odd governance structure. A non-profit board will keep control of its for-profit arm.
That is just as well, for going with the flow of investor enthusiasm leaves the CFO more time to tackle her other responsibility. And when it comes to charting a path to profits, the former Oxford University rower is paddling upstream.
For OpenAI, as for any startup, making money involves a series of steps: attract and retain brainboxes, have them create something clever, turn that something into a marketable product, sell more and more of that product while minimising costs until cashflow turns positive. Despite defections, including of several co-founders, OpenAI remains a talent magnet. The cleverness of its tech is indisputable. Mr Altman's claim that the latest o3 model, with an enhanced ability to reason, displays 'genius-level intelligence' should be taken with a pinch of salt—but only a pinch.
It is at the next stage that Ms Friar wades into problems. To see why, consider OpenAI's two more richly valued startup cousins. ByteDance's recommendation algorithm, which makes TikTok and its Chinese sister app the time sinks that they are, may be a bit more addictive than when it debuted in 2016. SpaceX's rockets are bigger, more reliable and cheaper than at its first successful launch in 2008. But neither underlying technology has dramatically changed; any additions are, like SpaceX's Starlink satellite internet, complementary. This stability has enabled both firms to build products and, in time, business models around them. Especially for ByteDance, these are lucrative. Last year the social-media titan turned a net profit of $33bn on sales of $155bn.
The reason OpenAI will struggle to follow suit is precisely what excites its backers—the sheer pace of AI innovation. It would be one thing for advances to be frequent. The challenge for Ms Friar is that they also frequently upend her firm's economics.
Some of the disruption comes from OpenAI's rivals. In January a Chinese startup called DeepSeek came out of nowhere with a model that was almost as clever as OpenAI's flagship but required many fewer power-hungry chips to train and use it. DeepSeek also made its code freely available to all and sundry, lowering barriers to entry into advanced model-making. This has eroded OpenAI's competitive advantage at the cutting edge, which it had maintained thanks to access to oodles of computing power courtesy of Microsoft, its big-tech partner. It also constrains its ability to keep raising prices for using its models, which can run to as much as $200 a month per licence.
Competition is, of course, tech's Schumpeterian lifeblood. Nothing stops OpenAI from making its models more efficient, including by adopting some of DeepSeek's ideas. The trouble is that the economics are changing in more fundamental ways too.
Compare o3 with GPT-4, the model that powered ChatGPT in 2023, and take energy use as a proxy for cost. OpenAI is cagey about its numbers. But according to estimates, for every $1 in training costs, GPT-4 would cost around $4 a year to run, based on OpenAI's current level of traffic. For o3, whose reasoning relies on more computing in the post-training 'inference' phase, the ratio could be as high as one to 100.
Confounding variables
These ballooning operating costs explain OpenAI's mounting losses. Despite tripling its sales to $3.7bn in 2024, it lost perhaps $5bn (excluding stock-based compensation). This year it expects revenue to triple again, to $13bn, and inference costs to grow at the same rate, to $6bn. A shifting cost structure also makes it hard to price products and plan budgets. A fixed subscription fee that made sense in the age of GPT-4 looks unviable for o3. You could try keeping subscriptions for older, dumber versions and add a variable usage fee for inference-heavy reasoning. But how many people will pay anything for an obsolete technology? And how long until the next model forces another complete rethink?
Any projections for revenue and costs beyond the next few months rest on heroic assumptions. OpenAI's forecast of $125bn in sales and $12bn in cashflow in 2029 might as well be pulled out of a hat. Not because it is too rosy; because it feigns certitude. The same goes for its $300bn price tag: an ungodly sum by startup standards but a trifle next to the $1.4trn in shareholder value Microsoft has created since teaming up with OpenAi in late 2022. This gap may make it easier for Ms Friar to marshal more capital. Yet it also highlights the uncertainty around what her company is truly worth—and the scale of her bookkeeping challenge.
Subscribers to The Economist can sign up to our Opinion newsletter, which brings together the best of our leaders, columns, guest essays and reader correspondence.
Get 360° coverage—from daily headlines to 100 year archives.
