OpenAI began 2024 with a positive surprise. It posted an annual revenue of $1.6 billion, much higher than expected. That does come with an important downside: it turns out to be nowhere near enough to be profitable, and neither is the projected revenue for 2024. If the tide won’t turn, OpenAI is staring at a startling bankruptcy before the year is out.
OpenAI’s costs are astronomical. The ChatGPT maker is on track to spend $7 billion by 2024 on training models like GPT-4o. On top of that, its annual personnel costs are $1.5 billion. The bright spot is that annual sales for 2024 are expected to be between $3.5 billion and $4.5 billion. That amount is significantly higher than the booked annual revenue for 2023 ($1.6 billion).
Daily running of ChatGPT, which costs about 350,000 servers with multiple Nvidia A100 chips, is at nearly full capacity, reports The Information. Despite a large discount from Microsoft, cloud costs for OpenAI add up to $700,000 each day. The revenue from paid ChatGPT subscriptions more than offsets that on a yearly basis, but the constant training of and R&D into new AI models is creating a financial strain that OpenAI cannot sustain. Even the $11 billion in total investments into the AI startup is rapidly going up in smoke.
Competition killing
This is not the first time OpenAI has been in dire financial straits. Despite a major deal with Microsoft, several media outlets reported in August 2023 that the ChatGPT maker’s fate was hanging by a thread. A tumultuous period followed in the winter of 2023-24 when CEO Sam Altman was fired and soon reinstated. With that, the financial story surrounding OpenAI receded into the background, until now.
New reporting on OpenAI’s dire financial situation comes at a time when the AI race is in full swing. After Anthropic presented the strong Claude 3.5 Sonnet, Meta followed with the open-source model Llama 3.1 and Mistral launched with Large 2. OpenAI itself is not sitting still and has started the rollout of its long-awaited AI search engine: SearchGPT. With regard to this service, it is also unclear how OpenAI plans to commercialize it. One possibility is that the company will try to capitalize on advertising revenue for the first time by displaying ads among the AI-generated answers.
SearchGPT in particular is challenging the incumbent Google Search, with the idea being that generative AI is taking the well-known internet search functionality to the next level. Google itself launched a half-baked AI Overviews feature within the search engine that recommended glue on pizza, among other things, something that suggests there’s an opening for alternatives if AI search is enticing enough.
Tip: OpenAI launches most extreme AI search engine: SearchGPT
A relatively mature AI search engine already exists, anyhow: Perplexity. That company does however rely on LLMs from other parties, a weakness that OpenAI can exploit.
Why in the cloud?
It is somewhat startling that OpenAI is spending so much money on Azure. After all, it has access to the most advanced hardware within its own corporate walls, as a visit by Nvidia CEO Jensen Huang in April suggested. The bottom line is that only a hyperscaler like Microsoft appears to have the compute and capital needed for hosting state-of-the-art AI training. Running ChatGPT for inferencing isn’t cheap either, but those costs are low enough to compensate with enough paying users. However, the user base is actually stagnating, a little over a year after ChatGPT had launched as the fastest-growing app of all time. In addition, the service regularly suffers from outages and users complain of performance degradation since the launch of GPT-4 in March 2023.
The ongoing AI battle is not only depleting the power grid, water resources and chip production, but also OpenAI’s wallet. Without competitors, this would not have been a problem; in that scenario, AI training would already have been less of a priority by now. With Sam Altman’s company chasing Artificial General Intelligence at all costs and having to come up with new models to keep up, it’s up to outside parties to keep OpenAI in the race, though.
Tip: Ilya Sutskever wants to do what OpenAI can’t: to develop superintelligence securely
It will remain a precarious situation because OpenAI lacks the ‘moat’ it requires. There are no switching costs for organizations and consumers as everyone simply charges $20 a month for the paid service and API access is largely similar, free tokens are plentiful for most users and the feature set found in ChatGPT, DALL-E 3 or SearchGPT is barely distinguishable from the AI alternatives. More problematic is that advanced AI training is virtually impossible to get any cheaper, since new LLMs tend to require more parameters, more training time and more data, not less.
Also read: OpenAI makes mini version of powerful GPT-4o available