A promise of generative artificial intelligence (AI) is its potential to cut costs.
Unfortunately, the realities of the unit economics inherent to building and running foundational AI models might result in some of even the most prominent platforms going broke themselves.
This, as OpenAI — the organization responsible for turning generative AI into a household name with the launch of its ChatGPT product last fall — may reportedly go bankrupt by 2024 if it continues to burn cash at its current rate.
The platform famously received a $10 billion investment from Microsoft, and has received other venture funding, but those billions only go so far when dealing with the sophisticated, high-cost computing requirements and burgeoning scale of generative AI’s change-the-game capabilities.
OpenAI’s total losses for 2022 were reportedly more than half a billion dollars ($540 million), and the firm spends up to $700,000 a day maintaining its underlying infrastructure and server costs, even as its user popularity drops.
The high cost of building and running an AI model is not a problem unique to OpenAI.
Google researchers have estimated that their own AI ambitions take up around 10% to 15% of the company’s energy usage each year, representing roughly the same amount of annual electricity use as the city of Atlanta — not a cheap bill to foot by any means.
Those numbers and their associated costs are likely similar for other firms now turning their focus to AI, like Meta and Microsoft.
Analysts estimate that Microsoft’s Bing AI chatbot, which is powered by an OpenAI large language model (LLM), needs at least $4 billion of infrastructure just to do its job.
Even well-funded AI startup Anthropic recently raised another $100 million this week (Aug. 14), presumably to help defray the skyrocketing cost of being in the AI business.
Read more: Costs Are Top of Mind as LLMs Upgrade Chatbots for Industry-Specific Use Cases
The high cost of AI is an uncomfortable reality, one that is driven primarily by the computing power an AI model requires. Some estimates place the cost of a single ChatGPT query at 1,000 times that of the same question asked of a normal Google search, making the margins for AI applications significantly smaller than other software-as-a-service (SaaS) solutions.
While the world is still somewhat in the initial development stages of the AI economy, where companies such as OpenAI, Meta, Elon Musk’s xAI, and more are seeking first and foremost to generate public interest and capture customers for their platforms, the sheer size of the costs required to compete in the sector make it such that the world’s biggest companies have an already baked-in advantage — their balance sheets.
But even those balance sheets are taking a hit as firms burn through billions training and bringing AI models to market.
Significant capital investment, industry-leading technical expertise, and above all, intensively expensive computing infrastructure built atop rows of increasingly-scarce GPUs are all needed to establish and maintain generative AI models, much less jockey for position with some of the largest and most valuable businesses in human history.
Training generative AI requires either owning or renting time on hardware, significant data storage needs, and intensive energy consumption, a structural cost that sharply diverges from the unit economics of previous computing and technological booms.
Hosting a webpage or running an app costs peanuts compared the daily cost of running an LLM or other AI service.
Compounding the matter is that AI models must be retrained regularly in order to remain aware of current events — for example, any model trained on data up through January 2022 (like ChatGPT-3) wouldn’t be aware of many events now in the past, including the start of the Russia-Ukraine conflict.
Despite all this, the generative AI industry itself is expected to grow to $1.3 trillion by 2032.
See also: Peeking Under the Hood of AI’s High-Octane Technical Needs
As AI solutions become more popular and commercialized the incredibly power-hungry, incredibly expensive, and incredibly specialized chips powering them are becoming harder to find and more costly to buy.
Top-of-the-line GPUs cost around $10,000 each.
NVIDIA, which has a near chokehold on the market for AI chips, has seen its products become both scarce and sought after.
That’s because firms need tens of thousands of GPUs to train their models, a first-step investment that generally runs in the hundreds of millions of dollars.
OpenAI’s ChatGPT-4 was likely trained on somewhere between 10,000 to 25,000 of NVIDIA’s A100 chips, and might need even more if it ever plans to launch ChatGPT-5.
Meta has stockpiled around 21,000 A100 chips, while Tesla counts roughly 7,000 in its arsenal.
China’s five largest tech firms collectively placed a $5 billion chip order this summer, hoping to build up their own foundational architectures in order to compete with Western tech companies.
While the best things in life are free, that simply doesn’t hold true in the business world — and tech firms’ sizable investments today are likely to pay off in spades tomorrow. As long as they don’t run out of cash before they get there.