The AI industry is going through a correction. Last week, Nvidia announced its financial results, and the whole world held its breath. If it fell short of expectations, it would all be over. If it beat them, we would be back on track. In the end, the results exceeded estimates, but the stock still fell after an initial rise. Many things about the AI boom seem suspicious. The way money is fueling the investment frenzy is quite questionable, and it has become a meme in which the same $1.4 trillion check (approximately CZK 32.2 trillion) circulates among a small group of participants. We can call it vendor financing, but it does not look good.
Risks for the Major Players
The greatest risk is borne by hyperscalers such as Microsoft, Amazon, and Oracle, as well as new cloud players such as Nebius and CoreWeave. They are caught in the middle between chip suppliers such as Nvidia and buyers of computing capacity such as OpenAI. They have no choice but to buy actual chips from Nvidia and hope there will be sustained demand—and therefore revenue exceeding costs—from those purchasing computing power, enabling them to meet their obligations. If not, the buyers will leave, reduce their commitments (or go bankrupt), Nvidia will already have sold the GPUs, and the hyperscalers will be left with billions of dollars and gigawatts of unused capacity that is rapidly losing value due to the short lifespan of GPUs.
Growing Doubts About Sustainability
Concern about the AI industry's (un)sustainability is growing. Sam Altman lost his composure on a podcast when the host asked him how OpenAI, with revenue of $13 billion (approximately CZK 299 billion), could make commitments totaling $1.4 trillion (approximately CZK 32.2 trillion) with companies such as AMD, Nvidia, and Microsoft. His nervous and irritated response did not go unnoticed and was interpreted as a sign that revenue is insufficient to support such investments. A few days later, OpenAI CFO Sarah Friar said she would support a federal guarantee for chip investments to make it easier to finance major investments in AI computing capacity, using words such as “guarantee” and “support.” Friar and Altman later walked back and clarified this comment, but it was not a good week for OpenAI.
The problem is not just revenue—whether $13 billion (approximately CZK 299 billion) is enough to support commitments of $1.4 trillion (approximately CZK 32.2 trillion)... it is also about margins. AI is amazing, but it may be too cheap to be true. It feels like the early days of Uber, when calling an Uber was so much cheaper than taking a taxi that you wondered whether owning a car made any sense.
A Real-World Example: GitHub Copilot
Take GitHub Copilot. The Pro version costs $10 (approximately CZK 230) per month and comes with a limit of 300 requests. Additional requests cost $0.04 (approximately CZK 0.92) each, so 300 × $0.04 = $12 (approximately CZK 276). GitHub is selling $12 worth of value (approximately CZK 276) for $10 (approximately CZK 230). We do not know the margins on that $0.04 (approximately CZK 0.92), but they could conceivably be negative, which would push the actual value even higher for the same $10 (approximately CZK 230).
Overall, it appears that all challengers, such as OpenAI, Anthropic, and Cursor, are subsidizing demand with negative margins. Google was caught off guard by the AI boom and took some time to recover, but now it is returning in full force, launching Gemini 3 as well as an IDE. Unlike the challengers, Google has deep pockets thanks to other highly profitable revenue streams (Search, YouTube, GCP) and is better prepared to play the negative-margin game.
AI as SaaS and PaaS
If the challengers do not want to be crushed by the avalanche of costs, something has to change. AI is both SaaS (software as a service) and PaaS (platform as a service), for both the consumer and enterprise markets. Developers offer SaaS applications such as ChatGPT, Sora, Codex, and Cloud Code, as well as PaaS in the form of APIs for large language models (LLMs) and traditional storage and computing.
The standard pricing model for SaaS is a subscription, both for consumers and enterprises. A subscription is essentially an entitlement to consume a certain amount of resources. With consumer SaaS such as Netflix, the entitlement covers the entire catalog of movies. With tools such as ChatGPT, the entitlement varies depending on the model used. If it is very cheap and several generations old, the entitlement may even be unlimited (meaning it is too cheap to meter). If the model is more expensive (for example, reasoning models) or cutting-edge, the entitlement is limited to a certain number of requests. Treating all user requests the same regardless of the model can be disastrous, as Augment Code discovered:
“A model based on user messages is not sustainable for Augment Code as a business. For example, over the past 30 days, a user on our $250 Max plan (approximately CZK 5,750) made 335 requests per hour, every hour, for 30 days, and is approaching $15,000 (approximately CZK 345,000) per month in costs for Augment Code. This kind of usage is not inherently wrong, but as a business, we must price our service according to our costs.”
AI Costs: Tokens as the Key Factor
The subscription/entitlement model works best when costs (COGS) grow more slowly than the user base. Netflix is the best example. Its costs consist of its catalog of movies and TV shows. The more users it gains, the more content it needs to satisfy the diverse preferences of its growing user base. However, the same movie can be watched by more users, so costs grow more slowly than the user base.
In AI products, however, the costs are tokens. Every user request consumes brand-new, unique tokens required for the LLM. There is no sublinear growth here. The more users there are, the higher the costs. That is why AI SaaS products are addressing this with usage-based pricing, which is becoming increasingly common.
For now, it is not purely usage-based. It is still a combination of an upfront monthly payment with a certain entitlement, plus an overage mechanism if needed. This is a typical model for enterprise SaaS, where companies buy “seats” for their users and pay “overage” fees for anything extra.
A purely usage-based model is more typical of enterprise PaaS—that is, cloud accounts where you pay only for storage and computing, nothing more and nothing less. This would require a major shift in mindset among consumers, who are used to paying a fixed monthly fee for digital services such as Netflix (with frequent price increases, it must be said), rather than paying something akin to a utility bill.
A purely usage-based model would allow AI SaaS providers to pass costs plus a margin on to users without the risky entitlement game, and it appears to be the simplest way out. End users will probably face higher prices or have to reduce their consumption of AI products, which will be difficult because we have become dependent on them.
Options for Reducing Costs
I do not know whether AI SaaS has many other options for reducing costs. I see efforts to optimize token usage, such as introducing an “auto mode,” where the user does not select the model; instead, the SaaS does so automatically and tries to use cheaper models whenever possible. I also wonder whether tokens can be reused. AI products feel highly personalized and tailored to each user request, but I think there is considerable overlap in what users enter (at least in the AI chatbot space), so perhaps there is room to “cache” requests as in search and reuse previous responses.
The AI industry is at a critical juncture. Either companies will continue burning money to provide amazing products that are too cheap to be true, or they will find a way to pass those costs on to end users while retaining them.
Source: betterthanrandom.substack.com
