I do wonder about the viability of the AI model market. OpenAI has projected that its business will burn through a staggering $115 billion by 2029, significantly higher than previous estimates. While revenue growth from the ChatGPT service is accelerating, the company faces escalating computing costs and data center expenses, leading to an expected cash burn of over $8 billion this year alone. To put this in perspective, OpenAI’s projected cash burn for the next few years is more than double earlier forecasts, with anticipated expenditures of $35 billion in 2027 and $45 billion in 2028. Despite these daunting figures, the company expects to generate approximately $13 billion in total revenue this year, driven largely by ChatGPT, which is projected to bring in $10 billion. Investors remain optimistic, with major firms valuing OpenAI at $500 billion, nearly twice its valuation six months ago.
A recent article from ZDNet emphasizes that artificial intelligence does not possess true reasoning capabilities, countering claims that generative AI can think like humans. The article highlights that AI systems, particularly large language models, are often described as “black boxes,” where their internal processes remain largely opaque to even their creators. Research from Arizona State University suggests that what is termed “chain-of-thought reasoning” in these models is merely sophisticated pattern matching, not genuine logical inference. The authors of this study warn against over-reliance on these systems, noting that they can produce plausible but logically flawed reasoning, potentially leading to misguided confidence in their outputs.
OpenAI itself put up a blog about why models hallucinate. Despite advancements in models like GPT-5, which has reduced the occurrence of these errors, hallucinations remain prevalent due to current evaluation practices that incentivize guessing over acknowledging uncertainty. For instance, research indicates that models are often rewarded for guessing, which leads to a higher rate of inaccuracies. OpenAI suggests that adjusting evaluation metrics to penalize confident errors more heavily than uncertainty could mitigate this issue. The findings stress that hallucinations are not an inevitable flaw but rather a result of how models are trained and evaluated.,
The New Stack urges tech professionals to take the lead in ensuring the safety of generative artificial intelligence, as concerns about copyright and ethical implications grow. With the rise of AI-generated content, issues such as potential copyright infringement and the exploitation of creative professionals are becoming increasingly prevalent. For instance, the AI company Anthropic recently faced a substantial copyright class action lawsuit, agreeing to pay $1.5 billion for illegally downloading and storing copyrighted works. Furthermore, the ethical treatment of data workers involved in AI training is under scrutiny, with many individuals exposed to harmful content for minimal pay.
Why do we care?
Let’s be real—OpenAI lighting $115 billion on fire by 2029 isn’t just a jaw-dropping number. It’s a stability problem. Even if you’re pulling in $13 billion a year, those numbers don’t line up. And when vendors face that kind of burn, they don’t collapse quietly—they change pricing, lock down APIs, cut features, or pivot business models fast. That volatility flows right onto your customers.
So if you’re betting your business on one vendor’s API, you’d better have a backup plan. Diversify, test alternatives, and build resilience into your stack. Otherwise you risk your service delivery being tied to a vendor that could move the goalposts overnight.
And let’s kill this myth—AI doesn’t “reason.” It’s pattern matching. That’s why hallucinations happen. OpenAI even admits their systems get rewarded for confidently guessing wrong. That’s not intelligence—that’s a feature of the system.
Then there’s the legal and ethical drag. Anthropic just agreed to pay $1.5 billion for copyright violations—and ironically, that settlement could entrench incumbents, who can afford billion-dollar mistakes. Smaller competitors can’t. That locks the market in a few hands, raising costs and limiting choice.
Our job isn’t to hype AI—it’s to help customers make it work safely. That means: define what success looks like, expect the errors, and watch the legal and vendor stability side. Don’t let your clients get blindsided chasing shiny AI promises.