The Looming Retail Investor Trap

The Wall Street Journal published a story yesterday that wasn't surprising to me, but it’s making some waves on social media. In short, OpenAI has missed its target of 1 billion weekly active users, missed multiple monthly revenue goals, and its CFO has raised internal concerns about whether the company can afford its compute commitments if growth doesn’t accelerate.

Newsflash: It can’t. The growth targets have been unreasonable for quite some time, and the math doesn’t add up. I’ve been banging this drum for a while now.

“Remarkable” is probably the wrong word to use for a company chasing a $1 trillion valuation while posting numbers like this.

Lies, Damn Lies, and ARR

While the press likes to report whatever these companies say like it’s gospel, I prefer my reality to be backed by data. The financials are murky, but the actual recognized revenues for 2025 seem to be about $13.1B, not the $20B in annual recurring revenue that the hype-bros pointed to.

The crude way to calculate ARR is to cherry-pick a timespan where signups or revenue look good, then multiply that by 12 (or 52, if they’re cherry-picking a week) and report it. But that’s not real revenue.

The reality is that OpenAI is still heavily subsidizing its products, its monthly user growth is significantly down, and daily active sessions have dropped across the board. And to be clear, it’s not that these users have left the market, because Anthropic and Google have been picking up users while OpenAI has been declining.

OpenAI also has an estimated $600B in future spending commitments for compute. When, not if, they can’t make good on those commitments, it will ripple through the industry and possibly take down Oracle and some data center providers like CoreWeave. They are in a position where it is becoming more difficult to raise money, and debt will either be too expensive or flat-out unavailable. They need an out, and an IPO is one way to get them there.

LLMs are a Commodity

If you take a moment to step back from the hype, you will quickly realize that LLMs are a commodity. There’s a lot of noise coming out of China, because DeepSeek and others have made it clear that there’s nothing proprietary about “predicting the next tokens”. Even if you believe the claims about model distillation being used to subsidize Chinese models, there’s no viable way to stop it from happening.

On the user side, the pay rate is only 3-5%. Think about that. OpenAI was hyped up as the fastest-growing app in history, but 95+% of users don’t pay a dime. As prices increase and limits are reduced, the value you prop will be even worse. I’m sorry, but if you can’t get people to pay $10-20 for the utility of these products today, what makes you think they’ll buy in at 9 times that amount?

One of the bright spaces is that these tools are used as coding assistants. And as an experienced software developer, I can tip my hat to the utility of these tools because a lot of coding is pattern-based. We’ve had predictive assistance in the form of code autocompletion for decades, and in many ways, this is the next evolution.

However, professional developers build things modularly, they build for change, and most IDEs are capable of switching AI providers with a simple settings change and a new API key. There is literally zero moat. The future is smaller, specialized models that run locally or on colocated servers, not compute provided by the big providers. Even at a modest $300/month, I could instead buy a dedicated device that would run open source models for the cost of electricity. Meanwhile, token-maxing AI-bros are bragging about hundreds of dollars per day in spending.

The Competition Problem

OpenAI is competing for the same customers as Google, Microsoft, and, to some extent, Apple (Apple will end up as the dominant provider for local AI, mark my words). All of these companies have profitable businesses and can, for the most part, self-finance their AI investments. This is a big problem for OpenAI:

• They already have infrastructure and customer loyalty in place.
• In Apple and Google’s cases, they both have the ability to produce and build custom chips.
• Neither Apple nor Google needs AI revenue to survive. They just need it to be “good enough” and “convenient enough” that their customers aren’t willing to go outside.

On the China side, DeepSeek commoditized r1 in roughly six months. Leading providers are forced to be on a “training treadmill” to stay ahead, which means, you guessed it, more cash burn. The point here is important enough to put it in a block quote:

Models are depreciating assets. In an IPO, you will be buying them at peak valuation.

The Infrastructure Problem

There isn’t enough power to support all of the data center buildouts required to hit targets. That’s a fact. And, it’s also a fact that data center projects are struggling with funding, and nearly half of US data center buildouts are facing delays or cancellations.

For companies that have been hyping up scale and compute as the secret sauce for super intelligence (it isn’t, but whatever), this should be devastating to their revenue and user projections, but from what I’ve seen, it hasn’t made a blip. OpenAI has scaled back on some of its commitments, but I’ve yet to see a realistic financial model from any of these privately held firms.

Even Nvidia, the current king of circular funding arrangements, has hesitated and is walking back some of its funding commitments. And, don’t even get me started on the GPUs. Many of them are sitting in warehouses, waiting to be deployed. The next generation of GPUs will be out before some of them see any use, and differences in cost and compute between old and new will be present, though we can’t predict efficiency and power until we see the specs. What it does mean is that anyone late to the install party will be competing with others who bought in later and will have access to better hardware.

That change to 7-year depreciation for 18- to 36-month-lifespan GPUs is going to do some real damage to books when they get written down.

The Cracks are Showing

I feel like I have to keep reminding people about what hasn’t worked, because the list is getting pretty long:

• Sora was supposed to reshape video. But the Disney partnership went nowhere. Users didn’t stick around, with a net zero retention rate after 60 days.
• The partnership with Walmart failed, with Walmart choosing to use their own, more specialized models instead of outsourcing to OpenAI, a lesson that more enterprises should be taking note of.
• ChatGPT study is dead.
• The GPT store is effectively dead: there’s no real marketplace, no ecosystem.
• Adult content is probably lucrative, but I don’t think they have the will to go all-in on it.
• Data centers are being rejected on a local level by the people who live in those areas.
• Public sentiment is very opposed to their products, and regulations are likely coming.
• We’ve been 6 months from replacing developers and other white-collar workers for 3 years, and yet LLMs can’t even run a McDonald’s drive-through window.

There’s a lot of “flooding the zone” going on in tech and politics, so sometimes it’s helpful to take a step back and look at how wrong these people are. It is being marketed as inevitable, but so were cryptocurrencies, blockchain, and the metaverse.

The Political Side

I hate politics. I love debating policy, but damn, I hate politics. We have to address it, though, because it matters in this context.

The big players are US companies. OpenAI has strong ties to the current administration. You know, the administration that is deeply unpopular domestically and globally. The world is fracturing into economic blocs, and I don’t think that being “close to the White House” is an asset these days. Many governments, especially in the EU, are correctly seeing US tech companies as a risk and are correctly starting plans to reduce or remove that reliance.

The other reason I need to bring this up is that I think it’s the only scenario where OpenAI survives:

• The US continues its pathway to being a tech oligarchy.
• OpenAI maintains its position as a favored operator in that structure, becoming a critical “national security asset”.
• Chinese companies and open-source models get banned, forcing the US economy into a locked-in dependency on a small cluster of Silicon Valley companies.
• US regulators are constrained, allowing these firms to continue profiting from bad behavior with limited or no liability for harms. (see social media)

If you think that’s where we’re heading, then go ahead and load up. But, call it what it is: a political and regulatory policy bet, not a technology bet.

It’s a Trap

This IPO needs to succeed to bail out the Valley. It’s the liquidity event for a bunch of boneheaded investors carrying enormous paper valuations from an era of free money.

There is intense pressure to make it succeed, and that pressure will be transmitted to you, the retail investor, in the form of even more breathless coverage, wild, unsubstantiated claims, and crypto-bros-turned-AI-hypesters finding their new religion in AI.

You will have a hard time convincing me that letting Elon Musk, Peter Thiel, OpenAI, and Oracle go broke would be a net negative on the long-term prospects of our world.