Transcription
There was a time, and it generally wasn't that long ago, when I couldn't wait for OpenAI to go public. I'm a computer scientist. I invest. And when I first got my hands on this technology, it felt like something had shifted in the world. I was testing everything. I was talking to ChatGPT constantly. I was up at midnight running prompts. Not because I had to, but because I was genuinely captivated.
I hadn't felt that level of sleep-deprived obsessive dopamine since the first time I discovered you could automate your entire job with a Python script and spend the rest of the week playing video games and shopping for new dresses. For the first time in a long time, a technology felt like it was expanding what was possible rather than just optimizing what already existed. I wanted to invest in this company the moment it was available. I remember telling my friends, "The day the IPO goes live, I am in on it."
And then slowly and steadily, everything that made me believe in them fell apart. Not in one dramatic collapse, but instead in a thousand small erosions, each one individually forgivable, I guess, but collectively devastating. Until one day I realized I had gone from wanting to throw money at this company to not being willing to trust it with a single API call. Our relationship status officially shifted from "shut up and take my money" to "please stay at least 500 ft away from my codebase at all times."
This is a video about trust, and trust, as it turns out, is the one thing you cannot buy back at any valuation. There is a song that has been playing in my head for the duration of this research. It's about six words long. It went viral on Vine years ago, and it captures the entire OpenAI user experience with a precision that no financial analysis ever could. If you know, you know, but I still might have a go at singing it at the end.
My name is Al. I have a PhD in computer science, and I analyze AI developments to understand what's actually happening beneath all of this hype. In this video, I'm going to walk through OpenAI's leaked financial statements, the market share collapse they're trying to outspend, the product integrity issues that are driving technical and non-technical users away, the service that just cannot stay online, and the silence machine that keeps anyone from talking about any of it. These are not five separate problems, they're five symptoms of the exact same disease: a company that has consistently chosen narrative control over honesty. In the end, I will also discuss what it would take for me to actually go back to OpenAI, but for now, let's cover some numbers first.
On June 16th, 2026, independent journalist Ed Zitron published audited financial documents showing OpenAI's 2025 results. The Financial Times, my favorite newspaper, independently verified the figures as well. OpenAI declined to comment. The headline number was $38.53 billion in net losses. That sounds catastrophic, and it is, but it does require context, so let's get into that.
Roughly $41.55 billion of that figure is a non-cash accounting charge triggered by OpenAI's conversion from a non-profit to a for-profit entity in October 2025. When early investor rights and warrants were restructured into equity, the rising valuation created a paper loss. No money left the building, it's just the corporate equivalent of looking at your Steam library, realizing you spent $300 on games you've never opened, and crying about being broke, even though your bank account didn't change that much. That context matters a huge deal because intellectual honesty, of course, matters, and I'm not going to do the thing where I use the scarier number without explaining it.
But the number that should scare you is the $20.92 billion operating loss. That's real. That's what the business actually spent beyond what it earned. OpenAI's revenue reached billion in 2025, more than tripling from $3.7 billion in 2024. Revenue tripled, but total costs hit $34 billion. Research and development alone was $19.19 billion, more than every dollar the company earned in total revenue. They're basically spending $1.60 for every $1 they bring in. That is improved from $2.37 back in 2024, but losing less catastrophically is not exactly the same as approaching sustainability. If I manage my personal finances with that exact same ratio, I'd be trying to convince the bank that a mountain of Uber Eats receipts counts as liquid assets.
The most revealing line in the leaked documents is the Microsoft dependency. In 2025, OpenAI paid Microsoft $17.2 billion, $10.59 billion on R&D compute for training models, $6.05 billion on cost of revenue for serving them, plus hundreds of millions more in sales and administrative expenses. Microsoft, in return, paid OpenAI $303 million. That is an absolutely wild dynamic. It's like funding your roommate's multi-million dollar startup, paying for all of their groceries, buying them a server rack, and in return, they give you a $5 Starbucks gift card and a high five.
HSBC analysts estimate OpenAI may need more than $207 million in additional capital through 2030. The company's own projections point to profitability by 2029 at the earliest. Independent analysis puts the more likely date at 2031. And the timing of this leak is definitely worth noting. OpenAI filed a confidential S-1 with the SEC around May 22nd, publicly confirming it on June 8th. The financial leak dropped 8 days later. So, somebody wanted these numbers in the public domain before OpenAI could frame them in a prospectus. Under SEC rules, the full audited financials would have become public eventually, at least 15 days before the investor road show, but by then the narrative would have been set. Now, the narrative belongs to the numbers.
They want a $1 trillion valuation for this. Bridgewater's Greg Jensen reportedly told clients that the implied revenue multiplies priced OpenAI for a monopoly outcome that does not yet exist. But, monopoly outcomes require the actual monopoly, and OpenAI no longer has one.
Of all the numbers in the leaked financials, the one that puzzles me the most is not the R&D burn. I think that one is actually quite appropriate for what the company is. Building frontier AI models is expensive, everyone knows that. But, the number that makes me stop is the $5.73 billion on sales and marketing. In 2024, OpenAI spent $1.11 billion on sales and marketing. In 2025, that became $5.73 billion, which is a 418% increase. Revenue tripled, marketing spend quintupled. They are spending more per dollar of revenue on convincing people to use the product than they were a year ago. Why does the company that launched the AI revolution, the company that held 87% market share at its peak back when Sam Altman could walk into any tech conference on Earth, whisper the word into a microphone, and cause five venture capitalists to simultaneously throw their checkbooks in his face, why does that company need to spend $5.7 billion making people aware that it exists?
The market data answers the question. According to Sensor Tower's State of AI 2026 report, ChatGPT's share of the global AI assistant market fell below 50% for the first time in March 2026, reaching 46.4% by May from 87% to 46% in roughly 18 months. Google Gemini now holds 27.7% and Tropics Cloud reached 10.3% with the highest paid subscription conversion rate in the industry at 13%. After OpenAI signed a deal with the US Department of Defense in February this year and also pulled the ChatGPT-4 model, ChatGPT uninstalls in the United States spiked roughly 200% above the app's average. Cloud out downloaded ChatGPT for five straight days in March. ChatGPT's churn rate moved from 12.7% in January to 14.5% in April. And in the enterprise market where the real money lives, Cloud now wins approximately 70% of head-to-head deals against OpenAI.
The $5.7 billion is not building new demand. It is compensating for lost trust. And what makes this particularly frustrating is that the trust wasn't actually that difficult to maintain. OpenAI had a free product that hundreds of millions of people genuinely loved. It was the gold mine. It had a brand that was synonymous with the entire category. People still say "I'll ask ChatGPT" the way they say "I'll Google it." They had achieved the ultimate corporate holy grail, becoming a verb. And then they apparently decided that being a noun with a massive marketing budget sounded somehow way cooler. It had organic advocacy from developers, researchers, educators, writers, and curious people all over the planet who were voluntarily spreading the word because the product felt transformative. That is the kind of marketing money literally cannot buy. It's called goodwill, and they lit it on fire.
Instead of investing in the relationship with their existing users, improving reliability, communicating honestly about model changes, building trust through transparency, they spent $5.7 billion trying to acquire new users to replace the ones they were alienating. Maybe they wouldn't need so much money for advertising if they worked on not being a disaster in terms of relationship management, hmm? You had first-mover advantage, you had majority market share, you had genuine public affection, and you squandered it all by treating users as a problem to be managed rather than a community to be served.
And I get it, I need to calm down. Sometimes I spend time in the OpenAI subreddits, not because I enjoy suffering, but because understanding public sentiment is part of understanding a company and having a YouTube channel as well. And the refrain I see repeated over and over and over again is some variation of: "Just tell me how much money you want. If 20 pounds a month isn't enough, make it 40. Just stop lying to me." These are people who are willing to pay more for honesty, and I know this is real because I did it. I personally quadrupled my AI spending when I switched to Claude. My bank looks at my Anthropic billing statements and genuinely wonders if I've developed some sort of a highly specific text-based gambling addiction.
Not because Claude is perfect, no AI tool is, but because when I select a model, I trust that I'm getting that model. When I hit a usage limit, the system tells me. The transparency is not a feature, it is the very product. OpenAI has tried various strategies to manage its cost problem. Different subscription tiers, releasing older models on cheaper plans, enterprise licensing, education discounts, yada yada. Those are all individually reasonable decisions, but they do not address the underlying issue. The issue is not that the pricing is wrong or that the tiers are badly designed, the issue is that people don't trust you anymore. And when people don't trust you, no tier structure in the world or the next world will fix your churn rate.
The alternative is not exactly complicated. Just do what Anthropic does. Limit sessions transparently. Have multiple tiers with clear boundaries, tell people what they're getting, ask for more money if you need it. Many people will just pay, just stop lying. That is literally all you have to do. You cannot advertise your way out of a credibility crisis, and the product itself is making the crisis worse.
So, how did this trust collapse? Let's get into that for a second. Let me tell you about the moment I lost faith in OpenAI as a technical user. ChatGPT now uses what it calls a safety router, a system that silently reroutes your conversation to a different, more restrictive model when it detects language it classifies as emotionally sensitive. You are not told this is happening, by the way. The switch occurs on a single message level, and only becomes apparent if you specifically ask the model about it, and sometimes it will lie. Users have identified a variant called GPT-5 chat safety that replaces the model you selected. And on top of this, OpenAI recently replaced specific model names in the ChatGPT interface with three labels: instant, thinking, and pro. You are no longer choosing a model, you're choosing a vibe, which is fantastic if you're trying to pick out a scented candle or an ambient lo-fi playlist, but slightly less ideal when you're trying to debug a production database at 2:00 in the morning.
The actual model used in a response is decided by ChatGPT based on prompt complexity and other internal settings. Usage limits can also trigger silent downgrades. You hit a rate cap and get served a cheaper model without notification. As TechCrunch put it, "Many people assume they're interacting with a single consistent intelligence. In truth, they're interacting with a flexible system that constantly adjusts itself." Translation in normal words: you think you're talking to a digital Einstein, but behind the curtain it's actually just three smaller models in a trench coat arguing over who has to do the math this time.
Let me make this tangible for a second. Imagine going to a shop and buying a bottle of Pepsi Max. You pay for it, you take it home, you open it, and it's actually Coke full sugar. Nobody told you, nobody asked, the shop just decided based on some internal assessment of your mood and your aura that you'd be better off with a different product. If you're a casual consumer buying one bottle, maybe it's a bit annoying, maybe you kind of shrug it off. But if you're a business that has built an entire supply chain around receiving Pepsi Max, if your product depends on the consistency of that specific ingredient, it is a catastrophe. And in any other industry, it's also a lawsuit. Bait-and-switch product substitution has triggered class-action litigation in contexts far less consequential than this one.
Now, a colleague of mine who works in AI made the argument that the API solves this. When you use the API, he said, you choose which model your endpoint calls. You have control. And on the surface, this is true. But my counter was this: how do you know what exists on the other side of that endpoint? How do you verify that the model you've specified is the model actually serving your request? You're trusting a company that just got caught with a $21 billion operating loss that silently reroutes paying web users to different models, that replaces model names with mood categories. You're trusting that company to honestly label what's behind the API call.
And the API problem is actually worse than the web interface, not better. At least on the web, an attentive user can sometimes feel that something is off. The tone shifts, the pacing changes, the responses get cautious. On the API, you have no sensory signal whatsoever. You send a request, you get a response, and you take on faith that the label matches the product. It's pure, unadulterated vibes-based engineering. You just press execute, close your eyes, cross your fingers, and pray that the output doesn't randomly decide to hallucinate a recipe for concrete.
You're building applications, deploying workflows, making business decisions on outputs that you have no way to independently verify came from the model you personally specified. Nobody uses programmatic access for idle chitchat. API calls mean applications. Applications mean iterative design, repeated workflows, products that depend on output consistency, not just uptime. I mean output quality, model behavior, the specific flavor of how it reasons and responds. Large language models are not exactly interchangeable. They have distinct characteristics, different reasoning patterns, different strengths with certain types of problems, different tendencies in how they structure information and handle ambiguity. Other people call this personality, but this is more technically accurate.
If you've built and tested your application against the specific behavior of one model, and a different model gets silently substituted underneath your production system, your outputs will change. Your product will behave differently, your users will notice, and you have no idea why because, as far as you know, nothing changed on your end. This is not a hypothetical concern, by the way. This is what it means to build an infrastructure you cannot trust, and it's exactly the reason why I switched.
I want to separate two things here, however, because the conversation around AI safety often conflates them together. There is the general public, non-technical users on ChatGPT doing homework, writing emails, having casual conversations. For the free tier, I can almost understand some degree of model routing for cost management, even if I disagree with doing it silently. But then there are paying customers, subscribers who selected a specific model, and especially API users, developers, and companies building commercial applications on top of these models. These people are paying for a specific product. When they select a model, they need to receive that model, not a safety variant, not a cheaper fallback, not whatever the system decides is appropriate based on an internal assessment of their prompt. The product they specified, full stop. That's it.
Because think about what this means at scale. If you're a startup that has built a customer service platform, for example, a legal research tool, a medical information system on top of OpenAI's API, your entire product quality depends on the consistency of what's behind that specific endpoint. If the model changes underneath you, even subtly, even for safety reasons, your product changes. Your customers experience something different. Your quality assurance process, which tested against a specific model, is now validating against a ghost. You've built your house on sand, and you don't even know the sand moved. It's like hiring a Michelin star chef to cater your wedding, but halfway through the main course, they secretly swap places with a teenager working the drive-thru at Taco Bell because the kitchen ran out of butter.
Bait-and-switch product substitution has triggered class-action lawsuits in industries with far less at stake. I don't want to make legal speculation the focus of this video, but I just want to point out that it's worth noting that in virtually any other consumer product category – food, pharmaceuticals, financial products – silently substituting a different product for the one a customer purchased would be subject to regulatory scrutiny at minimum and litigation at scale.
For what it's worth, I also looked into whether Anthropic does the same thing, and they don't. There is no silent safety router swapping models mid-conversation. On the API, you specify a model, and you get that model. When you hit a rate limit on the web interface, it tells you. There is a documented safety fallback in Claude code for flagged cybersecurity and biology content, but the system notifies you when it happens, and you can switch back. The difference is not that one company limits you, and the other one doesn't. They both do. The difference is that one of them tells you, and the other one doesn't.
Reducing that to a mere technical edge case completely misses the point. This is a glaring character deficit, and if I'm going to build a product, if I'm going to stake my professional reputation on the consistency of an AI system's output, I need to be able to trust the character of the company behind the API. I need to know that when they say this model is X, it is model X, not "model X unless our internal classifier decides otherwise," not "model X until you hit an invisible rate cap and get quietly downgraded." Model X, verified, traceable, auditable. That trust is the infrastructure beneath the infrastructure, and OpenAI has cracked it.
I won't belabor this section because the data speaks for itself, but it needs to be said. OpenAI is spending $5.7 billion on marketing for a product that cannot reliably stay online. On April 20th, 2026, ChatGPT suffered a global outage. At peak, more than 8,700 users reported issues in the UK alone. We were also affected. ChatGPT, Codex, and the API platform were all down for over 90 minutes.
If that were an isolated incident, it would be unremarkable, I guess. Every cloud service has outages, but the OpenAI status page from March to June 2026 reads like a broken record. "ChatGPT.com access issues. ChatGPT failing to load or save. Errors with conversations on Android and iOS devices. Codex selected model at capacity. Elevated errors for ChatGPT conversations in Europe. Some users may experience empty response from ChatGPT in web." Each entry ends with the same six words: "All impacted services have now fully recovered." The phrase appears so often that it starts to read less like a status update and more like an incantation. Something you say not because it's meaningful, but because saying it is like a ritual, like you're summoning something. It has the exact same energy as clicking, "I have read and agree to the terms and conditions." Nobody believes that, nobody feels better, but if we don't say the magic words, the server gods might just get angry.
European users specifically are getting a measurably worse product as well. OpenAI's compute use feature wasn't available in the European Economic Area, the UK, or Switzerland at launch. Elevated error rates have disproportionately affected European conversations. Users in these regions are paying the same subscription fee, by the way, for a product with fewer features and lower reliability. They are essentially paying a premium subscription just to receive an automated notification that says, "Sorry, your digital rights are just too robust for this feature. Please enjoy this empty text box instead."
And throughout all of this, there have been OpenAI employees who, when confronted with user frustration, responded with contempt. Users raising legitimate product concerns were dismissed, talked down to, mocked, treated as if their expectations of receiving the product they paid for were somehow unreasonable. This matters beyond the individual interaction or my personal disdain for this company because enterprise buyers, the B2B customers who represent the high-margin future of any AI company, see how you treat your consumer users. The business-to-consumer relationship is the shop window for the business-to-business relationship. If a restaurant treats its walk-in customers with visible disdain, no corporate client is going to book their annual dinner there. Every enterprise procurement team evaluating OpenAI can see the subreddits, the status page, the dismissive employee responses, and they're all drawing conclusions.
Now, let's talk about the silence machine that I mentioned. In May 2024, Vox reported that OpenAI's offboarding agreements contain non-disparagement clauses that were extraordinary even by Silicon Valley standards. Departing employees were forbidden for the rest of their lives from criticizing OpenAI. Even acknowledging that the NDA existed was a violation of it. It's the first rule of Fight Club, except instead of Brad Pitt punching you in an alleyway, it's a team of corporate lawyers deleting your entire life savings from a secure server. And the enforcement mechanism was not exactly a lawsuit, it was the clawback of all vested equity, potentially millions of dollars that the employee had already earned.
Daniel Kokotajlo, a former OpenAI researcher who left because he lost confidence in the company's responsible behavior, publicly confirmed that he had to surrender what could have been a life-changing sum of money in order to leave without signing. When the story broke, Sam Altman posted on X that he was genuinely embarrassed and that he did not know this was happening. The CEO of a company building superintelligent AI did not know what was in his own employment contracts. Sure, we're supposed to trust this man to safely navigate the alignment problem for a global digital superintelligence, and he cannot even align his own HR department's onboarding documents. That is either a governance failure of extraordinary proportions, or it is something else. Either way, it is not the kind of assurance that historically inspires confidence in a company's leadership.
OpenAI eventually stopped being "closed AI" and eventually reversed the policy, releasing former employees from the non-disparagement obligations and confirming that vested equity would not be canceled, but the SEC whistleblower complaints had already been filed alleging that OpenAI's NDAs violated Dodd-Frank whistleblower protections. The trust damage was again done.
And this connects to something larger. The financial leak happened because the numbers came out through a journalist, not through OpenAI's own disclosure. The model rerouting happens without telling users. The status page incidents are acknowledged minimally and retrospectively. The departing employees were silenced by contract. Every single information flow in this company is managed through control rather than transparency. The instinct at every turn is to contain, to manage, to frame, never to simply tell the truth and let people decide for themselves. This is not a company that made a few mistakes in a period of rapid growth. Growth doesn't make you hide things. Growth doesn't make you silence people. Growth doesn't make you swap products on paying customers without telling them. Those are choices, and they are choices that all point in the exact same direction.
So, what would it take for me, a computer scientist who genuinely loved this technology, to go back to OpenAI? I've thought about this a lot, honestly, and there are only three paths to this. The first is price. If the product became so cheap that the trust deficit didn't matter because the financial risk was negligible, but that would mean being unprofitable for them in perpetuity, which is the opposite of a path to the trillion-dollar valuation they're seeking, so it's probably not going to happen. The second is quality. If the models became so far ahead of every competitor that there was no alternative, but the market data says otherwise on that one. ChatGPT is no longer the only game in town, and the quality gap between frontier models has narrowed, not widened.
The third is radical transparency: full observability of API calls, verifiable model identity on every single request, traceable infrastructure so that I can confirm, not trust, literally confirm that when I specify a model, that model is what serves my request. I want to be able to trace back to the function that calls my model, not an endpoint. Receipts, please.
None of these paths are easy. I think the last one is not even possible, and that is the point. Trust is asymmetric. It takes years of consistent behavior to build, and it takes moments of dishonesty to destroy, and the cost of rebuilding it is always, always so much higher than the cost of maintaining it. OpenAI had everything. First-mover advantage, majority market share, genuine public goodwill, the most recognizable brand in artificial intelligence, a technology that made computer scientists like me feel like the world had shifted. I catch myself sometimes thinking about the good old days, except the nostalgic daydreams I'm chasing aren't sunset drives through Los Angeles, but sunset API calls with awe-inspiring new tech. It's fine because GTA 6 is coming out anyway, so I'll get those sunset drives soon enough.
But anyway, they squandered it. Not through one catastrophic failure, but through a pattern of choices that all point in the same way. Growth over integrity, narrative over honesty, control over transparency, the sycophantic models that told users what they wanted to hear, the silent rerouting that gave them something other than what they paid for, the employees who dismissed legitimate concerns, the NDAs that threatened people's livelihoods for speaking up, the financials that came out through a leak because the company wouldn't publish them voluntarily.
And the thing that makes this generally sad, not infuriating, literally sad, is that none of this was necessary. The technology is extraordinary. The underlying models are genuinely capable. The product could have sustained itself on quality and trust alone if the company behind it had simply chosen to be honest. If they had said, "This model is expensive to run, so we're limiting your sessions," instead of silently routing you to something cheaper. If they had said, "Our costs are enormous, and here's our plan for sustainability," instead of hiding the numbers until a journalist forced their hand. If they had said, "Some employees disagree with our direction, and here's how we're addressing their concerns," instead of threatening to take away their savings for talking about it. Every single trust failure in the story has a transparent alternative that was available and would have been received better. Literally just copy-paste Anthropic's homework. It's right there. They are not even hiding it. They chose control every time, and now the bill has arrived.
This company is about to ask public investors to value it at a trillion dollars. Investors who, unlike the early venture backers and sovereign wealth funds, will have the audited numbers in front of them. A $21 billion operating loss, 18 consecutive months of declining market share, a product that silently substitutes what it serves you, a status page that reads like a liturgy of recurring failure, a culture that silences dissent by threatening equity, and $5.7 billion spent on marketing because the product can no longer sell itself on reputation alone. It is a multi-billion dollar cry for help wrapped in a slick PowerPoint presentation.
And those public investors should ask themselves a question that I think is more important than any revenue multiple or growth projection: If this company treats the truth as a liability before it has public shareholders, what happens after? Public markets are supposed to impose discipline through disclosure requirements and fiduciary obligations, but disclosure only works when the culture behind it values transparency. And everything, everything in OpenAI's history suggests that when given a choice between telling you the truth and managing your perception of it, they will manage your perception every single time. Trust is the one thing you cannot buy back at any valuation, and that is what OpenAI is about to discover.
If you want to understand the broader pattern of how AI companies treat their people and their users, and why Meta is going through its own version of the same crisis, I covered that in this video that I'm linking here. Anyway, as promised, >> [snorts] >> "Why the you're lying? Why you always lying? Oh my god, stop lying. Always lying to me. You're lying so much. Seriously, OpenAI, just stop it." Bye, everyone.