Monologue: AI Hits A Wall

OpenAI and executive departures

0:00This is an iHeart Podcast. Guaranteed human. Amazon Health AI presents Painful Thoughts. I, um, I can't stop scratching my downtown. Yeah, but I'm not itching to go downtown and tell a receptionist I'm here to talk about my downtown. Some things you'd rather type than say out loud. There's no question too embarrassing for Amazon Health AI. Chat your symptoms and get virtual care 24-7.

0:30Healthcare just got less painful.

0:35Hey, it's Kelly Rowland. You may not know this, but I have eczema. So I get how it can steal your time. But why let eczema take over when you can talk to your doctor about EBCLIS? EBCLIS Lubricizumab LBKZ, a 250 milligram per two milliliter injection, is a prescription medicine used to treat adults and children 12 years of age and older who weigh at least 88 pounds or 40 kilograms with moderate to severe eczema. Also called atopic dermatitis that is not well controlled with prescription therapies used on the skin or topicals or who cannot use topical therapies.

1:06EBCLIS can be used with or without topical corticosteroids. Don't use if you are allergic to EBCLIS. Allergic reactions can occur that can be severe. Eye problems can occur. Tell your doctor if you have new or worsening eye problems. You should not receive a live vaccine when treated with EBCLIS. Before starting EBCLIS, tell your doctor if you have a parasitic infection. Paid partnership with Lilly. Respect your time. Ask your doctor about EBCLIS and visit ebglis.com. Or call 1-800-LILLY-RX or 1-800-545-5979.

1:35Hey there, I'm Kat. I'm a college athlete. I was diagnosed with ankylosing spondylitis, which caused back pain and stiffness every day. With Cosentix, I'm able to stay active. Cosentix Secukinumab is prescribed for adults with active psoriatic arthritis, ankylosing spondylitis, or non-radiographic axial spondyloarthritis. Available in both IV infusion and self-injection. Don't use if allergic to Cosentix. Get checked for TB before starting. Increased risk of infections and lowered ability to fight them may occur.

2:07Like TB or other serious bacterial, fungal, or viral infections. Some are fatal. Tell your doctor if you have an infection or symptoms like fevers, sweats, chills, muscle aches, or cough. Had a vaccine or plan to. Or if IBD symptoms develop or worsen. Serious allergic reactions and severe eczema-like skin reactions may occur. Learn more at 1-844-COSENTIX or cosentix.com. Move and feel better. Ask a rheumatologist about Cosentix. Roll out! Pros.

2:37Find the right flooring for the job. In Lowe's newly redesigned flooring showroom. Shop our expanded assortment of trusted brands and get up to 35% off style selections. Stain Master and Stain Master Pet Protect special order carpet. And while you're here, save $100 on a DeWalt 20-volt max 6-amp hour 2-pack of batteries. Now just $179. Get more of what you need faster and save more where it counts from the ground up. Our best lineup is here at Lowe's. Ballot through 826 while supplies last. Selection barriers by location.

3:07Call Zone Media Hello and welcome to this week's Better Offline monologue. I'm your host, Ed Zitron.

3:19Better Offline.

3:21It's going to be a long week after a long month after a long few years, so I'm recording this monologue a little earlier than usual, on Tuesday to be specific, which means I don't have thoughts on NVIDIA's earnings and won't have them unless, of course, things go really haywire. Now I've said that, that sounds possible, but I think it's likely the no-IT-loads-refused cash dump will continue at least for another quarter there. In any case, we have a far more pressing problem. The deterioration of OpenAI and the AI industry hitting a wall. Last week, OpenAI completed the $7 billion employee share sale, but unlike most employee

3:56share sales, OpenAI used its own cash to buy its employee stock at the previously raised valuation of $865 billion. A few days later, both Chief Operating Officer and former CFO Brad Lightcap and Chief Revenue Officer Denise Dresser left the company. Dresser had been at the company less than a year, and likely walked away from tens of millions of dollars' worth of unvested OpenAI shares. In April, she said that she had, and I quote CNBC, never seen this level of conviction spread so quickly and consistently within the enterprise

4:27industry, though it appears that said conviction didn't spread far enough to keep her at a company that definitely plans to go public sometime in 2027. Unless, of course, it's not quite so definite anymore. Well, Anthropic has dramatically accelerated its plans to go public, with Bloomberg reporting that it is aiming for a September or October listing, though we're yet to see any reporting on its actual S1. Now, this is a problem for OpenAI in many ways, but let's get through all the news first. At the very end of last week, Bloomberg also reported that OpenAI was on track to hit $40

5:00billion in annualized revenue, $7 billion in annualized run rate less than Anthropic had claimed it hit in May. And yes, you heard that right. I said on track, which does not mean it hit that number, and no, it did not define run rate, because Bloomberg neglected to do so either, and, well, who needs definitions when you're just talking about bullshit, right? It could mean the last four weeks multiplied by 12, or 13, or really anything. We truly don't know. Everyone seems to use a goddamn different definition of run rate.

5:30Anyway, then on Monday, Bloomberg reported that Anthropic had hit $65 billion in annualized run rate, which had defined, I shit you not, as a metric that projects full year revenue from a shorter period with no further explanation. Who fucking cares, right? It's only about what's ostensibly meant to be the largest AI company in the world with $300 billion worth of commitments. No need to actually know what you're talking about. No need to say anything further about a market moving number. No need to do that.

Revenue claims and market anxiety

6:01Why fucking bother, right? Just put the story in the bag, the hype in the bag. Check, please. We wouldn't really say that at a grocery store. Anyway, in other words, OpenAI had taken four months to not quite catch up with Anthropic, only for Anthropic to quickly drop that the race wasn't even close. Regardless, AI boosters immediately got a little anxious, because many had been convinced by anonymous sources and questionable data blocks like Yip It that annualized run rates were somewhere between $70 billion and $100 billion. I guess the voices in their heads weren't reliable.

6:31Did they try sacrificing an animal? I don't know what these fucking people do. It isn't read, it's just listening to the voices that say, sir, yes, $100 billion annualized by end of year. Only the worthy will be able to invest in this IPO, yes. Anyway, said anxiety is spread across the AI industry, with even the most mold-poisoned boosters now fretting that both Anthropic and OpenAI's revenues had slowed down, mostly because they have. I'll also add that these numbers, which come from the end of July, are still from the era of token maxing, where companies were encouraging employees to burn as many tokens as possible

7:04with no regard for costs, meaning that growth will invariably slow from here as enterprises cap spending, as Walmart and Stripe and multiple others have done. A day after Anthropic crept in its coffee, OpenAI made the incredible decision too, and I quote, pause some frontier reinforcement learning training to ensure it can meet the appropriate alignment. A complex way of saying that OpenAI will stop training its latest models for at least the next few weeks to increase safety, connected to the recent incident where an OpenAI model in an incorrectly configured server was able to hack Hugging Face, which is less of a case

7:36of it going rogue and more of a case of OpenAI being bad at its job. If you tell software to do something, it's going to try and do it. Oh, it isn't lined right. You programmed the software poorly. You're wrong here. This isn't the magic of AI. This is the tragic of OpenAI. It's unclear what OpenAI hope to achieve here, rather than vaguely saying, we care about safety, but from everything I'm seeing on social media and in my inbox, everybody is interpreting this as OpenAI has stopped training new models, which is also the truth, albeit only they claim for two weeks.

8:07Maybe. If you believe them. This is a really fucking stupid idea, by the way. Or at least was a really fucking stupid idea to announce at a time when everybody has noticed that OpenAI's revenue growth has decelerated. New models and features are effectively the only way for these companies to keep growing, and OpenAI is far behind Anthropic, which makes the idea of not doing anything new, or much new, all the more questionable. OpenAI will, of course, try and spin this as its new models having incredible new abilities it must prepare for. Oh, God, are you going to see?

8:37Look at anyone saying that, and you just look at them, you can see the three-finger thing from Inglourious Basterds, like, yep, there's someone who doesn't give a shit. There's someone who will scarf down the marketing slot when fed to them.

Compute limits and industry fallout

8:48Mmm, yummy. Anyway, this is all very ruminant, but the word on the street is that this is OpenAI hitting a wall and running out of compute to serve inference and train new models, with one person suggesting that Anthropic could follow suit. This does make sense for Anthropic, who got so desperate that it's renting capacity from Elon Musk's SpaceX and potentially even Meta, but I'm surprised to hear it about OpenAI, who boasted earlier in the year about its massive compute advantage over its competitors. And again, it's also very possible that compute capacity is just not coming online very fast.

9:21Sources at Oracle tell me that Stargate Abilene, which was meant to be finished in June, or maybe the beginning of the year, or maybe the end of the year, it's really unclear at this point, only has three out of eight buildings fully operational, with a fourth in some sort of strange limbo, where it's sort of ready but not quite, isn't clear if it's generating any money, and you need money to pay for goods and services, and it's called revenue, not sure any of these companies fully understand business, but anyway, this could also be a way for OpenAI to save money. Per the information, OpenAI spent $8.6 billion on research and development, so model training,

9:55in the first quarter of 2026, a number that likely increased in Q2 with the launch of its GPT-5.6 Sol, Lunar, and Cyber models. And also, all the compute it was spending accidentally hacking HuggingFace. I swear that's like a felony, right? Like, we don't really have, like, laws anymore, I realize, but anyway, in any case, if Anthropic and OpenAI have truly hit the limits of compute, there's very little to can accelerate it coming

10:26online, as every AI data center is the equivalent of cramming an entire city's power, usually spread over a billion or billions or more square feet into around a million square feet, built in the space of two years with constantly increasing costs around memory, specialist talent, specialist materials, and the oil necessary for the gas turbines and the trucks to bring all this shit to the plot of land. And all of this is happening as revenue growth decelerates. If OpenAI and Anthropic can't or won't train new models, that means they'll effectively become model wrapper companies, building products on top of Claude or GPT models in the hopes

10:59that they can juice further revenue growth, which is not going to help very much, given that basically every company that builds models, uh, businesses on top of their models is both unprofitable and can barely reach 90 million a month in revenue, which wouldn't even cover a day's worth of inference if it were profitable, which it would not be. I'll add that if Anthropic beats OpenAI to IPO, it's unlikely OpenAI will be able to float, or at least have anything approaching an IPO that anyone would actually want to invest in. While both are unprofitable, horribly so, Anthropic has the revenue advantage and slightly better

11:32unit economics, which is all it needs to make Sam Altman's public offering untenable. I think it's time to start having a reasonable discussion about the potential collapse of OpenAI, and the ramifications for counterparties like Google, Microsoft, Amazon, CoreWeave, Cerebrus, and of course, Oracle, and all associated venture capitalists. In particular, Thrive, who has billions of dollars tied up in this company. And of course, SoftBank, which needs that liquidity, otherwise, well, they're triple fucked. If you're a journalist listening to this, you need to start seriously thinking about and

12:03writing about this, because the current state of this company suggests that it has few viable paths to avoid perdition. It can't raise much more money. 90% of its last funding round came from NVIDIA, Amazon, and SoftBank, none of which are really going to do a return trip. SoftBank literally can't afford it. Venture capitalists have never offered more than $12 billion of any OpenAI round, and even a We Are The World telethon isn't going to be enough to keep this pig inflated. What a disgusting way of referring to a company. Anyway, continuing. There is not really a viable path for OpenAI here.

12:35I don't know what they do, but it's kind of looking like they're going to die, as I wrote in my newsletter this week, and I'll put a link to that in the notes. Either way, I'm off on a flight tomorrow. It's a crazy week. You'll know why soon. It's all good stuff. I love you all. Thank you for listening. I'll be back next week with the wonderful Cal Newport. Catch you then.

13:02This is an iHeart Podcast. Guaranteed Human.