The Best Model Loses
The Weekend Leverage, July 19th
Last week, I asked for cool side projects that y’all have been building lately. I was inundated with responses. There were, of course, many of the classic vibe coding projects like a personal to-do app. However, what I found most invigorating was responses like Keith’s.
Keith works in corporate strategy in Boston and isn’t a programmer. Despite that, over the last few months he’s turned a Mac mini, a Raspberry Pi, and an old Google Pixel into a personal software empire. I mean, just look at all the stuff he’s built.
There’s a “Library” that resurfaces everything he’s ever saved and makes suggestions based on his location and schedule.. So he’ll get podcasts at the dog park and movies on movie night. There is a map of every restaurant his friends have recommended, an AI gym coach that builds workout plans for him, a music rec engine, and a daily briefing bot, all on top of hardware he is hacking together. Again, this dude isn’t a programmer! He’s just doing stuff.
Of course, there are lots of software products that have similar features to Keith’s empire. But his builds exemplify the spiritual value of AI: it is giving people the permission and tools to build things they want. No longer do people have to acquiesce to the whims of the residents of Palo Alto. We can make digital things for ourselves, ex nihilo.
I believe that creation and stewardship are divine pursuits, spiritual urges manifesting in hammer and nails and hacking and welding and writing and making. We build not because we desire ownership, but because we crave independence. That a twenty-bucks-a-month subscription unlocks this feeling and capability is remarkable. Here’s to the Keiths of the world.
In that spirit of doing stuff rather than taking anyone’s word for it, I spent this week building my own experiments—starting with the question everyone’s been asking me.
Are the new models actually spooky, or is that just good marketing? The federal government held up Anthropic’s Fable and put GPT-5.6 Sol through 2 weeks of review before letting the rest of us near it. That is either a real signal or a great PR accident (or perhaps both). I believe that this tier of models is genuinely novel and important, but you shouldn’t just take my word for it. This article gives you two prompt guides on how you can test the models yourself. Read here.
Is Thinking Machines’ new model the most expensive ad of all time? In 2004, Chanel paid $58.5 million (inflation adjusted) for a 3-minute-video of Nicole Kidman fleeing paparazzi. The Baz Luhrmann directed commercial holds the record for the most expensive television advertisement ever made. On Wednesday, Thinking Machines maybe, sorta, kinda broke that record. Inkling, their new 975B-parameter open-weights model, is free to download, and Thinking Machines makes money on it only when someone runs it through their API or tunes it on their hardware. It mostly exists to sell Tinker, the fine-tuning platform where models can be customized. The company concedes in its own launch post that Inkling is “not the strongest overall model available today, open or closed.”
Thinking Machines didn’t publish the cost to make Inkling, but Emad Mostaque pegs Inkling’s pretraining run at $10M-$20M based on the disclosed specs and GB300 rental rates, and that figure covers pretraining compute only. It excludes the roughly 200-person team and every failed run, which are the costs that actually make this stuff expensive. For comparison, GPT-4 cost more than $100M if you include those categories of cost.
I’m sure if Thinking Machines had a preference, they would loooove to charge frontier prices for a frontier model, but OpenAI and Anthropic now operate at a compute scale Thinking Machines can’t match. So, they are going for the AI-native version of an open-source strategy. The comparison I kept hearing from investors was Databricks; the company that open-sourced Spark while building a managed platform on top. That version of open-sourcing has brought it to a $188B valuation this week.
Please don’t do this comparison! It is dumb!
Spark was built largely by Berkeley academics, so Databricks got its founding asset for roughly zilch (they’ve since paid plenty of engineers to maintain it, but the origination cost rounds to zero). More importantly, open-source software is a compounding asset. Community contributions make active projects more valuable every single year. Thinking Machines giving their model away is not the same! The community can quantize Inkling, port it, and tune it, but nobody outside the building can make the base model smarter. The training run is done. Open-source software appreciates. Open weight AI models mostly depreciate, and they do so quickly. Exactly 1 day later, Moonshot announced Kimi K3, with full weights promised by July 27, benchmarking well above Inkling and within reach of the Fable and Sol frontier. Inkling held the open-weights spotlight for about 24 hours. Now, Thinking Machines has to re-buy the ad at frontier prices when they train their next model.
The one thing keeping this from being pure cash incineration is that depreciation is also the revenue model. Fine-tunes don’t port across base models, so when Inkling 2 ships, every customer with a tuned checkpoint has to re-tune on Tinker. And to be fair, Thinking Machines built things into Inkling that the (stronger) Chinese models don’t offer, like a trained-in thinking-effort dial and native audio fine-tuning. I don’t think Thinking Machines’ strategy is dumb or illogical. My point is that it is tough to have a business model that requires you to create such expensive assets that depreciate so rapidly.
If you look at YouTube comments for the Chanel ad, people are still leaving remarks calling it the best ad of all time, 22 years later. I don’t think anyone will even be thinking about Inkling in 22 months.
Meta Enters the GPU Slumlord Business. On Monday, Meta announced it was building one of the most expensive single structures in the history of American capitalism, and the market could not have cared less. The company said in a blog post that Hyperion, its data center supercluster in Richland Parish, Louisiana, would expand to 5 gigawatts and cost more than $50 billion. Follow the repricing. When construction began in December 2024, the estimate was $10 billion. Then it was $27 billion. Now it is $50 billion, a 5x increase in under 2 years. In response, the market just did this emoji ¯\_(ツ)_/¯
The stock closed down 1.86%, roughly in line with a Nasdaq that fell 1.55% on an ugly chip-led selloff.
It’s interesting to contrast what previous capex announcements have done. When Meta guided 2026 capex to as much as $145 billion in April, more than double 2025’s $72 billion, the stock sank 7%. It then slid for another 2 months to a June 25 bottom of $542.87. What changed Zuck’s fortune was ironically, nothing to do with the product. On July 1, Bloomberg reported Meta was building a cloud unit to sell excess compute (yay new revenue streams!) and the stock jumped 10.1% intraday. On July 10, reports that Meta’s custom chip Iris starts manufacturing in September sent shares up another 6% (yay more verticalization to help margin pressure!). Add it up and the stock rallied 21% in 3 weeks on rumors of a landlord business.
On Friday, we finally got some tenant news. The New York Times reported that Anthropic is in early talks to lease compute from Meta in a deal worth up to $10 billion over 2 years, with monthly payments and an early exit for both sides. The market has changed from looking at Meta as an AI lab burning $145 billion to a burgeoning neocloud with the world’s best ad business attached.
Still, I think this business is far less stable than other commentators seem to think. Anthropic’s other landlord relationship, the SpaceX Colossus deal reportedly worth about $45 billion over 3 years at $1.25 billion a month, can have either party terminate with 90 days’ notice, and Musk himself described it as a 180-day lease with mutual 90-day cancellation after that. The largest compute contracts in history carry the commitment level of a month-to-month apartment lease. Given the Meta talks already include monthly payments and an early exit for both sides, expect the same structure here. What a wacky market when we have $50 billion buildings being supported by 90-day relationships.
AI gets rejected from baseball. MLB just made a mid-season rule change to ban generative AI in the dugout. Since 2016, every team has had league-issued iPads with a custom third tab where clubs could install their own apps. About a third of teams have installed AI chatbots that recommend decisions typically reserved for managers — substitutions, and even pitch calls.
When we think about the role of deployable, cheap intelligence, the thing everyone worries about is jobs. Which is fair! However, I think the short to medium term impacts will be subtler and more pervasive. Intelligence too cheap to meter is impossible to turn away, and it creeps its way in, subtly making everything more competitive. It makes the pitches harder to hit, it makes it harder for this newsletter to stand out.
Go see the Odyssey this week. I will admit to enjoying but not loving the Nolan movies. This, however, is easily his best film. It is a wonderful meld of his blockbuster sensibilities with artistic vision. The cinematography is particularly good, and the set designs are otherworldly. Not a perfect film, but one that deserves the hype! If you can, I would highly recommend seeing this on a film projector, digital will short change the incredible color grading of the movie.
Go and be kind this week,
Evan
Sponsorships
We are now accepting sponsors for the Q3 ‘26. If you are interested in reaching my audience of 35K+ founders, investors, and senior tech executives, send me an email at team@gettheleverage.com.