Transcription
I don't think anyone has picked up on the connection between these five stories.
First, Meta launched a consumer gaming app where you type a prompt and get a playable game. Then, Meta is also building a cloud business to sell its spare AI compute, and of course the stock moved on that news. Meanwhile, Reuters got hold of a Meta town hall where Zuck admitted agent development hasn't accelerated the way that he expected. OpenAI meanwhile floated giving the US government a 5% stake in their company worth about $42 billion at today's valuation. And finally, a sandwich chain, Jersey Mike's, the one with Danny DeVito in the ads, they filed for an IPO that mentions AI 22 times.
So, you have a gaming app, a cloud business, a confession, an equity offer, and a Subway sandwich shop. Those don't look like they belong in the same story, but they do. And if you're still watching the old AI scoreboard, the one about who shipped the best model in the past month, you're going to miss this story. Here's the shift that is happening across the industry right now.
So, first the old scoreboard. For about 2 years, the AI industry ran on a single question: Who has the best model? Every earnings call, every launch, every leak got scored against that. And the logic went like this: Models are the product, so whoever owns the best model wins, so spend whatever it takes to own the best model. And that logic justified the largest capital buildout in tech history. The big five hyperscalers are on track to spend north of $600 billion on capex this year, up roughly a third from last year, and most of it is AI infra. And to be clear, the model race worked. It produced real capability jumps, clear leaders. We just spent the spring and early summer watching frontier models get so good at cybersecurity work that the federal government has started staggering releases, right?
But here's the thing about a race. When the leaders start visibly changing what they're competing on, the race is starting to change. And last week, for the first time, you could see that emerge on a bunch of new fronts at once, from Meta, from OpenAI, and from the market itself. And by the way, I'll give you a bonus at the end about Anthropic, because that's coming in, too.
So, let's start with Meta. Meta quietly launched a consumer app. No announcement, no keynote. It just showed up on the App Store and Google Play at the end of June, and TechCrunch was able to spot it. You type a prompt like "puzzle game about matching colors", and the app generates a playable mini-game in a few seconds. Meta calls them Gizmos. There's a social feed where you browse what other people made, and the games respond to touch and tilt. You can use the camera. The whole thing came out of Meta's acquisition of the Vibe coding gaming platform Gizmo earlier this year.
Okay, next story from Meta, the cloud. Bloomberg reported that Meta is standing up a business to sell access to its excess AI compute, that means all of its models, as well, to outside customers. There's an internal group called Meta Compute that oversees the infra buildout, and the plan puts Meta in competition with AWS and Azure and Google Cloud. Analysts put Meta's total spend this year at around $145 billion. And the market's reaction is very interesting. Meta stock jumped on the report. Investors looked at "we're going to rent out our spare GPUs", and said, "let's get more of that."
Reuters reported that Zuck told employees that AI agent development of the last 4 months, quote, "hasn't accelerated in the way we expected." Now, let's hold these three stories together and discover the connection. A company that believes model ownership was the whole game would not do any of what we've just described. If the model's everything, you keep every GPU for training, and you'd never sell a competitor your compute. If agents are 12 months from transforming everything, you don't ship a whimsical prompt to Game Joy. You wait for the agents.
Meta is doing the opposite. It's monetizing infrastructure directly. It's experimenting with consumer surfaces that generate engagement today with small models it has already, and it's being really honest internally that their agent timeline has kind of indefinitely slipped. That's not a retreat. The money's still flowing at full volume. It's a company redrawing the map. The compute is an asset class now. The consumer surface is a distribution play, and the model's just one layer in the stack for that.
And here's where it gets more interesting, because while Meta was redrawing the infrastructure and distribution layers, OpenAI was making a move on a layer that people don't realize is a layer. On the same day that Pocket hit the news for Meta, reports landed that OpenAI has been discussing giving the US government roughly a 5% equity stake. Those same reports indicated that there are similar conversations going on with other major model makers. Now, at the $852 billion valuation from their March round, that would be a $42.5 billion stake. I'm sure it's higher now.
The mechanics matter here. This is not a sale. The reported structure is a donation of shares to a public wealth fund, an idea OpenAI first sketched in a policy paper back in April. The model is similar to the Alaska Permanent Fund, which pays every Alaskan an annual dividend from oil revenue. And Altman reportedly wants this to apply to every leading US lab, Anthropic, Google, Meta, all paying 5% into the same vehicle. The talks are very early, and a real version of this needs Congress. Very much treat this as a proposal.
Now, notice this surface just days after Washington told OpenAI to slow down ChatGPT 5.6. Under the executive order signed in June, the government gets up to 30 days of pre-release access to the most capable models. And the commerce secretary reportedly called Altman personally to warn against releasing without approval. So, ChatGPT 5.6 is sitting in a limited preview for trusted partners instead of shipping worldwide. The government also has already done this to Anthropic's Methos, and Fable was part of this story, too. And in the background, Senator Bernie Sanders introduced a bill in June that would take 50% of the AI company stock for the government into a sovereign wealth fund. Not 5, 50.
So, read the 5% offer for what it is. It's a bid to set the terms of a government AI stock before someone else sets them. The frontier labs have hit a point where the binding constraint on their business is not capability and actually it's not even compute. It's permission. Washington can now delay a flagship release with a phone call and in that world political alignment is not corporate social responsibility, right? It's just infrastructure to get your model out. Open AI is trying to buy regulatory headroom the way meta is monetizing data centers. Both of them are investing in a layer of the game that wouldn't officially exist on the old which model is best scoreboard.
Which raises the obvious question. Has anyone else noticed that the scoreboard changed? And last week the market answered that really loudly. CNBC ran a live segment on July 2nd titled AI's three big narrative violations. Analysts walking through long held assumptions about the AI trade that have now broken. And the first violation they led with is the idea that you have to own the model. And their exhibit A was the same meta compute story we just talked about. When the company that spent more than almost anyone else to own frontier model starts selling the picks and shovels instead, the market's going to update. The point isn't which analyst said what. The point is that a phrase like the model is the moat has moved from being a Silicon Valley contrarian take into being mainstream into being a CNBC chyron. The repricing of this story is happening right now on the street as the street figures out where AI value actually goes.
And that brings me, yes, I can't believe I'm saying this, to Jersey Mike's. The sandwich chain filed for its IPO on July 2nd under the ticker JMKE. It's Blackstone backed, it's $724 million in revenue last year. It's a genuinely solid business. And the filing mentions artificial intelligence 22 times. And now TechCrunch had a great headline for this. Jersey Mike's IPO illustrates how bad the AI hype has become and it's getting treated as a joke, right? But I want you to see what it actually tells you. When a Subway sandwich shop's lawyers decide that their S1 needs 22 AI references. That is a yardstick for how cheap capital has become around AI. The froth hasn't disappeared just because the market has gotten more sophisticated about models. Instead, it's kind of migrated outward into anything that can plausibly attach the letters A and I to a growth story.
Ultimately, this is the tension that defines the moment. The market is getting smarter about the AI core, hence the conversation on CNBC, but it's getting sillier about the edges because the money hasn't left the market. It's just searching for returns. So, Meta is redrawing their map, opening eyes, buying political cover. There's hype about sandwiches. This is the larger picture. The model race isn't over, but it's not the only race in town anymore. And this week really called out a bunch of other AI tracks that are visible.
If you count the layers that were actively contested in just the last few days, there's a whole infrastructure piece where both Meta and now SpaceX and xAI are competing over renting out massive quantities of infrastructure. There's a distribution layer where there's yet another bet like Pocket out there trying to get consumers to get interested in building games, building anything, not just consuming. That's a huge bet. And there's a political layer, right? And that's going to get increasingly contested as we move into the midterms and the general election cycle. It'll be full of debate about equity stakes, about pre-release model review, about safety, about staggered launches, about the public's legitimate stake in artificial intelligence and what that means.
Notice what's happening here. It's not that people are spending less. CapEx is up roughly a third. No one is saying they're scaling back CapEx. Enterprise budgets on AI are holding or growing despite lots of complaints. The signal is that the biggest spenders out there are figuring out how to allocate their capital to maximize value. And so they're moving into distribution. They're moving into infrastructure margins. They're moving frankly into the political permissions game.
So, Anthropic just got Fable 5 back. They mended fences a bit with the government, but that's not the story here. And they're continuing to hyper-focus on the enterprise use case. And that is why they are investing so heavily not just in their models, but also in the ability to deploy those models in harness in companies. And this brings me back to what I talked about last week, Claude Tag. It brings me back to talking about the forward deployed engineers Anthropic is investing in. They, too, are playing a different game. They're playing the distribution game for enterprise, and they're playing it very, very seriously because if they can build a sticky harness around the enterprise, they've won an incredibly resilient revenue source as enterprises put more and more and more into intelligence to gain leverage and frankly reinvent the firm. So, that's the Anthropic piece. And that one didn't make the news this week, but sometimes the quiet stuff doesn't and it still tells the same story.
So, where does this leave us? I would invite you to begin thinking about what it means to have a guaranteed pipeline of increasing capability for models, whether you're running a company or whether you're an individual working at a company, or whether frankly you're a solopreneur and a builder. You're going to get these models getting better and better. Fable 5 will not be the last one. There'll be new ones from Open AI. There'll be new ones from Anthropic. Maybe Google will get back in the game there. Who knows? Maybe Meta will ship something.
At the same time, that is not where the game is going. The game is going to how these models successfully integrated into society. Because if you put all of these pieces together, it's a return on AI in society conversation. That's why Open AI is talking to the government. That's why Anthropic is getting deeply involved in enterprise. That's why Meta is figuring out how to get consumers from consuming to building. All of these things are indicators that we are wrestling with the early stages of the society transition that follows like a wake behind the initial AI innovation wave.
We should ironically expect exactly what we saw from Jersey Mike's in a world where AI is changing a bunch of stuff and people don't quite know where to look and they don't quite know what real change looks like. And it makes sense in that world to have an S-1 filled with 22 AI references. You'll see more of stuff like that and you'll also see more of real legitimate experiments trying to help people figure out how they actually use AI in ways that are useful. I think Poke and Clicky are both out there on the more serious productivity side, less good consumer game side, and there'll be more. Think about where you are in this ripple of societal change that's happening and figure out how to surf it. I'll see you next time.