AI Companies Are Betting on Physics
From chips to robotics to household data collection, AI companies are finally learning that they can not escape the constraints of physics.
1. Coinbase hits the token hangover
Coinbase CEO (Brian Armstrong) provided 5 approaches they used to cut AI spending nearly in half: make AI costs visible, route tasks to cheaper models, cache more, keep context lean, and use open-weight models where they are good enough.
After months of tokenmaxxing, people are finally coming to their senses: this was never going to be sustainable if every internal workflow became an all-you-can-eat token buffet.
Armstrong’s advice is basically the grown-up phase of AI adoption: stop sending every tiny task to the most expensive model in the room.
But this wouldn’t be worth our time, would it?
What I want you to pay attention to here is the use of open-weight models. More and more, you’d see how people rely on open-weight ones more than the commercial ones, knowing that 99% of the day-to-day can be resolved by a good enough model, whereas only a small portion of companies and teams can justify their most advanced AI usage.
Then the next question would be, where does this leave the frontier AI labs?
2. Meta hits Google’s capacity wall
Google reportedly limited Meta’s use of Gemini after Meta asked for more capacity than Google could provide. The reported cap slowed some internal Meta AI projects and forced teams to ration token usage.
This is the same logic I wrote about in the Cursor and SpaceX piece (it’s coming out soon! waiting for the audio to be processed), just with different actors.
There it was SpaceX and Anthropic. Here it is Google and Meta.
Across the entire supply chain, the clean profit pool keeps showing up layers below the model: chips, data centers, power, cloud capacity.
Meta can do all the spending, hiring, and shipping of the product as they want. But if it does not control enough of the right capacity, it still ends up begging Google for room on the machine.
3. OpenAI built a chip
OpenAI and Broadcom introduced Jalapeño, a custom inference chip built to run AI models more cheaply and efficiently.
→ OpenAI
Software companies are all discovering physics at the same time!!
Models are nice, but power, cooling, chips, and supply chains decide who can serve them to billions of people. The proof is in Nvidia’s P&L: $75.2 billion of data-center revenue in one quarter, with roughly 75% gross margin.
That is the shovel business now OpenAI wants a part of.
It’s clear from the announcement that Altman wants to do more with this new chip than just serve OpenAI; they also aim to sell it to other AI labs. Would others buy into it? That’s an entirely different question.
4. ChatGPT wants ad money
OpenAI has been pitching marketers on ads inside ChatGPT and reportedly said around 20% of ChatGPT queries have direct commercial intent. That follows years of Sam Altman sounding uncomfortable with ads, including calling ads plus AI uniquely unsettling and describing ads as a last-resort business model.
So, last resort.
Not solid proof, but enough to see how desperate Altman is.
Let me spell this out, given that many enterprises have now switched to using Anthropic by default, and free users keep burning compute, then ChatGPT’s commercial-intent queries start looking less like a privacy problem and more like life-saving money.
The company also says the ads will be clean, labeled, and separate from answers.
However, once the assistant becomes a shopping surface, the pressure is obvious: the closer you are to the purchase decision, the more valuable the ad slot becomes.
So it isn’t a question of whether Altman will take his own promises seriously, but of when he will cross the line into being so close to users that it could feel like an infringement of personal space and an abuse of their knowledge about you.
5. Government becomes an AI margin risk
OpenAI is reportedly staging access to GPT-5.6 after a request from the US government, with preview access limited to a small group of trusted partners. Just the week after, Anthropic was subject to an export ban by the US government, which suspended access to Fable 5 and Mythos 5 for foreign nationals.
The Anthropic ban was not a one-off, weird event.
It was the first clean example of a new business risk: frontier AI access can be switched off by directive.
This is where enterprise AI gets fragile. A company can spend months wiring a frontier model into workflows, training teams, and selling clients on the new process.
Then one directive arrives, and poof, the workflow is either useless or in need of a redesign. So now you need to factor regulation into your cost estimation.
But that’s not the worst… If you are calculating AI profitability without a government-intervention line, your spreadsheet is missing the thing that can break the whole deployment.
I did a deep-dive piece on exactly this topic. You will want to understand the risk at every level. The work will be released to paid subscribers (and later to free subscribers) very soon.
6. Meta wants a prediction market
Meta is reportedly building Arena, a standalone prediction-market app, think Polymarket.
This is a smart move if you look at Meta’s actual asset: it already knows what people watch, share, argue about, and click via all their social media and chat platforms.
A prediction market adds a different layer: what people think will happen, with a price attached. Think of this as cashing in an opinion poll.
7. Midjourney moves into physics
Midjourney is moving from image generation into a full-body ultrasound scanner for wellness and preventive health use.
A wellness scanner is a very strange leap for an image company, until you see the broader pattern.
Again, the story is that AI companies are moving into physics.
Software margins are ugly when every answer costs compute.
Google is building robotics models, OpenAI has circled robotics, and everyone is trying to escape the same trap: a chatbot can sound useful, but the real money is in tasks that touch the physical world.
Somewhere, it needs bodies, rooms, motion, symptoms, objects, and all the messy real-world data that never lived on the internet.
8. Consulting’s AI dynamic
Clients are pushing consulting firms toward outcome-based AI fees, while consultancies deepen partnerships with frontier model providers.
It makes sense for AI labs to want to work with consultants because they’d get more enterprise seats, more usage, and more workflows built around your model.
But after the Anthropic export-control episode, the risk is obvious.
If a consultancy designed a client workflow on the latest frontier model, it is one directive away from being useless or spending months rebuilding the process.
The deeper the model sits inside the workflow, the more a government order can break the client promise overnight. This makes outcome-based pricing a catastrophe for consulting firms.
Every time they send a team in, w/o getting paid due to an executive order like this one, it's wasted time and hours not billed.
9. A cleaner is also now a data collector
Someone is offering free apartment cleaning in New York City if customers let cleaners record first-person footage of the work. The footage is used to train AI and robotics systems on household tasks.
→ BBC
Free cleaning is not generosity, but data acquisition with a mop.
So see this as a version of AI moving into the real world, just in a different way.
Robots need footage of people wiping counters, folding clothes, opening drawers, and navigating awkward apartments.
This is to solve the same problem that every robotics company faces: there is no internet-scale database of human motion in homes. And to make physical AI work, companies need cameras in the places where physical life actually happens.
You get a clean apartment, they get a training set.
Comment and tell me Is this a trade-off you’d make?