AI Notes: One Falls, Several Go All In

The number worth remembering this week isn't a benchmark score. It's $20 million.

As reported on October 13, Character.AI filed for bankruptcy, and Disney bought part of its assets for roughly that amount. A few years ago it was one of the brightest names in generative AI, with engagement numbers everyone envied. On the liquidation table, all of it was worth a rounding error.

Users are not a moat

Plenty of people will read this as "the AI bubble is starting to burst." I don't. What it exposes is a more basic problem in consumer AI: the cost structure.

Internet products have a marginal cost close to zero. Another user costs almost nothing, so building traffic first and monetising later works. AI products are different: every extra line a user types burns more compute in the back end. The longer people use it, the more you lose, unless each use maps to revenue. Character.AI's users were exactly the ones who chatted most and longest, and its model wasn't exclusive; they could chat just as well somewhere else.

So the $20 million price is the answer in itself: users with no paying relationship are worth almost nothing on an asset list. Models can be rented and interfaces copied. The only thing that can really be valued is who keeps paying you. The buyer took the characters and the content, not a business.

Apple is after permission, not intelligence

The same day, Apple's "Welcome Home" event laid out a full home lineup: the $299 HomeView hub, HomePod mini 2, a new Apple TV 4K, a rebuilt Siri, and locks, cameras, doorbells and thermostats.

Most of the discussion was about whether the new Siri is smart enough. I don't think that's the point. The bottleneck for AI in the home was never whether it understands you; it's whether it can act for you: unlock the door, change the temperature, check who's outside. Those actions need permissions on devices, and permissions live in hardware.

That permission layer is exactly what Apple is laying down. Once a lock and a camera are installed in a home, replacing them costs far more than switching chat apps. Siri being a few months behind can be fixed; once someone else holds the right to act, it's very hard to take back. Put it next to Character.AI and the contrast is striking: one had attention but neither a paying relationship nor the right to act; the other has both, and intelligence is the part it can afford to add later.

Models are becoming fast-depreciating assets

This week OpenAI pulled GPT-5.5 from ChatGPT, Work and Codex in favour of GPT-5.6 Sol, and set January 14, 2027 as the retirement date for the GPT-4o and GPT-4.1 APIs, with gpt-6-luna as the successor. Google opened Gemini 4 Argon to AI Pro users and AI Mode in Search, with 1 million tokens of output and the top spot on the Vals Index.

On the surface it's routine turnover. Underneath there are two implications.

For those downstream, retirement is a form of pricing power. When the upstream provider decides a model goes offline, everyone downstream has to re-test, re-tune and re-cost. A company built on someone else's model has part of its gross margin written into that provider's product roadmap, and it doesn't get a say.

For the providers themselves, the problem is bigger. A flagship model's lead now lasts months, while the data centers that train and run it depreciate over fifteen or twenty years. One side of the balance sheet holds models with ever-shorter lives; the other holds buildings and power contracts with very long ones. In finance, that kind of maturity mismatch is one of the structures you watch most closely.

Safety is turning from a cost into an asset

Anthropic opened its fifth APAC office in Singapore, signed an AI safety MOU with Singapore's Ministry of Digital Development and Information, and held pre-IPO investor briefings in the same stretch.

Safety spending used to be treated as a cost: the part that slows launches and adds compliance. For a company heading to public markets, its meaning changes. What public markets fear most is tail risk they can't price, and a safety agreement signed with a government prices some of the regulatory risk in advance. Governments and regulated industries are also exactly the customers who pay most reliably and switch vendors most slowly.

Anthropic has turned a cost into part of its valuation, and into a sales channel.

Humanoid robots: shipments are not demand

IDC counts nearly 25,000 humanoid robots shipped in the first half of 2026, up 432%, with China taking more than 95%. But 69% went to research, education and exhibitions, not factories.

What these numbers reveal is the buying logic. Research, education and exhibition budgets ask "do we need to have one," not "when does it pay back." Factories buy equipment on how fast it earns its money back. The first kind of demand arrives fast and leaves fast; the second arrives slowly but, once it comes, comes at scale. Humanoid robots will really take off on the day factories start running this purchase through a payback calculation.

The same question

Add Meta's new frontier model, a Microsoft regulatory filing (data-center capacity going from about 12GW to more than 38GW by 2032) and ASML's AI-driven Q3 beat, and the outline of the week is clear.

The application layer is clearing out and the infrastructure layer is doubling down. But both ends are asking the same question: who pays, how much, and for how long. Character.AI couldn't answer it, so it was liquidated. Apple and Anthropic are locking in the people who pay, in different ways. And those tens of gigawatts are a bet that someone will keep paying for the next fifteen years to support models that get replaced every few months.

Nobody can say yet whether that math works. All I'm sure of is that, starting this week, the market will be asking the question louder.