Three things crossed my desk this morning, and they all tell the same story. The AI business is hitting the part of the ride where reality starts sending invoices. Here is what happened and what I make of it.
Gemini 3.5 Pro Is Still More Slideshow Than Software
Google rolled out Gemini 3.5 Pro at I/O back in May with a big promise: a 2 million token context window and a new Deep Think reasoning mode. General availability was supposed to land this month.
Here we are, and as of a few days ago the thing was still locked to a handful of Vertex enterprise customers. It has not shown up in the regular Gemini app or the studio where the rest of us actually work.
The 2 million token window is no joke. That is double what Flash gives you and the biggest of any shipping frontier model. Feed it a whole codebase or a stack of contracts and it is supposed to keep up. Google has floated pricing around $15 and $60 per million tokens, so it is not cheap either.
But a model you cannot use is just a press release with good lighting. I have watched too many launch dates slide to get worked up before I can type into the thing myself. Show me the API key and then we will talk. Until then it is a real pretty maybe.
Claude Fable 5 Just Peeled Off the Free Sticker
Starting today, Claude Fable 5 is no longer bundled into the Pro, Max, Team, and Enterprise seats for free. You want it now, you pay for it on top of what you already hand over every month.
This is the same playbook every single time. New model drops, everybody gets a free taste, and once folks are building on it the meter quietly switches on. The preview was the trial. Today is the bill landing in the mailbox.
I am not even mad about it. Compute costs real money and somebody has to cover the power company. What gets under my skin is the little dance where they call it included, get you hooked, then slide it behind the counter. Just price the thing on day one and let me make my own call.
The Real AI Bottleneck Is the Electric Bill
While everybody argues about chips and model sizes, the quieter number is power. Microsoft has bolted on more than 4 gigawatts of new capacity in the last year and a half. CoreWeave is chasing 1.7 gigawatts by the end of the year.
The folks who actually run the grid are saying, plain as day, that they cannot keep up at this pace without some help from the regulators. We are putting up data centers faster than we are putting up the power to feed them.
Out here in Texas we know exactly what happens when the grid gets stretched too thin, and it is no fun for anybody. The next stretch of the AI race is not about who has the smartest model. It is about who can actually plug it in.
So watch the power deals, not just the demo days. That is where this thing gets decided.
That is the read from my porch. The hype is loud, but the constraints are louder if you know where to listen. More tomorrow.