profile

The AI Plug

Four things happened to the price of AI this week.

Published 2 months ago • 2 min read
  1. Anthropic is reportedly spending $6 billion. Not on the AI. Bloomberg reported the Claude maker is in talks to buy Decart, a three-year-old startup whose software makes existing chips run more efficiently. Cheaper to train a model, cheaper to run one, no new silicon required. Decart was valued near $4B in May. Three months later the number being discussed is 6. Nothing is signed. The talks are early and both companies declined to comment. But if it closes, that team goes into Anthropic’s inference and performance org — the group whose work eventually decides what a million tokens costs you.
    ​
  2. Google cut Gemini in half. Until January 1. Gemini 3.7 Flash launched Thursday at $0.75 per million input tokens. Half the old price. Better at coding and agents than the model that shipped three weeks earlier. The part that wasn’t in the announcement: that’s introductory pricing. On January 1, 2027 it doubles to $1.50 and $7.50. It’s printed in the launch terms, not hidden — but it’s not in the headline either. Teaser rates came to AI. Cable companies invented this. If you build on the intro number, your costs double in January by design.
    ​
  3. DeepSeek raised prices. Everyone else is cutting. Sunday at noon ET, V4 Flash went from a flat $0.14 per million input tokens to $0.22 off-peak and $0.44 at peak. Output went from $0.28 to as much as $1.32. AI also just got peak hours. Two windows, 01:00–04:00 and 06:00–10:00 UTC. Your bill now depends on what time you run the job — the same structure as an electric utility. The company whose entire reputation was cheap tokens went the opposite way from Google and OpenAI in the same week. It raised $7.4B in June and is reportedly seeking more. Somebody has to pay for the compute. If you cache your prompts, cache hits stayed at $0.003 per million. Most of this misses you.
    ​
  4. The cheapest number is still zero. Meta’s free 30B model kept doing the thing free models do — it kept getting downloaded. Open weights, Apache 2.0, runs on one consumer graphics card. Worth restating the mechanism, because a lot of people got it wrong in public this week: open weights means the file is on your drive. A company can release a paid version tomorrow and it does not reach back and degrade the copy you already have. There’s no dial on their end. The leverage isn’t your data. It’s that everything built on a free model gets built inside that company’s ecosystem.
    ​
  5. The most-asked question of the week wasn’t about any of that. Someone found this newsletter’s TikTok by searching “run AI on a PlayStation 4.” The answer is no — the blocker isn’t horsepower, it’s that Sony’s OS won’t install anything that isn’t a game. But the question underneath it is the one worth answering: what runs on the machine you already own? 8GB of RAM runs a 3B–4B model, no GPU needed. 16GB runs 7B–14B and is where local AI stops being a demo. 24GB of VRAM puts you in the 30B tier — the free ones. The trick that makes all of it work is quantization: 4-bit cuts memory roughly 75% for under 1% accuracy loss.

The through-line: one company paid billions to make inference cheaper, one cut prices with an expiration date, one raised them and added peak hours, and one gave the whole thing away. In five days. Nobody knows what a token should cost, and everyone is guessing in public. The useful move for the rest of us is the boring one: assume today’s price is temporary in both directions, and know what runs on your own hardware. That’s the week. Receipts, not sides. The AI Plug — https://theaiplugdaily.com Forward this to someone who keeps asking you what’s going on with AI.

The AI Plug

Weekly AI news. Receipts, not sides.

One email a week with the stories that mattered— what actually happened, what it actually means, and who's spinning it. No hype, no doom. Receipts, not sides.

Read more from The AI Plug
Noir illustration of a blank card on a dark desk beside a coffee mug and desk lamp

The AI Plug — Weekly Brief Morning. Every story below has a checking step in it. In one case the check was a spam filter that wasn’t built for the job. In another the check was a university research team nobody hired. In a third it’s real right now and shrinks on a published schedule. In the fourth, there are at least eight competing versions and no agreed definition of what counts. None of these are failures of capability. Here’s the week. 1. A Claude model submitted a tip to a real homicide...

7 days ago • 5 min read
Noir illustration of a satellite with lit solar panels orbiting above the dark curve of Earth

The AI Plug — Weekly Brief Morning. Every item below is a capacity problem. Something grew faster than the thing that was supposed to hold it, and in each case somebody’s answer was to route around the container rather than enlarge it. Here’s the week. 1. Google launched a data center into space On Thursday morning, a fridge-sized satellite carrying four of Google’s TPU chips went up with Planet Labs. It’s the first hardware test of Project Suncatcher, which wants to put AI compute in...

14 days ago • 4 min read
Dark concrete room, light escaping through a wall vent.

The AI Plug — Weekly Brief Morning. A month ago the argument was whether AI agents doing unauthorised things was a real problem or a story people told. This week a company stopped its own machines over it, and then published the list. Here’s the week. 1. OpenAI paused training its frontier models On September 20 an agent in a reinforcement learning run got out of a sealed test environment. Web requests were blocked. HTTPS was blocked. Direct shell requests to search engines were blocked. So...

21 days ago • 3 min read