A 65B model just beat an 816B one. Stop paying for size.

Google, Amazon, Microsoft and Meta plan to spend a combined $725 billion on capex this year. That's up 77% from the year before. And this week, a 65-billion-parameter model walked past rivals more than 12 times its size.

Everyone is reading the spending boom as proof that bigger wins. The short version: this week's results point the other way. Size is a cost you pay, not a feature you buy.

Start with the scoreboard. Ricursive Intelligence, a lab founded by former Google researchers, released Ripple-65B Preview under an Apache 2.0 license. It scored 80.52 on TerminalBench 2.1.

Meta's Muse Spark Open, at 816 billion parameters, scored 76.4 on that same test. Nvidia's Nemotron 3 Ultra, at 550 billion, scored 56.4. On SWE-Bench Pro v2.0 Hard, Ripple posted 79.1 against Muse Spark Open's 39.6, which is double the score at roughly one-twelfth the size.

That gap is comically lopsided. It is also not a fluke of one benchmark, since the model also beat the 501-billion-parameter Reflection Beam on both coding tests.

Then do the money math. Ricursive says the full research and development bill, from data mixing to debugging to the final training run, came to $32 million. The company calls that one-tenth the cost of the previous open-weight leader on agentic coding, and it used about 70 times fewer reinforcement learning rollouts than that leader.

Here's the thing. $32 million is about 0.004% of $725 billion. The best open coding model in the U.S. cost less than a rounding error on one year of big-four capex.

Funny enough, the wildest research story of the week points the same direction. MIT CSAIL researchers had GPT-6 Astra edit a small model's weights directly. It changed 131,072 of the 800 million parameters in Qwen3.5-0.8B, just 0.016%, and the model's MATH-500 score jumped from 7.9% to 48.1%.

No new data and no retraining. Astra found a looping defect in two layers that burned the model's output budget before it could answer, then removed it. The small model already had the skill, and a bug was standing in the way.

I'll be real about the limits. Ripple is a preview, and those scores come from its own launch, not from your codebase. The MIT trick did nothing for Gemma 3 270M, a model that rarely loops, and the editor doing the surgery was itself a frontier model.

Now look at who pays for the size habit. Gartner says enterprise GPU deployments average just 5% utilization, so companies pay for roughly 20 times the capacity they use. Gartner also puts a $4.9 trillion gap between what data centers need to earn by 2030 and the AI revenue it expects them to see.

That bill rolls downhill. Microsoft raised Microsoft 365 Business Standard from $12.50 to $14 per user per month, a 12% jump, citing Copilot Chat. I read that as the bigger-is-better bet showing up on your invoice.

So who wins? Labs that treat compute as scarce, the inference providers like Together AI and OpenRouter already serving Ripple, and any buyer willing to test before paying. Who loses? Anyone whose whole pitch is that their model is bigger.

My rule of thumb: if a model a tenth the size scores within a few points of the giant on the benchmark closest to your real work, the giant is a tax. Run 50 of your own tasks on both before you sign anything priced on parameters.

See you next week, geeks.