Can Harvey Tenet be a moat if it has the shelf life of an onion?

Can Harvey Tenet be a moat if it has the shelf life of an onion?

Some of you knew this was coming 😇

ICYMI: Harvey trained a legal model (Tenet) based on Kimi K3 👉🏽 https://lnkd.in/d3yPq-Wc

First, does owning a model make you profitable?

Let's run the numbers...

K3 runs 104B active params and needs ~1,000 B200s to serve 14.6 Trillion tokens a month to 1,500 customers = roughly $41M annually

Compute costs are more manageable than Fable 5 API costs but..

Assuming your model remains more valuable than a frontier..

And that may only last 43 days = shelf life of an onion 🧅

Moonshot and Alibaba alone shipped 14 open models in 559 days. One every 43. In 2026, one in every 27 days (see chart below)

So essentially, you're on a training treadmill after each model release

In short: the cure may be worse than the disease

Tokenomics is tricky

Enjoy The Weeknd 🎧 | 20 comments on LinkedIn