Since the beginning of this week, we have been running an experiment. We put GLM 5.2, an open-weight model, on two B300 nodes in our own infrastructure and started measuring whether it can replace the expensive Anthropic licenses we are currently paying for. The early signal is interesting.
One internal data point made the curve concrete. A colleague ran a comparison on a fairly large story this week. Opus 4.8 via the Anthropic API: 40 dollars, 90 minutes. The same story on our self-hosted GLM 5.2: about 30 minutes, roughly 4 dollars in GPU cost - on hardware that other workloads could be sharing in parallel without anyone really noticing. That is a single sample, it ignores the human cost of running the infrastructure, and it is not a benchmark. But it is also an hour of developer time saved and an order-of-magnitude cost difference on one task. Even rented, even early, the gap is already visible.
What I did not expect, is the secondary realization. The hardware needed to run models locally is extremely expensive today - but that will change over time. The category of "tasks a model that fits on affordable local hardware can do well enough" is already non-trivial today and will only grow. Code completion. Routine summarization. Boilerplate generation. The things we currently pay a cloud vendor for on a per-token basis are precisely the things local hardware will eat first.
The vendor is probably already aware of this curve. I bet that's why the contract lengths currently on offer are not generous: Three-year terms lock in revenue against an exit ramp the vendor can see clearly. That's an option transfer from buyer to seller, dressed up as a 5% discount.
Most enterprise procurement is still negotiating on price per token, volume commits, and uplift caps. That is the wrong axis. The right axis should be in contract length and termination rights. A two percent improvement on a three-year deal is worse than the same price on a twelve-month deal you can walk away from when the local model is good enough.
The vendors offering the friendliest discounts in exchange for longer terms are not being friendly. They are pricing the option you are giving up. The B300s in our rack are one signal of many that the option is worth more than they want you to think.
Thoughts? Find me on Bluesky.