Wink Pings

Small teams are still waiting for custom models to get cheaper, while big companies are already paying first

Across more than 20 seed- and Series A-stage AI startups, most do not touch custom models. Once the three layers of cost—upfront, maintenance, and serving—are laid out, the math doesn't work out. As prices fall, simple tasks and complex long-horizon tasks will come in one after the other.

Currently, only a small subset of companies using AI are investing in custom models. On X, Su gave a range: mainly the leading large AI companies in their categories, such as Harvey, Cognition, and DoorDash. Many of them have bought custom post-training services from Applied Compute, Trajectory, Baseten, and Fireworks.

Custom models refer to post-training on open-weight models to tune quality and cost for one's own tasks. Post-training usually does not mean training a foundation model from scratch; a common approach is to continue training an open-weight model.

Things are much colder on the small-company side. Su spoke with more than 20 seed- and Series A-stage AI startups across multiple sectors, and most believe it isn't worth it. Breaking down the math: upfront requires curating data, building evaluations, and training the first custom model; maintenance requires updating as foundation models are upgraded; serving requires paying inference costs. There are two kinds of benefit: post-training a smaller open-weight model can reduce cost or latency, provided utilization is good enough; and quality can improve on specific tasks, most noticeably on complex, long-horizon tasks.

Most small teams feel that savings plus quality are not clearly greater than upfront plus maintenance plus serving. Companies buying post-training now are either chasing very large quality gains or have inference costs so high that savings can cover the cost.

X user “大v(没互关就是在cd)” added a variable in a reply: task stability. If workflows change every month, cheap training may not buy much useful life. Break-even requires calculating real request volume and also how often the evaluation set needs to be rebuilt.

Su's judgment is that these costs will fall. As AI training AI becomes more capable, upfront and maintenance will become more efficient; AC is building a self-distillation product, and Trajectory has launched self-serve; serving costs can be improved by intelligently pooling requests across GPUs to raise utilization. Once prices fall, seed- and Series A-stage teams will also enter. Simple tasks will first look at cost savings; complex long-horizon tasks can capture both quality and cost, making both sides worthwhile. Harvey and Cognition got there first because they can afford it. Everyone else waits for the price.

The barrier to custom models is ongoing maintenance and utilization; training once is only the entry ticket.

Image from the original post:
image

The original post is here.

发布时间: 2026-11-18 09:26