Old valuation debates in the history of large models

The most valuable insight from this article is that early Chinese LLMs focused on scaling up size but long failed to answer "what should it become?" Investors feel uncomfortable hearing this because many projects' valuation logic rests on a default premise: scarcity of capability + massive parameters = valuable. When demand isn't defined, scarcity is just self-congratulation from the supply side.

I see many AI and robotics projects annually. Founders often put model capabilities on the first slide. Now, I ask first: What is the valuation logic? Where is the ceiling for this track? History has demonstrated: Small models failing means tweaking parameters and rerunning; failures at the hundred-billion parameter level involve real GPUs, electricity bills, and financing pressure. Compute power being pushed from background to foreground isn't rhetoric; it's accounting.

But what truly changes the valuation logic isn't getting bigger, but models becoming replicable, compressible, and open. The GLM team opened weights, code, and training logs, stuffing 130 billion parameters into cheaper machines. Open weights turn "I'm smarter than you" into a public asset; low-cost deployment turns "I can afford to be expensive" into a cost issue. The model layer shifts from a scarce entry ticket to an efficiency competition.

After it gets cheap, where should money look?

"Models getting cheaper" is often framed as good news. Yes, but not entirely. Cheapness lowers the barrier for customers to experiment, but also lowers gross margin protection. No matter how large API call volumes are, without scenario stickiness, it's just a traffic business. No matter how heavy private deployment is, without subsequent O&M and process transformation, it's a one-time sale. Customers don't buy parameters; they buy results. Enterprises won't pay long-term for model capabilities; they'll only pay for hiring fewer people, making fewer errors, waiting less for tickets, and stopping lines less often.

So I focus on three lines. Task Data: Can we access data that others can't but customers are willing to give? Process Closed Loop: Can the model embed into legacy systems, changing job manuals, approval chains, and on-site operations? Contextual Barriers: As models homogenize, whoever understands the user's long-term state better has more pricing power. I recently wrote about mobile AI memory; looking at edge-side memory and evaluation benchmarks now, I feel competition is shifting from "whose model is stronger" to "who remembers better and uses it more stably."

Agents also warrant caution. They sound like gateways, but if they're just wrapping models into automated clicks without industry know-how, fault tolerance, or cost accounting, they're just demos. Truly investable Agents often aren't flashy; they look like dirty work: connecting to ERPs, reading tickets, verifying fields, handling exceptions, turning ten minutes of human effort into ten seconds in the system, while being auditable, accountable, and repurchasable.

Looking back at this chapter, the technology curve first rewards the bold, then punishes those who only tell big stories. Hundred-billion, trillion, 130-billion, open weights, low-cost deployment—all ultimately come down to one question: Who is willing to pay for it, and will they still pay next year? The true valuation inflection point in the brief history of LLMs isn't larger parameters, but after capabilities become cheap, who can fit the model into a billable, repurchasable, replaceable business closed loop.