Token Factories with SOE Funding: Value Depends on Cost-per-kWh Metrics

Let's lay out some data first. Compute cluster utilization has been pushed from an industry average of 50% to over 90%, Token production capacity increased by 2-3x with equivalent hardware, comprehensive electricity costs dropped by 20%. This round raised hundreds of millions of RMB, led by CRRC Capital, with China Mobile's fund having previously provided an exclusive investment at the hundred-million-RMB level. The actual controller, Liu Haifeng, holds about 64.84%. Putting these numbers together, it shows CAS Brain (Zhongke Leinao) has bundled compute, power, and model output into a single business model.

It's now called a "Token Factory," where the core question becomes: How many effective Tokens can one kilowatt-hour produce?

In the short term, for companies taking money from central SOEs and telecom operators, the first reaction should be to look at who is paying the bill. The most awkward part of AI infra is that model companies, cloud vendors, and intelligent computing centers all claim their compute is cheap, but when customers actually run things, costs are still influenced by factors beyond just GPUs. Over the past month, using APIs and open-source models, and recently trying AI agents in Docker, my most intuitive feeling is that repeated calls per task, context accumulation, and tool failure retries mean the final bill gets stuck more on scheduling and engineering waste than raw compute. If an integrated compute-power platform can truly raise utilization from 50% to over 90%, the savings include not just electricity, but also depreciation of idle cards, server room cooling, network scheduling, and ops manpower. This math is very attractive to central/state-owned enterprises, local intelligent computing centers, and energy groups.

But we need to throw some cold water on the short term too. The 2-3x Token capacity increase and 20% drop in electricity costs are quantitative metrics given by the vendor when launching the platform. Whether these land in real business depends on the workload. Short chats, long documents, code agents, video understanding, multi-turn tool calls—the token consumption structures are completely different. Cache hit rates, whether context compression is implemented, if the model supports parallel decoding, and whether network latency and storage throughput can keep up—all affect the final numbers. When writing reviews, I hate seeing pretty metrics leading directly to conclusions because good benchmark scores don't mean daily usability is smooth. If the Token Factory only talks about standard test sets, customers might find that the electricity saved is eaten back up by complex agents once they're inside.

Long-term, this is more interesting. In the past, when discussing AI costs, the default assumption was buying/renting cards and calling models. CAS Brain's path puts energy scheduling, compute clusters, and model services into one system, even treating "intelligence per kWh" as a metric. If this direction works, AI infrastructure will look more like grid, transportation, or manufacturing businesses. Central SOE industrial capital entering suggests they care about compute-power scheduling, local data retention, and integrating model services into industries like energy, transport, and manufacturing. With CRRC Capital, China Mobile, and local industrial funds pooling money, the signal is clear: they want a controllable, deliverable foundation that can enter the industrial side.

However, there are long-term risks. Once Token output becomes a standard metric, the industry easily shifts from "who can train models" to "who controls cheap power and cheap compute." This is good for application developers, as model call costs may continue to drop. But for infrastructure providers, it might be another barrier. Token definitions, metering controls, and valid intelligence value judgments need transparent rules. If this isn't transparent, it could become a new black box. Trying agents and sandboxes lately, I feel permissions, logs, and metering are becoming more important than capability itself. A system that runs tasks autonomously—if it only shows you results and not the process—you won't dare connect it no matter how low the price. If the Token Factory wants to sell "intelligence per kWh" in the future, it must first make the ledger clear, distinguishing effective output from failed retries, with defined standards for cache hits and network waste. Without an auditable ledger, so-called factories risk becoming concept packaging.

This funding round itself isn't rare; what's rare is pushing AI infra from compute leasing into the intersection of energy efficiency and model output. Short-term, it helps local intelligent computing centers and industrial clients solve utilization, electricity cost, and delivery pressure issues. Long-term, whether it can truly become a Token Factory depends on if those metrics hold up in complex tasks. Funding amounts and launch event numbers are just references; the value lies in whether one kilowatt-hour can stably and auditably convert into effective Tokens.