Luoge Data CTO Liu Xuepeng: Edge AI Customers Only Pay for the 'Last 100 Meters'
Liu Xuepeng, CTO of data-engineering firm Luoge Data, proposed at an ecosystem conference hosted by edge-AI chip company Houmo Intelligence in Beijing: models solve problems a kilometer away, compute solves problems five hundred meters away, but customers only pay for the last 100 meters. Privacy concerns in elderly care, classified-data red lines at Chinese central state-owned enterprises, and the impossibility of sending fiscal data to the cloud together show that the commercialization barrier for edge AI is not model capability but data trust and delivery quality.
Liu Xuepeng, CTO of Luoge Data, offered a judgment during this roundtable: 'Models solve problems a kilometer away; compute solves problems five hundred meters away; but customers are only willing to pay for the last 100 meters.' This is not an abstract metaphor. This year his team deployed solutions for numerous Chinese central state-owned enterprises and listed companies, and he observed a recurring pattern—enterprises do not doubt AI's usefulness; they do not know where to place AI in order to dare to use it.
The fiscal and tax scenario is the most typical. Large volumes of invoice, financial-report and order data inherently cannot be sent to cloud models for inference; they must remain proprietary. Liu Xuepeng summarized this need as 'private-domain exclusivity': AI will become an enterprise's digital expert or digital employee, and this kind of 'employee' that touches core business must be hired into the company's internal systems rather than outsourced to an external cloud service. Once a model begins reading your most critical operational data, the data boundary is no longer an add-on option but a prerequisite for commercial deployment.
The security threshold is even harder in party-government and classified-data domains. Yao Ling, deputy general manager of Kylin Software, serves customers in key positions within these sectors. Her judgment is blunt: as long as an AI tool is deployed on the internet, classified data must never touch it. She gave a concrete example—you think renaming a file makes it safe? A large model can perfectly piece together fragments uploaded today and tomorrow into a complete classified document. So for customers in sensitive positions, local security is non-negotiable.
Huang Junshi, founder and CEO of Liaoqu Intelligent Technology, told a more everyday version. His team deployed a 'Digital Family Member' system at a Shanghai elderly-care facility, where it chats with residents, reminds them to take medication, and helps guard against fraud. What an elderly person says each day and how their health is doing are extremely sensitive information that cannot casually be uploaded to the cloud; the system must run around the clock, yet the device cannot hum. At a site like this, privacy, compute, power consumption and noise are no longer technical metrics—they are the conditions under which residents are willing to use the system.
Chen Zhanling, CMO of Yinpu Computing, added a frequently overlooked judgment from the integrator's perspective: saving tokens is not the customer's primary pain point—customer experience is. How long until the first character appears, whether the machine is quiet, whether the system stutters, whether it can run continuously and stably—these seemingly trivial issues determine whether a product is used long-term or shoved into a cabinet after the demo ends. He mentioned that his team even considers adding liquid-cooling mechanisms to reduce noise.
But the 'last 100 meters' is difficult not merely because of cramming a large model into a small device. Liu Xuepeng listed three engineering barriers: data integration—a model entering an enterprise is like a consultant arriving on site; without business data, nothing can be done; interaction design—enterprise users find it hard to accept pure conversational logic, so AI must be seamlessly embedded into workflows they already know; stability—a model is fundamentally probabilistic, but consolidating financial statements requires extremely precise matching calculations, demanding extensive engineering techniques as a safety net. He summarized: in the past, building software meant delivering a system; deploying AI today is more like training an omnipotent genius into a diligent, safe and trustworthy dedicated expert.
Chen Zhongming, chief engineer of AI systems at Houmo Intelligence, provided the underlying constraint from the chip level: compute, bandwidth and power consumption form an 'impossible triangle.' What Houmo has done best so far is solve the power-consumption problem; the next goal is to break this triangle. He outlined two tracks—the M series pushes compute upward, and a separate line uses more advanced process and packaging technology to increase bandwidth.
It should be noted that this roundtable took place at Houmo Intelligence's own ecosystem-partner conference, and all five participants are Houmo's ecosystem partners or employees; their judgments serve the commercial promotion of edge-AI infrastructure. Liu Xuepeng's prediction at the close of the discussion—that in China, AI will truly land with 'edge as primary, cloud as supplementary,' where the edge is a person's or enterprise's dedicated brain while the cloud handles knowledge flow—is his personal business judgment. The original article itself notes: 'Whether this judgment is right will become clear in time.'
Nevertheless, the 'last 100 meters' framework itself carries transferable value beyond a single vendor. Any team trying to move AI from demo to production will encounter the same problem: customers do not pay for your model parameters—they pay for data that stays home, devices that do not make noise, and systems that do not cause trouble. As AI moves from 'able to chat' to 'able to work,' it must ingest your most core, most closely held data. At that point, where you place compute is no longer merely a technology-selection question but a trust question.