Lenovo Sets Up Zhangjiakou Base to Ease Corporate AI ROI Concerns

Stock News11-10

After nearly four years of rapid large-model development, enterprise software and corporate intelligent transformation are encountering a "productivity stagnation": almost every mid-to-large enterprise has built a large-model demonstration demo, yet very few have truly embedded it into core business processes and let AI take over actual production tasks.

The root cause holding back technology deployment is not that model parameter scales are insufficient, but that enterprises face a dual anxiety before the commercialization ledger — chaotic scheduling of the computing foundation leads to idle and wasted hardware, while the invocation side frequently encounters unpredictable "Token assassins" and hallucination risks.

Enterprises can neither clearly calculate how much cost reduction and efficiency improvement large models can bring, nor afford the uncontrollable marginal cost of inference. On November 10, at the Lenovo Wanquan Hybrid Token Factory launch event held in Zhangjiakou, Hebei, this problem troubling the industry was brought into the open.

Lenovo Group Ltd. (00992) did not continue the hardware manufacturer's traditional emphasis on chip card counts or purely theoretical floating-point computing power (PFLOPS). Instead, it fully converted the entire system into industrial product indicators aligned with corporate financial calculations: comprehensive cost per million tokens optimized by 70%, token output per second increased 5-fold, and token yield rate exceeding 99.9%. The delivery logic of large-model services is being pushed into a standardized framework of a "manufacturing assembly line."

Enterprise AI Trapped in the Financial Ledger

In the current B2B market, "+AI" is easy, but crossing over to true "AI+" is extremely difficult. So-called "+AI" treats ready-made general large models as plug-in tools, dabbling only in certain lightweight auxiliary scenarios such as weekly report generation and preliminary code completion; "AI+," by contrast, uses intelligent capabilities to reconstruct the underlying business logic of R&D, production, supply chains, and customer service.

This leap has yet to occur on a large scale, and the core reason is the natural disconnect between the current computing delivery model and enterprises' strict financial budgets. On one hand is the "heavy-asset sunk cost" of computing procurement. To support possible business peaks, some enterprises blindly purchase high-performance computing cards or exclusive computing pools, but during daily troughs, high cabinet rental and electricity expenses continue to devour profits, leaving hardware resources idle for long periods.

On the other hand is the "black-box cost" of public cloud invocation. General APIs are billed by invocation volume, and for complex business chains such as long-text inference and multimodal image-video generation, Token consumption rises exponentially, with invocation fees often spiraling out of control when business scales up, eroding enterprises' already thin profit margins.

Even more fatal is uncertainty. Unlike the deterministic code output of traditional IT systems, large models are inherently probabilistic. Once a model outputs low-quality tokens, logical hallucinations, or triggers compliance red lines in a core business chain, enterprises must not only pay a premium for ineffective computation but also spend several times the business labor to correct and clean it up.

If large-model deployment cannot solve the predictability of costs and the certainty of delivery results, enterprises' investment in AI will forever remain at the "innovation trial-and-error" stage and cannot form a sustainable positive ROI cycle.

Putting "Probabilistic Experiments" into "Industrial Standardized Production Lines"

The path to breaking the deadlock that Lenovo offered at this launch event is to directly dismantle the traditional IT hardware-centric delivery dimension and abstract large-model computing into a standard "manufacturing factory." In the first mother factory landed in Zhangjiakou, its physical foundation is a 15,000P heterogeneous computing cluster and an approximately 100MW intelligent computing center, but Lenovo's factory acceptance criteria are: annual token output exceeding 20 quadrillion, peak output of 200 million tokens per second, and daily support for 100,000 concurrent users. Computing power is restored to "industrial raw material," while the "finished product" enterprises truly need is high-quality business Tokens.

Under this logic, Lenovo offered two rather striking industrial delivery benchmarks: 70% cost optimization and 99.9% yield rate. The 70% comprehensive cost optimization addresses the biggest threshold for enterprise large-model deployment — the initially high single-inference cost. Through heterogeneous computing management, lightweight hybrid model routing, and direct green power connection with local PUE below 1.1 in Zhangjiakou, the factory carried out three-level optimization from underlying energy and hardware scheduling to upper-layer inference, substantially lowering the comprehensive cost per million tokens. This reduction directly touches the ROI break-even point for most traditional industries to begin large-scale AI substitution.

The 99.9% token yield rate marks the first systematic introduction of the manufacturing core indicator "yield rate" into the measurement of large-model output. Large models are not frightening; what is frightening is an excessively high "defect rate" that stalls downstream production lines. The so-called "token yield rate" means setting up a filter at the inference hub level, immediately removing low-quality, hallucinatory, non-compliant, and ineffective computational content at the output end, ensuring that Tokens delivered to business systems have industrial-grade usability.

From competing over peak computing power to competing over "yield rate" and "factory cost," the competition rules of AI infrastructure have been pulled back from mystical algorithm contests to the lean management logic of classic manufacturing.

"Three-Layer Powertrain": Heterogeneous Mixed Deployment and Assembly-Line Configuration

Driving large-model operating costs down to 30% of the original while maintaining a high yield rate does not rely simply on hardware stacking, but on a complete powertrain architecture similar to automobile manufacturing. In the system Lenovo built, enterprise intelligence is deconstructed into a three-layer system: the base engine (Wentian AI infrastructure), the central transmission (Wanquan Hybrid Token Factory and xCloud intelligent cloud), and the terminal drive system (Qingtian and Baiying solutions).

The traditional intelligent computing center's biggest pain point is the ecological fragmentation among different chip vendors and computing architectures, making heterogeneous computing difficult to coordinate and schedule. The Wentian system uses a unified heterogeneous scheduling platform to let chip resources of different architectures form a combined computing force, and official data show its comprehensive computing efficiency can increase by more than 30%.

The central system serves as the scheduling hub (Token Hub) of the entire production line. Facing specific enterprise business instructions, it no longer mechanically throws requests directly to expensive hundred-billion-level large models, but dynamically routes them according to task complexity to edge small models, dedicated fine-tuned models, or private/public clusters, achieving inference acceleration of more than 5-fold.

The terminal drive system injects cleaned and tested high-quality tokens into specific business agents, directly driving real tasks such as R&D code writing, supply chain demand forecasting, intelligent customer service, and industrial inspection. The essence of this architecture is to transform large models from an "expensive general superbrain" into a "combination of specialized workstations" on an assembly line, where different workstations match chips and models of different costs, thereby achieving the highest efficiency of unit computing output globally.

The Commercial Payback Ledger in Vertically Heavy Scenarios

Whether the assembly-line logic can run through must ultimately be tested by output results in heavily vertical industrial scenarios. In the cultural media and entertainment going-global sector, the hybrid token factory for film and micro-short dramas landed by Lenovo and Hunan Broadcasting in Changsha provides a highly representative sample. Film and television production, especially today's booming micro-short dramas, generally faces long creative cycles, expensive post-production, and a lack of industrial standards for multimodal generation.

According to disclosed cooperation plans, the token factory brings script creation, storyboard design, video generation, and final finished-film output into production-line operations through hybrid routing and inference acceleration of full-modal tokens including text, images, video, and rendering. Quantitative results show: content production costs reduced by more than 50%, production cycles shortened by 60%, monthly finished-film output reaching more than 1,200 hours, and normal support for online collaboration by more than 10,000 creators.

From the quantitative production of micro-short drama content to the transformation of heavy industry in North China and the reconstruction of advanced manufacturing processes, when complex AI generation can be broken down into quantifiable token assembly lines with calculable cycles, enterprise decision-makers will truly dare to shift budgets from traditional IT maintenance to deep intelligent reconstruction.

The large-model industry has passed the first stage where "concepts, rankings, and card counts" determined victory, and is irreversibly sliding into the second half where "unit delivery cost, stable output efficiency, and ROI realization" are the main battlefield. For anxious enterprises, large models never needed to be enshrined on a technological altar; only when computing power is truly die-cast into standardized, low-cost, high-yield industrial assembly lines does the flywheel of enterprise intelligentization truly begin to turn.

Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.

Comments

We need your insight to fill this gap
Leave a comment