Author | Wei Rong Editor | Bao Yonggang
If different hardware solutions from various vendors deliver the same 1 million Tokens, which one is actually more cost-effective?
This question likely cannot be answered solely by looking at a GPU price list.
In the past, customers rented GPUs based on quantity and duration, making bills relatively clear. Today, an increasing number of customers are purchasing Token services directly. Hardware prices, model execution speed, and actual Token output all impact the final cost.
To make matters more complex, calculating costs is only the first step. Model adaptation, resource scheduling, and inference optimization also determine how many Tokens a batch of chips can ultimately deliver.
As the compute business shifts from renting GPUs to delivering Tokens, who can accurately calculate resource inputs, improve delivery efficiency, and transform this capability into a sustainable business?
For independent intelligent computing cloud companies like DataCanvas, establishing a unified metering and settlement system across different chips, and then improving delivery efficiency through proprietary scheduling systems, is becoming a key competitive advantage against major cloud providers and traditional compute rental firms.
From GPU Card-Hours to Token Delivery: The Compute Business Needs a New Settlement Layer
Data from the National Data Administration shows that China's daily average Token call volume exceeded 130 trillion in March 2026, representing a growth of over a thousand times compared to 100 billion in early 2024.
While demand has grown rapidly, some intelligent computing centers are struggling with how to absorb their compute capacity.
In the past, large model training was a significant source of demand in the compute rental market. Customers needed continuous use of large-scale clusters, long-term rental of resources ranging from thousands to tens of thousands of cards, allowing them to secure relatively stable supply and pricing.
Today, an increasing share of compute demand comes from inference and AI applications. Customers may only need to deploy a single model or elastically call Token services based on business traffic. For these customers, bearing fixed resource costs while handling self-maintenance is less attractive. In contrast, flexibly purchasing compute services based on actual business volume reduces idle costs and eliminates the hassle of managing hardware.
As transaction methods change, the calculations required by compute operators have also changed. Previously, billing was based on GPU card-hours, recording how much resources were rented and for how long. Now, when customers purchase Token services, operators must calculate how many Tokens these resources actually delivered and what the cost per million Tokens is.
This calculation is particularly difficult in a market where multiple types of hardware coexist.
IDC data indicates that in the first half of 2026, shipments of AI accelerators for data centers in China reached approximately 2.008 million units. Domestic manufacturers held the majority share, while overseas manufacturers accounted for about one-third.
Different hardware architectures result in varying computational performance and Token production efficiency. Comparing only hardware rental prices makes it difficult to judge which solution has a lower actual delivery cost.
The 2026 "White Paper on the Development of China's Token Factories" also notes that metering and billing rules for token factories have not yet formed an industry consensus, and there is a lack of unified capacity benchmarking and cost estimation methods across different chips.
This is precisely the problem DataCanvas aims to solve. Their approach is to first establish a unified compute metering standard for different chips.
Consequently, DataCanvas introduced the concept of "One Unit of Compute." DataCanvas defines1 DCU as running 312 TFLOPS of FP16 theoretical peak compute power for one hour. Different chips can convert their resource consumption according to this standard.
With a unified metering unit, DataCanvas further proposed the "Dual Ledger of DCU and Token": the DCU ledger records the amount of compute resources invested, while the Token ledger records the amount of Tokens finally delivered.
Simplistically, completing the delivery task of 1 million Tokens using different chips may consume varying amounts of DCUs. By combining respective resource costs, DataCanvas can compare the unit Token delivery cost of different solutions, providing a basis for resource allocation and service quoting.
An intelligent computing cloud company managing multiple compute centers simultaneously, using chips from different vendors, and serving customers with diverse needs, must continuously compare the usage costs and delivery benefits of various resources.
The Dual Ledger of DCU and Token enables DataCanvas to manage these businesses under a unified metering unit, comparing the operational performance of different chips and resource pools, providing the foundational conditions for scaled operations.
Calculating Clearly Does Not Mean Using Well: Heterogeneous Compute Must Also Cross the Efficiency Barrier
With a unified metering standard, DataCanvas can compare the input costs of different compute resources. However, to truly reduce Token delivery costs, these compute resources must operate more efficiently.
Especially in the stage of large model inference, compute efficiency increasingly depends on the scheduling and optimization capabilities of software systems.
When using the same GPU to run a large model, not all time during answer generation is spent on effective computation. Differences in computational load, input length, and output speed across requests mean that data reading, computation waiting, and redundant processing can slow down actual execution speed.
These losses eventually reflect in Token delivery costs. If software can reduce waiting and redundant computation, the same hardware can complete more tasks.
DataCanvas integrates compute scheduling and model execution optimization within the same system via Alaya NeW OS.
Alaya NeW OS covers stages such as compute scheduling, model training, model inference, and data processing. Through a Serverless architecture, Alaya NeW OS separates customer computing tasks from underlying hardware management and can uniformly schedule various mainstream domestic and international hardware architectures.
In the inference phase, DataCanvas further employs technologies such as PD separation, MTP acceleration, and KV cache reuse to improve the actual computational efficiency of chips.
Specifically, PD separation runs the input processing and answer generation phases separately, allowing different phases to use more suitable compute resources; MTP acceleration improves generation speed in certain scenarios by predicting multiple Tokens at once; KV cache reuse reduces redundant computation of existing context, saving compute and storage overhead.
These technologies are not simply stacked functions but respectively address issues of resource allocation, generation speed, and redundant computation during inference. From allocating computing tasks to optimizing Token generation, Alaya NeW OS handles efficiency issues across multiple stages simultaneously, enabling DataCanvas to deliver more Tokens with the same hardware resources in certain inference scenarios.
This software capability has also received recognition from third-party evaluations. Alaya NeW OS is the first AI system in China to pass evaluations in four major domains by the China Academy of Information and Communications Technology (CAICT): compute scheduling, model training, model inference, and data processing.
Coverage of these four domains means that DataCanvas can handle various tasks ranging from hardware resource allocation to model execution and data processing on the same technical foundation. Customers do not need to seek separate technical solutions for each stage, reducing deployment and system integration complexity.
In May 2026, Fang Lei, Founder and Chairman of DataCanvas, stated in an interview that after a manufacturing enterprise connected to DataCanvas's Token Factory, the comprehensive compute cost of its AI quality inspection system decreased by 55%.
In another project for an industrial software enterprise, DataCanvas used distributed inference clusters and heterogeneous compute scheduling to increase inference efficiency by 12 times and reduce hardware procurement costs by 54%.
For DataCanvas, the value of Alaya NeW OS lies in the fact that this technology can serve more projects and continue to optimize.
For example, after adapting to one type of chip, new resources with the same architecture can leverage existing achievements; mature inference optimization solutions can also be promoted to customers with similar needs. DataCanvas reduces duplicate development costs and spreads software investment across more businesses.
This reusable software capability allows DataCanvas to create additional service value beyond hardware resources through scheduling efficiency and delivery quality, gradually forming a technical premium distinct from traditional compute rentals.
Operational data already reflects changes in resource usage efficiency. Reports indicate that DataCanvas's compute resource utilization rate increased from 56.0% in 2025 to 79.6% in the first half of 2026, a rise of 23.6 percentage points.
This provides a technical foundation for DataCanvas to establish an independent market position between major cloud providers and traditional compute rental firms.
Ecosystem Neutrality Opens the Market: Capital Efficiency Supports Growth
By 2026, the global NeoCloud sector has moved past the stage where "holding hardware alone guaranteed profit."
The intelligent computing cloud market continues to grow rapidly. IDC data shows that the domestic intelligent computing cloud infrastructure market size reached CNY 48.67 billion in 2025, a year-on-year increase of 128%. Meanwhile, model call demand surged by approximately 16 times, bringing sustained market space for intelligent computing cloud services.
However, growth itself no longer automatically equates to a good business. The yardstick for capital assessment of top targets is changing: shallow indicators such as compute scale and hardware benchmarks are gradually losing effectiveness. Stable delivery capability, Total Cost of Ownership (TCO), Token service efficiency, and the engineering maturity of AI infrastructure are becoming important factors in measuring long-term corporate value.
Citi Research, while optimistic about the evolution of NeoCloud platforms from compute suppliers to complete AI platforms, pointed out that capital markets are beginning to pay more attention to financing costs, delivery capabilities, segment growth rates, and technical reserves. Some investment institutions believe that vendors with long-term investment value need not only technical capabilities but also stable customer structures, healthy financial models, and standardized enterprise-grade delivery capabilities.
These requirements ultimately fall back on the actual operation of compute centers. Whether hardware can be put into use on time and whether compute can be continuously delivered will affect corporate operational performance.
Among the many factors affecting NeoCloud operational efficiency, electricity is becoming an increasingly critical link.
The new bottleneck for NeoClouds is no longer just about acquiring hardware, but whether electricity can be stably supplied and how long it takes to connect to the grid. It is not uncommon for a large compute center to wait three to five years from decision-making to actual power connection.
Queuing for the grid is harder to expedite than hardware lead times.
Therefore, power supply, networking, and disaster recovery—operational aspects that once seemed insignificant—determine whether the theoretical performance of hardware can be converted into stable production compute.
To address this core challenge of electricity, DataCanvas incorporates compute-power coordination into compute operations. By coordinating compute and power scheduling, it can reduce the electricity cost per Token by more than 30%.
In terms of compute resources, DataCanvas ensures stable supply through3 to 8-year long-term irrevocable agreements. This model emphasizes "control rather than ownership." Even without holding all assets, DataCanvas can rely on its software platform to uniformly schedule resources, transforming them into schedulable, mobile, and continuously deliverable compute services.
While stable resource supply solves the question of where compute comes from, how these resources are operated and what services are provided to customers determines the business model of NeoCloud enterprises.
Global NeoCloud enterprises are exploring different development paths. Some focus on compute resource supply, others extend into model services, and some further integrate AI engineering and data center operational capabilities. DataCanvas has chosen a comprehensive operational route, transforming compute of different architectures into sustainably deliverable services through heterogeneous scheduling, training-inference collaboration, and underlying resource management.
Similar explorations have appeared in the overseas NeoCloud market. In July 2026, leading vendor Nebius launched a new infrastructure cooperation model where partners bear infrastructure and hardware investments, while Nebius handles system architecture, cloud platform software, and customer services, integrating cooperative resources into its own AI cloud platform.
Nebius believes that this asset-light cooperation model can expand service scale with minimal new capital investment and create higher-margin revenue sources.
DataCanvas also values the combination of software operational capabilities and capital efficiency. By ensuring compute supply through long-term resource agreements and relying on Alaya NeW OS for unified scheduling, model execution, and service delivery, DataCanvas can apply software capabilities to more compute centers and customer projects, improving the output efficiency of existing resources while expanding business scale.
For independent NeoCloud enterprises, growth cannot rely solely on constantly increasing hardware investment; it must also achieve higher output from operating existing resources. Hardware investment determines compute scale, while software capabilities determine the commercial value of resources. This is also the key to DataCanvas improving capital efficiency.
However, whether software operations can truly improve capital efficiency ultimately returns to business scale and operational returns.
At the end of September, DataCanvas officially submitted its listing application to the HKEX. The prospectus showed that in the first half of 2026, DataCanvas's intelligent computing cloud service revenue reached CNY 154 million, approximately 5.2 times that of the same period last year; the number of customers increased from 42 to 84.
More noteworthy is the improvement in profitability. In the first half of 2026, excluding non-cash impacts such as share-based payments and preferred stock valuation changes, DataCanvas's adjusted net loss (Non-IFRS) narrowed by 75.2% year-on-year. Simultaneously, the gross margin of its core intelligent computing cloud services turned positive. As revenue scale expands, fixed costs incurred in earlier periods are effectively diluted, demonstrating the emerging scale effects and profit elasticity of intelligent computing cloud services.
For NeoCloud enterprises securing resource supply through long-term cooperation, the viability of the business model hinges on whether business revenue can continuously cover resource and service costs. The simultaneous occurrence of revenue growth and gross margin improvement for DataCanvas indicates that business expansion has begun to translate into operational returns, providing preliminary validation for its independent NeoCloud model.
Major cloud providers possess complete ecosystems, while underlying rental firms provide hardware resources. DataCanvas is forming an operational system spanning resource supply, software scheduling, and cost control that does not rely on a single hardware ecosystem.
As models, chips, and industry applications further differentiate, this ability to collaborate across ecosystems and continuously optimize compute operations will become an important foundation for independent NeoCloud enterprises to expand business scale and build long-term competitive advantages.