At present, the domestic AI compute market is showing a deeply divided reality: compute centers are being built continuously across the country, yet a large amount of assets sit idle, while at the same time large model companies are struggling with a severe shortage of high-end compute power.
During the 2026 Sustainable Global Leaders Conference, Biren Tech (BIREN TECH) Senior Vice President of Strategic Planning Wei Ming attended the forum "Embracing the Token New Economy: Activating Existing Compute Power and Solidifying the Foundation of the AI Industry" and shared his latest insights on revitalizing existing compute power.
In a post-conference conversation, Wei Ming pointed out that although the domestic chip compute industry does face "structural idling," demand for high-end compute is equally "real and present." In his view, as the open-source ecosystem and AI Agents continue to mature, software-hardware collaboration is opening up room for domestic chips to grow. The Token economy breaks away from the traditional cloud rental logic of selling resources by monthly or annual subscription, and through refined billing it gives rise to a diverse Token factory business model spanning big companies and small teams, opening a brand-new commercialization window for domestic compute chips.
The following is the full text of the interview, with some abridgements.
"Compute idling is a fact, but high-end compute is also in short supply"
Question: As an industry insider, do you think the current contradiction of idle compute stems from compute resources not being fully utilized, or from weak effective demand for AI applications?
Wei Ming: Biren Tech (BIREN TECH) has been established for seven years, and domestic chips have gone through a development stage from "usable" to "easy to use." Idle compute is not simply weak demand; there are multiple practical reasons behind it. In the early days, compute center construction across regions lacked coordinated planning. After compute centers were built, operations and maintenance capabilities failed to keep pace, and insufficient software-hardware collaboration constrained the release of compute efficiency. Driven by factors such as Moore's Law, chips iterate rapidly, and early products had shortcomings in compute density, interconnect speed and memory bandwidth, making it difficult to meet continuously upgrading application needs, which to some extent also caused compute idling. However, positive changes have already emerged. In the past, domestic chips needed to adapt to many models one by one, which took a long time; now the model ecosystem is becoming more focused. The rise of AI Agents has further compressed the time cost of chip adaptation. At present, the market presents a very divided situation: on one side there is idle compute, and on the other side high-end compute is extremely scarce. The number one problem facing large model companies at this stage is insufficient supply of high-end compute. Fortunately, the capabilities of domestic high-end chips have risen and can take on this kind of business. There are also technical solutions, relying on approaches such as PD separation to pair domestic cards with strong compute but weaker VRAM with high-VRAM hardware under coordinated scheduling, thereby revitalizing existing idle compute. The problem of idle compute is expected to improve gradually, and we are very confident about the large-scale deployment of domestic chips. From the perspective of the industry as a whole, the hard power of domestic chips continues to improve, completing an evolution from nothing to something and from weak to strong. The software ecosystem is also being repaired, open-source models reduce adaptation workload, and the supporting software stack is becoming increasingly mature. AI Coding tools further shorten the engineering deployment cycle, feeding the saved costs back into AI services and completing a commercial closed loop. The real surge in demand is supported by actual data: some large model companies saw API call volume and revenue in just one or two months exceed the entire previous year, and paying users continue to grow. Strong market demand brings a valuable development window for domestic chips, the domestic industry chain continues to grow, and the compute resource system will become further standardized.
Question: Domestic chips are already close to or even ahead of foreign counterparts in inference scenarios, so where do we mainly lag in training scenarios?
Wei Ming: We need to look at them separately. In inference scenarios, domestic chips are already close to foreign levels, and in some indicators and cost-effectiveness they can even surpass them. But the gap in training scenarios is still fairly obvious. Large model training requires high-end training cards, and limited by multiple conditions such as process technology and supply chain, domestic training card products still need time for continuous iteration to gradually narrow the gap.
"Hardware sets the ceiling, but the software ecosystem determines how much performance is released"
Question: You mentioned that the shortcomings of domestic chip performance also need to be narrowed through software-hardware integration and a better software ecosystem. How exactly will that be done?
Wei Ming: First, do software-hardware co-design at the chip design stage. From the very beginning of chip design, bring hardware and software into unified planning and balance the two from the source. Second, at the software level, embrace open-source communities and increase open-source efforts. With the power of the community, make software iteration faster. Third, build an advanced toolchain to improve software adaptation efficiency and respond more quickly to customers' actual needs. Fourth, work together with multiple partners to push things forward. Chipmakers cannot do it alone; we must coordinate with ecosystem partners such as server makers and large model companies. When new business opportunities appear, multiple parties collaborate to land solutions as quickly as possible.
Question: Software-hardware plus ecosystem collaboration aims at sustainable compute supply and the best balance of cost and efficiency. So what is the current cost-effectiveness of domestic chips compared with foreign ones?
Wei Ming: Cost-effectiveness is dynamic; different models and different application scenarios produce different results. Domestic chips have already achieved commercial deployment in many industries. When customers choose domestic products, on the one hand there is national policy guidance, and on the other hand domestic chips have already demonstrated cost-effectiveness with real commercial cases, which is very critical. Hardware design determines the upper limit of system performance, but the software ecosystem determines how much of the hardware performance you can release. One of our core tasks now is to polish the software ecosystem and fully unleash the potential of the hardware. The software ecosystems of overseas vendors were not built overnight; Nvidia's CUDA is the result of years of accumulation and continuous tuning by a huge number of developers. We also need to do the same solid work: refine the chip hardware well while also building a good software stack. Working with server and upstream and downstream partners to meet customer needs is something that requires time to accumulate. This year Biren Tech (BIREN TECH) will invest a great deal of energy in the ecosystem, build a developer community, provide toolkit downloads on our official website, and open a cloud environment so developers can directly test chip capabilities online. Through these methods, customers can understand and get started with our products faster and gradually narrow the gap with leading overseas companies.
"Traditional cloud services charge monthly or annually, while the Token model has finer granularity"
Question: Everyone is now talking about the Token economy model. What is the core difference between it and the traditional cloud rental model? The cloud itself is also elastic and billed on demand. Is Token just a repackaged new concept or is there a genuine essential difference?
Wei Ming: Traditional cloud rental sells hardware resources: servers, storage, machine usage time. In many cases, an entire machine is rented to customers by the month or by the year. After the machine is rented to you, it runs at full load during busy daytime business and sits idle at night, so a large amount of compute is wasted. The Token model sells AI output capability, billing by the number of input and output tokens, with finer granularity and greater flexibility. First, there is a difference in billing granularity. Traditional cloud rental finds it hard to rent for very short periods such as one or two minutes; most are monthly or annual subscriptions. Token billing is based on actual invocation volume, with clear pricing for how much input costs and how much output costs, and you pay for what you use. Second, it lowers the threshold for use. With traditional server rental, enterprises need IT staff to install drivers and debug environments, so the barrier to entry is high. With the Token/API model, you can directly call an interface and pay to get results, shielding away the underlying complex technical barriers. Third, tiered pricing becomes possible. Prices can be differentiated according to business latency and quality requirements. For example, if response speed requirements are extremely high, you pay a higher price; if you can accept slower or later-returned results, you can use a lower price. Behind this, compute cards of different performance levels are scheduled to provide differentiated business models for different customers.
Question: In the market there are Token factories run by big companies, and also Token factories built by small teams that buy machines themselves and deploy open-source models. What are the differences between these two business types? What do they mean for the chip industry?
Wei Ming: With many open-source models now available, big companies can build Token factories to sell services externally, and many small teams can also purchase a few machines, download open-source models and provide Token services externally. There are mainly three differences between the two. 1. Price: small teams often quote lower prices and attract customers through price. 2. Model iteration: the foundation is open-source models, but for subsequent model iteration and upgrades, big companies' official versions will continue to update and iterate; small teams may not be able to keep up with the latest version. 3. Compute scale and service quality: small teams have limited compute scale and may also cut corners on model performance to compress costs, resulting in uneven output quality. So the market will stratify: those pursuing stable, high-quality service can choose big companies' Token factories; those pursuing extreme cost-effectiveness can choose small team solutions. For us as a chipmaker this is favorable. There are more market customers, and regardless of whether service providers are large or small, they all need high-quality compute chips. We can provide hardware to different types of customers and help them realize business value.
Comments