A Sea Change in the Computing Industry: From “Stacking GPUs” to “Selling Tokens”

In the final week of July 2026, the computing industry is witnessing a profound business‑model transformation. Against a backdrop of persistently rising prices for high‑end computing power and severe supply shortages, leading players are moving beyond simple hardware rental. They are now exploring a brand‑new paradigm: the “Token Factory” model—shifting from selling hardware to selling services, and from billing by GPU hours to billing by the number of AI‑generated tokens delivered. This shift signals that the competitive battleground is moving from “how many GPUs you own” to “how much intelligence you can efficiently produce,” and the entire industry stands at the threshold of a transition from a resource‑based economy to a service‑based one.

1. Token Factory: A Paradigm Shift in Computing Business Models

A “Token Factory” refers to a computing service where providers no longer use the number of GPUs or computation hours as the primary billing unit, but instead anchor pricing on the number of tokens (the smallest units of text processed by a model) generated by AI models. This model is known as TaaS (Token‑as‑a‑Service). At its core, it abstracts computing power into a measurable, tradable unit of intelligence—much like electricity evolved from “renting a generator” to “paying per kilowatt‑hour.”

To date, more than a dozen listed companies—including Runjian Co., Ltd., Hongxin Electronics, Xingyun Technology, Chaoxun Technology, and Nanwei Software—have publicly announced that they are exploring or implementing Token Factory businesses. At the just‑concluded 2026 World Artificial Intelligence Conference (WAIC), the “Token Factory” became a buzzword at many computing‑firm booths, with numerous companies showcasing their token‑based billing platforms and scheduling systems; industry consensus is rapidly solidifying.

Xingyun Technology is a typical example of this transition. On the evening of July 29, Xingyun announced that its previously signed large‑scale computing services contract had not only been expanded and repriced, but also now included “fixed service fees related to token revenue.” Specifically, the contract with a leading long‑context large language model provider increased its computing service units from 128 to 256, pushing the total contract value from approximately RMB 1 billion to RMB 3.053 billion—a staggering 201% increase. At the same time, the monthly service fee per unit rose by 21% to 36%. The announcement specifically noted that the new token‑based fees would be settled in a tiered manner based on actual tokens generated, meaning that Xingyun’s revenue structure is shifting from pure resource leasing to “base compute fee plus performance‑based sharing,” greatly enhancing its business flexibility.

Another representative player, Runjian Co., unveiled the 2.0 version of its “Runjian Smart Computing Token Factory” during WAIC, claiming to have integrated over 50,000 heterogeneous GPUs and launched a developer package with token‑based billing, offering low‑barrier access for small and medium‑sized model teams. Hongxin Electronics, in partnership with several domestic chipmakers, has built a tokenised compute pool based on chips from Enflame and Cambricon, aiming to carve out a differentiated service capability on the path of domestic substitution.

2. The Price Surge: Market Realities of Supply‑Demand Imbalance

The emergence of the Token Factory model is rooted in a persistent shortage of high‑end computing power and continuously climbing prices, creating a seller’s market. This price wave has hit both consumer‑grade graphics cards and the rental market, with an intensity and duration rarely seen in recent years.

On the consumer side, the graphics card market is experiencing a veritable “price storm.” Following price‑hike notices from major GPU manufacturers (NVIDIA and AMD), brand factories have imposed full stock‑holding policies, suspending bulk shipments, leading to rapidly depleting channel inventories. Multiple merchants at Shenzhen’s Huaqiangbei electronics market reported that “prices have surged about 30% over the past two days.” The RTX 5070, which retailed for around RMB 4,500‑5,000 last year, is now selling for RMB 6,000‑6,500; the RTX 5080 has jumped from around RMB 8,000 to nearly RMB 12,000. “It’s basically a new price every day, with an average daily increase of RMB 500 to 1,000.” Some high‑end models have even become “priceless but unavailable,” with gamers and AI enthusiasts complaining bitterly while scalpers and hoarders thrive.

On the rental side, compute‑rental prices have also continued to rise. According to industry analytics, the combined spot and contract rental price for NVIDIA B200 GPUs has risen from roughly USD 4.0 per GPU‑hour in August 2025 to about USD 5.0‑5.2 per GPU‑hour in July 2026, a cumulative increase of 25‑30%. One‑year contract rates for H100 have also recovered from their late‑2025 lows, now stabilizing at around USD 2.8‑3.0 per GPU‑hour. Even in the domestic Chinese rental market, prices for compliant models like A800 and H800 have risen 15‑20% month‑on‑month, with long‑term contracts requiring queues months in advance.

The root cause of these price hikes is the “suction effect” of AI compute demand on upstream capacity. The three major memory makers—Samsung, SK Hynix, and Micron—have shifted 70‑90% of their advanced packaging and wafer capacity to higher‑margin HBM (High Bandwidth Memory). HBM yields significantly higher profits than consumer‑grade GDDR and is primarily supplied to data‑centre clients like NVIDIA and AMD. As manufacturers prioritise HBM orders, GDDR capacity for gaming graphics cards has shrunk sharply, driving up memory costs. These cost pressures have cascaded down the supply chain, ultimately inflating retail GPU prices. Additionally, the global AI “big‑model” race has entered a white‑hot phase, with top players scaling clusters from thousands to tens of thousands or even hundreds of thousands of GPUs, further exacerbating supply‑demand tensions.

3. From “Stacking Cards” to “Systems”: The New Logic of Competition

The rise of the Token Factory is inseparable from the overarching upgrade in competitive logic within the compute industry. The era of simply adding more GPUs to boost raw power—the “stacking cards” approach—is over; system‑level optimisation is now the decisive factor.

WAIC 2026 sent a clear signal: domestic Chinese computing has fully moved beyond the initial “stacking” stage and entered a new phase of system‑level competition in high‑speed interconnect, software stacks, and cluster scheduling. Huawei’s Ascend 950 supernode made its debut as a live system, featuring 1,024 Ascend NPU cards and, through its proprietary HCCS high‑speed interconnect, delivering 1 EFLOPS of FP8 compute power, with a cluster linear acceleration ratio exceeding 0.92—close to the theoretical optimum. Sugon showcased its all‑domestic 100,000‑card AI supercluster “Sugon 8000,” employing proprietary distributed storage and liquid cooling with a PUE as low as 1.08. Muxi Co. unveiled its Xijing S600 supernode, which supports horizontal scaling to ten‑thousand‑card clusters and is compatible with PyTorch 2.0 and above.

These demonstrations show that industry evaluation criteria are no longer limited to peak FLOPS or memory bandwidth of a single chip; they now extend to cross‑node communication latency, fault‑recovery time, automatic parallelisation strategies, compiler optimisation efficiency, and other system‑level capabilities. As a senior architect in the field put it: “In the future, as domestic GPUs flood the market, hardware supply will gradually become adequate. Leading compute companies are shifting their competitive focus to software‑system capabilities like scheduling and inference optimisation. Whoever can produce more tokens from the same hardware will hold the pricing power.”

This shift also aligns with the inherent demands of the “Token Factory”—token output efficiency depends not only on GPU floating‑point speed but also on the entire system’s memory bandwidth, network latency, scheduling algorithms, and model‑parallelism strategies. Therefore, improvements in system capabilities directly translate into token output per unit cost, forming the technical foundation for the business model’s viability.

4. From Selling Resources to Selling Services: A Reconstruction of Industry Value

At the heart of the Token Factory model is the elevation of computing power from a “resource” to a “service”—a transformation that is reconfiguring value distribution across the entire industry chain.

In the past few years, competitive advantage hinged on “owning GPUs.” Many investors treated GPUs as the new “digital real estate,” pouring capital into purchasing cards and building data centres, with even “compute speculators” hoarding inventory. This resource‑based model was straightforward but suffered from diminishing marginal returns and was vulnerable to hardware depreciation and technological obsolescence.

Today, leading firms are upgrading their service offerings from a one‑off “compute sale” to a more profitable and sticky TaaS model. What does this mean? Computing power is no longer just a pile of graphics cards; it is a complete production system spanning hardware drivers, middleware schedulers, model optimisation, and application interfaces. Tokens, as the universal semantic unit connecting underlying compute with upper‑layer applications, are evolving into the core “currency” for measuring AI productivity.

For downstream customers, token‑based billing lowers the barrier to entry. Developers need not worry about how many GPUs or which models are underneath—they simply pay for the actual tokens consumed, much like paying for water usage. This is particularly beneficial for small and medium‑sized enterprises and individual developers, allowing them to participate in AI innovation with more flexible costs. At the same time, token billing naturally ties to inference throughput, incentivising providers to continually optimise system efficiency, creating a virtuous flywheel: higher efficiency → lower cost → more customers → more data → faster optimisation.

Of course, this transition still faces real‑world challenges. First, there is no unified pricing standard for tokens; different models and tasks yield vastly different “value per token,” making fair and transparent pricing a difficult industry problem. Second, an estimated 15‑20% of high‑performance GPUs globally remain “unboxed and unpowered”—idle capacity that has yet to be effectively utilised; moving from hoarding to productive output will take time. Third, cross‑vendor and cross‑region interconnectivity and standardised scheduling remain technical hurdles.

Nevertheless, the direction is clear: competition in computing is shifting from “how many GPUs you have” to “how many tokens you can produce.” In the future, whoever can output the most tokens at the lowest cost and highest efficiency will secure a central position in the new AI computing infrastructure.

5. Outlook: Towards Universal Access and AI Democratisation

In the final week of July 2026, the computing industry has reached a new inflection point. The Token Factory is not merely a business‑model innovation; it is a milestone marking the transition from the “hardware era” to the “service era” of computing. It heralds a future where computing power, like water and electricity, becomes a metered, tradable, and universally accessible public utility.

When compute is consumed by the token, the cost of AI models becomes transparent and controllable. Enterprises and individuals can subscribe to intelligence on demand, much like cloud services. This will dramatically lower the barriers to AI application innovation, spurring more specialised models and applications across diverse domains, and genuinely embedding AI into every industry and household. Moreover, the Token Factory could alleviate waste by finely scheduling and billing, directing resources toward high‑value tasks and improving overall societal compute efficiency.

Looking ahead, as chip manufacturing advances, packaging technologies mature, and system software continues to improve, the cost per token will decline further. At that point, AI compute will no longer be an exclusive luxury for tech giants but a fundamental infrastructure accessible to everyone. This revolution—from stacking GPUs to selling tokens—has only just begun, and its far‑reaching effects will unfold over the coming years.