The Zhitong Finance App learned that the wave of price increases for GPU cloud services is rewriting the negotiation pattern in the global computing power industry? Recently, Nebius announced a 20% price increase for the entire GPU cloud service line starting October 1 — this is the second round of price increases in the year. The B300 has a cumulative increase of 56%. Customer demand has been visible for more than 24 months, and they are even willing to bid for Blackwell's computing power at a premium; the backlog of CoreWeave orders has exceeded 100 billion US dollars, and the scale of contracted electricity continues to rise. Hwang In-hoon bluntly stated, “The core constraint is not demand, but production capacity.” When “if you have a card, you have pricing power” became a consensus, the capital market began to re-examine the question: Who are the players who really “definitively benefit” in this round of hashrate bullish market?
In China, Xunze Technology (03317), known as the “first token share”, gave a different answer from Neocloud overseas. Contrary to the traditional computing power rental path of “build a computer room first, then find a customer”, Xunce first holds high-quality customers and high-quality orders, and then builds computing power according to the order. This “order first, construction follows” style of play made TokenCloud's one-stop AI model training and estimation platform stand at a unique starting line from the beginning.
Driven by customer demand, quality orders in hand
What is the computing power rental industry most afraid of? It's not that you can't get a card, it's that the computer room is built but there are no customers — huge depreciation and idle waiting to be rented is the number one killer of all asset-heavy computing power platforms. Overseas Neocloud companies are still reliable, and large manufacturers hedge risks, and a large number of computing power players who follow the trend are facing the hidden concern of “being built and idle.”
Schuntzer's path is exactly the opposite. Prior to the launch of TokenCloud, Xunce's TokenOS operating system had completed real-time tokenization of enterprise multi-source heterogeneous data, and was implemented on a large scale in 11 high-value industries, including finance, telecommunications, electricity, energy, high-end manufacturing, and biomedicine. This means that before developing computing power, Xunce already had a large number of real, paid enterprise customers, and the intensity of AI usage continues to rise.
Comparing the classic script of “Taobao Tmall + Alibaba Cloud”, the e-commerce platform (TokenOS) has accumulated a large number of high-quality merchants and real transaction needs. Cloud Computing (TokenCloud) is an infrastructure to handle these needs, and there is almost no need to acquire customers from zero. Behind every unit of TokenCloud's computing power construction, it corresponds to an actual order that has been signed and paid for. In the words of the industry, every unit is profitable to build, and there is no idle waiting to be rented.
What's more important is the quality of the order. At a time when AI computing power is scarce from “narratives to reports,” supply-side bargaining power is being reversed, and high-quality terms with a long cycle and high unit price have become scarce resources. Xunze's customers are enterprise-level To-B customers — accounts, budgets, and willingness to renew contracts in high-barrier industries such as finance, telecommunications, and energy, far better than “retail investors” in the loose rent market. In the first half of the year, the company's gross margin remained high at 60.1%, ARR grew rapidly, and ARPU jumped from 1.64 million yuan to 5.56 million yuan, all of which are direct footnotes on order quality. When Nebius' on-demand price became the industry's pricing anchor and went all the way up, the value of this batch of “high cost, long cycle, and good terms” Cloud orders in Xunce's hands was only further amplified.
The cloud utilization rate is close to 100%, and the capital cost advantage is strong
Computing power is a “utilization rate+capital cost” business. For the same H100 computing power card, the business model with a utilization rate of 90% and a utilization rate of 40% is very different; for the same card purchase loan, the interest rate directly determines the break-even point. In these two dimensions, Xunce has built moats that are difficult for peers to replicate.
Let's take a look at cloud usage rate first. The construction logic of traditional computing power leasing companies is “gambling” — betting on future demand, betting on technology routes, and betting on customer composition. The built clusters are either vacant and waiting or being promoted at a reduced price. According to the reverse planning and construction of the highest quality orders, Xunce accurately matches the scale of computing power with actual demand, and superimposes TokenCloud's own full-link acceleration capability — changing the acceleration card from “waiting for calculation” to “continuous calculation” through model distillation and reasoning optimization — its Cloud utilization rate is close to 100%. This means that with assets of the same size, the number of tokens and revenue generated by Xunze is far higher than that of peers, and unit depreciation costs are diluted to the extreme. In 2026, when the supply of computing power is in short supply and price increases are commonplace, the “computing power asset running at full capacity” itself is a money printer.
Let's look at the cost of capital. Computing power-heavy asset expansion is extremely dependent on leverage, and slight differences in financing interest rates will expand into a huge area of profitability over a period of several years. The advantages of Xunce come from both sides: First, a high-quality and stable order itself is the best credit certificate — predictable cash flow makes banks willing to give extremely low-cost loans, which is the same logic that overseas Nebius uses “deployed GPU+ contract cash flow” as collateral to obtain low-cost capital; second, the support and support of shareholder backgrounds further reduces the overall cost of capital. Low-interest capital purchases computing power assets, high utilization rates produce high returns, and high-quality orders feed back credit — the quick funding chain forms a positive cycle where costs continue to decline.
Ecological card slot bound to a domestic GPU, and a self-reinforcing positive cycle
If orders and funding are “techniques,” then supply relationships and ecological flywheels are the “ways” of quick action.
The global shortage of computing power is essentially a supply bottleneck. Huang Renxun expects chip sales to double in 2027, but he still emphasizes that production capacity cannot keep up; Intel CEO Chen Liwu admits that CPU supply can only meet about 50% of demand. In the context of restrictions on imported high-end GPUs, domestic computing power has become the definitive direction, and “who can get a stable and priority supply of domestic GPUs” will hold the ticket to this game.
Xunce's card position is a forward-looking one: since June of this year, the company has successively reached deep strategic cooperation with almost all leading domestic GPU manufacturers such as Mu Xi, Tianshu Zhixin, and Bizao, and is moving further towards joint research and development — jointly building a training and inference chip and platform ecosystem for vertical industries and enterprise AI. This is not a simple procurement relationship; it is a deep binding of “computing power+data”: TokenCloud can unify heterogeneous computing power such as GPUs, NPUs, and FPGAs across architectures to achieve intelligent scheduling on a scale of 100 cards to 10,000 cards, paving the way for large-scale implementation of domestic chips in highly sensitive industries such as finance, telecommunications, and energy. For domestic GPU manufacturers, Xunce is not only a stable major customer, but also the best partner to verify the actual value of chips and polish the software ecosystem — the stability and priority of the supply relationship, thus obtaining institutional guarantees.
All of these advantages eventually converged into a self-reinforcing positive flywheel: TokenCloud's enterprise-level customers naturally precipitated data governance and tokenization requirements, driving TokenOS's revenue and gross profit increase; TokenOS refined data capabilities and scenario understanding in more industries, which in turn helped TokenCloud attract better customers, sign better orders, and support the construction of higher utilization computing power. TokenCloud and TokenOS are not two isolated products, but rather two sides of a deep collaboration on the AI to B circuit — every positive interaction reinforces barriers that are difficult for opponents to overcome.
The reversal of bargaining power in the supply of computing power is no longer a narrative, but a reality written into Nebius's price increase letter, CoreWeave's 100 billion order book, and Xunce's earnings report. However, they are also on the cusp, and the talents of players vary widely: some people build cards based on concepts and gamble heavily; others build on orders with high-quality orders, and every cent of capital expenditure corresponds to a definite return. What makes Xunze TokenCloud unique is that it is not a “start from scratch” gamble, but rather a computing power map that naturally grows from TokenOS's accumulated high-quality customers and needs — order first, full load operation, low interest capital, stable supply, and flywheel rotation. When the tide recedes, those who can stay on the table have always been players who “make money by building every unit”.