Chinese storage giant considering IPO at potential $100+ billion valuation; simultaneously, major domestic tech firms openly competing for and securing thousands of NVIDIA B300 chips, bypassing prior discretion.

In the past, large companies typically concealed their true identities when making purchases through intermediary channels. Today, an increasing number of projects are signing contracts directly, with much less effort to obscure their identities.

Meanwhile, B300 chips are being rapidly acquired by leading customers. An internet giant in East China has procured over 2,000 units, while a technology leader in North China has also acquired around one to two thousand units. Just a handful of companies are consuming a substantial proportion of market supply.

The listing processes of China's two leading storage manufacturers continue to attract capital market attention. According to reports, one storage giant's IPO valuation is expected to be raised from several hundred billion yuan to the trillion-yuan level, while the listing timeline may be delayed compared to previous expectations.

Multiple industry insiders stated they have not received clear confirmation of the valuation increase, but generally believe that a further valuation adjustment would not be surprising. As valuations rise, existing shareholders' equity value may face revaluation, and an enterprise of this scale entering the A-share market could also put pressure on the market capitalization of existing semiconductor-listed companies.

A Shanghai AI chip company that completed IPO guidance late last year has shown no new progress over the past six months. It is said the company previously did not obtain the "green channel" approval, encountering some challenges in its listing process. However, backed by a high-level leadership visit last year, it has now obtained the IPO "green channel," with the listing process showing signs of turnaround. It is understood that investors are currently in discussions with the AI chip company regarding pre-IPO round financing.

Following storage price increases, price hikes have begun to propagate throughout the entire AI industry chain. Recently, NVIDIA 5090 rental prices have increased by approximately 20%, and at the same time, CDN prices have also risen by close to 20%.

However, the underlying logic differs for the two. The 5090 price increase stems primarily from demand growth—as inference and Agent applications develop rapidly, high-end GPUs face persistent supply shortages. CDN price increases, by contrast, are driven significantly by pressure on low-cost PCDN resources.

Industry insiders note that CDN handles data distribution for models, images, video, and other content, serving as critical network infrastructure connecting users to computing centers. Over recent years, PCDN has continuously driven down bandwidth costs through edge nodes and P2P technology. As telecom operators strengthen governance, low-cost supply has diminished, and CDN prices have resumed their upward trajectory.

Multiple industry practitioners believe that model adaptation itself no longer represents a core barrier; what truly differentiates companies is performance optimization capability. Even after achieving comparable adaptation, performance differences between different chip manufacturers remain significant. It is understood that a certain unified ecosystem internally maintains a detailed scoring system and results, but due to sensitivity among some manufacturers regarding horizontal comparisons, only "pass" and "fail" categories are disclosed externally.

Some employees joined Zijing Unlimited, a joint venture between JD.com and a domestic chip manufacturer, while other key members chose to start new ventures. Among them, a new company co-founded by core employee Jeffrey Xu was established in May of this year, focusing on AI inference chips and software-hardware systems. According to industry insiders, the company secured over 100 million yuan in funding shortly after its establishment.

Some AI chip companies have chosen to absorb costs and maintain product pricing, arguing that the company remains in an unprofitable market expansion phase where orders and customers are more important than profits.

However, some companies have paid the price for this approach. An AI chip company that had nearly finalized an intelligent computing project saw it fall through as storage price increases caused cost fluctuations. The project party requested both sides share the additional costs, but the company took a firm stance refusing to compromise, ultimately losing the order.

Some suppliers revealed that overdue payments from certain large customers have become routine. More unusually, some conglomerates previously required cash payment only and prohibited the use of acceptance notes. However, the reality has shifted to "no cash available, therefore unable to pay." When suppliers proactively offered to accept notes, the customers refused, citing "non-compliance with regulations." Ultimately, suppliers could neither obtain cash nor acceptance notes, forced to continue waiting. "What if we can't get cash payment for a year?" has become a common challenge facing many suppliers.

According to industry insiders, the new product is primarily targeted at large-scale training cluster scenarios, with a focus on competing with NVIDIA's H100 platform, with single-card pricing reaching tens of thousands of yuan. Unlike other industry companies that frequently hold product launches, this company has maintained a consistently low-key approach.

According to a chip industry insider, computing power operators' customers mostly come from operations or software backgrounds and lack a foundation for evaluating hardware, making it difficult to assess the value of different chips based on tokens alone. Consequently, these customers often rely on computing power operators to provide hardware selection solutions.

In the past, computing power operators used almost exclusively NVIDIA products. Today, to meet customers' diverse price-to-performance needs, operators have begun proactively testing domestic chips to form practical comparisons with NVIDIA products.

However, testing is one thing; actual deployment still faces obstacles. Internet giants possess strong engineering teams capable of completing adaptation and optimization independently, while most central and state-owned enterprise customers' data centers have staff skilled only in basic operations and maintenance, lacking chip adaptation capabilities. Even if operators have verified that a certain domestic chip works, these customers still hesitate to adopt it—because if problems arise, they lack the technical capability to handle them.

Consequently, central and state-owned enterprise customers have chosen a "safe" path: blindly copying the configurations of industry leaders like the State Grid, directly deploying high-end chips. Take a certain supply chain group as an example—its business data consists primarily of logistics and instructions, and deploying 7B or 11B models would be sufficient, yet it follows suit with high-end chips, resulting in extremely poor price-to-performance.

A Taiwan-based motherboard manufacturer has introduced a DDR5 single-channel motherboard, directly eliminating one sub-channel to reduce chip requirements. Some industrial control manufacturers have re-enabled DDR3 support, prioritizing supply chain solutions. Meanwhile, other manufacturers have released dual-mode DDR4/DDR5 compatible designs, preserving customer switching flexibility.

According to practitioners' calculations, transmission latency over 100 kilometers is approximately 1 millisecond, and over 10,000 kilometers only about 0.1 seconds. For large models with typical first-token response times of 200-300 milliseconds, the additional latency introduced by cross-region scheduling is not as substantial as one might imagine.