Ahead of DeepSeek R2 1200B's Release, Tech Giants Are Rethinking Their AI Compute Strategies
Updated on:June-16th-2025
Recommendation
The next generation of intelligent computing chip DeepSeek R2 1200B is about to bring about industry changes, and major manufacturers have taken the lead in layout.
Core content:
1. DeepSeek R2 1200B parameters have doubled, which has a far-reaching impact, and major manufacturers have taken action
2. The technical VP of the major manufacturer has completed internal alignment, H series equipment is on sale, and B series computing power resources are being intensively purchased
3. In order to cope with the scale of R2, super node products are advanced, PCIE is abandoned, and domestic AI chip manufacturers are accelerating their layout
Yang Fangxian
Founder of 53A/Most Valuable Expert of Tencent Cloud (TVP)
The troops have not moved yet (DeepSeek R2 1200B has not been shipped), but the food and grass have been prepared first (large manufacturers have restructured the computing power layout of intelligent computing centers) . Large manufacturers have already grasped the trends of R2 1200B and have made arrangements in advance!
Special note: Although R2 1200B has not been released, it is highly credible based on the cross-verification of information from multiple channels, including actions of major manufacturers, renderings by foreign media, and computing power encyclopedia!
Compared with R1, the total parameters and activation parameters of Deepseek R2 are nearly doubled , which has a profound impact on intelligent computing centers. Leading manufacturers have already deployed their troops and taken action. They have more information and are taking adaptive countermeasures :
1) Large companies deploying their forces: The technical VPs of the top large companies have basically completed the internal technical backbone alignment meeting for R2 this week, and the major companies have already taken action. Friends with resources can look for their meeting minutes to understand the content.
2) Big sale of H equipment: After retiring A800 equipment, major manufacturers put H100/H800 on the shelves and started selling them off . For example, in several provinces such as Jiangsu, Sichuan, and Qinghai, second-hand equipment and H equipment were sold off after the holidays.
Used H series resources within 6 to 24 months are being sold off, with prices ranging from 1.4 million to 1.8 million. Cluster sizes range from 64 to 256 units. Verify capital - inspect goods - test - take away. The process is similar to buying and selling real estate . You can trade and take away at any time!
80G video memory is the original sin . H100/800 video memory 80GB*8=640G video memory. Two 1280G video memory units are still unable to cope with the 1200B native full-blood version. At least 3 units are needed, and the cost-effectiveness is instantly not there.
1200B R2, the best solution is to use one B200 or two H200/H20 141G, and at least two 96G*8 are needed;
3) B series scheduling action: The purchasing departments of major manufacturers have been intensively meeting with their own exclusive computing power suppliers this week to discuss B series computing power resources, construction planning and supply plans; B200 is the perfect match for R2. Technical architects of major Internet manufacturers generally believe that B200 and 1200B are a perfect match. The price-performance ratio of the 192*8 inference cluster at FP8 precision is unbeatable;
4) Supernode products in advance: To cope with the situation that the scale parameters and activation parameters of R2 are doubled compared to R1, in order to better solve the requirements of multi-core communication, reduce latency, and computing power stability. Canceling the PCIE interface and adopting the onboard AI node solution has become the choice of various companies. PCIE has become a shackle that restricts global computing power products. Accelerating the abandonment of PCIE is the common wish of global AI chip manufacturers. NVIDIA launched NVLink Fusion, and domestic AI chips of various countries have integrated 2D on-chip mesh to accelerate the pace of abandoning PCIE;
Domestic AI chip manufacturers have begun to design their own brand of "super node" solutions, and major domestic AI chip manufacturers have given priority to commercial-grade super node product plans!
The super node solution is not only beneficial for training, but also more cost-effective and has lower operating costs than traditional cluster solutions when used for large language model inference business!
Each AI chip company will develop its own supernode solution. The quality of the supernode depends largely on the capabilities of the AI chip. Other communication chips are basically purchased and are not much different.
In the next cycle, the equipment to be replaced in the intelligent computing center is the supernode, not just the card.
