Don't just focus on Nvidia, this market is driving new demand for liquid cooling!
01. Introduction to ASIC chips
As the AI craze sweeps the globe, Nvidia, as the global chip giant, maintains a dominant position and currently holds 80% of the market share for AI chips. However, Nvidia's main competitors, chip manufacturers such as AMD HW, are rapidly rising to seize Nvidia's AI chip market. In addition to the mainstream chip manufacturers competing for the AI chip market mentioned above, self-developed ASIC chips represented by companies such as Microsoft and Google are also being launched one after another.
What is an ASIC chip? ASIC (Application Specific Integrated Circuit) is a chip designed specifically for a particular application. Previously, ASIC's main application scenarios were blockchain and user mining equipment. With the increasing demand for AI chips from AI, the advantages of ASIC have gradually attracted industry attention and begun to be applied on a large scale. The main feature of ASIC is customized high energy efficiency. Compared with general-purpose processors (CPU, GPU), ASIC has higher efficiency, lower power consumption, and higher performance when performing specific tasks.
02. CSP giants compete to break up with Nvidia, with AISC chip market growing at an annual rate of 50%
Driven by low-cost AI models such as Deepseek, the demand for AI computing power has gradually shifted from training computing power to inference computing power in recent times. ASICs represented by Google TPU have performance and lower power consumption in the field of AI inference that are not inferior to Nvidia GPUs, and are expected to replace GPUs in the AI inference field.
At present, large-scale cloud service providers are beginning a lengthy "breakup" process with Nvidia. According to industry reports, the procurement of application specific integrated circuits is expected to continue to rise at a compound annual growth rate of 50%, with this growth mainly coming from companies such as Microsoft, Google, and Amazon Web Services, which are accelerating their ASIC chip research and development. Meanwhile, according to reports, Nvidia's core cloud computing customers are still continuing to purchase its hardware, but are also intensifying their orders for ASIC hardware and booking TSMC's production capacity.
Nvidia's server hardware accounts for 80% of the AI and cloud service industry market, and its Blackwell architecture GB200 GPU is rapidly entering major data centers worldwide. However, as each Nvidia chip is priced at up to $70000 to $80000, and the cost of configuring the entire GB200 server ranges from $1.8 million to $3 million, it is not surprising that these enterprise customers are seeking to break free from their dependence on Nvidia.
Although completely getting rid of Nvidia will be a long-term goal - after all, Nvidia not only dominates the hardware field, but its closed CUDA software ecosystem has also built a strong moat - these major CSP giants have already embarked on the path of hardware self-reliance, taking the first step through self-developed ASIC design, while continuing to purchase Nvidia products in the short term.
03. Technology giants accelerate ASIC layout, power consumption skyrockets, benefiting liquid cooling
Overseas manufacturers are accelerating the layout of ASIC chips and liquid cooling solutions. It is expected that the expansion of the ASIC market will further increase the demand for liquid cooling, in order to reduce dependence on general-purpose GPUs and lower computing power costs. Google AWS、Meta， Global cloud service providers such as Baidu are accelerating their AI ASIC layout.
Then let's take a look at the chip performance of the top few ASIC chip manufacturers in terms of shipment. At present, the most active manufacturers in promoting ASIC chips are Google, Microsoft, and Amazon, with fast iteration speed, almost on par with Nvidia. With one new product per year, chip performance is rapidly improving. Google's seventh generation TPU Ironwood has performance comparable to Nvidia's B200, while chip power consumption is also soaring. Google's TPU has a power consumption of over 600W per chip this year, with a cluster power of up to 10MW, and Microsoft's Maia 100 chip has a power consumption of up to 700-900W.
Google: Google has started using liquid cooling solutions for its third-generation TPU, and the seventh generation TPU Ironwood, released in 2025, can support up to 10MW level liquid cooled cabinets.
Amazon: In December 2024, Amazon released the Trainium 2 chip, which offers a 30% -40% higher cost performance ratio than GPUs for instances based on Trainium 2. At the same time, it announced that the new generation of AI servers will use liquid cooling solutions.
Microsoft: The Maia 100 chip released by Microsoft uses TSMC CoWoS-S and adopts TSMC 5nm chip. The chip is equipped with 64GB HBM2E, even with 500MB of L1/L2 cache, and has 12 400GbE network bandwidths. In addition, the thermal design power consumption (TDP) of this chip also reaches 700W, which is comparable to the performance of H100.
In terms of liquid cooling, Microsoft is using a device called "Sidekick", which can be placed next to the Maia 100 server rack to deliver cold liquid to the cold plate connected to the Maia 100 chip, then send the hot liquid back to the "Sidekick" to remove heat, and send the liquid back to the rack again.
According to SemiAnalysis, Maia 100， The chips will be deployed in customized racks and clusters called Ares. These servers are deployed in standard 19 inch or OCP sizes, and are reportedly "wider", with Ares only offering liquid cooled configurations.
Microsoft also pointed out that it is currently developing microfluidic cooling technology. Microfluidic technology is still in its emerging stage, which integrates tiny fluid channels into chip design, embeds liquid cooling inside the chip, and directly delivers the coolant to the processor, thereby achieving cooling inside the silicon wafer.
