Today I watched a video of the data center of Jane Street, a globally renowned quantitative trading company, and I felt the real sense of the next phase, where the section that requires AI is likely to rotate towards power and liquid cooling! 🧐

Although everyone has been hotly discussing the recent comparison image of Nvidia’s VR200 and GB300, which involves storage, MLCC, PCB, etc. In fact, they have all risen once already, and the extent is not small, with valuations not cheap either.

Liquid cooling and power, as the most important links in data centers, must not have any failures; otherwise, the $7.8 million VR200 could be written off on the spot, and no one can afford that cost, so their importance and irreplaceability will be even higher. The key is that these companies are not too expensive at the moment, and the safety cushion is sufficient!

In the video, those neatly arranged GB300 cabinets have a peak power consumption of 140 kilowatts, which cannot be handled by air cooling at all. The next generation, Vera Rubin Ultra, directly reaches 600 kilowatts per rack, equivalent to the electricity usage of 500 American households, all packed into a 7-foot-tall iron cabinet. This leap in upgrade essentially reaches the physical limit.

Now, the entire Wall Street quantitative institutions and big Silicon Valley companies are frantically expanding AI data centers. At Microsoft's Q1 earnings call, Nadella said they added 1GW of capacity in a single quarter, aiming to double that within two years. Google and Meta’s combined Capex for 2026 exceeds $70 billion. And one-third of that money is not for buying GPUs, but for purchasing transformers, bus ducts, liquid cooling systems, backup generators, etc.

What’s even more interesting about liquid cooling is that Jane Street mentioned something quite funny; previously, when doing operations and maintenance, they just checked if the fans were spinning, but now they have to worry every day about servers growing moss. 90% of the heat in the data center is transferred away by the liquid in the cold plates, and to prevent micron-level gaps from being blocked by bacteria, the water must be filtered to 25 microns and mixed with 25% propylene glycol.

Jane Street's technical director also stated a hard truth; hardware is indeed expensive, but their biggest fear is the opportunity cost. Due to the extreme scarcity of computational power, various trading teams inside are queuing up to grab quotas. If they miscalculated the power load and the circuit breaker trips, the training tasks would roll back, and the missed trading strategies and real money would be the most fatal.

This logic applies not only to quantitative institutions but also to major companies; the opportunity cost is far greater than the hardware cost. Ensuring the safety and reliability of long-term operations and maintenance is the most critical thing. Among them, some liquid cooling operation companies are also gradually evolving into a lifetime binding SaaS model similar to an elevator business, which could become the potential invisible cash flow king in AI service types.

I will publish another tweet later detailing these types of listed companies.

👇 The original video is in the comments section!

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。