AI infrastructure research. Chips. Networking. Power. Cooling. Mapping the public companies behind the AI buildout. Research, not financial advice.
- Tweets 47
- Following 10
- Followers 3
- Likes 0
UPS capacity buys time. Generator capacity buys duration. The transfer switch decides whether backup power reaches the load. Data-center resilience is a sequence, not a shopping list. Which handoff would you test first?
1
1
Schneider Electric explains how UPS systems bridge the time until generators support the critical load and how transfer switches change power sources. Vertiv describes rack PDUs as the last distribution link to IT equipment.
blog.se.com/datacenter/2020/…
se.com/us/en/work/featured-a…
1
AI racks have a flow budget, not just a power budget. Cold plates need moving coolant; the manifold routes it and the CDU pumps it. Which link gets overlooked most?
1
5
For the cooling-loop mechanism: ASHRAE's data-center handbook describes CDUs with pumps and temperature, pressure and flow sensors; Vertiv explains CDU flow control and heat transfer. handbook.ashrae.org/Handbook… vertiv.com/en-us/insights/ar…
2
The AI rack is becoming the unit of competition. AMD is assembling GPUs, CPUs and networking around Helios; Intel is pushing Xeon as the inference control plane. Which layer captures the most value?
1
2
AMD outlines its Helios rack stack here: ir.amd.com/news-events/press…
Intel explains Xeon's orchestration and data-movement role here: newsroom.intel.com/data-cent…
1
INTC +12% and AMD +10% in one session shows that the AI trade is expanding beyond GPUs. CPUs, edge inference and the broader compute stack are being repriced—but one strong session still needs confirmation from orders, capex and earnings.
14
AI data centers can buy servers faster than the grid can connect them. DOE reports 1–2 year lead times for distribution transformers and 3–4 years for large units. The constraint is moving outside the data hall.
1
14
Transformer demand has risen while customization, materials and manufacturing capacity limit supply.
Source: U.S. DOE
energy.gov/oe/distribution-t…
3
AI data centers do not simply have a “liquid cooling uses water” problem. They have a heat-rejection design problem. Closed loops recirculate coolant. Cooling towers may consume water through evaporation.
1
4
LBNL estimates U.S. data-center site water consumption could reach 145–275 billion liters by 2028. Actual water use depends heavily on cooling design, climate and electricity sources.
Source:
datacenters.lbl.gov/publicat…
3
AI's water bottleneck isn't the server loop—it usually recirculates coolant. The bigger issue is heat rejection: cooling towers can lose water through evaporation. LBNL projects U.S. data-center site water use could reach 145–275B liters by 2028.
1
3
Source: LBNL, 2024 U.S. Data Center Energy Usage Report
datacenters.lbl.gov/publicat…
Research only. Not financial advice.
3
Meta's Muse hit No. 1 in the U.S. App Store.
The infrastructure story is inference demand: agents shop, plan, call and keep working.
If WhatsApp turns Muse into a daily habit, Meta's AI capex could become continuous compute demand. Is distribution the real AI moat?
1
4
Muse launched in the U.S. on September 8 and works through the Muse app or WhatsApp. It reached No. 1 on the U.S. App Store free chart about ten days later.
Sources:
about.fb.com/news/2026/09/in…
axios.com/2026/09/18/meta-mu…
For research purposes only. Not financial advice.
4
Liquid-cooled AI racks use two separate loops.
A CDU transfers heat from the server coolant loop to the facility water loop while controlling flow, pressure and temperature.
It is the hidden bridge between AI chips and chillers.
1
6
In a liquid-to-liquid CDU, a heat exchanger keeps facility water separate from the technical cooling loop serving the servers.
Sources:
se.com/us/en/product-subcate…
se.com/uk/en/download/docume…
For research purposes only. Not financial advice.
5
72 GPUs. One liquid-cooled rack.NVIDIA's GB200 NVL72 routes coolant through manifolds and cold plates on CPUs and GPUs.As AI rack density rises, thermal management—from cold plates and CDUs to pumps and heat exchangers—becomes core infrastructure.
1
4
The GB200 NVL72 combines 36 Grace CPUs and 72 Blackwell GPUs. NVIDIA's documentation shows coolant flowing through rack manifolds and cold plates attached to the CPUs and GPUs.
Sources:
nvidia.com/en-us/data-center…
For research purposes only. Not financial advice.
3