Physix Frontier
Data Center Supply Chain: Assembling a Dashboard with New Tools
As someone researching data center supply chains, I tried combining Kingspan announcements, compute network panels, OCR, and Excel to look at the risks hidden in power, building materials, and public opinion amidst the US AI data center boom. I previously wrote "Money is starting to flow from compute to bricks," and this time I broke that phrase down into several fields.
The question is actually straightforward. News mentions that US data center equipment imports reached $653 billion in 2025, doubling compared to 2020, with about $580 billion being computing hardware. On the other hand, US polls show roughly three-quarters of people oppose new data centers, and social platforms have spotted suspected Chinese astroturfing accounts amplifying disputes over electricity bills, noise, and land use. I wanted to know if these things could be organized into a panel that allows for continuous monitoring, rather than scrolling past today and forgetting tomorrow.
What I pieced together
The solution isn't complex. I only encountered Kingspan a day ago, so I wouldn't claim familiarity; I just used it as an example to see how a company making building envelope systems is being pushed by data center demand. Simply put, envelope systems include exterior walls, roofs, insulation, etc. I clicked through their investor relations page and saw bond issuance announcements and performance outlooks repeatedly mentioning "data center." So I picked out fields like capital expenditure, customer types, and financing costs.
I've been using the Compute Network [panel/tool] for about a month and am roughly familiar with it. It shows metrics like regional power, rack load, and cooling requirements. I exported a CSV and selected a few data center clusters. BMC is also new to me; I only understand what a server management controller is and can't really query hardware logs yet. I've been testing OCR these past few days—turning text in images into copyable text—and threw CNBC screenshots and Kingspan PDFs into it. I've used Excel for a week to merge the fields.
Here's a pitfall. OCR works okay for clean text, but when PDFs have charts, footnotes, or two-column layouts, numbers and dates get mixed up. I tried converting a paragraph about import amounts; the $653 billion was recognized fine, but the date in the footnote jumped to the next line. Later, I processed clean PDFs and images separately, which stabilized the results. When merging in Excel, date columns turned into text. I mentioned this issue in another post last week, and manually converting formats took some time.
How effective was it?
The final result isn't pretty, but it's useful. Left side is companies, middle is metrics, right side is risk labels. Metrics include power gaps, equipment sources, financing costs, and community opposition levels. For now, I've categorized risk labels into three types: hardware dependency, power & cooling, and public opinion & policy.
The surprise was that breaking down abstract risks from news revealed supply chain issues aren't just about chips. Computing hardware is indeed huge, but Kingspan issuing bonds indicates data center expansion has entered the phase of building, retrofitting, and locking in financing costs. Power and cooling aren't just model parameters; they directly translate into community opposition, electricity price disputes, and approval cycles. News says large data centers are being built in Abilene, Texas, and satellite images of Gainesville, Virginia data centers were reported by Chinese official media. Putting these together, it's no longer just business news.
The downsides are obvious too. This dashboard relies on me manually supplementing evidence and can't automatically judge how big the "China risk" actually is.
The Compute Network panel focuses on operational metrics, Kingspan announcements focus on finance, and news focuses on public opinion—the timeframes don't align. I'm new to BMC, so it's not connected yet. Cleaning data with OCR and Excel isn't easy either; if the data is dirty, none of the upfront effort can be saved.
The conclusion is: it depends. If you just want to skim a news headline, this panel isn't necessary. If you research physical AI, humanoid robots, or AI infrastructure and want to understand power, cooling, building materials, bonds, and public opinion beyond the models, then this step-by-step assembly is worth it. It's unfriendly to beginners because you'll likely get stuck on field explanations and cleaning.
Next, I want to spend a few more days with Kingspan and BMC to add financing costs and server O&M entry points. I'll export more regions from the Compute Network to see if power gaps correlate with community opposition.
📌 This article is compiled from CNBC Tech. Original source: https://www.cnbc.com/2026/09/03/us-ai-data-centers-china-supply-chain.html
All rights reserved by the original authors. This is a compilation and independent analysis based on public reports.
