Physix FrontierMaking Money with AI Data Centers: Beginners Can Verify This Themselves
As someone who does literature reviews and model evaluations, I tried making a mini knowledge graph using three sentences of public material. A knowledge graph turns nouns into nodes and relationships into edges. For example, Samsung is a node, AI chip demand is a node, and the middle says "drives." When guiding students through literature reviews, I often have them draw this small diagram first.
This time I used the MangoBoost report as material to compare two approaches.
The pure manual table approach involves creating a new local file, e.g., mangoboost-graph.md, opening it in Notepad, and listing columns: Entity, Type, Relationship, Evidence. Extracting three sentences: MangoBoost makes AI data center efficiency solutions; a type of chip called DPU supports cost efficiency; companies backed by Samsung profit from AI data center demand. DPU can be explained as a chip inserted into servers to handle miscellaneous tasks for data centers. After filling it out manually, there are about 5 nodes and 4 edges: MangoBoost, Samsung, AI Data Center Demand, DPU, Efficiency. In my testing, it took about ten minutes to fill out—accurate but slow.
The other approach was letting Claude extract for me. I've been using Claude for small tasks for about a month; it's only suitable for small tasks. I pasted the three sentences into the chat box and typed: "Based only on the sentences I gave, generate a text table. Columns are Entity, Type, Relationship, Evidence. Do not infer equity, revenue, or performance." After hitting send, the interface directly produced a text table. This time I got about 6 entities and 5 relationships, where 1 relationship incorrectly translated "backed" as "invested." I changed it back to "supports," then copied the table into the local file. The boundary of this experimental design is clear: the model is responsible for organizing, not judging.
Putting the two routes together, I treated the manual table as a reference to see if the model over-extracted. The manual table had 5 entities; the AI table had 6. The extra one was usually "cost efficiency," which isn't wrong, but it's best merged into "efficiency," otherwise the graph gets scattered. Regarding relationships, the manual version wrote "drives," "provides," "optimizes," while the AI version wrote "makes profitable," "reduces costs." The latter sounds more like news conclusions and needs confirmation against the original text.
Beginners can create a new file named mangoboost-graph.md, paste the three sentences, and build a four-column table. First, manually fill in 5 rows to ensure you haven't been led astray by the model, then use Claude to generate a version, writing differences into the Evidence column, e.g., the original text only says "supports," not "invests."
The easiest mistake is translating "backed" directly as "investment," which differs significantly in Chinese. Another pitfall is attributing percentages to the wrong subject. The material mentions Samsung's stock rose about 220% in the past six months, Samsung Electronics rose about 130%, and some reports say Q2 profits are expected to grow 1800%. These numbers come from different companies or different metrics and cannot be mixed into "MangoBoost profits grew 1800%." Add a line next to every number specifying who grew, when, and what metric.
Another pitfall is the model treating "DPU reduces power consumption" as a verified fact. The claims in the material are vendor assertions; dataset bias needs consideration. When writing the table, add the word "claimed," otherwise the evaluation isn't clean.
Next, you could try pulling in companies mentioned in news like Samsung, SK hynix, AMD, and Dell to see who depends on whom in the AI infrastructure chain. To understand AI data center news, drawing out the nodes and edges first is more useful than memorizing terminology.
📌 This article is compiled from Bloomberg Tech, original text: https://www.bloomberg.com/news/articles/2026-09-09/samsung-backed-mangoboost-makes-profit-on-ai-data-center-demand
Copyright belongs to the original author; this is a compilation and independent analysis based on public reports.
