Breaking a compute news story into a fact checklist

I compared Codex and a search agent, running them through actual tasks. Conclusion: Worth it, especially for breaking down a seemingly huge compute news headline into a verifiable fact checklist. Here, Codex refers to an AI assistant that reads prompts and organizes text, while the search agent refers to a search tool that automatically finds web pages and summarizes sources.

This news is worth noting. Having covered the AI sector for years, I've seen plenty of headlines with big numbers. IT Home reposted Bloomberg news on September 11 stating that Microsoft plans to increase global data center capacity to over 38GW by 2032, currently around 12GW. 38 is more than three times 12. The report's scope includes Microsoft's owned and leased data centers, excluding compute rented from neocloud providers like CoreWeave. Currently, out of Microsoft's ~12GW, only ~2GW is used for AI chips; by 2032, AI-dedicated compute is expected to account for roughly one-third of total capacity.

There are two terms here that beginners might find confusing. GW is a power unit; roughly understood as how much simultaneous power load these server rooms can carry. Data center capacity isn't just about how many servers are placed; it's more about whether electricity and space can sustainably feed the machines. Neocloud providers are clouds specifically renting GPU compute to AI companies, like CoreWeave mentioned in the report.

The tutorial task is to create a compute news fact checklist. Tools required: a text editor, Codex, and a search agent. I've used Codex for 1 month and the search agent heavily in the last week. The former excels at structuring messy text; the latter excels at finding sources, but neither replaces human verification.

First, create a new file. Right-click on the desktop blank area, select New Text File, name it Compute_News_Fact_Checklist.md. Double-click to open, and you have a draft sheet you can edit repeatedly.

Second step: Copy the original summary. Open IT Home, type "Microsoft 38GW data center" in the top search box. The summary area is usually right below the title; I've used IT Home for two weeks and know this well. Click the first title, copy the summary text, and paste it into the file. The content should include 2032, 38GW, 12GW, owned and leased, excluding CoreWeave. If these numbers are missing, you clicked into a secondary retelling; switch articles.

Third step: Let Codex break it down first. Open Codex, start a new conversation. Paste the following into the input box.

Please break down the following news into three columns: Confirmed Facts, Scope Explanation, and Items Needing Further Investigation. Do not add numbers not present in the news. Original text: Microsoft plans to increase global data center capacity to over 38GW by 2032, currently around 12GW. Statistical scope includes owned and leased data centers, excluding compute rented from neocloud providers like CoreWeave. Currently, out of ~12GW, only ~2GW is used for AI chips; by 2032, AI-dedicated compute is expected to account for roughly one-third of total capacity.

This way you get a set of short sentences instead of one whole paragraph of news.

Fourth step: Verify with the search agent. Open the search agent, write "Microsoft 38GW data center Bloomberg 2032" in the input box. Require it to only list source titles and key numbers, not summarize into opinions. After seeing results, open at least two pages to confirm consistency on 38GW, 12GW, and excluding CoreWeave. Most Chinese drafts come from the same Bloomberg message, not multiple independent pieces of evidence.

Fifth step: Fill results into a table. Return to the text file, insert a Markdown table (arranged with vertical bars and dashes). You can type it manually or let Codex generate the table and copy it.

Item	Info Given by News	My Verified Risk
Target Capacity	Over 38GW by 2032	This is a plan, not built yet
Current Capacity	~12GW	Must distinguish owned, leased, neocloud
Multiple Relationship	More than 3x	Readers easily mistake plan for status quo
AI Compute	Currently ~2GW, future ~1/3	Total capacity growing doesn't mean AI server rooms immediately dominate
Exclusions	Excludes neocloud rented compute like CoreWeave	Readers easily count all AI compute as Microsoft's own

Finally, write one plain sentence. At the bottom of the file, write: "This news means Microsoft wants to scale up the power size of server rooms it controls itself, reducing bottlenecks from compute shortages." Don't use words like "Microsoft completely dominates." If someone asks you what 38GW is, you can explain it in three sentences.

Three common pitfalls when doing this for the first time. First pitfall: Mistaking plans for current status. 38GW is a 2032 target; currently ~12GW, so add time context to every number in the facts column. Second pitfall: Ignoring exclusions. If you only remember "Microsoft 38GW," you might mistakenly include compute rented from CoreWeave, so write a separate exclusion line. Third pitfall: Trusting the search agent's summary. It might write "sources say" as "Microsoft confirmed." For any words like confirmed, announced, official, you must go back to the original text to check for a subject.

My testing shows Codex's organizing ability is more stable than the search agent, but the search agent finds sources faster. Using both together yields good efficiency. The underlying trouble is that AI tools excel at compressing information, and humans easily mistake compressed info for complete info. Making this checklist is about stretching it back out.

After learning this, next try breaking down the August piece on Microsoft's self-developed chip mass production, to see where chip ramp-up, data center capacity, and capex fit in the chain respectively.