On August 28, 2026, X announced that it had investigated inauthentic accounts suspected of ties to China that were involved in influence operations, and identified a "bot farm" of roughly 200,000 accounts. Of these, 200 had posted about US AI data centers and power policy. As a ratio, that's 0.1%, or one in a thousand. The 200,000 figure indicates the scale of the prepared account infrastructure, but X's public statement offers no way to gauge how far the posts actually reached or what effect they had on public opinion.
In June, OpenAI had also published a report called "Data Center Bandwagon" targeting the same issue. What OpenAI documented was the use of generative AI to produce posts, build personas posing as Americans, attach sensational commentary to legitimate news, and sustain operations across multiple social platforms over the long term. Separating the scale revealed by this latest disclosure from the influence that remains undisclosed helps clarify exactly what information operations are automating in the AI era.
Only 0.1% of the 200,000 Accounts Posted About AI and Power
According to X's Global Government Affairs team, the 200 accounts posted "in ways that could manipulate legitimate debate over US AI and energy policy." The content included claims that AI data centers are driving up household electricity bills and straining the power grid, as well as AI-generated cartoons depicting operators profiting at ordinary citizens' expense.
Four screenshots were attached as examples. Accounts using English-sounding names and anime-style profile pictures linked to power-market articles tagged with "#capacityauction." The linked articles were real, but the accompanying cartoons and comments divided the story into data-center operators as beneficiaries and ordinary households as victims. The fact that a link points to genuine reporting does not guarantee that the poster's identity or the added commentary is accurate.
However, X did not link all 200,000 accounts to the AI-power debate. What the remaining 199,800 accounts were doing has not been disclosed. Beyond the observation period and total post count, X has not revealed impression numbers or reactions from genuine users. Even X's own wording stops short of saying the debate was actually swayed—only that the posts were made "in ways that could" manipulate it.
The same caution applies to the China connection. X described the accounts as "inauthentic accounts suspected of ties to China," without indicating direction by the Chinese government or operation by a single organization. The 200,000 figure represents the size of the detected network, not the number of people actually influenced.
What "Inauthentic" Covers: Automation, Impersonation, and Coordination
X's own Authenticity policy does not confine "inauthentic accounts" to a single category. The policy lists, as separate items, automated or scripted accounts that violate developer rules, fabricated personas, impersonation of others, and coordinated amplification across multiple accounts. It covers both fake accounts operated by humans and semi-automated operations.
Therefore, the term "bot farm" that X used does not necessarily mean all 200,000 accounts were run purely by software. Whether generative AI produced the text and images, whether automated logins or bulk management tools were used, and whether the posting itself was automated are separate questions. X has not broken down the numbers by type this time.
Detection and permanent suspension also need to be treated separately. At the end of its post, X explained its general policy of suspending accounts that violate the Authenticity policy. However, it did not explicitly state that all roughly 200,000 identified accounts were suspended. The measures set out in the same policy range from phone-number or puzzle verification, to limiting a post's reach, temporary feature restrictions, profile corrections, and suspension. The public statement alone does not support the claim that "all 200,000 accounts were removed at once."
Automated Logins and Fabricated Personas
OpenAI's June report on Data Center Bandwagon offers useful context for understanding these four examples. The account cluster OpenAI examined issued instructions to ChatGPT in Simplified Chinese while requesting output in both English and Chinese, connecting via VPN. OpenAI assessed that the operators were most likely a social-media operations team at a private tech company whose clients included provincial-level Chinese government bodies—not a confirmed case of direct direction from the central government or an intelligence agency.
The operators had the AI generate text posing as people from varied backgrounds—American workers, students, mothers, office employees, investors. They linked to legitimate reporting on power-market capacity auctions and attached comments and images claiming that ordinary households were bearing the cost of AI data centers. The examples X released this time share the same combination of persona names, capacity auctions, and household burden narratives. Because the two companies have not jointly confirmed this to be the same operation, whether it was run by the same network remains unconfirmed.
ChatGPT's use expanded from producing posts to managing operations. According to OpenAI, the operators requested code to automate logins and interaction management across multiple social platforms, and also used it to extract usernames and format links. Their operational notes described a policy of cultivating personas that appeared to be real users by posting about daily life, having accounts react to one another, and keeping spare accounts in reserve. It even included ideas for splitting up activity to make it harder for platforms to detect coordination.
In these usage records, generative AI functioned both as a tool for constructing arguments and as an operational component continuously supplying fake personas and material. Even so, OpenAI stated it found no evidence that the operation achieved meaningful reach beyond its own sphere of activity. The capacity to carry out an operation and the actual result of moving people must be measured separately.
2019 and 2020: Distinguishing Core Accounts from Amplification Networks
Even in China-linked operations from the Twitter era, the majority of total accounts were not actively posting. When Twitter disclosed an operation targeting the Hong Kong protests in 2019, it distinguished between 936 accounts that were actively engaged and a much larger spam network of roughly 200,000 accounts that were suspended before they could substantially act. Only the 936 accounts confirmed to be active were included in the public archive.
In its 2020 disclosure, Twitter split a China-linked network into a core of 23,750 accounts and roughly 150,000 amplification accounts. It made the core data available to researchers and explained that most of the amplification accounts had few followers and generated little engagement. That kind of breakdown makes it possible to separately evaluate the layer that produces content, the layer that inflates the numbers, and the reserve inventory.
| Official disclosure | Large-scale network | Core with documented activity | Published impact/data |
|---|---|---|---|
| 2019, Hong Kong-related | ~200,000 accounts | 936 accounts | Core archive published; large-scale network suspended before substantial activity |
| 2020, PRC-related | ~150,000 amplification accounts | 23,750 accounts | Core published; amplification accounts had low followers/low engagement |
| 2026, this case | ~200,000 accounts | 200 AI/power-related accounts | Four examples presented; full list, engagement data, and detection method undisclosed |
The round figure of "roughly 200,000" matches between 2019 and this latest case, but there is no evidence they represent the same network. What the comparison shows is that the earlier version of the company distinguished core accounts from amplification layers and released data that third parties could examine, whereas this time it published only its conclusions and four examples. Assessing the true scale requires at minimum the same kind of breakdown.
Why Target Real Power Constraints?
The power issue chosen for Data Center Bandwagon is not a fabricated grievance. PJM, which operates the regional transmission grid covering the eastern United States, secured 134,479 MW in its capacity auction for the 2027–2028 delivery year but fell 6,623 MW short of its reliability requirement. The price was set at the cap of $333.44/MW-day approved by the Federal Energy Regulatory Commission, and PJM has explained that large data-center loads continue to be added to demand forecasts.
Berkeley Lab's 2025 Update also presented a reference case in which US data centers would consume 649 TWh by 2030, or 11.8% of the nation's electricity. Its sensitivity analysis ranges from 521 to 843 TWh, or 9.5% to 15.3% of the national total. There is a legitimate basis for concern that delays in expanding generation and transmission capacity would put pressure on capacity markets and grid operations.
At the same time, who actually bears the cost of capacity-market prices, and to what extent, varies with regional rate structures and regulatory decisions. Self-supply arrangements and bilateral contracts also affect what customers actually pay. PJM itself has explained that not all customers pay the same clearing price. And the 649 TWh figure is a measure of electricity consumption, not a rate of increase in household bills. What this operation did was insert fabricated personas into a genuine constraint and compress cause and cost-bearing into a single narrative.
Whether this latest batch of roughly 200,000 accounts can be called a successful information operation depends on whether X can separate the posting layer from the amplification layer and disclose impression counts and reactions from genuine users. It will also need to explain what evidence underpinned its determination of ties to China, and how many accounts were suspended, restricted, or subjected to verification. If that data is made public, it will become possible to distinguish between an operation that merely assembled a large stock of accounts and one that actually moved US AI and power policy debate.