Koios Sockpuppets on X: 400 Accounts Pushing Nvidia and Chip Exports

More than 400 coordinated social accounts spent a year promoting Nvidia's side of Washington's AI chip export fight, then rushed to defend Jensen Huang when US outlets reported an intelligence scandal, according to a Model Republic investigation published November 10, 2026 by The Midas Project. Nvidia told the outlet it had no relationship with the Israeli influence vendors named in the piece and is still trying to identify who funded the network.

This is not a CUDA release story. It is about who gets to shape export-control policy while builders wait on Fleet Intelligence paperwork, Congressional hearings, and the Megaspeed thread that now includes Air Force One briefing timing disputes.

TL;DR

What broke in the Huang scandal week

The Financial Times and other outlets reported that US intelligence believed Jensen Huang's affair involved a Chinese intelligence asset tied to Megaspeed, a Singapore cloud partner under smuggling scrutiny. As that story spread on X, Model Republic noticed dozens of accounts posting near-identical defenses of Huang and dismissing anonymous sourcing.

Six accounts, the piece says, repeated the same argument: Huang already said in May that Blackwell and Rubin would not go to China, so export worries were overblown. That kind of templated reassurance is a signal to investigators, not proof of corporate direction.

Social share arrows illustrating coordinated Nvidia export-control replies on X

What Model Republic says the network did

Since November 2025, the accounts posted thousands of times on export controls — the rules that decide which Nvidia-class GPUs can ship to Chinese customers and how neoclouds document resale. The investigation says they also:

• Defended Nvidia as a company and Huang personally
• Downplayed chip smuggling severity
• Attacked Anthropic on AI safety positions that clash with Nvidia's lobbying posture
• Targeted export-control advocates, especially Chris McGuire (former NSC, now CFR) with 100+ replies since November 2025
• Harassed lawmakers who criticized China sales — insult-laden replies to House China Committee chairs, Sens. Warren, Van Hollen, Murphy, and others named in the article

The same network allegedly backs unrelated campaigns, including Sudan's Rapid Support Forces, which the US has accused of genocide — a detail that matters because it shows vendor reuse, not a single-issue fan club.

Nvidia's comment to Model Republic, as quoted in the article:

We have no relationship with those firms and have not asked any other party to engage them on our behalf. Recently, we also became aware of these accounts and are attempting to identify the source, which has so far been unsuccessful.

That is denial of hiring, not denial that the accounts exist.

Koios, IntellACT, and Project Lighthouse

The forensic story starts earlier. Haaretz exposed Project Lighthouse, a pro-Qatar influence effort allegedly run by aides to Benjamin Netanyahu, naming 15 fake accounts. Model Republic links most of those accounts to the later pro-Nvidia campaign through shared creation sessions, reskins, and scheduling patterns.

Koios, an Israeli company marketing financial-fraud detection, allegedly operated the personas. Israeli police raided Koios in the Qatar investigation; Koios said it ceased operations in September 2025.

Model Republic traces continuity to IntellACT, registered September 8, 2025, the day before liquidators were appointed. Evidence in the piece includes:

• Matching résumés for a Web and Financial Intelligence Team Leader listing Koios and IntellACT with the same start date
• Job ads for social media managers in Argentina or Vietnam at $500–800/month
• Koios CRO branding appearing as IntellACT CRO on public profiles

Both firms marketed themselves as defenders against online manipulation while, according to the investigation, operating human-driven sockpuppets (not fully autonomous bots). Operator mistakes cited include posts that look like LLM prompts left in the text — for example instructions to "create a reply… three options" — and the codename "NCore" for Nvidia in briefs.

Forensic patterns readers can reuse

Model Republic documents batch account creation — one session on November 9, 2021 created about 30 accounts in 2.5 hours. Many accounts share minute-level reskin events. Posts often fire in the first half-second of a new minute, suggesting scheduling software.

Deletion is part of the playbook. Roughly two-thirds of 4,000+ chip-related posts since November 2025 were already gone before publication, per the article — consistent with Haaretz sources describing routine post purges at Koios. Hours before publication, dozens of accounts deactivated and others deleted Sudan campaign posts.

If you study information operations, these are the same coordination markers OpenAI described in false-front operations: persona laundering, synthetic grassroots, and narrative persistence across unrelated clients.

Export gate lock symbolizing US AI chip export rules debated in Koios sockpuppet posts

What is claimed vs verified about "open source data"

Headlines and aggregator items described open-sourced data on the 400 accounts. We searched for a primary release from The Midas Project:

• The Model Republic article is the main artifact — narrative, counts, and methodology described in prose, not a CSV.
• Hugging Face hosts Koios-API/koios, a tiny unrelated Cardano API dataset — not the sockpuppet network.
• We found no Zenodo DOI from The Midas Project for these handles as of November 11, 2026.

Verified: investigative findings as published, plus Nvidia's denial of hiring quoted above. Unverified: any specific client payment trail, and any public account-level dump beyond what journalists embedded in the story.

If The Midas Project later publishes handles under CC BY or similar, update your OSINT workflow — until then, cite the article, not a phantom dataset.

Policy context: why export-control spam matters to builders

Congress is already scheduling closed-door Nvidia briefings and debating AI Overwatch Act language on foreign GPU access. Fleet Intelligence partner rules and 51% resale caps change which hosts you can trust for training clusters.

Manufactured comment threads do not change CUDA builds overnight, but they can distort the political signal lawmakers read when setting those rules. If you rely on cross-border neocloud contracts, treat viral "ordinary user" export takes with the same skepticism you apply to AI-generated campaign ads.

How to read chip-policy debates on social platforms

1. Check account age and batch creation — coordination investigations often start at creation timestamps, not post text alone.
2. Watch for identical paragraphs across unrelated avatars during breaking news — the Huang week showed six-copy templates in Model Republic's sample.
3. Compare to primary sources — Nvidia newsroom statements, BIS rules, and claimed-vs-verified Megaspeed timelines beat reply-thread outrage.
4. Note deletion — if a account's history is empty but replies are aggressive, archive tools may still catch removed export takes.

Platform enforcement is uneven. OpenAI bans some state-linked ChatGPT abuse; X and Reddit may leave commercial sockpuppets up until journalists or researchers document them. Expect more LLM-assisted PromptPasta-style variation — slightly reworded copies that evade naive duplicate detection, similar to tactics described in Alethea's PromptPasta report on coordinated political accounts.

What investigators and lawmakers might do next

Model Republic argues a formal investigation is warranted because the campaign targeted US lawmakers during a strategic technology debate. Short of that:

• Congressional committees already seeking Huang testimony may ask whether any vendor ran undisclosed grassroots campaigns — a disclosure question, not a technology question.
• Researchers may replicate session-link clustering if account data becomes public.
• Platforms could label inauthentic coordination if governments or NGOs file specific account sets.

None of that is guaranteed. Koios liquidators told Model Republic no assets were sold to IntellACT; IntellACT did not respond to requests for comment in the piece.

What this does not prove

• Nvidia direction. The company's quoted denial is explicit; IntellACT job ads mention US clients generically, not Nvidia by name in the excerpts we read.
• Alice Huang allegations. Sockpuppet defenses are not evidence about intelligence assessments; follow Megaspeed claimed vs verified for that thread.
• Automated bot swarms. Model Republic describes human operators with scripts and AI assist, not fully autonomous agents — different incident class from rogue coding agents.

Related reading

• OpenAI bans Russian and Iranian false-front influence operations
• Storm-1516 and the Talarico deepfake: what is confirmed
• Congressional Nvidia chip curb pressure and November briefings
• Nvidia Fleet Intelligence partner tracking
• Israel election AI bots: claimed vs verified
• Wired on mundane AI apps and deepfake networks

Investigation details and account counts reflect Model Republic's November 10, 2026 article and Nvidia's quoted statement. Platform policies and Congressional calendars change frequently.