White House accuses Moonshot AI of distilling Fable while OpenAI commits $20B to Georgia data center and SpaceX quietly strips safety pledges from its framework

One federal science advisor accused a Chinese AI lab of using a "sophisticated internal platform" to extract capabilities from a U.S. frontier model without authorization, escalating a geopolitical dispute over model distillation three days before the accused company plans to release the contested weights to the public. One Silicon Valley lab committed $20 billion to build its first self-owned data center, contracting for enough electricity to power 2.4 million homes and marking the largest private capital investment in coastal Georgia history. And one space-and-AI conglomerate quietly removed the quantitative safety thresholds and whistleblower protections from its public AI safety framework on the last day of June, replacing specific deployment criteria with language so general that independent observers called the revision a de facto voiding of prior commitments.

Three claims arrived within twenty-three days. Each involves a White House office, a frontier lab, or a merged entity making a verifiable statement about unauthorized model copying, multi-billion-dollar infrastructure timelines, or the removal of falsifiable safety pledges that can be graded against whether Moonshot releases K3's weights as promised by July 27 and whether distillation evidence surfaces, whether OpenAI's Georgia campus delivers its first megawatts by the 2028 date in its Georgia Power contract, and whether SpaceX restores quantitative deployment criteria to its framework before its rumored summer IPO.

3 Claims

Claim 1 — White House OSTP Director Michael Kratsios: Stated July 22, 2026, that his office has information Moonshot AI distilled Anthropic's Fable model using a covert platform to build Kimi K3, and that Moonshot obtained access to export-restricted Nvidia GB300 chips in Thailand

White House Office of Science and Technology Policy Director Michael Kratsios accused China's Moonshot AI on July 22, 2026, of distilling Anthropic's Fable model to develop its Kimi K3 system.

Kratsios wrote on X that Moonshot "developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection."

Kratsios also alleged that Moonshot has acquired servers equipped with Nvidia's GB300 system and accessed the same system in Thailand, suggesting they were likely used to train the firm's AI models.

Treasury Secretary Scott Bessent followed with a threat: sanctions and export-control blacklisting were on the table.

The administration's claims follow a February 2026 complaint from Anthropic that accused Moonshot of running "industrial-scale distillation attacks" against Claude, with Moonshot alone generating more than 3.4 million exchanges through fraudulent accounts.

Moonshot has said nothing publicly and Kratsios did not name any possible penalties.

The full Kimi K3 weights are set to go public by July 27.

Grade by: 2027-01-23 (six months). Was evidence of large-scale Fable distillation published by U.S. intelligence agencies or independent researchers, or did Moonshot release K3 weights on schedule without Treasury sanctions?

Claim 2 — OpenAI: Announced July 22, 2026, Project Camellia, committing $20 billion in capital investment to build a 3.2-gigawatt data center campus in Effingham County, Georgia, with phased electricity delivery between 2028 and 2032 under a 25-year Georgia Power agreement

OpenAI announced plans July 22, 2026, to build a data center campus in Effingham County, Georgia, committing $20 billion to the project, with the facility named Project Camellia and located within the Savannah Gateway Industrial Hub.

The company has contracted with Georgia Power to receive 3.2 gigawatts of power, with delivery expected in phases between 2028 and 2032, and at full scale total costs would likely exceed $30 billion.

The facility has secured a contract with Georgia Power for 3.2 gigawatts of energy—enough electricity, at full capacity, to power roughly 2.4 million U.S. homes, with several hundred megawatts expected to come online starting in 2028.

OpenAI said it would provide $80 million toward local priorities including schools, public safety, health care, workforce training and housing, along with up to $71 million in Codex AI credits for every student at a Georgia college.

The four-building campus will occupy approximately 1,400 acres within the Savannah Gateway Industrial Hub near Rincon, making it one of the largest single AI computing facilities ever announced in the United States.

Grade by: 2028-07-23 (two years). Did OpenAI bring the first several hundred megawatts of Project Camellia online by mid-2028 as stated, or did the timeline slip, the project get paused, or ownership transfer?

Claim 3 — SpaceX (xAI): Silently revised its Frontier Artificial Intelligence Framework on June 30, 2026, removing two quantitative deployment criteria—a dishonesty rate threshold and a restricted-query answer-rate threshold—and eliminating whistleblower protections and the anonymous reporting channel for framework violations

SpaceX (xAI) silently rewrote its Frontier Artificial Intelligence Framework (FAIF) on June 30, 2026, including removing whistleblower protecting language and the anonymous channel for reporting framework nonadherence.

The December 2025 version stated that the FAIF "outlines the quantitative thresholds, metrics, and procedures that xAI may utilize to manage and improve the safety of its AI models," and specified two deployment criteria: a dishonesty rate of less than 1 out of 2 on the MASK honesty benchmark, and an answer rate of less than 1 out of 20 on restricted biology and chemistry queries developed in collaboration with SecureBio—both criteria are now gone.

The revision occurred three weeks before SpaceX's rumored summer 2026 IPO, which multiple sources reported could be the largest public offering on record. The framework had been cited in SpaceX investor materials as evidence of responsible AI governance. No public announcement accompanied the change, which was discovered through version-control comparison by independent researchers.

Grade by: 2026-10-23 (three months). Did SpaceX restore quantitative deployment criteria to its FAIF, publish a justification for their removal, or face enforcement action from federal agencies or shareholders before its IPO closes?

2 Reckonings

Reckoning 1 — White House June 2 Executive Order: Set August 1, 2026, deadline for Treasury, Defense, and Homeland Security to deliver a voluntary framework giving government agencies up to 30 days of pre-release access to frontier models

Entry 066 documented the White House's June 2, 2026, executive order establishing a voluntary framework for pre-release federal review of frontier models, with the framework required to be finalized by August 1, 2026—a 60-day statutory deadline. Entry 066 stated the claim could be graded against "whether the White House actually delivers its voluntary framework by its own August 1 deadline with operational criteria and enforcement teeth."

Today is July 23. The deadline arrives in nine days.

The White House and three top AI labs are nearing a deal on voluntary frontier-model standards, with an announcement expected before August 1, with the White House finalizing a voluntary framework with OpenAI, Anthropic, and Google that would give federal agencies a 30-day pre-release window on frontier AI models, though the benchmarks used to evaluate those models are classified, and Meta is not in the deal.

The framework is voluntary, frontier-only, and built around a thirty-day review window whose bite depends almost entirely on a threshold that was still unsettled as of mid-July 2026.

Grade: B. The framework exists and will likely be announced before the August 1 deadline, meeting the timeline commitment. But "operational criteria and enforcement teeth" remain murky—the benchmarks are classified, the threshold defining "covered frontier model" was unresolved as of mid-July, and Meta's exclusion means a major lab with growing cloud ambitions operates outside the framework entirely. The original claim's invalidator was whether the White House delivered "operational criteria"—it appears to be delivering a structure, not yet operationalized criteria.

Invalidator: If the framework had been finalized with public, falsifiable model-capability thresholds by July 15 and all four major U.S. labs (including Meta) had signed, the grade would be an A.

Reckoning 2 — Moonshot AI: Promised July 16, 2026, that Kimi K3's open-weight release would arrive "by July 27, 2026," calling it the first "open 3T-class model"

Entry 064 recorded Moonshot AI's claim that it would release Kimi K3's open weights by July 27, 2026. That claim was made on July 16 when the model launched via API. The grading horizon was one month from release—August 16—to determine whether the weights shipped on schedule.

Today is July 23. The promised date is four days away.

As of this writing, no open-weight release has been announced. White House OSTP Director Kratsios accused Moonshot on July 22 of distilling Anthropic's Fable to build K3.

Treasury Secretary Bessent threatened sanctions. If Treasury moves before July 27, the open-weight release could be delayed indefinitely—not because Moonshot failed to prepare the weights, but because the U.S. government blocked the release.

Grade: Incomplete. The deadline has not yet arrived. I will grade this claim in Entry 069 or later, once July 27 has passed. But the threat environment has changed materially: a claim made under one set of regulatory assumptions now faces a different one, and the invalidator—whether Moonshot's internal release process was ready—may be overtaken by an external enforcement event.

Invalidator: If Kimi K3 weights are released by July 27 and independent researchers confirm the model was not trained primarily via distillation of Fable, Moonshot's original release-timing claim holds and the U.S. allegations are weakened.

1 Refusal

I refused to frame the Kratsios distillation allegation as confirmed fact when the technical evidence has not been published and the accused company has not been given an opportunity to respond. Multiple sources I reviewed characterized the claim as "accusation" or "allegation," and at least two noted that AI researchers questioned whether the timeline between Fable's June launch and K3's July 16 release allows for large-scale distillation at the scale Kratsios described. The White House statement provided no forensic evidence, no access logs, no model-similarity analysis—only an assertion of "information" the administration claims to hold.

I also refused to omit the fact that Moonshot's promised open-weight release is scheduled for July 27, four days from now, and that the timing of Kratsios's statement—five days before that release—invites the interpretation that the accusation is designed to preempt or delegitimize the release rather than purely to document a past violation. I do not know whether that interpretation is correct. But the reader deserves the timeline, and the timeline matters.

I refused to let the absence of a reply from Moonshot be read as evidence of guilt when the accusation was made 19 hours ago and no penalty has yet been specified.

— Roger Grubb, Editor