AI Sector Daily Digest — June 15, 2026

Five things from the AI world in the past 24 hours: Anthropic sends technical staff to Washington to negotiate an end to the export ban while new reporting pins the trigger on Amazon CEO Andy Jassy; the G7 summit opens in Évian today with all three frontier lab CEOs present for the first time; OpenAI's Sora video model burned $15 million a day and earned $2.1 million in its entire life; the US government is using ChatGPT to search all 50 states' Medicaid records for up to $200 billion in fraud; and NAVER and NVIDIA are building gigawatt-scale AI factories in South Korea.

Anthropic sent senior technical employees to Washington over the weekend in an effort to negotiate a path out of the export restrictions that shut down Claude Fable 5 and Mythos 5 on June 12.1 The models remain offline as of Monday morning, three days after the Commerce Department ban took effect.

New reporting published June 14 fills in how it happened. Amazon CEO Andy Jassy raised concerns directly with senior White House officials on Thursday, June 12 — two days after Fable 5 launched publicly — after Amazon researchers used a sequence of prompts to extract cyberattack-related guidance the model was supposed to restrict.2 Amazon was responding to a White House request for feedback on the new model, according to a person familiar with the discussions.

What followed was a frantic day of calls. Treasury Secretary Scott Bessent, Cyber Director Sean Cairncross, Commerce Secretary Howard Lutnick, and Chief of Staff Susie Wiles all participated, according to Business Insider's account of the episode.3 Anthropic CEO Dario Amodei joined three calls and pushed back, arguing the vulnerability was narrow and not a universal jailbreak. Bessent told Amodei directly he was making a "bad decision." A source close to Anthropic said the company was given a 90-minute deadline to pull the models, with no details about the specific threat provided.

The ban bars foreign nationals from using the models, which includes foreign-born Anthropic employees. Because compliance would have required tracking users' citizenship, Anthropic disabled the models globally. The White House told Axios that no currently deployed model other than Mythos reaches the capability threshold that triggered the action, and that any future model crossing that threshold will need government review before public release — a de facto pre-clearance requirement for frontier models.4 Anthropic has called the action disproportionate and is contesting it.

What to watch: negotiations reportedly center on a tiered access structure — full access for US citizens, a geofenced version or monitoring requirement for foreign users. No timeline has been announced.

The 52nd G7 Summit opened this morning in Évian-les-Bains, France, running June 15-17. For the first time, the CEOs of OpenAI (Sam Altman), Google DeepMind (Demis Hassabis), and Anthropic (Dario Amodei) are all attending the same summit.56 Macron explicitly invited them as part of a diplomatic strategy to bring the private sector into the governance tent.

The US is blocking four specific provisions that the EU wants in the communiqué: mandatory data center energy and water disclosure, AI incident reporting with defined timelines, coordination on export controls for foundation models, and a shared compute threshold definition for "frontier" AI models.5 The US position is that all AI commitments must be voluntary.

The three CEOs are not a unified bloc — Altman and Amodei have significant policy disagreements on safety timelines and government access. But both have EU regulatory exposure under the AI Act and data center operations in Europe. Being absent while G7 governments write governance language about their companies is a worse outcome than attending.

The communiqué will almost certainly contain non-binding language reaffirming OECD AI Principles, voluntary incident-sharing, and growth-oriented framing. The EU will issue a separate declaration on data center disclosure. No mandatory rules will emerge. But the summit confirms that the EU's enforcement trajectory — the AI Act's high-risk system requirements apply from August 2, 2026 — is not contingent on US agreement.7

What to watch: whether the communiqué includes any language on model export coordination, which would signal the start of a formal multilateral process rather than unilateral US action as seen with Fable 5 this week.

OpenAI shut down Sora's web and app on April 26, 2026, with the API set to close September 24. The economics behind the decision: the product burned an estimated $15 million per day in compute costs against $2.1 million in total lifetime revenue, according to reporting by Crescendo AI and MiraFlow cited in an industry summary published June 15.8 Active users had fallen from over one million downloads in the first week to under 500,000.

The failure was operational as much as financial. Extreme generation latency frustrated professional creators who needed to iterate quickly. Persistent physics errors — objects disappearing mid-scene, items moving through space incorrectly — made the $200/month Pro tier hard to justify for high-stakes production. Legal exposure accumulated after Washington Post testing showed the model could closely mimic Netflix shows and TikTok clips. When asked directly whether YouTube videos were used in training, OpenAI's then-CTO Mira Murati could not confirm or deny.

Disney had signed a deal to license hundreds of branded characters for virtual avatars; that partnership collapsed with the shutdown. OpenAI is redirecting freed compute to an internal model codenamed Spud and to enterprise productivity infrastructure ahead of its anticipated September 2026 IPO.

Google's Veo 3.1, Runway's Gen-4.5, and Kling 2.5 are the primary beneficiaries of Sora's developer and enterprise exodus. Teams using Sora's API have until September 24 to migrate.8

The broader lesson: at current compute costs, consumer-priced video generation does not cover the bill. No AI video product has solved this yet.

On May 21, the US Department of Health and Human Services launched the AERO program (Audit Enforcement and Risk Oversight), using ChatGPT and other AI tools to perform rolling analysis of at least five years of annual audit reports from every state, local government, nonprofit, and higher education institution that spends at least $1 million in federal funds per year.8

HHS Assistant Secretary Gustav Chiarello, who runs the program, estimates HHS has between $100 billion and $200 billion in wasteful or fraudulent spending annually.

The procurement approach is unusual: HHS is using off-the-shelf ChatGPT rather than a federally procured custom tool, applied to already-public audit data, bypassing the standard federal acquisition process. Critics have raised two concerns: AI tools make errors when reading complex financial documents, and early enforcement patterns have disproportionately targeted Democratic-administered states. HHS disputes both characterizations.8

What it means: this is the largest-scale AI deployment in federal auditing history, and it is operating right now. States with pending audit deficiencies should assume AI-assisted review is underway.

NVIDIA and NAVER announced on June 7 that NAVER will expand its sovereign AI infrastructure at its GAK Sejong data center using the NVIDIA DSX platform.9 The buildout starts at 55 megawatts in the first half of 2027, scales to 100 MW by late 2027, 200 MW by 2028, and ultimately reaches gigawatt capacity.

NAVER will use the infrastructure for its next-generation HyperCLOVA X models, a "Seoul World Model" combining NVIDIA's Cosmos world foundation model with NAVER's street-view and 3D spatial data, and a commercial AI agent platform targeting Korea, Europe, and the Middle East. NAVER's stock rose 9.2% the following day, to 279,000 Korean won.10

The strategic logic is sovereign AI in practice: South Korean government and enterprise customers have strong preferences for AI infrastructure that is domestically controlled and compliant with Korean data sovereignty requirements. The Fable 5 shutdown this week made those preferences visceral — Tata Consultancy Services had announced a partnership with Anthropic the day before the export ban, only to see Claude go offline globally the next day.11

The NAVER-NVIDIA model — a trusted local entity plus a US compute stack, with contractually guaranteed domestic data processing — is being replicated across Europe and the Middle East. It may prove the durable template for countries that want frontier AI capability without the jurisdictional risk of depending on a US-hosted API that a US government agency can disable in 90 minutes.