SK Telecom invested $100 million in Anthropic, helped plan a multilingual model, and still, nearly three years later, reportedly faced losing access on White House instructions.

The deal assumed companies controlled access

In August 2023, Anthropic raised $100 million from SK Telecom, and the two companies said they aimed to build a telco-focused multilingual large language model. The partnership reflected the operating assumption of the period: capital and technical collaboration could secure frontier-model capability through private agreement.

Anthropic’s 2023 raise from SK Telecom

The structure was bilateral: SK Telecom supplied capital and joined a model plan; Anthropic supplied frontier-model capability. Whether that collaboration continued appeared to rest with the two companies.

SK Telecom followed the same commercial logic in February 2024 when it partnered with Perplexity to gain access to proprietary models and agreed to offer Perplexity’s paid AI search product to users for free. A model could be obtained through partnership, incorporated into a service, and distributed to customers. The permission boundary was the contract.

Sovereign ambition did not replace foreign partnerships

By January 2026, a team from SK Telecom had advanced alongside teams from LG and Upstage in South Korea’s competition to develop the country’s first sovereign AI model. That did not erase the carrier’s foreign partnerships. It added another role: SK Telecom was simultaneously an investor in a model lab, a buyer of proprietary-model access, a distributor of AI services, and a participant in sovereign-model development.

The first three phases could coexist: foreign investment and proprietary access did not preclude sovereign ambition. The June report changed the classification. It treated access not as the settled output of a partnership, but as a permission that could be withdrawn from above it.

The veto moved above the contract

The June 18 report is specific but not confirmed: sources said the White House ordered Anthropic to revoke SK Telecom’s access because of alleged ties to China. The evidence does not establish those ties, show that Anthropic completed a revocation, or establish a formal policy covering other customers.

But if the reported order is accurate, its structure is clear. Anthropic may choose a partner, accept its capital, and plan a model with it, while the U.S. government retains the practical ability to countermand the resulting access. Frontier-model distribution is no longer only a company-to-company decision: the contract opens the gate, but national-security judgment can close it.

South Korea’s status as a U.S. ally makes the distinction sharper, not weaker. This was not reported as a blanket restriction on South Korean access. The stated concern was alleged ties to China. Eligibility therefore no longer follows only the customer’s home country or the supplier’s willingness to serve it; it can follow the customer’s wider network of relationships.

The reported follow-on move involving Mythos 5 extends the same logic from a named customer toward a named model. The report does not specify the restriction’s scope or mechanism, so it cannot support a broader claim about who lost access or under what rule. Its sequence still matters: first a carrier relationship was reportedly subjected to state review, then a model was reportedly restricted. The object being governed was no longer only the partnership. It was the capability moving through it.

A reported veto is not a regime

Neither reported action is confirmed policy. One source-based account cannot establish a standing export-control system, and SK Telecom’s history supplies real counterevidence to any claim that cross-border AI partnerships had already become impossible. Its investment in Anthropic, joint-model plan, and later Perplexity partnership remained commercially viable across nearly three years.

The reversal is narrower and more consequential. In 2023, commercial agreement was treated as the governing condition for access. In the June reports, it becomes necessary but insufficient. A separate geopolitical judgment can override it, even when the customer comes from an allied country and after capital and model-development plans have crossed the boundary.

That is conditional model access, not a general prohibition. The model remains commercially available until another layer decides that this customer, relationship, or capability falls outside the permitted path. Companies can negotiate terms they control, but they cannot conclusively settle permissions held elsewhere.

Centralization turned the supplier into a checkpoint

The two episodes used the same supplier relationship in opposite directions. In 2023, it carried capital and a shared model ambition across a border. In June 2026, the reported order used Anthropic’s control over access to reach back across that border.

The earlier partnership was not false; the governing question changed. The first structure asked whether two companies could work together. The second asked whether the capability they shared remained acceptable to U.S. national-security authorities. The same centralized supplier relationship that made access commercially negotiable also made intervention structurally simple: one instruction to the supplier could reach a customer across the partnership boundary.

In 2023, $100 million and a multilingual model plan put the permission boundary between two companies. By June 18, 2026, sources placed that boundary at the White House.