Anthropic’s Clash With Trump Administration Strengthens Case for ‘Crypto x AI’

This alert was originally sent directly to clients of Galaxy Trading and Galaxy Asset Management on June 15, 2026. Trade or invest with Galaxy to receive the most timely research directly in your inbox.

What Happened

On June 9, Anthropic released Claude Fable 5, the first publicly available model for its newest Mythos class (see more in-depth coverage from our weekly newsletter here). On Friday, the U.S. government issued an export control directive ordering Anthropic to suspend Fable 5 and Mythos 5 access for all foreign nationals. That evening, Anthropic disclosed the order and disabled Fable 5 and Mythos access to all customers.

According to Anthropic’s statement, the order “did not provide specific details of its national security concerns. Our understanding is that the government believes it has become aware of a model of bypassing, or 'jailbreaking' Fable 5.” Anthropic goes on to say it was previously aware of these jailbreaking methods and that other models like GPT 5.5 have the same level of capabilities, that no universal jailbreak method exists, and that while it believes the government should have the power to block unsafe deployments, the process must be “transparent, fair, clear, and grounded in technical facts.” Subsequent reporting by presidential advisor David Sacks and media outlets indicates that Amazon (an Anthropic investor, but also arguably a competitor) may have identified the jailbreak and advocated for the U.S. government to implement the controls, and that the government disagrees with Anthropic’s characterizations of the jailbreak’s severity.

The incident has provoked strong reactions given the implications for future accessibility to frontier models and the urgent need for clear regulations on AI policy. It is likely just the beginning of what will become a much more consequential debate in the years to come as the power of frontier labs like Anthropic and OpenAI become more controversial.

Crypto AI Tokens Outperform

Immediately following Anthropic’s disclosure, crypto tokens in the AI vertical, especially those related to decentralized training, rallied. The strong price performance was largely driven by speculators who viewed Friday’s events as a tailwind for decentralized AI projects that attempt to mitigate interference in model development and accessibility.

Over the past year, Galaxy Research has covered the convergence of crypto and AI in depth. Most relevant to Friday’s developments, in September we published “Decentralized AI Training: Architectures, Opportunities, and Challenges,” highlighting the progress of leading projects in making decentralized training a reality. As noted in that report, we expected these projects to receive greater attention “if centralized incumbents try to stifle innovation by keeping weights closed or injecting unwelcome alignment biases.” Developments since Friday might well be such an exogenous catalyst.

On X (formerly Twitter), commentators quickly framed the event as validation of the need for decentralized and permissionless AI, pointing to the structural risk it exposed. Access to a centralized model can be cut by a third party with little notice or recourse. While short-term price moves should be treated as just that, short-term, ensuring the ability to train, deploy, and access models is exactly what a large cohort of crypto projects including Bittensor, Chutes, Dolphin, Pluralis, Nous Research, Covenant, Macrocosmos, and Gensyn aim to do.

Since we published our piece on decentralized training in September, several new successful training runs have demonstrated continued progress in enabling the development of frontier models through decentralized and permissionless mechanisms. Pluralis is running an 8 billion-parameter open training run that can be observed in real time here. While promising, these projects still face significant technical limitations in reaching parity with frontier and open-source providers. Additionally, these projects still face many of the same headwinds laid out in our last piece, including figuring out proper incentivization/monetization mechanisms and the risk of becoming a regulatory target for facilitating unsupervised model development and deployment.

Beyond the decentralized training space, a much wider spectrum of projects has emerged in recent years targeting other growing AI verticals. We have written extensively on the emergence of agentic payments, zero-human companies, and agentic capital markets. In recent months, inference solutions such as Venice, founded by longtime crypto entrepreneur Erik Voorhees, have made progress toward product-market-fit thanks to their differentiated privacy offerings and shifting AI model usage favorable to providers of cost-sensitive open-source models.

Friday’s drama showed that access to a closed frontier model can be revoked by a third party who isn't the buyer, in hours, with no recourse. The first-order hedge against that risk is open weights, which require no crypto at all. A team can self-host Llama, DeepSeek, or GLM and remove the off switch. Decentralized AI projects that incorporate crypto rails have additional protections. In designs like Pluralis's Protocol Models, no participant ever holds the complete weight set, so no party is left who can realign, alter, or revoke the model after the fact. Venice’s privacy offerings shield users from having their personal information harvested. Broadly, crypto and AI projects leverage blockchains and crypto tokens to tackle issues like censorship resistance, permissionless access, privacy, and capital formation/monetization. That distinction is what separates a bet on open-source AI broadly from a bet on the crypto-native training networks specifically.

In terms of real adoption, inference access layers are likely to be the primary beneficiaries rather than training networks that remain orders of magnitude behind the frontier. Decentralized training is a multi-year development, and it pays off only if these networks close that capability gap while solving the incentive, verification, and distribution problems that have nothing to do with Friday's news. What Anthropic’s run-in with Washington does do is make clear that decentralized training is not just an ideological experiment and it gives the open ecosystem its clearest external catalyst to date.

Broader Implications

The implications of Friday's export controls go far beyond crypto. This will likely be remembered as a defining moment in the erosion of frontier labs' independence from government interference. The clash is not surprising. Anthropic already had a high-profile showdown with the Department of Defense at the beginning of this year. Alongside the release of Fable, Anthropic's CEO, Dario Amodei, also published a piece describing an urgent need for policy to catch up with the pace of AI development. And two of the most prescient pieces on the direction of AI development, "Situational Awareness" and ai-2027.com, both forecast that frontier labs would inevitably come into conflict with the state and end up under some form of government control, whether through forced consolidation, oversight committees, or defense-contracting arrangements rather than outright nationalization.

If a competitor can both flag a rival's model as a security concern and sit at the table where trusted partners are chosen, safety review can become a competitive instrument.

Additionally, on June 2, the Trump administration signed an executive order, “Promoting Advanced Artificial Intelligence Innovation and Security,” that created a new designation called a "covered frontier model," defined through a classified NSA process that benchmarks a model's cyber capabilities. The order set up a “voluntary” framework under which developers would give the government early access to covered models before release and help select the "trusted partners" who receive early access. Notably, Section 3(c) states that nothing in the order creates a mandatory licensing or preclearance requirement to release a model. The order's "trusted partners" provision is where it intersects with the reporting on Amazon. The framework gives the government discretion, in collaboration with industry, to decide which firms get early access to covered models. If a competitor can both flag a rival's model as a security concern and sit at the table where trusted partners are chosen, safety review can become a competitive instrument.

The developments also occur in the broader context of a growing clampdown on AI token spending by major consumers and rising use of open-source models. In a note from Citadel Securities last week, analysts mapped shifting "Tokenomics," pointing to Amazon pulling its token leaderboard, Microsoft cancelling Claude Code subscriptions, and a run of outsized token bills. A recent decline in the Silicon Data LLM Token Expenditure Index, a benchmark for the effective price of LLM usage, shows users shifting to cheaper models wherever they don't need frontier ones.

The correction underway is being implemented through routing: Reserve use of frontier models for the 10%-20% of work that needs planning intelligence and send the rest to open models. Coupled with these regulatory developments, that shift could put a structural ceiling on closed-frontier demand. These two pressures point the same direction. From above, the government is asserting the right to decide who runs a frontier model. From below, the largest buyers are trimming frontier consumption and routing everything that doesn't require frontier capability to cheaper open-weight models. The first makes closed-model access politically contingent. The second makes it economically optional for most workloads.

Taken together, they weaken the premise that a handful of closed frontier APIs will serve as the default substrate for building with AI. While those APIs will no doubt remain some of the largest and most critical providers for inference, open-weight model providers and the cost-competitive inference layer that serves them are increasingly benefitting from reduced token spends. The question now is whether that advantage extends to permissionless, crypto-native infrastructure. However, these networks will first need to approach open-source and frontier capability without becoming the regulatory target the closed labs just became.

The episode also sets a precedent worth watching as frontier models are starting to be treated as export-controlled technology, closer to munitions than software, which points to higher compliance costs, slower releases, and more selective deployment across the leading labs. Each of those effects raises the cost of operating at the closed frontier, and each widens the same opening that open-weight and decentralized alternatives are positioned to fill.

Closing Thoughts

On Friday, Galaxy's Head of Firmwide Research Alex Thorn published a piece on bitcoin's four-year cycle arguing it may not have bottomed yet. Whatever the cause of a downturn, these are the stretches when the verticals that drive the next cycle get re-underwritten, before the rest of the market is looking. Most of that attention today is going to the convergence of crypto and TradFi, and to projects like Hyperliquid that have held their fundamentals through the drawdown.

In my own view, crypto x AI projects remain one of the most overlooked verticals most likely to benefit from narrative and real fundamental adoption in the years ahead. Venice has pulled some attention back to the space, but as the market map above shows, it sits inside a far larger surface area including decentralized training and inference, agentic payments, compute markets, and the early agentic capital markets we've covered elsewhere.

The projects that come out of bear markets strongest are the ones that use crypto's settlement rails to build permissionless markets that couldn't exist any other way. Decentralized training is one of those markets. So are private/permissionless inference and agentic payments. Rather than moving old activity onchain, each creates new economic activity, which is what turns a narrative into adoption.

Friday was a preview of the conditions these projects were built for. By the time that's widely understood, the quiet part of the cycle will be over.