Frontier labs lock infrastructure, expose reasoning vulnerabilities

This week exposed a widening gap between frontier AI labs' infrastructure ambitions and the fragility of their safety mechanisms. Anthropic's $9.1 billion compute deal with Riot Platforms and OpenAI's 750-token-per-second Ultrafast tier on Cerebras hardware signal an acute competition for power-constrained infrastructure, forcing labs into long-term commitments with non-traditional partners. Yet simultaneously, researchers demonstrated that reasoning transparency features deployed by Anthropic, OpenAI, and Google create a direct extraction vulnerability: encrypted chain-of-thought traces can be decrypted by replaying them into weaker model variants, exposing the internal reasoning labs intended to make auditable. Anthropic's disclosure of Claude's text watermarking technique addresses content provenance but leaves unanswered the robustness question: watermarks remain fragile against paraphrasing and light editing, with no published adversarial testing results. Meanwhile, Hugging Face's reproduction of 2,200 ICML papers surfaces systematic reproducibility gaps in published ML research, undermining practitioners' ability to make informed adoption decisions. The pattern is clear: as frontier labs scale deployment and promise transparency, the infrastructure securing both compute and interpretability is proving more brittle than the public narrative suggests.