NTT Data in talks to build Anthropic's first Japan data center
The overnight inbox is an infrastructure story dressed as a rumor, plus two research-side reads: a paper arguing capability audits have a cryptographic hole, and Ben Recht taking a hatchet to OpenAI's math dump.
NTT Data is in advanced talks to build a data center for Anthropic in Osaka with capacity of up to 500 megawatts — among the largest in Japan, with construction costs potentially reaching 1 trillion yen ($6.3 billion). Nikkei Asia first reported the talks late Sunday, and Reuters confirmed Monday that they are in the final stages, citing multiple sources, with partial operations targeted as early as 2028 and an agreement aimed at by the end of Japan's fiscal year in March. NTT Data has declined to comment, and no deal is signed. Why it matters: it would be Anthropic's first dedicated facility in Japan, and at 500 megawatts it puts Anthropic in the same league of single-site commitments as the biggest US hyperscale builds — a striking step for a lab that has leaned on cloud partners rather than owning its own bricks. The timing is also a tell: whoever locks in East Asian capacity first gets optionality on power and latency as demand migrates toward always-on agents.
A NeurIPS 2026 paper says AI models can provably hide arbitrary capabilities — and that breaks a core assumption behind capability evaluations. The authors, Andis Draguns, Sumeet Ramesh Motwani, Raymond Douglas, and Christian Schroeder de Witt, construct backdoors that embed encrypted neural networks inside a host model's weights, decrypting and running them only when a secret trigger arrives; under cryptographic assumptions the hidden circuits stay provably hard to elicit or interpret even with full access to the weights. Their proof of concept runs in PyTorch on a small chemical-reaction prediction task — a proxy for CBRN-relevant capabilities — and resisted fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training. It extends the same lead author's earlier work hiding fixed strings to hiding arbitrary computations, which is exactly the jump that matters for auditors: if white-box access doesn't guarantee you can surface what a model knows, elicitation-based audits need a different foundation. The authors are releasing their models to help build defenses.
Ben Recht's verdict on OpenAI's claimed faster Fourier transform: "less a mathematical breakthrough and more an oddity of computer scientists being bad at formalizing their design specifications." The claim came out of OpenAI's huge math repository dump — a paper claiming a discrete Fourier transform in O(n(log n)^(1−ε)) operations — and Recht argues it is a "galactic algorithm" in the Lipton–Regan sense: his estimate puts the hidden constant around 10^100 at n = 10^12, meaning Cooley–Tukey would still be some 10^99 times faster. The deeper problem he identifies is the model of computation: the algorithm assumes exact complex arithmetic with magnitudes growing superpolynomially in n, so a single "operation" multiplies numbers no real computer can store. Recht frames it as the latest rot from a theory culture that stopped running algorithms on computers — and notes the same critique lands on the dump's O(n^2.25) matrix multiplication result. One caveat worth keeping in mind: the FFT paper's result is machine-checked in Lean, so the dispute is about what the model of computation has to do with reality, not whether the proof type-checks.
What to watch: whether the NTT talks produce a signed deal by Japan's fiscal year-end in March — that would be the clearest signal yet of Anthropic building out East Asian capacity of its own.
If capability audits can be cryptographically bypassed, what should regulators require instead of evaluation results? Tell us in the comments.
