Powering Nvidia's AI Chips with a Next-Gen Nuclear Reactor — Valar Atomics Achieves First US Demonstration
California-based nuclear startup Valar Atomics has successfully demonstrated running Nvidia Blackwell AI chips on power from its next-generation reactor, the Ward 250. This marks the first time a next-generation reactor has been used to power AI chips in the United States.
The demonstration took place at the company's facility in Utah, where electricity generated by the reactor was used to temporarily host a website. While the amount of power generated remains "just a trickle," the two companies have reportedly agreed to jointly explore developing nuclear-powered AI systems.
As AI data center power consumption skyrockets worldwide, clean electricity from next-generation reactors has emerged as one of the most-watched solutions. Though small in scale, this demonstration is a significant step toward proving its feasibility.
Oracle Details the Risk Factors Behind Its AI Investments
Oracle has laid out the risk factors around its large-scale AI investments in detail in official documents. Its investor disclosures present multiple scenarios in which the company's bet on the AI market fails to play out as expected.
Specifically, the major risks cited include uncertainty over demand for its massive AI infrastructure investments, intensifying price competition with rivals, and shifts in the regulatory environment. It is relatively rare for an enterprise cloud provider to disclose the downside risks of its AI investments this candidly.
As concerns about an AI "bubble" grow in parts of the market, the fact that a major player itself officially recognizes these risks should serve as a healthy reality check for the entire industry.
Cloudflare Expands Its AI Traffic Control Option to All Customers
Cloudflare has extended its AI bot traffic control option for website operators to all customers. The feature, previously offered on a limited basis, is now open to all users, including those on the free plan.
Site operators can now allow, deny, or conditionally control AI crawlers' access to their content. With concerns mounting over AI companies harvesting content without permission, the move puts the choice back in the hands of content creators.
Cloudflare has been championing the slogan "Your Site, your rules," taking a clear stance on content sovereignty in the AI era.
Community Developer Single-Handedly Scales Gemma4-31B to 44B
A developer has drawn attention on Reddit's LocalLLaMA community after independently extending Google's Gemma4-31B model into a 44B-parameter (88-layer) model.
The developer adopted identity initialization based on the LLaMA Pro approach and added a Gemma4-specific `layer_scalar` modification to progressively scale the model from 60 to 80 layers, and then to 88. The model was then fine-tuned on Korean legal and STEM data, reportedly achieving good results in specific domains.
The report also includes interesting findings, such as that duplicating full-attention layers contributed more than duplicating sliding-attention layers. As a community-driven effort to scale a large model independently, it stands as an example of the potential of open-source AI.
"Senior SWE Bench" — A New Benchmark for Realistic Engineering Tasks
A new AI benchmark for software engineers, "Senior SWE Bench," has been released. Unlike existing benchmarks, its distinguishing feature is a focus on realistic feature-implementation tasks with incompletely defined specifications.
In real-world development, situations requiring the ability to "read between the lines" are far more common than tasks with perfectly clear requirements. This benchmark aims to evaluate AI agents' capabilities under such realistic conditions.
It is drawing attention as a new yardstick in coding AI evaluation — one that measures whether these tools are actually usable on the job.
Anthropic Japan Sales Rep Shares Practical AI Agent Workflows
At the Anthropic booth during AWS Summit Japan 2026, Shaboz Ibragimov, a sales representative at the company's Japanese subsidiary, presented how he uses AI agents in his own work.
He shared practical insights, including specific workflows and business processes that became more efficient thanks to AI agents. Having an AI provider's own employee publicly show "how they actually use it" offers a useful reference for other companies.
At the same time, a survey on AI adoption among managers was also released. While it made visible how much AI usage is affecting performance evaluations and promotions, it also brought issues into sharp relief, such as managers' own insufficient use of AI and a lack of proper training.
---