Nvidia's Next-Gen Rack "Kyber NVL144" Delayed by More Than a Year, Rubin Ultra Canceled

Nvidia's next-generation AI server rack, the Kyber NVL144, has been delayed by more than a year due to circuit board manufacturing issues, slipping to 2028, according to a report from research firm SemiAnalysis. The impact of the delay has been far-reaching, with Asian suppliers recording stock drops of up to double-digit percentages.

In addition, the higher-performance Rubin Ultra variant has reportedly been canceled. These delays could open an opportunity for AMD and Google to compete in the AI chip market. For data center operators that depend heavily on Nvidia's hardware roadmap, the search for alternatives may soon become unavoidable.

Reddit Takes On a New Battle Against AI Spam

Reddit is facing a new spam problem. Cases have surged in which brands post stealth marketing content on Reddit in the hope of getting their products and services mentioned by AI chatbots like ChatGPT and Gemini.

OpenAI and Alphabet have content licensing deals with Reddit, and their AI tools draw information from Reddit's forums to generate answers. As a platform frequently cited by AI chatbots as a "trusted source of information," Reddit has become a battleground for companies looking to boost their visibility with consumers through AI.

Reddit has laid out a policy of keeping the platform "authentic and safe in the AI era" and is stepping up efforts to detect and eliminate this kind of spam using its own AI technology. As AI becomes the mainstream way people search for information, how platforms safeguard accuracy is emerging as a challenge for the entire information ecosystem going forward.

Fable Submits First AI-Generated GPU Kernel to KernelBench

Fable, an AI research lab, has submitted the first AI-generated GPU kernel to the KernelBench-Mega benchmark. According to the benchmark's maintainers and the official leaderboard, the kernel is "the first real (and fastest) megakernel."

The result shows that AI systems are getting better at foundational AI research tasks—such as kernel optimization. In his Import AI newsletter, Jack Clark described it as a sign of "the beginning of an RSI (recursive self-improvement) loop." If AI becomes able to improve the very code it runs on, the entire development process could accelerate.

That said, this is a single result at this point and does not immediately translate into general-purpose automation. Rather than fueling excessive expectations, it should be viewed as steady progress.

Big Tech Flips from "AI Job Destruction" Warnings to Optimism

For the past few years, Big Tech companies have repeatedly warned about the potential for massive AI-driven job destruction. Recently, however, their stance has reportedly taken a sudden turn.

A number of tech executives have shifted to the view that AI will actually create jobs and give rise to new forms of work. Goldman Sachs has also published an analysis of "AI's impact on the labor market," presenting the perspective that it should be understood as a restructuring of tasks rather than simple job loss.

Some see this reversal as driven by companies' desire to ease pressure from regulators. To figure out what is really happening, we will need to track industry-wide hiring data and productivity metrics over the medium to long term.

"Bored of AI" — Consumer AI Fatigue Becomes Visible

Voices complaining of "AI fatigue" are growing louder. One technology blogger published a post titled "I'm just so bored of AI," which drew a huge response on Hacker News.

The post expresses discomfort with AI features being carelessly bolted onto every product, along with frustration at AI being consumed as a buzzword rather than for its real value. Some analyses suggest that brands championing "AI-first" are losing consumer support, and the limits of the "AI label" as a marketing strategy are beginning to show.

Meanwhile, analyses claiming that "the AI Big Crunch has begun" have also emerged, with some questioning the sustainability of the AI boom. Rather than denying the technical progress itself, the most reasonable view is that we are in a transition period from excessive expectations to practical application.

tftf: An Ultra-Lightweight Tool for Handling Giant Models That Won't Fit in VRAM

An open-source tool called tftf (Transforming Transformers) for working with giant transformer models has been released to the local LLM community.

Rather than loading the entire model into memory, the tool processes it tensor by tensor, enabling operations such as LoRA merges and format conversions even on models that exceed available VRAM and RAM. It is drawing attention as an approach that solves a practical pain point for developers doing fine-tuning, especially when working with very large models.

The developer admits that "almost all of the coding was done by AI" but released it as a tool they hope will be useful to the community. As running models in local environments advances, the importance of lightweight tools like this is expected to keep growing.

---