NVIDIA's Upcoming "Vera" CPU to Use SK Hynix Memory — Major Partnership Expansion Ahead of the Second Half
NVIDIA CEO Jensen Huang had dinner in Seoul on June 7 with SK Group Chairman Chey Tae-won, SK Hynix CEO Kwak Noh-jung, and SK Telecom executives, and made some intriguing remarks at the press conference that followed.
He revealed that "Vera", the new CPU NVIDIA is developing, will use SK Hynix DRAM. "The Vera CPU is a revolutionary CPU, and it will also use SK Hynix DRAM," he said. Vera is a CPU of NVIDIA's own design and is expected to be positioned as the successor to the existing Grace CPU.
Huang went on to say, "Our business with SK Hynix has been huge this year, but we are preparing for something even bigger in the second half of this year and into next year," suggesting that the two companies' partnership is evolving into a deep relationship that goes well beyond simple memory supply.
These remarks signal that NVIDIA increasingly recognizes the importance of memory partnerships within the AI semiconductor ecosystem, and they are expected to be a tailwind for SK Hynix's HBM (high bandwidth memory) business as well.
Bloomberg Warns of an "AI Death Spiral" — The Threat to the Internet
Bloomberg has released a video titled "Why an AI Death Spiral Threatens the Internet." It points to the danger of a "death spiral": AI-generated content spreads rapidly across the internet, and AIs ingest each other's outputs as training data, progressively degrading content quality.
The problem of declining web traffic as AI search goes mainstream has already been widely noted, but the video appears to focus on the more fundamental impact on the internet as a whole caused by the feedback loop of "AI output becoming AI input."
arxiv Paper Analyzes LLM Inference Price Evolution and Market Competition
A paper titled "Price Evolution, Production Frontiers, and Market Competition in LLM Inference" has been published on arxiv. It offers an academic analysis of price trends, production frontiers, and the structure of market competition in LLM (large language model) inference services, shedding light on the economic mechanisms behind the sharp decline in LLM API prices.
LLM inference costs have fallen dramatically over the past year, and the paper systematically lays out the technological progress and competitive factors behind this. It reportedly includes a discussion of where the rise of open-source models, advances in inference optimization, and price competition from new entrants will push the market's overall equilibrium.
Show HN: WibeOS — An OS Where an LLM Hallucinates Apps into Existence
A project called "WibeOS" was featured in Hacker News's Show HN section. The OS is built on an ambitious concept: when a user requests an app, the LLM generates (hallucinates) the application on the spot and runs it.
The project is available on GitHub, and the approach of integrating an LLM at the OS level makes for a fascinating experiment when thinking about future computing paradigms. At this point the project stands out more for the novelty of its concept than its practicality, but it deserves attention as a glimpse of what an AI-native OS could look like.
---