DeepSeek Pushes Forward with In-House AI Chip Development — The Wave of Homegrown Hardware
Chinese AI startup DeepSeek is developing its own AI chip, according to a Reuters report dated July 7. Bloomberg confirmed the report the same day, citing anonymous sources.
The chip under development is reportedly designed for **inference** rather than the training phase. In other words, the aim is to make it more efficient for already-trained models to generate responses to user queries. DeepSeek has drawn attention for offering high-performance models such as V4 at low cost, and this move can be seen as an attempt to cut inference costs even further at the hardware level.
As the U.S. continues to tighten export controls on AI technology to China, Chinese AI companies are accelerating efforts to reduce their dependence on NVIDIA chips. How far along DeepSeek's chip design is, along with mass production timelines and performance details, remains unknown at this point — but the strategy clearly points toward self-sufficiency in both software and hardware.
Amazon Returns to the U.S. Bond Market — Major Fundraising for AI Infrastructure Buildout
Amazon has re-entered the U.S. bond market to raise funds for building out AI infrastructure. According to Bloomberg's July 7 report, the company is raising capital through bond issuance to construct AI data centers and expand its computing infrastructure.
Amazon's appetite for AI investment is exceptionally strong, signaling the company's long-term commitment to the AI infrastructure race. Expanding AWS computing capacity is an essential investment to keep up with the explosive growth in generative AI demand. That said, capital raised through bond issuance comes with repayment obligations, so whether AI infrastructure generates the expected returns will be the key question going forward.
NVIDIA Releases "Nemotron-Labs-3-Puzzle-75B" — A Hybrid MoE+Mamba Design
NVIDIA has released a new large language model, "Nemotron-Labs-3-Puzzle-75B-A9B," on Hugging Face. Reported on July 7 in the LocalLLaMA community on Reddit, the model is based on Nemotron-3-Super-120B-A12B with a post-training compression framework called **Iterative Puzzle** applied.
The model's main features are as follows:
Hybrid architecture**: Adopts a design that alternates layers of Mamba, MoE (Mixture of Experts), and Attention
Parameter configuration**: 75B total parameters, 9B active parameters (A9B)
Purpose**: Improved efficiency for interactive inference, inference-heavy tasks, and long-context workloads
Dramatically improves inference efficiency while preserving the accuracy of the existing 120B model as much as possible
NVIDIA is best known as a GPU maker, but it is also taking a distinctive approach to model development. The Mamba+MoE design follows a recent trend, apparently aimed at balancing long-document processing with efficiency.
Savi's AI Anti-Scam App — Protecting Consumers from AI Voice-Clone Scams
According to a July 7 TechCrunch report, startup Savi has announced an app that protects consumers from realistic AI-driven scams, particularly **virtual kidnapping scams that demand ransom using AI-generated voices**. The company has raised $7 million (about 1 billion yen) in seed funding and released apps for iPhone and Android.
With advances in AI voice cloning technology, scams that mimic a family member's voice to demand ransom have become a real-world threat. Savi's app provides mechanisms to detect and prevent these AI scams and is drawing attention as a defense against the malicious use of AI technology.
While the benefits of AI continue to spread, countermeasures against its misuse have become urgent, and the fact that investment in startups in this space has begun is a telling sign.
AI Legal Startup Norm Raises Funds at a $1.2 Billion Valuation
According to Bloomberg, AI legal startup Norm has carried out a funding round at a valuation of $1.2 billion (about 180 billion yen). It is a symbolic example showing that AI is making serious inroads into the legal industry.
In the legal field, moves to use AI to streamline routine work such as contract review, legal research, and compliance checks are gaining ground. A valuation above $1 billion suggests that investors see significant growth potential in the legal AI market.
A 3M-Parameter Transformer Installs Unseen Rules at Inference Time — Fully Reproducible on a Single RTX 3090
In a study attracting attention in the LocalLLaMA community on Reddit, an independent researcher has unveiled a method that uses a transformer with just 3M parameters to install rules the model never learned at inference time.
The key to this research is a mechanism called a "**fast-weight memory bank**". By writing and reading vectors within the model's own forward pass, the approach — distinct from test-time training — reportedly enables the model to apply rules that were not included in its training data at inference time.
Notably, the research is **fully reproducible on a single RTX 3090**. It shows that independent researchers without large-scale computing resources can still explore innovative architectures — a symbolic example of how AI research is becoming democratized. A preprint is currently available on Zenodo, with an arXiv release also planned.
---