DeepSeek V4 Official Technical Report Released — Full Benchmark Picture Revealed
DeepSeek has released preview versions of the **DeepSeek-V4** series on HuggingFace, along with the complete technical report. The V4 series consists of two models: **Pro** (1.6T parameters, 49B active) and **Flash** (284B parameters, 13B active), both supporting a **1 million token context length**.
On the architecture front, the standout feature is a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). In the 1M token context setting, V4-Pro reduces **single-token inference FLOPs to 27% and KV cache to 10%** compared to V3.2. Training efficiency has also been improved through enhanced residual connections via Manifold-Constrained Hyper-Connections (mHC) and the adoption of the Muon optimizer.
The benchmark results are overwhelming for an open-source model, recording MMLU 90.1, LiveCodeBench 93.5, and a Codeforces rating of 3206. In coding in particular, it shows performance on par with Opus-4.6 and GPT-5.4, and it achieved 80.6% on SWE Verified for agentic tasks. That said, in areas such as factual accuracy on SimpleQA and long-context benchmarks, it still falls short of the closed-source state of the art.
While the basic specifications had already been reported in earlier coverage, this full report now reveals detailed performance comparisons across each domain. It is an achievement that significantly pushes forward the frontier of open-source LLMs.
Apple Agrees to $250 Million Settlement Over Undelivered AI Siri
Apple has agreed to a **$250 million (approximately 37 billion yen)** class action settlement for failing to deliver the promised AI-enhanced Siri for the iPhone. According to The Verge, Apple heavily promoted AI-powered Siri under its "Apple Intelligence" initiative in 2024, but the actual rollout was significantly delayed, leaving many users' expectations unmet.
The settlement covers users who purchased iPhones expecting AI features. Apple had put "a smarter Siri" front and center in its advertising and promotions, but implementation difficulties left many of the planned features undelivered.
This case is emblematic of a challenge facing the entire tech industry: leveraging AI features to drive product sales while the technical reality fails to keep up. In addition to paying the settlement, Apple may face calls for greater transparency around the delivery of future AI features.
US Power Infrastructure Limits Becoming a Bottleneck for AI Development — Guggenheim Chairman Warns
**Alan Schwartz**, Chairman of Guggenheim Partners, has warned that the limits of US power infrastructure are becoming a serious risk in the AI development race. In a Bloomberg Television interview, Schwartz noted that US electricity consumption, which had been flat for decades, is now surging at an unprecedented pace due to the electrification of the broader economy and **power demand from AI data centers**.
AI progress depends on expanding compute capacity, but the power supply supporting that compute is failing to keep up. While the data center construction boom continues, grid expansion and new power plant construction take years, and the supply-demand gap keeps widening.
Schwartz stressed that "investment in power infrastructure is essential to winning the AI race," suggesting that in the geopolitical competition over AI development leadership, delays in infrastructure buildout could prove fatal. Similar power constraints are becoming a key challenge for AI strategies in countries around the world, including Japan.
Super Micro Posts Strong Earnings — AI Server Demand Improves Margins
**Super Micro Computer** has announced strong quarterly results, showing improved profit margins on the back of robust AI server demand. For the quarter ending in June, the company guided to earnings of 65 to 79 cents per share excluding certain items, on revenue of $11 billion to $12.5 billion — above analyst expectations.
A major factor has been improved manufacturing cost management for AI servers: the company is building an efficient mass production system for servers equipped with expensive AI chips. Demand for AI servers with NVIDIA GPUs remains rock solid, and Super Micro's performance is being propelled by its position as one of the primary suppliers.
AI infrastructure demand goes hand in hand with data center power issues, and improved cost efficiency on the server side also contributes to lowering the total cost of ownership (TCO) of entire data centers. The AI hardware industry is expected to continue posting high growth.
Data Center M&A Accelerates — $30 Billion in Asia, Australia's Largest Deal
As investment in AI data centers accelerates worldwide, major M&A moves are coming one after another.
**Stack Infrastructure**, a data center company under **Blue Owl Capital**, is reportedly exploring a sale of all or part of its Asia operations (Australia, Japan, and Malaysia). According to Bloomberg, the deal is estimated at **more than $30 billion**.
Meanwhile, in Australia, **CDC Data Centres** has signed the country's largest data center contract, sending shares of its parent company Infratil soaring. CDC expects operating profit for fiscal 2027-28 to jump from the current roughly A$400 million to at least A$1 billion, underscoring the rapid growth of the data center business driven by AI demand.
Data center demand is inseparable from power constraints, and much of the investment tends to concentrate in regions with good access to renewable energy. The Asia-Pacific region is attracting particular attention, and increased data center investment is expected in Japan as well.