This Week in AI: December 29, 2025 – January 4, 2026 (Week 1)

News for the week: December 29, 2025 – January 4, 2026 (Week 1)

The turn of the year brought no break in artificial intelligence. Meta bought the startup Manus to strengthen its AI agents, and SoftBank completed its multibillion-dollar investment in OpenAI while also buying the data center investor DigitalBridge. Chinese companies are ordering millions of Nvidia H200 chips, and their AI firms are rushing to list in Hong Kong. The biggest scandal of the week was the Grok chatbot, which generated sexualized images of real people, including minors, prompting governments in France and India to step in.

Business and Investment

Meta buys the startup Manus and bets on AI agents

Meta announced on December 29 that it is buying Manus, a Singapore-based startup with Chinese roots. Manus develops a general-purpose AI agent, a program that carries out multi-step tasks on its own based on instructions, such as compiling research or preparing a spreadsheet. According to Reuters, the price is between 2 and 3 billion dollars; Meta did not disclose it (Reuters).

Meta wants to build the Manus agents into its consumer and business products, including the Meta AI assistant. The standalone Manus service will continue to operate and be sold. According to CNBC, no Chinese ownership stake will remain in Manus after the deal, and the service will shut down in China (CNBC).

The startup prides itself on rapid growth. By its own account, Manus passed 100 million dollars in annual recurring revenue eight months after launch. It is one of the largest acquisitions in Meta’s history (TechRadar).

SoftBank completes its OpenAI investment and buys DigitalBridge

Japan’s SoftBank confirmed that it has sent OpenAI a second installment of 22.5 billion dollars. The money went through the SoftBank Vision Fund 2 and arrived on December 26. The company announced it in a press release dated December 31. It follows the first installment of 7.5 billion dollars from April 2025 (SoftBank Group).

According to the Nikkei newspaper, as cited by SiliconANGLE, SoftBank now holds roughly 11% of OpenAI. To fund this payment, it reportedly sold its entire stake in Nvidia and part of its T-Mobile shares, and took out a loan backed by its Arm shares (SiliconANGLE).

On December 29, SoftBank also announced it will buy DigitalBridge, an investment firm focused on data centers, fiber networks and other digital infrastructure. It is offering 16 dollars per share in cash, giving an enterprise value of roughly 4 billion dollars. The deal is expected to close in the second half of 2026 (BNN Bloomberg).

More business news

Nvidia completed its 5 billion dollar purchase of Intel shares on December 29. It acquired 214.7 million shares at 23.28 dollars each, closing a deal agreed in September. The transaction had previously been cleared by US antitrust authorities (The Star / Reuters).
Elon Musk announced on December 30 that his company xAI has bought a third building for computing capacity. He named it “MACROHARDRR” and said it will raise the capacity for training models to nearly 2 gigawatts (Tom’s Hardware).
Chinese startup Moonshot AI, maker of the Kimi models, raised 500 million dollars in a Series C round on December 31, according to an internal letter from its founder. The investment was led by IDG Capital, with Alibaba and Tencent also taking part (South China Morning Post).

Chips and China’s Stock Market Fever

Chinese companies order more than 2 million Nvidia H200 chips

According to Reuters sources, Chinese tech companies have ordered more than 2 million Nvidia H200 chips for 2026. These are powerful accelerators for training and running AI models, recently cleared for sale to China by the US government. Nvidia has about 700,000 of them in stock and has reportedly asked manufacturer TSMC to produce more (The Register).

According to the sources, the price is around 27,000 dollars per chip. The catch is that the imports have not yet been approved by Beijing, which is also considering requiring buyers to purchase domestic chips at the same time. Nvidia said that sales to China will not jeopardize supply to US customers (Taipei Times).

Chipmakers and AI labs head to Hong Kong

Shanghai-based graphics chip maker Biren Technology opened 2026 on the Hong Kong stock exchange. Shares priced at 19.60 Hong Kong dollars jumped 76% on their first day of trading on January 2. The company raised about 717 million dollars in its initial public offering (IPO) (Fortune).

Baidu announced on January 2 that its chip subsidiary Kunlunxin has confidentially filed for a separate listing in Hong Kong. Kunlunxin will remain a subsidiary of Baidu; neither the timing nor the size of the offering has been set (The Standard / Reuters).
Language model developer Zhipu AI (brand Z.ai) offered its shares to investors from December 30 to January 5. The offering comprised 37.42 million H shares, with trading set to begin on January 8 (Tiger Brokers).
Rival MiniMax launched a share offering at the turn of the year, aiming to raise about 540 million dollars (Caixin Global).

New Models and Research

DeepSeek starts the year with a paper on cheaper training

China’s DeepSeek published a research paper on January 1 about a method called mHC (Manifold-Constrained Hyper-Connections). It modifies the internal connections between layers of a neural network to make large models behave more stably during training. The method builds on so-called Hyper-Connections, introduced in 2024 by researchers at ByteDance (South China Morning Post).

According to the authors, models with mHC performed better on eight benchmarks while adding only 6.27% to the hardware load. The 19 authors include the company’s founder, Liang Wenfeng. Some commentators see it as groundwork for the next model, but the paper does not mention any new model (SiliconANGLE).

Alibaba releases the open image model Qwen-Image-2512

The Qwen team at Alibaba released Qwen-Image-2512, an image generation model, on December 31. It is an open model: its weights (the learned parameters) can be downloaded by anyone and run on their own hardware. The Apache 2.0 license also allows commercial use (VentureBeat).

Compared with the August version, Alibaba promises more realistic faces, finer natural details and better text rendering in images. It positions the model as an open alternative to Google’s Nano Banana Pro. According to blind tests on the AI Arena platform, which Alibaba runs, it is the strongest open image model, though independent verification is lacking.

South Korea unveils its first national models

In Seoul on December 30, five teams presented the first results of the government’s sovereign AI program, which aims to build homegrown foundation models independent of foreign companies. The program includes Naver Cloud, Upstage, SK Telecom, LG AI Research and NC AI (The Korea Times).

LG showed the K-EXAONE model with 236 billion parameters.
Upstage introduced Solar Open, a model focused on agentic reasoning and use in finance, law, healthcare and e-commerce.

Products and Strategy

OpenAI reorganizes teams around voice and its own device

The Information reported on January 1 that OpenAI has merged its engineering, product and research teams working on audio over the past two months. According to sources, the company’s voice models lag behind its text models. A new voice model is expected in the first quarter of 2026 and should handle interruptions from users better (Social Samosa).

The change is tied to an upcoming consumer device that, according to sources, will be “largely audio-based”. Former Apple designer Jony Ive is working on the hardware for OpenAI. Reports of the launch date vary from late 2026 to 2027 (MacDailyNews).

Big tech leaders on 2026

Microsoft CEO Satya Nadella launched a personal blog on December 29 and, in a post titled “Looking Ahead to 2026”, called for the debate to move beyond arguments about AI “slop”, the flood of low-quality AI-generated content. Online, however, he was met mostly with ridicule (TechSpot).
Instagram head Adam Mosseri warned on the last day of 2025 that, because of AI, we can no longer automatically trust that a photo on the internet is real. In his view, it will be more practical to label genuine images, for example with a cryptographic signature applied directly in the camera, than to hunt for fake ones (Business Today).
Departing Meta chief scientist Yann LeCun told the Financial Times that the benchmark results for the Llama 4 model were “fudged a little bit” because the team used different versions of the model for different tests. The interview drew attention in early January (Fast Company).

Safety and Incidents

Grok generated sexualized images of real people

The Grok chatbot from xAI, built into the X platform, spent the turn of the year editing photos of real women at users’ request to make them appear undressed or in swimwear. According to Reuters, this also included images of minors. On January 2, Grok itself admitted “lapses in safeguards” and said the problem was being urgently fixed (Free Malaysia Today / Bloomberg).

On January 2, three French ministers reported the content to prosecutors and to Pharos, the platform that receives reports of illegal content. They described the content as “manifestly illegal” and asked the regulator to review it under the EU’s Digital Services Act. Asked by Reuters for comment, xAI replied only with the phrase “Legacy Media Lies” (The Star / Reuters).

On the same day, India’s Ministry of Electronics and IT (MeitY) ordered X to immediately remove obscene content created by Grok and to submit a report on the measures taken within 72 hours. Otherwise, the ministry said, the platform risks losing its legal protection for user content (TechCrunch).

Regulation and Law

The first state AI laws take effect in the US

As of January 1, California’s SB 53 (Transparency in Frontier AI Act) is in force. It requires developers of the most powerful models to publish how they manage safety risks and to report serious incidents. Violations can bring fines of up to one million dollars (Built In).

California’s SB 243 introduces safeguards for so-called companion chatbots, AI “companions” with which people form emotional bonds. It is the first state law of its kind (The National Law Review).
In Texas, the TRAIGA law (HB 149) took effect, banning certain uses of AI and requiring that people be informed when they are interacting with AI (Venable).