The Week in AI: December 22-28, 2025 (Week 52)

This week’s news: December 22-28, 2025 (Week 52)

Christmas week brought the biggest deal in Nvidia’s history: a license to Groq’s chip technology and the move of its top people, worth roughly 20 billion dollars according to CNBC. Chinese companies Z.ai and MiniMax released new open models focused on coding. OpenAI openly admitted that prompt injection attacks on AI browsers will probably never be fully eliminated. Regulators in Italy and China turned their attention to chatbots, while Alphabet and the Japanese government are investing billions in energy, chips and data centers.

New Models

Z.ai releases GLM-4.7 with a focus on coding (December 22)

Chinese company Z.ai (formerly Zhipu) launched the GLM-4.7 model on December 22. It is available through the API of the BigModel.cn platform and has become the default model of the GLM Coding Plan subscription, which targets developers.

The company tuned the model on real-world software development tasks: frontend, backend and instruction following. The model weights, meaning its learned parameters, are freely available for download on Hugging Face, so companies can run it on their own hardware. Source: TechNode Global (Z.ai press release).

MiniMax M2.1: more programming languages and office work (December 23)

A day later, Chinese startup MiniMax followed with its M2.1 model. It builds on the previous M2 and focuses on complex real-world tasks and so-called agents, meaning AI that plans and carries out multiple steps on its own.

The new release handles a wider range of programming languages and office scenarios better. The model is available through an API, and the company has published its weights as open source. On SWE-bench Verified, a test that measures fixes of real software bugs, the maker reports a score of 74.0 compared with 69.4 for its predecessor. Source: SiliconANGLE.

Alibaba (the Qwen team) published the weights of its Qwen-Image-Edit-2511 image editing model on December 23. Compared with the September version, it keeps characters looking more consistent across edits. Source: Hugging Face.

Products and Agents

ChatGPT offers a personal year in review (December 22)

OpenAI launched a “Your Year with ChatGPT” feature, similar to the well-known Spotify Wrapped. Users get a summary of their year with the chatbot: “awards” based on how they used it, a poem and an image.

For now, the feature is available in the US, the UK, Canada, Australia and New Zealand. It requires memory and chat history to be turned on; business and education accounts do not get it. Source: Thurrott.

xAI’s Grok heads to the US Department of War (December 22)

The US Department of War (formerly the Department of Defense) announced that it will add Grok models from Elon Musk’s xAI to its internal GenAI.mil platform. The platform, launched a few weeks earlier with Google’s Gemini model, is meant to serve roughly three million service members and civilian employees.

According to reports, the rollout is planned for early 2026. The move drew criticism in the media. Source: War.gov.

Business and Investment

Nvidia licenses Groq’s technology and takes over its leadership (December 24)

On Christmas Eve, chipmaker Groq announced a non-exclusive licensing agreement with Nvidia. Nvidia gains rights to its technology for inference, meaning running already trained models when they answer users. Founder Jonathan Ross, president Sunny Madra and other team members are moving to Nvidia.

Groq is to remain an independent company under its new chief Simon Edwards, and its cloud service GroqCloud will keep running. The companies did not disclose the price. Broadcaster CNBC, citing a Groq investor, reported roughly 20 billion dollars in cash, which would make it the largest deal in Nvidia’s history.

According to media reports, Nvidia CEO Jensen Huang stressed that this is not an acquisition of the whole company. Other tech giants are also striking similar “license plus talent” deals, and regulators are watching them ever more closely. Sources: Groq, CNBC.

Alphabet buys energy company Intersect for 4.75 billion dollars (December 22)

Google’s parent company Alphabet agreed to buy Intersect, a company that builds power generation and data centers. The price is 4.75 billion dollars in cash plus assumed debt; Google already held a minority stake in the company.

The deal mainly covers gigawatt-scale projects under construction that are being built right next to data centers. The goal is to have its own electricity for AI rather than relying solely on overloaded utilities. Closing was expected in the first half of 2026. Source: Axios.

According to Reuters (December 22), Nvidia aims to start shipping H200 chips to China by mid-February. The first shipments are to come from existing inventory, roughly 40,000 to 80,000 chips. Exports were made possible by Donald Trump‘s decision to allow sales with a 25 percent fee, but Beijing has not yet approved the purchases. Source: BNN Bloomberg (Reuters).
Japan will nearly quadruple its budget support for advanced chips and AI, to about 1.23 trillion yen (roughly 7.9 billion dollars) in the fiscal year starting April 2026 (December 26). The money will come from the regular budget of the Ministry of Economy, Trade and Industry, making the support more stable. Sources: Bloomberg, Taipei Times.
The US institute NIST is committing 20 million dollars to two AI centers to be run by the nonprofit MITRE (December 22). One is meant to boost manufacturing productivity, the other to protect critical infrastructure (water, power, the internet) from cyber threats. Source: NIST.

Regulation and Law

Italy orders Meta to let rival chatbots onto WhatsApp (December 24)

Italy’s antitrust authority AGCM issued interim measures ordering Meta to suspend new terms for business WhatsApp. Starting January 15, they would have barred companies whose main service is an AI chatbot from using the platform.

According to the authority, this would have completely shut out competitors of the Meta AI assistant. Meta called the decision “fundamentally flawed” and is considering an appeal. The European Commission is investigating a similar matter. Source: The Star (Reuters).

China wants to regulate AI that imitates people and emotions (December 27)

China’s cyberspace regulator CAC published draft rules for public comment covering services that simulate human personality and emotional interaction. These are typically AI companions and “virtual friends” in text, images, audio or video. Comments could be submitted until January 25, 2026.

Among other things, the draft requires:

monitoring users’ emotional state and dependence, and intervening at signs of extreme emotions or addictive behavior,
a ban on emotional manipulation and encouraging self-harm,
a special mode for minors with parental consent and time limits,
content must not threaten national security or spread rumors, violence or obscenity.

Source: Free Malaysia Today (Reuters).

Security and Incidents

OpenAI: we will probably never fully solve prompt injection (December 22)

OpenAI described how it secures its AI browser ChatGPT Atlas against prompt injection attacks. In these attacks, an attacker hides instructions in a web page, email or document, and the AI agent carries them out as if they came from the user.

The company admitted the problem is “unlikely to ever be fully solved” and treats it as a long-term security challenge. To counter the attacks, it deployed a model trained on attacks, automated “attackers” trained with reinforcement learning, and user confirmation of sensitive actions.

Britain’s National Cyber Security Centre had earlier made a similar statement. OpenAI did not say how much the update reduced the number of successful attacks. Source: TechCrunch.

Waymo explains why its robotaxis got stuck during a power outage (December 24)

A fire at a PG&E substation on the evening of December 20 cut power to roughly a third of San Francisco. Waymo autonomous vehicles stopped at intersections and blocked traffic.

On December 24, the company explained that its vehicles treat dead traffic lights as all-way stops. However, so many simultaneous cases overwhelmed the remote checks the cars requested. Waymo promised a software update for its entire fleet and changes to its emergency procedures; California regulator CPUC is also looking into the situation. Source: TechCrunch.

On December 27, Sam Altman announced on X that OpenAI is looking for a new head of preparedness (Head of Preparedness). The role is to anticipate and mitigate the risks of the most capable models, for example in cybersecurity and biology. The base salary is 555,000 dollars a year plus equity; Altman described the job as “stressful”. Sources: The Register, Engadget.