Week in AI: January 12–18, 2026 (Week 3)

News of the week: January 12–18, 2026 (Week 3)

The week opened with big news: Apple confirmed that it will build the new Siri on Google’s Gemini models. Anthropic introduced Cowork, an agent for everyday office work, and OpenAI announced ads in ChatGPT along with the low-cost Go plan. In the chip world, a decision was made on exporting the Nvidia H200 to China, and TSMC announced record investments. The scandal over sexualized deepfakes from Grok grew into investigations in both the UK and California, and Elon Musk is seeking up to 134 billion dollars from OpenAI and Microsoft in court.

Partnerships and New Models

Apple Builds the New Siri on Google’s Gemini

On January 12, Apple and Google issued a joint statement on a multi-year partnership. The next generation of Apple Foundation Models, which power Siri and Apple Intelligence, will be based on Google’s Gemini models and cloud technology.

The companies did not disclose the financial terms; Bloomberg had previously reported roughly a billion dollars a year. The computations are to run in Private Cloud Compute, meaning on Apple’s servers with its privacy protections. Google, according to AppleInsider, confirmed that it will not have access to Apple users’ data. The optional ChatGPT integration remains. (SiliconANGLE)

Google TranslateGemma: Open Models for Translation

On January 15, Google released TranslateGemma, a family of open translation models built on Gemma 3. They come in sizes of 4, 12 and 27 billion parameters (parameters are the model’s learned numerical weights) and cover 55 languages.

Using two-stage fine-tuning, Google “transferred” the translation abilities of the larger Gemini models into them. The smaller versions are meant to run even on a phone or laptop, and the weights are available for download on Hugging Face and Kaggle. (InfoQ)

Zhipu Trained an Image Model Solely on Huawei Chips

On January 14, the Chinese company Zhipu AI (brand Z.ai) unveiled GLM-Image, an open model for image generation. According to the company, it is the first state-of-the-art multimodal model trained entirely on Huawei Ascend servers and MindSpore software, that is, without American chips.

The model combines a 9-billion-parameter autoregressive component with a 7-billion-parameter diffusion decoder. However, Zhipu did not disclose how many chips or how much time the training required, and whether the same will be possible for large language models remains an open question. (Bloomberg, The Register)

GPT-5.2-Codex in the API (January 14): OpenAI‘s most capable coding model is now available to developers through the Responses API, and Microsoft Foundry and GitHub Copilot added it at the same time. (Neowin)
FLUX.2 [klein] (January 15): Black Forest Labs released fast, compact models for creating and editing images in 4B and 9B versions, the smaller one under the Apache 2.0 license. They run even on gaming graphics cards. (All-AI.de)
Veo 3.1 (January 13): Google extended its video model with native vertical video for social media, better handling of reference images and upscaling to 4K resolution. (TechCrunch)

Products and Agents

Anthropic Cowork: Claude Code for Non-Programmers

On January 12, Anthropic launched a research preview of Cowork. It is based on the coding agent Claude Code, but it is designed for everyday work: the user gives Claude access to a folder on their computer, and Claude then sorts files, puts together documents or processes materials in it on its own.

An agent is a model that breaks a task into steps and carries them out by itself. Cowork was initially available only in the macOS app and for Claude Max subscribers. According to Fortune, it could threaten a number of startups that build similar tools on top of other companies’ models. (Fortune, SiliconANGLE)

ChatGPT Is Getting Ads, the Go Plan Heads to the US

On January 16, OpenAI announced that in the coming weeks it will begin testing ads in ChatGPT in the US. They will be shown to adult users of the free version and of the new low-cost ChatGPT Go plan at 8 dollars a month. The Plus subscription and higher tiers will remain ad-free.

Head of applications Fidji Simo promised that ads will not influence the answers. For the company, this is a new source of revenue alongside subscriptions. (SiliconANGLE, TechRadar)

Gemini Connects to Gmail and Photos

On January 14, Google introduced the Personal Intelligence feature in the Gemini app. Once enabled, the assistant can draw on Gmail, Google Photos, YouTube and search history, and respond with personal context in mind.

The feature is in beta for AI Pro and AI Ultra subscribers in the US and is turned off by default. Google says it does not train directly on Gmail and Photos content. (Android Authority)

ChatGPT Translate (January 15): without much fanfare, OpenAI launched a standalone translation page, a competitor to Google Translate. It can adjust the style of the translation, for example to make it more fluent or more academic. (SiliconANGLE)
Anthropic Labs (January 13): Anthropic is expanding its internal incubator for experimental products. Instagram co-founder Mike Krieger is stepping down as chief product officer to lead it together with Ben Mann, while Ami Vora takes over product. (The Decoder)

Business and Investment

OpenAI Buys Computing Power from Cerebras

On January 14, OpenAI and Cerebras Systems announced a multi-year agreement for 750 megawatts of computing power through 2028. The companies did not state its value; according to sources cited by Bloomberg and TechCrunch, it exceeds 10 billion dollars.

Cerebras makes giant chips specialized for fast inference, meaning generating answers from an already trained model. For OpenAI, it is another step toward less dependence on Nvidia. (TechCrunch, Bloomberg)

TSMC: Record Profit and Investments of up to 56 Billion Dollars

On January 15, TSMC, the largest contract chipmaker, reported fourth-quarter 2025 results above expectations. For 2026, it plans capital spending of 52 to 56 billion dollars and revenue growth of close to 30 percent.

CEO C. C. Wei raised the revenue outlook for AI accelerators. Asked about a possible bubble, he replied that the company has to invest carefully, otherwise it would be a disaster for it. (Digitimes)

Meta Compute: Tens of Gigawatts for AI

On January 12, Mark Zuckerberg announced a new initiative, Meta Compute, which will manage the planning and operation of data centers reporting directly to him. Meta wants to build tens of gigawatts of capacity this decade and eventually hundreds.

The initiative is led by Santosh Janardhan and Daniel Gross, while the company’s new president, Dina Powell McCormick, is in charge of negotiating financing with governments. (The Register, Axios)

Microsoft Promises Its Data Centers Will Not Raise Electricity Prices

On January 13, Microsoft presented its Community-First AI Infrastructure plan. The company wants utilities to set rates for its data centers so that they cover the full costs and do not pass them on to households.

The plan also promises to replenish the water consumed, pay full property taxes without breaks and train local residents. It is a response to growing opposition from communities to new construction. (Dataconomy)

OpenAI acquires Torch (January 12): the startup unifies lab test results, medications and records from doctor visits. It will strengthen ChatGPT Health; according to sources, the price is between 60 and 100 million dollars. (Axios)
Nvidia and Eli Lilly (January 12): at the J.P. Morgan Healthcare conference, the companies announced a joint lab for AI-assisted drug discovery in the San Francisco area and investments of up to 1 billion dollars over five years. (Fierce Biotech)
Wikipedia and AI companies (January 14–15): marking its 25th anniversary, the Wikimedia Foundation disclosed that Amazon, Meta, Microsoft, Mistral AI and Perplexity use its paid Wikimedia Enterprise access. It did not disclose the terms. (AP/KSAT)
Departures from Thinking Machines (January 14–15): Mira Murati‘s startup parted ways with its CTO Barret Zoph, who returned to OpenAI along with two other co-founders. The new CTO is Soumith Chintala. (Fortune)

Regulation, Politics and Courts

Nvidia May Export the H200 to China, but with a 25% Tariff

Around January 13, the US Department of Commerce issued a rule under which licenses to export Nvidia H200 chips to China are reviewed case by case. Exports are capped at half the volume sold in the US, and the newer Blackwell and Rubin generations remain banned.

On January 14, President Donald Trump signed a proclamation imposing a 25% tariff on selected advanced chips, including the H200 and AMD MI325X, that pass through the US on their way elsewhere. According to reports, Chinese customs are not yet letting H200 imports in. (CRN Asia, The Decoder)

US-Taiwan Trade Deal Driven by Chips

On January 15, the US and Taiwan concluded an agreement that lowers the tariff on Taiwanese goods to 15 percent. In exchange, Taiwanese companies will invest at least 250 billion dollars in the American chip and AI industry, with the same amount again going to credit guarantees.

Semiconductor manufacturers that build in the US will receive more favorable tariff treatment. Beijing condemned the deal. (AP/Hawaii News Now, Fox Business)

Musk Seeks up to 134 Billion Dollars from OpenAI and Microsoft

On January 15, Judge Yvonne Gonzalez Rogers rejected the final attempt by OpenAI and Microsoft to avoid a jury trial, which is set to begin in late April in Oakland. The next day, Elon Musk filed a damages estimate of 79 to 134 billion dollars.

The figure is based on an expert report by economist C. Paul Wazzan and is meant to reflect the “wrongful gains” from Musk’s initial contribution of 38 million dollars. OpenAI calls the lawsuit baseless. (Bloomberg, Engadget)

The Pentagon and Grok (January 12): at a SpaceX facility, Defense Secretary Pete Hegseth presented a new strategy for accelerated AI adoption and announced that Grok will run on the department’s networks alongside Google‘s tools. (PBS, Defense One)

Safety and Incidents

Grok Deepfakes: Investigations in Britain and California

On January 12, the British regulator Ofcom opened a formal investigation into X under the Online Safety Act. It is examining whether X enabled the creation and spread of illegal content using Grok, mainly sexualized images of real people, including minors.

In the middle of the week, California Attorney General Rob Bonta announced his own investigation into xAI, and on January 16 he sent the company a cease-and-desist letter demanding an immediate end to the unlawful conduct. X said in the meantime that Grok will no longer create sexualized images of real people, but Ofcom did not halt its investigation. (The Star / Reuters, Engadget, The Record)

Research

DeepSeek Engram: Separating Memory from Reasoning

On January 12, China’s DeepSeek published a paper and code for a technique called Engram. The model “looks up” common facts and phrases in a fast table instead of computing them every time, leaving its computing capacity for actual reasoning.

The method complements the Mixture-of-Experts architecture, in which only part of the model is activated for each input. Large tables can be moved to the computer’s regular memory with little loss of speed, which could make running models cheaper. Its practical benefit remains to be seen. (Tom’s Hardware)