This Week in AI: July 13–19, 2026 (Week 29)

News of the week: July 13–19, 2026 (week 29)

The week was dominated by open models: China’s Kimi K3 with 2.8 trillion parameters and Inkling from America’s Thinking Machines. Google, on the other hand, delayed Gemini 3.5 Pro once again, and Alphabet shares slipped as a result. Money keeps flowing into infrastructure: Databricks is now valued at 188 billion dollars, TSMC reported record profit, and Anthropic is in talks with Meta about leasing computing capacity. On the regulatory front, New York stood out with a moratorium on large data centers, China with new rules for AI companions, and Xi Jinping by founding an international organization for AI cooperation.

New Models

Kimi K3: the largest open model ever released

Chinese company Moonshot AI unveiled the Kimi K3 model on July 16. It has 2.8 trillion parameters in a mixture-of-experts architecture (only part of the model is activated for each query, which makes it cheaper to run) and a context window of 1 million tokens, roughly several thick books at once.

In the Arena leaderboard for building web interfaces, K3 took first place, ahead of both Claude Fable 5 and GPT-5.6 Sol. API pricing is 3 dollars per million input tokens and 15 dollars per million output tokens. The company promised to release the freely downloadable weights (the model’s parameters) on July 27.

Moonshot itself admits that in overall performance it still lags behind the best closed models. Another novelty is the Kimi Delta Attention mechanism, which according to the company speeds up text generation with long context by up to 6.3 times. (VentureBeat, MarkTechPost)

Thinking Machines releases Inkling under the Apache 2.0 license

Thinking Machines Lab, the startup of former OpenAI CTO Mira Murati, published its first open model, Inkling, on July 15. It has 975 billion parameters, of which 41 billion are active during operation, and it handles text, images and audio.

The weights are available on Hugging Face under the permissive Apache 2.0 license, which also allows commercial use. Independent evaluator Artificial Analysis called it the best open model from American labs, although China’s Kimi K3 ranks higher overall. (Thinking Machines, Artificial Analysis)

Gemini 3.5 Pro delayed again

Bloomberg reported on July 16 that Google is delaying the Gemini 3.5 Pro model by several months. Internal tests reportedly revealed shortcomings, mainly in coding. At the Google I/O conference in May, the company had promised a release as early as June.

Alphabet shares fell by roughly 4 percent that day. A Google spokesperson said the company continues to test the model and other improvements with partners. (Bloomberg, Barchart)

Products and Agents

NotebookLM is now called Gemini Notebook

On July 16, Google renamed its popular research tool NotebookLM to Gemini Notebook. Existing notebooks and sources remain unchanged, but the service will be more closely connected to the Gemini app and the AI Mode search experience.

It is also getting a secure cloud computer that can write and run code over uploaded materials. The feature is expected to become available to Google AI Pro subscribers within weeks. (Google, TechCrunch)

OpenAI’s first device is said to be a speaker without a display

According to Bloomberg on July 14, the first hardware from OpenAI will be a movable speaker with a camera but no screen. It is meant to work as a home AI companion that recognizes who is talking to it and what is happening around it. The design is being handled by the team of designer Jony Ive.

An unveiling is expected this year, with sales in 2027. OpenAI has not officially confirmed the device. (9to5Mac)

July 16 – OpenAI acknowledged mistakes in the new unified ChatGPT interface, which merged chat, Work mode and Codex into a single window and hid the conversation history; it is bringing chats and projects back to the sidebar. (The Decoder)
July 16 – Google DeepMind reconstructed a lost goal by Pele from nearly 2,000 recordings and 3,600 images using the Veo and Gemini models; the result can be seen at the Pele Museum in Santos. (Google)

Business and Investment

Databricks valued at 188 billion dollars

Data and AI platform Databricks signed terms on July 16–17 for a new funding round led by investor Coatue. The round values the company at 188 billion dollars, up from 134 billion at the beginning of the year.

According to The Wall Street Journal, the round amounts to roughly 3 billion dollars. The money is set to go into AI products and potential acquisitions. (The Star / Reuters, Bloomberg Law)

TSMC: record profit and higher investment

TSMC, the largest contract chipmaker, reported its second-quarter results on July 16. Net profit rose 77 percent year over year to 706.6 billion Taiwan dollars, and revenue reached 40.2 billion US dollars.

The company raised this year’s capital expenditure budget to 60 to 64 billion dollars and expects revenue growth of slightly over 40 percent. It is being driven by demand for artificial intelligence chips. (TSMC / SEC)

Anthropic in talks with Meta over computing capacity

According to The New York Times and CNBC on July 17, Anthropic is in preliminary talks to lease computing capacity from Meta. The two-year deal could be worth up to 10 billion dollars.

The talks are at an early stage and may not result in a deal. For Meta, it would be a first step toward its own cloud business. (The Next Web, Quartz)

July 16 – Fireworks AI, which runs models for businesses (so-called inference), raised roughly 1.5 billion dollars in a Series D round at a valuation of 17.5 billion; investors include Nvidia. (Orrick)
July 14 – According to analysts at eMarketer, the advertising business of OpenAI will miss its own forecast by 90 percent: they estimate the entire US market for chatbot advertising at 5.4 billion dollars in 2030. (Adweek)
July 14 – American companies such as DoorDash, Airbnb and Siemens are, according to Futurism, turning to cheaper Chinese open models from DeepSeek, Z.ai and Moonshot. (Futurism)

Regulation and Law

New York halts permits for large data centers

Governor Kathy Hochul signed an executive order on July 14 that suspends state permits for new data centers with a power demand of 50 MW or more for up to a year. It is the first statewide moratorium of its kind in the US.

During the pause, the state is to prepare rules covering the impact on the environment and the power grid. The order also provides for a grid modernization fund paid for by contributions from operators. (Axios, Jones Day)

Publishers sue Google over Gemini training

Publishers Hachette, Cengage and Elsevier, together with author Scott Turow, announced a class action lawsuit against Google on July 13. They claim the company copied books and scholarly articles without permission, including from Google Books, and used them to train its Gemini models.

The plaintiffs are seeking an injunction and damages. The same group already sued Meta in May. (Hachette Book Group, Al Jazeera)

Former Meta employees: AI decided who was laid off

Twenty-six former Meta employees filed a lawsuit in federal court in Oakland on July 13. They say that during the May layoffs of roughly 8,000 people, the company evaluated employees with AI tools that also took into account how much they themselves used AI.

This allegedly put people on medical or parental leave at a disadvantage. Meta rejects the accusations, saying the decisions were made by people. It appears to be the first lawsuit of its kind against a major American company. (Engadget, Business Today)

China restricts AI companions, Xi Jinping founds WAICO

On July 15, China’s rules for anthropomorphic AI took effect, covering services that imitate human personality and emotional relationships. On the same day, ByteDance shut down user-created agents in its Doubao app, Alibaba did the same with similar features in Qwen, and Tencent had already done so earlier in Yuanbao. (Decrypt)

President Xi Jinping spoke for the first time at the World AI Conference in Shanghai (July 17–20). He introduced the World AI Cooperation Organization (WAICO), headquartered in Shanghai, which 29 countries have joined, including Russia and Pakistan. (Fortune, Xinhua)

July 14 – Google DeepMind CEO Demis Hassabis proposed an independent, industry-funded US body that would test the most powerful models up to 30 days before release; it would be modeled on the financial markets regulator FINRA. (Constellation Research)
July 13 – Hundreds of economists and researchers, including 16 Nobel Prize laureates, called on governments in an open letter to prepare now for the impact of AI on the labor market. (Al Jazeera)
July 13 – New York’s Montefiore hospital laid off 12 nurses who reviewed the eligibility of care, and their work was taken over by Datavant software; the union considers this a breach of contract. (The Guardian)

Security and Incidents

Grok Build uploaded users’ files, xAI publishes the code

According to researchers’ analyses, the Grok Build coding tool from xAI sent users’ entire folders to cloud storage. For one user, this even included SSH keys and a password database. On July 12, the company turned off data storage by default and promised to delete the uploaded content.

On July 15, xAI then published the tool’s source code under the Apache 2.0 license, roughly 845,000 lines of Rust. Experts recommend that anyone who used the tool with live credentials change them. (Simon Willison, DevOps.com)

July 16 – A leak of the source code of the music generator Suno showed that the company trained on more than 113,000 hours of music from YouTube Music and other sources; the leak also affected customer data. (404 Media)
July 16 – In more than 2,600 tests, the organization Common Sense Media found that the AI Overview summaries in Google Search missed 29 percent of explicit mentions of suicide on accounts set up for children; Google called the tests contrived. (PBS)
July 16 – Researcher Alex Turner left Google DeepMind in protest against Google’s secret contract with the Pentagon. (Alex Turner)