AI in practice Archive

Jan 2, 2026

Baidu's AI chip division, Kunlunxin, has confidentially filed for an IPO in Hong Kong. The company submitted its application on January 1, setting the stage for a spin-off while keeping Kunlunxin under Baidu's umbrella. According to Reuters, a recent financing round valued the company at around $3 billion. The final size of the offering hasn't been determined yet.

Kunlunxin started as an internal unit back in 2012 and has primarily supplied chips to Baidu. Over the past two years, though, the company has been expanding its customer base beyond its parent company.

The IPO comes as China accelerates efforts to develop homegrown semiconductor alternatives in response to US export restrictions. Kunlunxin isn't alone in eyeing Hong Kong—other Chinese AI and chip companies, including MiniMax, Biren Technology, and OmniVision, are also pursuing listings on the exchange.

Comment Source: Reuters

Matthias Bastian

Jan 2, 2026

Short News

OpenAI has reportedly shifted production of its first AI hardware device from Luxshare to Foxconn. The company wants to avoid manufacturing in China, according to Taiwanese business newspaper Economic Daily News, citing supply chain sources. Instead, production will happen in Vietnam or the US.

The device is still in the design phase and could turn out to be a smart pen or portable audio gadget. It's expected to have a microphone and camera, letting users transfer handwritten notes directly to ChatGPT. The company is aiming for a launch in 2026 or 2027, and according to the newspaper, Foxconn would handle all OpenAI orders, from cloud infrastructure to consumer devices. OpenAI is calling the project "Gumdrop" internally.

The Financial Times reported on technical problems with the project back in October, including software bugs, privacy issues, and missing cloud infrastructure.

Comment Source: Economic Daily News

Matthias Bastian

Jan 1, 2026

Short News

OpenAI is building its planned ChatGPT hardware around conversation. To make that work, the company is pouring resources into improving its audio AI models, according to The Information. Over the past two months, OpenAI has combined several internal teams to focus on audio.

Right now, OpenAI's audio models can't match the accuracy and response speed of their text-based counterparts, according to current and former employees. A new audio model architecture in development aims to sound more natural and emotional, deliver more accurate answers, and handle real-time back-and-forth conversation. OpenAI is targeting a release in the first quarter of 2026. Kundan Kumar, a researcher the company recruited from Character.AI, is leading the effort.

The actual devices are likely a long way off. OpenAI is reportedly working on several products, including glasses and a smart speaker without a screen. Last year, the company acquired io, the startup cofounded by former Apple designer Jony Ive, for nearly $6.5 billion to help with development. The goal behind all this hardware is to build a "super AI assistant" that becomes as central to daily life as the smartphone.

Comment Source: The Information

Matthias Bastian

Jan 1, 2026

Short News

Chinese AI startup Moonshot AI, the company behind the Kimi chatbot, has closed a $500 million Series C funding round. The deal values the company at $4.3 billion, according to an exclusive report from LatePost. IDG led the round with $150 million, while existing investors including Alibaba, Tencent, and Wang Huiwen also participated.

CEO Yang Zhilin said in an internal letter that the company now holds more than $1.4 billion in cash. The funds will go toward expanding computing capacity and developing the K3 model. An insider says this cushion means Moonshot AI isn't rushing to go public, unlike Zhipu and MiniMax, two other Chinese AI startups pushing for IPOs. In September, Moonshot AI launched its "OK Computer" agent feature and a subscription model. According to Yang, paying users are growing 170 percent month over month.

This year, Moonshot AI made waves with its Kimi-K2-Thinking model, an open-source reasoning model that held its own against proprietary competitors.

Comment Source: LatePost

Matthias Bastian

Dec 31, 2025

Short News

Alibaba has released Qwen-Image-2512, an update to its text-to-image model. The company says the new version does a better job generating realistic images of people, with finer facial detail and less of the artificial look that plagued earlier versions.

The updated image model aims to eliminate the "plastic" look of its predecessor. | Image: Qwen

The display of text in images, for example, in infographics or presentations, has also been improved. Landscapes, animal fur, and other natural elements also come out with finer detail.

The new Qwen model handles text in images better than before. | Image: Qwen

In over 10,000 blind tests on Alibaba's AI Arena platform, Qwen-Image-2512 came in fourth overall, making it the top-ranked open-source model, according to Alibaba. It's up against other open models like HunyuanImage-3.0, Z-image, and Flux.2.

Qwen-Image-2512 is available on Hugging Face and ModelScope, and you can try it out through Qwen Chat. More details are in the Tech Report and on the blog.

Comment Source: Hugging Face / Qwen-Image-2512 | ModelScope / Qwen-Image-2512 | Qwen Chat / Demo | Qwen / Tech Report | Qwen / Blog

Matthias Bastian

Dec 31, 2025

Short News

OpenAI is handing out stock-based compensation averaging $1.5 million per employee, about 34 times what major tech companies typically paid before going public. According to financial data shown to investors and reported by the Wall Street Journal, the company's roughly 4,000 employees are getting more than any tech startup has ever offered. That's also more than seven times what Google paid back in 2003.

Share-based compensation as a percentage of annual revenue

Company	Share of turnover
OpenAI	46,2 %
Palantir	32,6 %
Alphabet (Google)	14,6 %
Meta (Facebook)	5,9 %
Average (18 tech companies)	~6 %

Sources: Wall Street Journal, Equilar analysis. All figures refer to the year before each company's IPO, except OpenAI, which uses 2025 projections.

The generous pay is meant to keep OpenAI ahead in the AI race, but it's also pushing up losses. Meta CEO Mark Zuckerberg has turned up the pressure by offering top employees packages worth hundreds of millions of dollars and poaching more than 20 OpenAI staffers. OpenAI expects its compensation costs to climb by $3 billion annually through 2030.

Comment Source: Wall Street Journal

Matthias Bastian

Dec 31, 2025

AI in practice

Chinese AI startups beat Silicon Valley to the public markets in a massive IPO wave

Matthias Bastian

Dec 31, 2025

Short News

Nvidia is reportedly in advanced negotiations to acquire Israeli AI start-up AI21 Labs, according to a report by Calcalist. The deal could be worth between two and three billion dollars. AI21 was valued at $1.4 billion during its last funding round in 2023, with both Google and Nvidia among its investors.

Similar to its recent Groq deal, Nvidia appears primarily interested in AI21's workforce. The company employs around 200 people, most of whom hold advanced degrees and have rare expertise in AI development. At the reported deal value, Nvidia would be paying roughly $10 to $15 million per employee.

AI21, based in Tel Aviv, was founded in 2017 by Professor Amnon Shashua, Professor Yoav Shoham, and Ori Goshen. The company builds large language models and offers software tools, including an agent builder. Its products target enterprise customers, with estimated annual revenue of around $50 million.

Comment Source: Calcalist

Matthias Bastian

Dec 31, 2025

Short News

Nvidia's China business is ramping up again. According to Reuters, the chip company is in talks with contract manufacturer TSMC to expand production of its H200 AI chips to meet surging demand from China. Chinese tech companies have ordered more than two million H200 chips for 2026, but Nvidia currently only has 700,000 units available.

Production at TSMC is set to begin in the second quarter of 2026, with Nvidia planning to sell the chips for around $27,000 each. ByteDance alone, TikTok's parent company, plans to spend around 100 billion yuan (about $14 billion) on Nvidia chips in 2026, according to the South China Morning Post.

The Trump administration recently cleared H200 chip exports to China with a 25 percent fee attached. However, China hasn't approved the imports yet. Officials there are still debating whether access to advanced foreign chips might slow down the country's domestic chip industry. One proposal under consideration would require every H200 order to include a certain percentage of chips made in China.

China is also working on a rule that would require domestic chip manufacturers to use at least 50 percent locally made equipment.

Comment Source: Reuters | South China Morning Post

Matthias Bastian

Dec 31, 2025

Short News

Microsoft CEO Satya Nadella has reportedly entered "founder mode." According to the Financial Times, Nadella has shaken up Microsoft's management to keep pace with Amazon, Google, and startups like Anthropic. "Satya is trying to demonstrate a sense of urgency," one executive said. "The goal is to get out of some of the structures that exist and make the route to him easier."

Nadella has brought in several new executives, including former Meta technology chief Jay Parikh. Mustafa Suleyman, co-founder of Google Deepmind, now leads Microsoft's in-house AI development with his own budget and salary structure. The move has created internal friction, but Microsoft is accepting that to stay competitive for AI talent.

Microsoft remains strong overall, especially thanks to its Azure cloud AI services, but faces tough challenges ahead. The company will lose exclusive rights to OpenAI's research and models in the early 2030s, and OpenAI is no longer required to use Microsoft's data centers. Meanwhile, Copilot's 150 million users trail Google's Gemini (650 million) and ChatGPT (800 million). Nadella has reportedly already stepped in to influence development directly.

Comment Source: Financial Times