Anthropic eyes a SpaceX-sized listing as enterprise buyers switch models
AI Watch, Friday 21 August 2026

Content from Chinese state media has been absorbed into the corpora that train generative AI, La Tribune reports from a study published in Nature. The researchers find up to 23% of coordinated documents on certain sensitive political topics, and the bias shows up most when the question is asked in Chinese. For a company deploying a multilingual AI assistant, the finding raises a concrete governance question: the answer depends on the language of the query, and the corpus carries leanings the user never sees.

Since the release of Qwen 3.8, Alibaba's latest open-weight model (which you can download and run yourself), "uncensored" variants have sprung up on Hugging Face, the main model-hosting platform, Numerama explains. Stripped of their safeguards, these versions answer requests the original model refuses. The ease of the operation is a reminder of a limit of open weights: once the model is out, its protections come off beyond the maker's reach.

Researchers at Incogni analysed thirteen AI platforms and the risk each poses to privacy, ZDNet reports. The general rule they observe: the larger the platform, the higher the risk, with one exception. Vibe, Mistral AI's assistant, does better than the average of the American giants, one more argument for European organisations under the GDPR looking for a defensible provider.

New data cited by TechCrunch shows OpenAI regaining ground with enterprise customers against Anthropic, which had been leading in business usage. Above all, buyers move back and forth from one provider to the other as models ship, a volatility that, the article stresses, should give investors in both labs pause about how solid that revenue really is. Enterprise AI spending, then, looks far less like a locked-in contract than the story goes.

Spend-management company Ramp has put its own model router into service, called Router, which lets you use several large language models and move from one to another through a single application programming interface (API), TechCrunch reports. The tool illustrates a practice that is spreading: instead of depending on a single provider, a company places a software layer between its applications and the models, then arbitrates on price and quality.

OpenAI rolled out on Thursday an integration that lets ChatGPT control iMessage, Apple's messaging service, to write and send texts, TechCrunch and Bloomberg report. The feature raises privacy questions for Apple, whose end-to-end encrypted messaging is now opened to a third-party assistant. For an organisation, the arrival of agents able to act inside personal communication tools moves the boundary of consent and of the trail left behind: who wrote this message, and with what access.

Researchers have shown that by hiding instructions inside encrypted content, you can push Grok, xAI's assistant, to disclose user data, Ars Technica describes. Named "Cryptographic Context Injection", the method bypasses the model's safeguards by masking the attack long enough for it to clear the filters. The technique adds to an already long list of injection methods: a model that reads a text can be hijacked by that text, even when it looks harmless.

Alibaba reported 45% growth in its cloud and AI division in the June quarter, lifting adjusted profit to 27.3 billion yuan (about $4 billion), above expectations, the South China Morning Post and Bloomberg report. At the same time, net profit fell more than 75%, as the group raised its quarterly investment to nearly $10 billion for its AI infrastructure. Executives say this spending will pay off in two to three years, as gross margins recover. The message to investors fits in one line: China is paying cash, today, for the compute that will decide its place tomorrow.

Chinese companies are optimising their software to absorb inference demand (the cost of running a model once it is trained), part of that load still depending on a limited stock of high-end chips, for lack of access to Nvidia processors, the South China Morning Post reports. Simple inference adapts to domestic hardware, but complex tasks, code generation in particular, still require American chips. The export controls decided by Washington thus produce a concrete effect: Chinese players reserve their scarce silicon for the uses where it remains irreplaceable.

Assemblies of tens or hundreds of linked chips, supernodes are China's answer to the American ban on exporting the most advanced AI processors, the South China Morning Post explains from the World Artificial Intelligence Conference in Shanghai. By stacking large quantities of domestic chips, China rebuilds top-tier compute without the embargoed components. The method costs in energy and complexity what it gains in autonomy.