When AI companies sell models and get scared, read between the lines

Over the past few days, people like OpenAI CEO Sam Altman, Microsoft CEO Satya Nadella and Google DeepMind’s Sir Demis Hassabis have sounded the drum of caution in their own ways. Some have taken a more picturesque route to craft the story, while others have told it directly. At first, the message may come across as a wave of unprecedented sense of corporate responsibility, but the narrative construction is not hidden.

Also read:’‘Primarily a PR story’: Cornell’s John Theakston breaks down the OpenAI narrative

It required OpenAI to retell a story that no one had noticed before, suggesting that its model underpinned how an agent entity caused an “unprecedented cyber incident” by gaining access to the servers of infrastructure AI platform Hugging Face. This is not a potential dystopian AI rebellion as is being reported, as this was part of an internal cybersecurity assessment benchmark, and cybersecurity AI agents are expected to figure out ways to break into systems.

John Theakston of Cornell University recently told OpenAI that much of the existential fear being spread is primarily a PR story. “It is important to read this story with an understanding of OpenAI’s narrative framework. This is primarily a public relations story promoted by OpenAI, part of the same messaging campaign that began with the announcement of GPT-2 in 2019,” he told HT. “The clear message of this campaign is that OpenAI’s technology is dangerous, but they want to clearly state that their technology is powerful and worthy of large investments, and privileged regulatory status.”

This has worked to some extent. On Thursday, days after OpenAI’s message, US lawmakers mandated a ‘kill switch’ that would give governments the option to shut down AI tools if they behave differently. Most companies provide cursory statements regarding the actual outline of the regulation.

Also read: Alibaba Quan and Baidu: A clear outline of Apple Intelligence recreated for China

On Friday, 25 tech companies, including Nvidia, Microsoft and Indeed Meta, banded together to draft a letter suggesting that regulation should not be done in a way that “will stifle innovation overseas”. Two sides of the same coin.

To understand the strategy, you have to look at the narrative the biggest AI labs are selling. By warning the world that Artificial General Intelligence (AGI) is imminent and potentially dangerous, companies like OpenAI, Google, and Microsoft accomplish two things. First, they signal to investors and engineers that they have world-changing power technology. Second, they position themselves as the only adult in the room who is able to manage it.

Sir Demis Hassabis, Nobel laureate as well as co-founder and CEO of Google DeepMind, recently called for a frontier AI watchdog. He is confident that the industry will come forward with spin-offs, including attracting high-quality technical talent and the necessary computing resources for large-scale testing in a regulated area.

“Right now, we are locked in an extremely intense, multilayered commercial and geopolitical race. While these competitive dynamics fuel rapid progress and accelerate incredible progress, progress at the border is outpacing our understanding of technology. No one in the world knows for sure what’s going to happen from here, and even experts disagree,” Hassabis wrote in an essay on X.

Also read: ‘AI logic is changing towards cost and trust’: Mehran Gul

At this time, there is no specific definition of the Frontier Model, just as there is no concrete definition of AGI. Nothing from the new narrative suggests that it is about saving humanity from the AI ​​attack AI companies want to bring to organizations around the world.

Microsoft CEO Satya Nadella on July 12 warned companies using AI about AI. You may ask why? “You essentially pay for intelligence twice, once with money, and then with something more valuable: You have to disclose proprietary knowledge to make that intelligence useful,” he said.

The warning has not been shared with the intention of controlling AI use, but there is duplicity in intention. First, a subtle way of asking companies not to use Chinese models, and second, to stick with the model of an AI company. While this is likely to benefit Microsoft because it has an enterprise business, it will hurt companies like OpenAI and Anthropic, which don’t have workplace tools and enterprise cloud customers.

Also read: ‘The future hasn’t been written yet’: Sir Demis Hassabis wants a watchdog for frontier AI

Mehran Gul, author of ‘The New Geography of Innovation’ and former World Economic Forum, UN advisor as well as Fulbright Scholar at Yale, does not agree with Nadella’s analogy. “My obvious objection to this is that if you have an open-ended model that can be run on premises, you own your own data, and the model is essentially just an engine that you are adding to your existing database without sharing the data with any Chinese supplier, how does that fear actually come into play?”, he told HT.

The lack of concrete response from US AI companies is puzzling, especially when it comes to model training and token usage costs. This is part of the economic reality, along with calculations about costs, infrastructure investment and conversations about ‘circular funding’.

Gul says, “I don’t quite understand why American companies couldn’t release lower-priced AI models. If Anthropic already has a high-end model, I don’t understand why it couldn’t take a loss on a model that competes on price with companies like DeepSeek and Zipu, and offer a hybrid model that makes switching so much easier.”

Also read: Apple is suing OpenAI for allegedly stealing trade secrets

Since DeepSeq took the AI ​​world by storm early last year, Chinese AI models have consistently demonstrated a cost advantage. In terms of estimated token cost, GLM-5.2 costs $1.40 per million input tokens and $4.40 per million output tokens – in comparison, similar usage in Anthropic’s Cloud Opus 4.8 will cost developers and enterprises $5 and $25, respectively.

Microsoft, Anthropic and other AI companies have been vocal about their fear of models emerging from Chinese frontier labs because of the performance and cost tradeoff. According to the 2026 Stanford HAI AI Index report, the performance gap between top US and Chinese AI models has narrowed to a very small 2.7%, with the US leading in total top-tier model volume and private investment, while China leads in research volume, patents and citations.

The West’s closed model approach argues that AI should be kept proprietary for security reasons. Nevertheless, Hugging Face CEO Clément Delangeau thanked Z.ai, saying that the Chinese model became “an important part of our defense” during the breach caused by OpenAI’s own model. One can see this as a real explanation of what happened or as a dose of skepticism, thus furthering the case for limiting access to Chinese models.

Also read: Innovation is everywhere, not opportunity: UN pushes for AI governance

Z.ai released GLM 5.2 in mid-June under the MIT License, a permissive open-source license that allows unrestricted commercial use. It has 753 billion parameters, which is a measure of the size and capability of an AI model. Alibaba, Tencent, Xiaomi and Moonshot AI are not the only ones facing competition.

Apple is reworking the Apple Intelligence suite for China. Alibaba’s QianWen (also known as Alibaba Tongyi Qianwen) will serve as the primary system engine for on-device and server-side AI functions. It is unclear at this time whether Alibaba provided Apple with an optimized model or whether it is the 27 billion parameter Qwen 3.6 that keeps all 27 billion parameters active, unlike Apple’s own 20 billion parameter sparse model.

“It’s important to understand that China’s AI race is being fought on use cases, not model size. Competitors like Huawei are pursuing hybrid, full-stack AI strategies spanning silicon, OS, and on-device models,” explains Tarun Pathak, research director at Counterpoint Research.

Also read: Gemini Spark, the real money fear of AI, and Tilly Norwood

Foreign companies partnering with domestic tech giants will have to comply with China’s strict data localization rules. Apple Intelligence, Samsung’s Galaxy AI, Xiaomi Pengpai AI/Mimo, Huawei Xiaoyi, Oppo AndisGPT, Vivo BlueOneDevice and Nubia on Doubao built in partnership with ByteDance are now marked by the required localization and often heavy compression of models for on-device and local computation.

This brings us back to the sudden appetite for regulation. When companies like OpenAI and Google push for regulation, it is mostly on terms they are likely to easily navigate. For example, if regulators mandate that models above a certain enumeration threshold require expensive licensing, auditing, and constant red-teaming (an adverse security practice), a typical AI garage startup is doomed before arrival. This serves to slow down the open-source community.

Take for example Meta’s latest ad campaign that launched this week. It clearly highlights real AI utility, and focuses more on the feeling of using AI. This underscores the AI ​​industry’s struggle to prove generative AI has a definitive consumer application beyond chatbots and coding assistants. Equally for enterprises, the bill for the use of AI is constantly proving to be higher than the wages of the humans they have removed from the workplace.

(Vishal Mathur is Technology Editor, Hindustan Times. When he doesn’t understand technology, he often searches for an elusive analog space in the digital world.)