The battle for the future of artificial intelligence is between open- and closed-weight models.Dado Ruvic/Reuters
In our last column, we made two arguments about which artificial intelligence companies will eventually emerge as losers.
First, OpenAI and Anthropic are the most financially fragile companies in the AI sector. That’s because they are single-product entities selling a commoditized product into a market with free open-source models, financed by recurring capital needs that assume a profitability they have not yet shown. Secondly, these are the most likely companies to go to zero just as they are the lining up for their IPOs.
Recent developments have vindicated us, and they underscore how Canadian investment portfolios are quietly being put at risk.
Why Meta’s Mark Zuckerberg is pushing the open approach to AI models
On July 24, Nvidia chief executive Jensen Huang published an open letter – titled “Open Weights and American AI Leadership” – put together by the chip giant and signed by 25 companies. Among them were Microsoft, Meta, IBM, Palantir, Cisco, Dell, Mistral, Hugging Face, Perplexity, ServiceNow, Andreessen Horowitz and Canada’s own Cohere.
Open-weight models are those which publicly release the enormous files of numbers that encode all the model has learned through training. Meta’s Llama, France’s Mistral, and China’s DeepSeek and Qwen are examples. They are free, and can be downloaded and run on your own machines at zero cost.
Proprietary models, including OpenAI’s GPT, Anthropic’s Claude and Google’s Gemini, are closed-weight. They can only be rented on the vendor’s servers, at the vendor’s price, with the vendor seeing every query.
None of the three companies selling access to the best closed models in the world – OpenAI, Anthropic and Google DeepMind – had signed Mr. Huang’s letter.
This shows which layer of the AI stack the industry is trying to commoditize.
The letter’s signatories were chipmakers, cloud platforms and enterprise-software companies who do not want frontier laboratories owning the operating relationship with their customers. Every one of them benefits if raw intelligence becomes cheap and interchangeable, because every one of them earns its money somewhere else in the chain.
Mr. Huang wrote: “If there is great AI, even if it is open and wherever it comes from, there will be more use, more Nvidia computers sold, more data centers built.” Nvidia organized the letter demanding openness in terms of models, while keeping its Compute Unified Device Architecture – the software that actually locks customers to its chips – proprietary. Everyone in this coalition is arguing for openness in someone else’s yard.
AI hyperscalers upend usual earnings stock swing pattern
This is the argument we made a few weeks ago. The industry seems to have collectively decided that the model layer is a going to be a commodity. That is to say, where margins go to die.
The numbers already show this clearly. Artificial Analysis, an independent benchmarking firm, puts every major model through the same battery of hard tests, including graduate-level science, competition mathematics, coding and multistep reasoning. The firm then rolls the results into a single score out of 100 that it calls the Intelligence Index. It is the closest thing the industry has to a standardized scoreboard, and it is public.
At the top sits the most capable proprietary model in the world, Anthropic’s Claude Opus 5, at 60.7. In fourth place, ahead of Claude Opus 4.8 and every version of ChatGPT, is Kimi K3 from the Beijing laboratory Moonshot AI, at 57.1. Only three per cent separates the best closed models that money can rent from the best free, open-weight model built in Beijing.
So, 25 American companies signed a letter in praise of open weights. It was published during a week where the open frontier was being set and reset in Shanghai, where Chinese President Xi Jinping committed his country’s AI ecosystem to open source in a keynote at the World AI Conference.
In sharp contrast, no U.S. lab has committed to releasing a frontier open-weight model. The leading American open release is Thinking Machines Lab’s Inkling, which arrived on July 15 and scores a measly 42 by Artificial Analysis.
Meanwhile, Washington has decided that the proprietary frontier is a strategic asset and has begun defending it as one, threatening sanctions against Moonshot and other Chinese labs.
This episode puts the American AI ecosystem in direct opposition to its own government. The Huang letter asks policy-makers not to conflate “legitimate model-development techniques” with misappropriation. That is the industry telling Washington to stop protecting the very thing it has decided to protect. The ecosystem is dead-set on commoditizing the model layer. The government is dead set on defending it.
Ironically, for many users, the strongest argument for open weights this year was made by the U.S. government, by accident. On June 12, the Commerce Department’s Bureau of Industry and Security sent Anthropic an unpublished letter imposing export controls on its two most capable models, Claude Fable 5 and Claude Mythos 5. Anthropic had to disable both for every foreign national on Earth – including its own foreign-national employees – and the models were pulled from European countries.
The restriction was withdrawn 18 days later, but this was long enough to prove something permanent, and to teach it to every chief technology officer outside the U.S. A closed model is a chokepoint that a foreign government can switch off without notice. An action intended to protect the American frontier turned out to be the most effective advertisement the open-weight camp has had all year.
All this was further underlined on Monday, when Meta CEO Mark Zuckerberg called to lower barriers in the U.S. for open-source AI models to compete with Chinese rivals. Meta just released a new open-weight model this week, called Muse Glimmer. In a statement, Mr. Zuckerberg said the U.S. needed to rethink its policies if domestic firms were to lead in open-weight models.
The conclusion, as we drew it a few weeks ago, now has better evidence behind it. The layer that is being deliberately commoditized is the one carrying the most capital and the highest valuations. The companies driving that commoditization are precisely the ones whose profits sit somewhere else.
Most Canadians own stakes in that layer involuntarily, through index funds held by pension plans and the default options in their retirement accounts. It’s unlikely they have been told which part of the stack their money is sitting in.
The question for anyone underwriting this cycle is not which model wins. It is who owns the capacity, the data, the workflows and the permissions surrounding whichever model happens to be winning this quarter. Those aspects – not the leaderboard – is where the money will be.
George Athanassakos is a professor of finance and holds the Ben Graham Chair in Value Investing at the Ivey Business School. His latest book is Value Investing: From Theory to Practice. Christian Dippel is an economist at the Ivey Business School.