The author is a Reuters Breakingviews columnist. The opinions expressed are his own.

By Felix Martin

LONDON, May 29 (Reuters Breakingviews) - OpenAI scored a famous victory over early backer Elon Musk last week when a California jury dismissed his attempt to unwind the artificial intelligence pioneer’s 2019 pivot from non-profit to private company status. Yet the legal triumph did not resolve the philosophical disagreement underlying the dispute. The financial implications for OpenAI and its investors are potentially existential.

The root of the long-running spat between Musk and Sam Altman, his OpenAI counterpart, was an age-old struggle between two competing software industry philosophies. On one side was OpenAI’s original embrace of open-source architecture, where products are freely available for users to download and run locally on their own devices. On the other was the more hard-nosed closed-source business model to which the ChatGPT developer and rival Anthropic have migrated: selling access to proprietary large language models on a metered basis, with the number-crunching done by massive and remote data centres.

OpenAI’s closed-source business model may have prevailed in court. Yet two days earlier and 8,500 thousand miles away a remarkable presentation by Singapore Foreign Minister Vivian Balakrishnan demonstrated why open-source AI is poised to become a bigger threat than ever in the global market. Balakrishnan explained that LLMs are now indisputably valuable for all kinds of knowledge work, while stressing that he speaks “not as an engineer – but as a practitioner with a day job.” For a government minister, however, security and privacy are paramount. Sending state secrets to be processed in the cloud by software owned and regulated in another jurisdiction, as required by the U.S. AI firms’ closed-source business model, is a non-starter.

Three technological revolutions have now converged to furnish a solution, as veteran investment strategist Michael Power has documented in a series of recent notes. The first is that the algorithms underpinning LLMs have become an order of magnitude more computationally efficient. The second is that cheap, consumer-grade chips from suppliers such as Apple AAPL.O, Xiaomi 1810.HK, and Huawei can provide the number-crunching oomph on a user's desktop or smartphone that previously required a centralised data centre running Nvidia's famous – and famously expensive – graphics chips. The third is that open-source LLMs are available from global tech behemoths such as Google, Alibaba, and DeepSeek. Power has coined an acronym to describe the radical new possibility opened up by these three revolutions: Build Your Own Data Centre, or BYODC.

Balakrishnan explained that he has done just that, building his own personal AI agent using off-the-shelf hardware and an open-source LLM. This has harnessed the miraculous productivity gains of AI without his data ever leaving his office. That has “turbocharged the pace at which things can be done”.

This discovery may seem relevant only for senior public officials with James Bond-style security requirements. Yet there are two reasons it has wider relevance.

The first is that data confidentiality and security is a bottleneck to adopting AI across a huge swath of mundane but high-value commercial applications. Lawyers, doctors, accountants and financial advisors are all subject to regulatory, fiduciary, and other compliance requirements which make cloud-based data-processing risky, if not illegal. Running an open-source model on local hardware solves this problem.

The second reason is that it’s cheaper. Using a closed-source, cloud-based LLM requires a monthly or annual subscription, with unpredictable metered charges for more intensive use on top. Moreover, today’s loss-leading prices won’t last. AI companies will eventually need to recover the costs of their epic investments.

The economics of the open-source architecture are starkly different. There is a one-off hardware expense. The software itself is free. Running costs, meanwhile, are the electricity required to power your machine.

The net result is that an open-source LLM architecture is drastically cheaper to run. UC Berkeley economics professor Brad DeLong, another early adopter, calculates that the switch cut the processing costs of a recent project from $3,000 to $31. Power estimates that the savings from setting up an open-source system for a medium-sized company would repay its capital costs in less than 18 months.

A common objection to open-source LLMs is that they are not as good as their closed-source peers. Yet the independent LMSYS Chatbot Arena Elo Rating Leaderboard ranks the performance of the best closed-source model – Anthropic’s Claude Opus 4.6 - as only 2% better than the top-ranked open-source alternative - Zhipu AI’s GLM 5.1. Besides, most commercial users probably don’t need to be at the cutting edge.

Another supposed obstacle to adoption is that building an open-source system is something only geeks can do. Generative AI itself has rendered that assumption obsolete. Technical knowledge is no longer required: you can just tell a chatbot what you want to do. As Balakrishnan says: “the barriers … have collapsed. The tools have already been invented. It is a matter of getting people to understand what tools are out there … and put themselves on a completely different trajectory.”

If he is right the AI industry is facing its own equally epochal disruption. Power warns that the shift to open-source models on local devices will turn the hundreds of billions of dollars being poured into data centres to support the dominant closed-source business model into stranded assets. This will be “the greatest example of disintermediation in US economic history”.

Any switch will doubtless not be so black-and-white. Companies and individuals will still have compelling reasons to fork out for closed-source models that are conveniently integrated into software suites like Microsoft Office. For now, the closed-source business model is delivering handsomely. Anthropic’s annual revenue run-rate has rocketed from $9bn at the end of 2025 to $45bn today, according to The Information. The company is now raising funds at a $900 billion valuation, the Financial Times reported, almost three times its price tag three months ago.

If the open-source models running on local devices really are ready for prime time, however, the days of such spectacular growth are numbered. Hardware, not software, will be the place to be. The suppliers of open-source AI-ready desktops, laptops, and mobile phones will be the winners.

Power asked Google’s Gemini model to sum that scenario up. “The PC revolution for the AI era” was its deadpan response. Investors don’t need reminding that most manufacturers of mainframe computing systems did not survive.

Follow @felixmwmartin on X.