• Credit controversy reignites debate over hyperscalers’ use of IP
• ‘Last mile’ effort built on brute compute no signal of AGI
• Corporates likely to reinforce data moats at OpenAI’s expense

OpenAI’s announcement this week that an internal model had solved one of humanity’s most difficult mathematical problems should have been a cause for celebration.

On the face of it, it solidifies the potential for groundbreaking advancement in both science and mathematics – only made possible through the use of hugely powerful frontier AI models.

In short, OpenAI says its models have solved the Navier-Stokes problem which, simplified, asks whether it can be mathematically proven that fluids like water and air will always move smoothly without forming unpredictable, broken-down motion under all possible realistic conditions. Apologies to any experts for what may be a gross oversimplification.

Navier-Stokes is one of the six Millennium Prize Problems published by the Clay Mathematics Institute to highlight the largest unsolved problems in the field. Solving it warrants a price of USD 1m.

However, the achievement has been mired in controversy, with accusations spiralling that OpenAI’s breakthrough may have been influenced by the work of two prominent mathematicians using its closed Codex systems. If this narrative spreads, observers may come to believe OpenAI’s achievement is a “last mile” effort achieved only with limitless compute capacity.

Exchanges between OpenAI and Tristan Buckmaster, one of the mathematicians working on the same problem, over the credit for the achievement, before OpenAI claimed it, have turned a celebration into a scandal.

Also, when analysing OpenAI’s claims, it might be right to be somewhat sceptical over the achievement that it is claiming.

The company deployed millions of dollars and over 10,000 autonomous agents to focus on a “forced” version of the Navier-Stokes problem. This introduced artificial external forces which induced an equation breakdown.

In the eyes of this layman, OpenAI did not solve the pure problem around unforced fluid dynamics which has long befuddled theoretical mathematicians; instead, it engineered a narrow edge case where brute-force computational power triggered a mathematically induced aberration.

The Clay Prize does allow for the equation to be solved using force, but really it’s somewhat of a technical victory achieved using resource and compute that was not imaginable when the rules of the prize were laid out. OpenAI has said it will not claim the prize.

It proves that mathematical equations can indeed be broken when forced with overwhelming computational power.

Superintelligence now the IPO story

Is the achievement hugely impressive? Yes. Is it an example of AGI or the dawn of superintelligence? Perhaps not yet.

To a sceptical eye, the entire episode could appear to be a marketing effort by OpenAI to increase the potential value of its stock in the minds of investors before a possible IPO in 2027.

The reasons for this are likely to be a growing fear that the expensive metered token model of the major hyperscalers, like OpenAI and Anthropic, is coming under considerable threat from open source models and local AI infrastructure.

This displacement risk became apparent at the end of August, when Nvidia revealed that 50% of its revenues in 2Q came from its non-hyperscaler revenue segment (ACIE), marking a shift to customers building their own internal AI infrastructure at a fragment of the cost of using Claude or ChatGPT.

Several businesses have already started to move towards building internal AI infrastructure that does not depend on hyperscalers like OpenAI or Anthropic in reaction to the SaaSpocalypse sell-off at the beginning of the year.

Not only is it far cheaper for a budget conscious C-Suite, but also businesses value their data and intellectual moats highly and are happy to have an AI that is good enough for their enterprise needs, rather than a model that can solve a Millennium Prize Problem.

Andrea Pignataro, CEO of ION Group (parent of Mergermarket) outlined the necessity of gating corporate IP and workflow patterns from hyperscalers in an essay in February, and Nvidia’s results suggest this thinking has spread widely across industries.

In a market where its enterprise revenues are under threat, OpenAI’s IPO pitch is that its models are progressing towards superintelligence, an almost priceless commodity, with groundbreaking potential in almost all fields of human achievement.

If viewed in that light, the Navier-Stokes announcement makes a lot of sense – and without the controversy over the ownership of the solution, it might still have landed exactly the way the company wanted it to, even if it’s hardly in the spirit of the rules as laid out by the Clay Institute.

However, the suggestion the company may have used the work of others in their training data could prove deeply harmful in the long-term, alongside the notion that OpenAI only started working on it when it got wind academics were about to solve it and that fierce competitor Anthropic might also have been striving to solve Millennium Prize problems using its models.

The reaction of some of the world’s most famous mathematicians has been consternation.

Perhaps the most harmful indictment in the hours following the announcement came from famed UCLA Mathematics professor Terence Tao, a previously vocal supporter of the use of AI in mathematics.

Tao wrote that even the rumour of someone working on an academic problem could now trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential.

“The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field,” Tao added.

If groundbreaking mathematical research and scientific endeavour also moves towards local AI infrastructure to guard against the risk of being intellectually pickpocketed, then hyperscaler use cases shrink even further at the same time their training data diminishes.

For all the talk of AI’s potential to achieve great feats, like the helping humanity cure cancer, there now conceivably is a world where this doesn’t involve hyperscalers.

In an interview with the BBC, Professor Chris Bakal, from the Institute of Cancer Research, London, and CEO of AI-powered oncology drug development player Sentinal4D, noted that an increasing amount of medical research was being done using local models.

“It is not scraped from the internet. It does not need a giant data centre to run,” he said in the BBC piece, “The future of medical AI will not belong to whoever builds the biggest computer. It will belong to whoever has the right measurements.

“That kind of prediction could cut years from the time it takes to develop new treatments. This is where AI delivers real benefit to patients.”

If OpenAI’s Navier-Stokes announcement prompts more researchers and scientists to favour local, closed models over sharing their data with hyperscalers, it would represent a pyrrhic victory – one that could yet harm its IPO prospects.