‘Something Has Gone Completely Wrong’ Pt. 2

“The basic view among enterprises in this country is I’m going to chillax and waste my time with tokens, I’m going to get no value, and they’re going to get my IP.”

— Alex Karp, CEO of Palantir

As I said on Thursday, until this week, Alex Karp and I would have agreed on little.

The eccentric boss of Palantir, a data aggregator turned war machine, is no stranger to the limelight.

Though he would balk at that characterisation — the data part, not the war part. As he’s said time and time again, ‘Palantir is not a data company’.

The second characterisation (war machine) is something they are far more comfortable with. Software that’s operationally incisive and functionally deadly is a big part of their brand. As Alex said in an earnings call:

‘Palantir is here to disrupt and make the institutions we work with the very best in the world, and when it’s necessary to scare enemies and, on occasion, kill them.’

Their Chief Technology Officer, Shyam Sankar, has even taken to the front pages of the Free Press to brag about joining the Army’s ‘Executive Innovation Corps’.

Source: Palantir

He joins technical and executive leaders from OpenAI and Meta who have also enlisted in the Army Reserve to help ‘fuse cutting-edge tech expertise with military innovation.’

In other words, to help accelerate the race to build autonomous weapons.

Swell.

But I digress. After all, I’m here to tell you about why Alex is right.

The Jig is Up

As I alluded to on Thursday, there’s an interesting gap between Palantir and other AI tech companies. And it has become far more telling in the past week.

Returning to the original quote at the top, in a CNBC interview two months ago, Alex made the case that something was wrong at the enterprise-level of AI.

According to Karp, companies have realised that the sucking sound they’ve been hearing isn’t the cost of tokens, but their data going out the back door.

As he said:

‘The complaint I’m hearing from every enterprise I work with is why would they [AI labs] get access to my data if they are going to build my alpha?’

Two months later, the sharpest example of that fear didn’t come from a Fortune 500 company. It came from an Aussie mathematician in a university office in New York.

A Million-Dollar Equation

The Navier-Stokes equations describe how fluids move. Like water through a pipe, or air over a wing. The maths behind it is basically aerodynamics.

Some engineers rely on these equations every day. But nobody has ever proved they always hold.

The open question is whether a smooth, well-behaved flow can suddenly spiral into infinity. Mathematicians call that a ‘blowup’.

It is one of seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000. Each carries a US$1 million reward. In 26 years, only one has been solved.

Tristan Buckmaster, an Aussie professor at NYU, has spent his career on the problem. Recently, he teamed up with another mathematician currently employed by Anthropic.

They ran their drafts through various AI models, mostly OpenAI’s Codex coding model.

Source: AlphaMatch

And last month, they got a key result. A subset of the solution. Edging closer to the prize.

Then Comes Ten Thousand (PR) Agents

By early September, rumours were circulating that someone connected to Anthropic had cracked a Millennium problem.

Those rumours reached OpenAI, which saw the marketing threat. After all, both companies are lining up for IPO — potentially as early as next month.

How would it look if you’re about to launch your latest AI model, and your competitor instead sweeps the headlines with a huge breakthrough?

So, they threw millions of dollars at the problem.

Over roughly 88 hours, OpenAI ran 10,000 AI agents of its cutting-edge unreleased model to crunch the problem.

As one researcher admitted, ‘an insane amount of compute had been used’.

And it bloody worked.

Source: x.com

OpenAI says it won’t be claiming the million dollars. I don’t think that was ever the point. On its face, this looked like a massive marketing coup for OpenAI.

But then came the wrinkle. That Aussie professor cried foul.

The professor’s suspicion is about the route. The line of attack he’d been working on for a year was obscure, and almost nobody else was using it.

Arriving at the same one in under a week, he argued, isn’t something you do by just handing AI the problem.

OpenAI, he suggested, used his Codex work.

OpenAI initially denied. He says that when he threatened to go public, an executive warned, ‘Why would you want to ruin your career’.

Things devolved from there, and I’ll skip the boring details. But a company preparing to sell itself to public markets needed a headline.

It got one, just not the one it wanted. For boardrooms, though, the line that matters is in OpenAI’s own statement:

‘The company cannot rule out that de-identified data from his use of its products helped improve its models.’

The Walled Garden Trade

Nothing improper has been proved. But ’cannot rule out’ is enough to make any company with intellectual property nervous.

Karp made this argument in July and was dismissed as a salesman talking his own book. He now has a case study. So what are the potential impacts?

If companies start holding back their most sensitive material, the edge moves. It goes to whoever can run a capable model inside the customer’s own walls.

That is Palantir’s entire pitch. More broadly, the same idea is called ‘sovereign AI’. It means keeping data, and the computers that process it, onshore and under local law.

As Nvidia’s CEO, Jensen Huang, said:

‘[Countries] should not be exporting flour to import bread. Every country needs its own sovereign AI – to produce intelligence rather than import it.’

Two small ASX names are trying to cover both questions between them.

SCX.ai [ASX:SCX] listed last month after raising $40 million. It runs AI models on specialised chips in local data centres, so client data stays in the country.

SCX contracted annual recurring revenue reached $6.5 million at the end of July. Paying customers grew from 13 to 49 in about two months.

Decidr AI Industries [ASX:DAI], another local AI minnow, named SCX as its first sovereign partner in May. Decidr has been buying tools that record how a company’s experts think, through guided sessions and meeting transcripts.

The aim is a private knowledge base that the customer owns and refines on local hardware.

Both are extremely speculative. SCX is early in a costly build-out and relies on one chip supplier. Decidr reports an annualised revenue run rate of $13.6 million, but it has posted a net loss of about $23 million.

Still, local AI could have a better chance than many believe if the tech behemoths continue to absorb the valuable IP of everyone in their orbit.

I still think Palantir is a dangerous company. But on this one, Karp read the room months before anyone else.