Days of future past

25 September 2026

Pattern recognition is an important aspect of venture capital. No company exists in a vacuum, everyone invests it with qualities they have seen before. This can be both blessing and curse for new companies. Frontier artificial intelligence labs benefit greatly from this phenomenon.

Shaped by the internet

Looking back at the prime company from the internet age – Google – it now appears grossly undervalued at its IPO in 2004, let alone as a venture investment in the nineties. Investors look at OpenAI and Anthropic with minds shaped by that era and dream of the transformational economics that appear to go hand in hand with groundbreaking technology companies.

One of the reasons software in particular is so highly valued, even before the age of artificial intelligence, is very simple – costs do not scale with revenues. Software can be replicated and distributed at almost no marginal cost leading to exceptional profitability for these companies. Evaluating frontier artificial intelligence labs is difficult precisely because of the unusual nature of the product. The bull case for these companies argues that their reach will extend far beyond software into every aspect of our lives and all other companies will be dependent on them. Other software companies may even cease to exist as their capabilities pale into insignificance, so trivial will it be for future models to replicate and eventually far exceed anything an existing software company can do.

This is a fully privatised version of the Linuxification of the economy thesis with a single, victorious company at its core. Valuing such a company would mean looking at the existing revenues of all the current software companies (more if Jevons paradox holds) then add on top what would be a system of private taxation for every other company in the economy, as they would not be able to function without access to frontier models. The company that achieves this position would have more power than any in history, with a valuation to match. There are certain assumptions underlying this bull case but they are far from impossible. Anyone who told you in 2004 that Google, valued at $23 billion at its IPO would be worth $4.2 trillion in 2026 would have been dismissed as an absurd optimist, yet here we are. With this example in mind, OpenAI and Anthropic reportedly seeking $2 trillion IPO valuations does not seem excessive, assuming one of them ends up in the position described above within the next twenty years – far more powerful than Google is today.

Luxury models

The question is whether they are actually Google. Frontier labs are unlike other software companies in that their costs scale with their revenues – more conventional economics. Also, their users are not paying the cost of their models. We are in the building dominance phase with heavy subsidies from investors to users, prices will have to go up at some point but to get all the way to profitability requires there to be no real alternative to the frontier models. And the problem facing frontier labs is that there is an alternative in the shape of capable open weight models. They do not like it, but it is reality.

Frontier labs are pouring resources into models that are capable of solving obscure problems in pure maths, much to the chagrin of mathematicians who are concerned that machine brute force undermines the delicate process of generating new insights. Not only is this bad in itself, but it is doubtful how attractive these capabilities are to the broader user base of LLMs who are primarily using them to write code.

Esoteric, high performance, status signalling products could therefore be a better comparison. Nobody really needs a Lamborghini, few can afford them, but manufacturing supercars is a viable business. Just not Google. Experienced software developers are finding they can get quality code from open weight models. The unit cost of artificial intelligence continues to collapse so lighter models are only going to get more capable than they are already. Another factor is the people in a position to utilise these advanced capabilities are the very mathematicians that the frontier labs have annoyed so greatly. Also, academics do not usually have the same resources as the average Lamborghini customer.

Open source alternatives

Which leaves us with a third, more sobering outcome from the labs’ perspective. Netscape, the first true internet company. Its founder Marc Andreesen was a computer science prodigy who created the first widely used web browser (Mosaic) with Eric Bina. Far superior to earlier browsers, Mosaic was developed at Bina’s workplace, the National Center for Supercomputing Applications (NCSA) at the University of Illinois Urbana-Champaign (the university Andreesen and Bina both attended). Mosaic spread rapidly and was a key factor in the popularisation of the World Wide Web in the early nineties, transforming the internet from the preserve of academics to the backbone of the world economy. Andreesen, Bina and some of the team that built Mosaic went on to start Netscape whose most recognisable product was Navigator, one of the earliest commercial web browsers and initially a great success.

The NCSA had another project around the same time as Mosaic called NCSA HTTPd, an early web server. Its progress stalled when a group of NCSA staff left to start Netscape. Eventually a new group picked up the work and built on the foundations laid by NCSA HTTPd to create Apache – an open source alternative to commercial server software and offering for free something Netscape wanted people to pay for. Apache took off and remains widely used whereas Netscape did not last the decade as an independent company.

In February 1996, less than a year after a fantastically successful IPO the previous summer, Netscape founder Marc Andreessen was pictured sitting barefoot upon a golden throne on the cover of Time Magazine. The following month Larry Page, then a Stanford PhD student, pointed the new web crawler he created at his university homepage. His mathematically elegant insight was that the graph structure of the web was far more important than the browser used to access it, laying the foundation for what became Google.