Bespoke Software Services

At a venture fund, I have reviewed startups (SaaS) built on top of frontier AI models, and I still continue to see the same pattern as a venture scout (you’d wonder if saaspocalypse was still a thing). But it’s not really a bad thing because I’ve become very convinced of just how transformative artificial intelligence (AI) will be to mankind and I believe we still have a long way to go. The founders building on top of frontier AI labs have gone on to showcase a broad range of use cases through the startups they have built. Some of them were about the same and I have unfortunately seen many of them.

It’s still phenomenal though. Building has become cheaper, and people get to work on their dream idea.

The approach is still interesting - frontier AI labs serve as the infrastructure layer while founders build solutions to those pain points. But doesn’t this mean the labs can observe broader patterns in demand? Of course, this does not mean that these labs have access to proprietary information, but successful use cases or startup solutions will provide market signals on where commercial opportunities might exist. But the question has already started shifting quite some time ago.

At what point will the frontier AI labs simply do what these founders are already doing?

The answer to that question has already started to unfold as the frontier AI labs are no longer competing on model intelligence alone. In fact, they are expanding into enterprise workflows through connectors, workplace integrations, agents and engineers who work directly with customers to implement AI solutions. The split between offering intelligence and providing an application is becoming increasingly challenging to maintain. For founders whose products offer nothing more than a convenient interface to an existing model, this is cause for concern.

After all, take into account the economics from a customer’s standpoint. Would you rather pay for a general-purpose AI subscription that supports multiple workflows or have separate subscriptions for each single-purpose AI application? It used to be simplifying invoices with a frontier AI lab as the infrastructure layer while at the moment, more services have been added - digital business cards, expense management, and so on so forth. I’ve seen about three startups doing the same thing at one event.

Ok, back to the question.

I’d say it truly depends on whether those specialised applications deliver enough additional value to justify the prices. I came up with the table below to visualise the cost and for comparison purposes. According to the monthly prices in the table, a meeting notes and marketing content application would cost a combined $99 per user, compared with $25 for a standard ChatGPT Business seat. That’s four times the price assuming the general-purpose platform can execute both the tasks and does not require minimum seat requirements.

The economics becomes more compelling when we take into account how quickly the underlying technology is becoming cheaper. Interestingly, the cost of AI inference has continued to decrease. Most recently, Epoch AI published a report titled “The Plunging Price of Thought” whereby their researchers estimated that the cost of achieving a fixed level of AI performance has decreased by approximately 47% per quarter since 2023, resulting in a 13-fold annual reduction. ¹ It goes to show how quickly access to AI capabilities is becoming more affordable albeit it not translating directly into an equivalent decline in subscription prices. That’s why for founders whose startups offer little differentiation, maintaining a pricing power could become very difficult.

Obviously the comparison laid out above assumes that the frontier AI labs can deliver the same quality of service as specialised applications, which isn’t always the case. The startups within a certain vertical that built their companies around proprietary dataset, complex workflows and integrations would make it difficult for general-purpose AI platform to replicate. This is exactly where the distinction matters because what is the point of building something without a meaningful differentiation?

This is exactly where the concept of platform envelopment comes to mind. Thomas Eisenmann, Marshall Van Alstyne, and Geoffrey Parker introduced the concept in their 2011 paper, Platform Envelopment. ² The concept essentially refers to the phenomenon whereby a provider penetrates a new platform market by bundling its services with a target’s platform to leverage shared user networks and bypass traditional innovation barriers. I believe the underlying economics are still relevant albeit the paper being published prior to the AI boom. The frontier AI labs already control the intelligence that powers most of these applications, and as they expand their product offerings, they are increasingly competing with businesses that were built on their infrastructure.

Let’s work with a hypothetical startup which develops an AI application that summarises meetings and generates follow-up tasks. It spends a couple of years acquiring customers, enhancing user experience and building integrations with workplace software. What do you think will happen when the underlying model introduces a similar functionality within an existing subscription? The startup is now left to convince existing customers that their services are still worth it. This can be quite frustrating for founders knowing something they spent years building may be replaced by the model that powers it.

This doesn’t mean founders should avoid building on frontier AI models especially when the cost will continue to reduce and capabilities will increase. Building on top of frontier AI models will still be extremely useful for MVPs. But the more important question will be where the actual value of the business resides - is it the underlying intelligence, which they don’t own or control, or in the proprietary data, workflow, distribution, integrations and customer relationships they have built? That just goes to show how capturing value and creating value are two very different things. An AI wrapper can deliver enormous value to customers, but that doesn’t mean that it can retain enough of that value to sustain attractive margins, especially when the underlying technology becomes increasingly accessible to competitors.

This brings me back to my original argument whereby several AI startups are paying frontier labs to power the very experiments that reveal which workflows customers are willing to pay for. When a commercially attractive opportunity arrives, the labs could potentially develop a similar functionality, bundle it into their existing offering and distribute it through their established customer base. As mentioned earlier, this wouldn’t necessarily translate into a successful replacement by the frontier AI labs, but it does raise an important question about moats.

I believe that the startups that will survive would have access to proprietary data, regulatory expertise, distribution advantages or customer relationships that will take quite long for the frontier AI labs to form. This notion should then propel founders to ask themselves if a specific problem that they want to solve has a defensible moat. At the end of the day, there is a massive difference between building a company using someone else’s infrastructure and building a company whose competitive advantage relies entirely on that infrastructure remaining one step behind. Once the frontier labs keep improving their models and capabilities, we shall know.