How to capture value in the AI gold rush

Startup leaders want to integrate AI into their products. Prospective founders want to build businesses on top of AI. Investors wish to generate alpha. But, it’s unclear who is best positioned to capture value as Large Language Models enter the market.

After reviewing Peter Thiel’s Characteristics of a Monopoly (from Zero to One) and Clayton Christensen’s The Innovator’s Dilemma, I believe the most valuable AI companies will be:

Many of today’s AI startups either focus on building differentiated and proprietary models or using out-of-the-box models to build SaaS products. Those using out-of-the-box models are failing to differentiate through technology, which will result in a mediocre and commoditised market. This same pressure may even push the creators of proprietary models, currently focused on providing services to startups, to eventually productise this technology themselves, offering solutions directly to businesses and consumers.

1. Proprietary technology

Technology is the best way to achieve and maintain leadership in a market. A product is best positioned to win if it can solve practical problems that other products cannot. Even with competitors, if your technology allows you to offer a dramatically better or cheaper solution, you have the best chance of dominating your market.

The most advantageous way to differentiate your technology as an AI startup is to create your own proprietary models. A startup with exclusive access to the best model for its use case has a fantastic technical moat. Unfortunately, this is the most expensive and high-risk way to build an AI startup. The expertise and resources required to create your own model are costly, and there is no guarantee your model will be better than an out-of-the-box solution.

This is why many AI startups outsource the most complex layers of their technology stack to third-party technology that is easy for others to access. They utilise APIs like OpenAI or open source projects like Stable Diffusion. This is an easy way to build an acceptable solution, but drawbacks exist. Notably, technology differentiation is difficult when your competitors have open access to your most important technology.

Another way to differentiate your technology is by augmenting the technology you are outsourcing. Many startups have used tactics like reinforcement learning to layer additional intelligence on top of out-of-the-box AI models. GitHub Co-pilot, for example, uses a GPT language model reinforced with code generation in mind. The downside of this approach is that you may have to reimplement your augmentations with each new model generation (like the upcoming GPT-4). Additionally, skeptics believe that GPT-4 without augmentations will likely outperform GPT-3 with augmentations, decimating any technology moats that depend on tactics like reinforcement learning. Organisations with exclusive access to large and valuable specialised datasets are best positioned to augment existing models. These organisations are able to train models in a differentiated way because their competitors cannot access the data they use for training (most models are trained on the public web, so it is reasonable to expect the quality of these models to eventually converge).

However, it is possible to build a technological moat for an AI company without the differentiated technology being the AI model itself. User experience may be the best pathway towards technical differentiation for many AI startups. One thing that is apparent when using generative AI is that experienced DALL-E or ChatGPT users can get much better results than inexperienced users. This is because the quality of the output depends on the quality of the prompts used. This gap in the outcome quality is an opportunity for user experience design to define market leaders. This has been the case in software-as-a-service for a long time. As cloud computing and web standards have improved over the past decade, technological moats in SaaS have diminished. As a result, many of the biggest winners through the history of software-as-a-service are less-capable products with better product onboarding and self-serviceability. The same will likely apply to B2B AI startups.

2. Economies of scale

Cost is the most significant factor for whether a new technology will help or disrupt currently dominant businesses.

New technologies that are dramatically cheaper than current solutions are best deployed by startups. These technologies tend to be worse than what exists but can satisfy niche customers because they are cheap, fast, and good enough. However, the technology improves and eventually becomes good enough for big customers. This is when startups begin to disrupt the dinosaurs in their industry.

Expensive new technologies are best deployed by incumbents who have achieved massive scale because they can absorb these costs. Even the most well-funded ventures are capital-poor in comparison to Big Tech. When a new technology is expensive to build and operate, economies of scale are challenging to achieve. This puts startups at a massive price disadvantage to big businesses willing to lose money in a market.

Significant capital is required to train a new AI model. It is also expensive to operate a model after it has been trained. This is why OpenAI’s APIs are costly in comparison to other cloud services. It is also why companies like Facebook and Google have had reduced profitability as they’ve come to depend more on AI. Each AI-powered operation has a high fixed cost, which means startups that rely on AI will struggle to achieve economies of scale compared to other software businesses.

The current state of AI looks a lot like what Clayton Christensen would call a sustaining technology. A sustaining technology is one whose value is likely to be captured by current Goliaths (maintaining their monopoly), as opposed to a disruptive technology that will usher in a new generation of Goliaths.

Successful technologies offer a solution with a dramatically better return on investment than what was previously available. Because AI products are more expensive to operate than traditional SaaS products, they must offer much better solutions to be worth the additional cost. If a SaaS solution already solves a problem cheaply, it is trivial for them to add a layer of intelligence and absorb the additional costs. This is why startups should not aim to disrupt existing software businesses by building similar products with a layer of AI.

Startups should instead focus on solving neglected problems that were previously impossible or incredibly expensive to solve with software. If a problem is currently unsolved or only solved through expensive (manual) effort, there is an opportunity for you to disrupt. While SaaS has digitised existing business operations, the best opportunity for AI startups is to completely automate them, because this is where cost savings will be greatest for customers. For example, many highly paid jobs are mostly research and synthesis (e.g., analysts, lawyers, even GPs). AI can do these jobs much cheaper than professionals.

Running models on device is another opportunity for startups. Many established SaaS companies are all-in on cloud computing. Running AI models in the cloud is expensive. But models like Stable Diffusion run great on mobile devices (Apple has even optimised their operating systems for Stable Diffusion). Native mobile apps that run AI operations on device will have dramatically lower costs than cloud-hosted AI apps. The output of these local models may be worse, but it is likely to improve. The on-device strategy sounds more like a disruptive technology (currently worse, but cheap and rapidly improving). If this ends up being the winning strategy, cloud-based SaaS startups will take time to pivot. This creates an opportunity for startups to disrupt.

Privacy and terms

I care about privacy as much as you do. I will only use your email address to send you this newsletter or to reach out to you directly, and you can unsubscribe at any time. I will not share, sell, or rent your email address to any third party, though I do store it the software I use to dispatch emails.

The information provided on this blog is for informational purposes only and should not be considered investment advice. The content on this blog is not a substitute for professional financial advice. The views and opinions expressed on this blog are solely those of the author and do not necessarily reflect the views of other organizations. The author makes no representations as to the accuracy, completeness, currentness, suitability, or validity of any information on this blog and will not be liable for any errors, omissions, or delays in this information or any losses, injuries, or damages arising from its use. The author may hold positions in the companies or products discussed on this blog. Always conduct your own research and consult a financial advisor before making any investment decisions.

Subscribe for advice

Free weekly advice covering product strategy, development operations, building teams and more.

More advice

Australia to quash angel investing

The Australian Government is about to make it nearly impossible for successful startup workers to reinvest their earnings into new startups. Let’s explore the upcoming changes and how they will affect startups, workers, and the Australian economy.

 
Stepping on toes

How much should competent people, confidently managing their responsibilities, meddle in the affairs of other teams they perceive to be dropping the ball?

 
Processes make inexperienced people wiser, and experienced people dumber

People hate process, but process is crucial to scaling a businesses. Today, we explore the difference between good and bad processes, and ways to ensure startups can benefit from standardisation, rather than suffer.