Is There a Place for Startups in the AI Revolution?
A Betteridge's Law upset
This was supposed to be a doom post. I wanted to rant about platforms capturing all the value of AI, the leverage of client bases, the leverage of capital, the waitlists, and all that. I wanted to title it “No Place For Hackers In The AI Revolution.”
But as I was typing up the post, I realized that I’m actually more excited to build than I’ve ever been.
So instead, I want to talk about why we should build for AI — despite everything.
The arrival of Internet startups
The Internet revolution favored startups.
It wasn’t just a new technology — it was a new platform. Everything you built on this new platform was 10x’ed by the access to the global computer network. You built an X on the Internet — and it skyrocketed: a store on the Internet, a CRM system on the Internet, a TV on the Internet.
The business model, however, usually stayed the same. You paid for the same thing — but over the Internet. Despite that, many incumbents couldn’t adapt their already successful business models to the Internet, and thus the startups flourished.
The subsequent mobile revolution was an extension of the Internet revolution. Putting the Internet into the palm of your hand increased the Internet penetration through time spent (rather than active users). It was the Internet, not the small computer, that drove the mobile revolution.
But again a lot of incumbents, both brick-and-mortar and online, didn’t leverage the new platform. Revolutionary mobile-native products were mostly created by startups: Twitter, WhatsApp, Instagram, Waze, Uber, Doordash.
However, it was during the mobile revolution that things started to change.
The undisruptable platforms
Some incumbents not only survived the mobile revolution but used it to outpace competitors. Google was already dominant in the search market — and cemented its leadership with Android (which it acquired early for a meager $50 mln). Facebook, Amazon, Netflix, PayPal, and Spotify successfully made the mobile switch. Instead of a “Youtube, but on mobile” startup we got Youtube on mobile.
It turned out that the arrival of a new platform couldn’t disrupt Internet-native companies with a defensible business model. If an incumbent has a customer base and high switching costs (like Salesforce) or favorable scale economies (like Netflix) or network economies (like Amazon) — launching on a new platform was not a big deal for them.
In 2022, Sam Altman talked about this in a chat with Elad Gil:
[Big tech] has become more powerful for longer and less beatable by startups than they are supposed to be. <...> We’ve had companies this big before if you look by the fraction of GDP, but we may never have had companies this powerful.
The case against AI startups
We now have the building blocks of the case against AI startups. It goes like this:
You can’t build your own model because incumbents will outspend you. If you build on OpenAI or other 3rd-party AI providers, you’ll get beaten by incumbents who will build their own specialized, better-performing models. Or they’ll just use OpenAI’s models and roll the same features out to their huge customer bases, leveraging switching costs, network effects, and scale economies.
As you look at the possible outcomes, you start to doubt yourself: is it even worth starting an AI startup in 2023?
There are 3 reasons why I think the answer is “yes,” and I’m going to go through them one by one:
The counter-positioning
The platform
The unknown
The counter-positioning
Counter-positioning is one of the strongest startup strategies, first described by Hamilton Helmer. It’s when a startup uses a new, superior business model that an incumbent can’t adopt because it would harm its existing business.
For example, Netflix’s original DVD-by-mail service disrupted Blockbuster’s $5 billion business by eliminating late fees. Blockbuster couldn’t get rid of late fees. They made up about half of Blockbuster’s income. Additionally, its brick-and-mortar stores would run out of popular titles if customers weren’t motivated to return by late fees. When Blockbuster finally got rid of late fees in 2005, the damage was so bad that it led to the departure of its CEO — and re-institution of late fees.
Blockbuster wasn’t slow or stupid when it didn’t remove the late fees, which was what customers wanted. Blockbuster was forced to cling to an old product that was less attractive to customers than Netflix. This is the core of counter-positioning: you come up with a model that would damage your competitor when it tries to imitate you.
The counter-positioning unlocked by AI is centered around AI substitution. B2B incumbents that add value to human work will be disrupted when humans are substituted by AI.
There are two main reasons for this.
First, the seat-based model doesn't translate to AI. For example, Salesforce bills large customers $150-$300 per seat per month. Almost all of this is the gross margin (74% in Q1 2023). It might look like Salesforce is well-positioned to create autonomous AI sales agents. But introducing AI sales agents means rapidly cannibalizing the super-profitable human seat licenses. To offset the compute cost for an AI sales agent Salesforce would need to dramatically raise price per seat/agent. This is a tough sell with a high damage potential.
Second, the AI substitution works against the interests of the decision-makers who originally decided to purchase the incumbent's service. For example, imagine how Intercom's offer to AI-substitute customer support agents will sound to a Head of Customer Support. "Rapidly dismantle the whole team that you built up over the years and exchange them for a black box with which you have zero competencies." Intercom’s and the client’s decision maker’s motivations are misaligned with AI substitution. Again, a tough sell with a lot of relationship damage potential.
Overcoming counter-positioning is extremely hard. The whole organization has to be re-aligned in favor of profit cannibalization for the sake of long-term survival. Feats like that are rare (Kindle vs. physical books for Amazon; iPhone vs. iPod for Apple).
The startups will attack the incumbents through counter-positioning. They will have to find entrant-friendly niches and the right decision-makers. But in the end, AI-enabled businesses will win — and with them, their AI providers.
The platform
With the introduction of plugins, ChatGPT became the first AI platform. OpenAI's demo shows plugins working together to book a table, decide on a recipe, and order ingredients — all in one chat. Essentially, AI platforms allow AIs to interact with each other to solve complex problems. AI platforms are built for AIs, not humans (think AIX instead of UX).
The AI-powered chat of the future is where AIs talk, not humans. AIs have natural language APIs, which means that you don’t need to create explicit integrations between AIs. In fact, even ChatGPT plugins can be used with AnthropicAI's Claude.
Apple and Google will likely add AI platforms on top of their mobile OSes. Apple, in particular, has the advantage of having a powerful onboard neural engine. Different apps will work together to solve complex user problems, even without explicit integrations.
This is a new domain space where incumbents have much less power. NLAPIs dramatically lower switching costs. User bases, which incumbents consider to be their moats, can become mixed ("Ask on Nextdoor who wants to buy my chair listed on eBay"). It's a huge potential for disruption for startups. The thing is — we don't know what exactly the disruption might look like.
The unknown
— You're limiting the universe to only things human could understand.
— Well you're limiting the universe by limiting the possibility of human understanding."Over The Garden Wall," "Into The Unknown"
Chat platforms filled with AIs chatting with each other to solve user problems. APIs in natural language. Generative AI agents substituting whole professions. AIX instead of UX.
We're entering the unknown. Just as it was impossible to imagine how exactly the Internet would upset industry dynamics, it's now impossible to know what opportunities, powers, and disruptions AI will create.
In fact, we're not entirely sure how to even build products for AI. Replit's Amjad Massed (hat tip to Fraser Kelton) describes how they are "testing by vibes" when building an AI assistant for developers:
We kind of check the vibes of the model, and if it passes the vibe check it goes into an A/B test to check the acceptance rate. If it inches upward, we're doing something well. If it's neutral, maybe we didn't do something useful. Would love to get more objective about it, but we haven't found a way other than building up more and more benchmarks over time.
Imagine Salesforce, Meta, or Google making a major product decision based on vibes — and being OK with that.
Eric Ries wrote that startups exist to learn. In the unknown, shipping quickly and optimizing for learning often beats economic power.
The best example of this, of course, is OpenAI itself.
It had Google, Meta, Salesforce, and Microsoft to compete with. The incumbents’ AI research labs were created before OpenAI; they were well-funded, and they attracted top talent. Meta created PyTorch, the most popular deep learning framework. Google invented the transformer technology, which backs OpenAI's GPT model. Microsoft owns 20% of AI patents issued from 2010 to 2018.
But it was OpenAI, a startup, that finally shipped a breakthrough AI product. And even after 4 months, Google hasn't produced even a runner-up product, despite an AI chat being a direct threat to their main revenue stream. The only competitor to OpenAI’s ChatGPT today is Claude, created by another startup, Anthropic.
Into the unknown
As we venture into the AI future with its unknowns, many strategic truths still hold.
It's probably best not to take on incumbents with huge user bases and network effects in a winner-takes-all market. Uber's, Meta's, and Google's core businesses are most likely all safe from AI disruption (unless, you know, the unknown).
AI can't tackle poor business models. A restaurant recommendation app is still most likely a bad business that can only exist as a part of a large incumbent's ecosystem.
An AI product probably won't overcome a strategically significant switching cost, even if it performs better. Selling an AI-powered ERP to SAP customers sounds like a bad idea.
But when you find a startup idea that promises a defense against an incumbent through counter-positioning, leveraging of AI platforms or something else entirely — be prepared that it will pull you in.
At least, that's what happened to me. I'm pretty sure I'm starting an AI startup.
I’m a product manager and a software engineer. Follow me at @sbichenko