November 2023
From: Brian and Tobias
Subject: The innovator's dilemma and ways to create value with AI
We’ve spent 2023 vacillating between enthusiastic awe and measured concern. It’s not that we’re torn between techno-utopian ideals and doomerist fears. As VCs, we’re thinking about where there’s opportunity. Two things are clear:
(1) AI is insanely powerful and will create tremendous value
(2) Value will accrue in complex ways and making investments is tricky
This newsletter explores The Innovator's Dilemma and Sustaining vs. Disruptive Innovations, both from Clay Christensen, in the context of AI. These two concepts provide a helpful framework for navigating this overwhelming landscape. Let’s dive in.
AI is the first potential “platform shift” Tobias and I have seen since becoming investors. It's the thing we’ve been waiting for. As a result, many have debated whether AI will be a sustaining innovation or a disruptive one, a concept Clayton Christensen used to explain why some innovations disproportionately benefit incumbents. Ben Thompson, inspired by Christensen, provides an easy framework for assessing the quality of an innovation through this lens: “if the innovation was sustaining, then incumbent companies became stronger; if it was disruptive then presumably startups captured most of the value.”
Thompson has argued that we’re in store for a sustaining innovation because incumbents – Microsoft, Apple, Google, Amazon, Facebook, etc. – will easily embed AI into their existing products. OpenAI and others have made AI consumable via a simple API, democratizing access to underlying foundation models. If open source continues to advance, then the value to incumbents should be all the greater. The hardest tech challenges in AI are no longer a barrier. Rather, data to fine-tune models and distribution to expose applications built on those models matter much more. Incumbents have clear advantages over startups on both fronts.
If you’re an investor, you’ve played out this future before. Incumbents are well-positioned to capture value across the obvious AI use cases:
- A startup that automates slide creation – why won’t Microsoft do this?
- A startup that automates and auto-completes illustration – why won’t Adobe do this?
- A startup that automates email writing and response – why won’t Google do this?
These incumbents are already doing all of the above, and the data bears out that incumbents are benefitting. Accel analyzed the market cap growth of the biggest tech companies from September 2022-September 2023, a good proxy for AI’s impact on valuation. There is no question it has been an accelerant for incumbents:
On the flip side, there are many smart folks arguing that AI is a disruptive innovation. Peter Fenton recently appeared on Patrick O'Shaughnessy’s Invest Like the Best podcast and made the argument that AI is a disruptive innovation because it is creating demand and pricing disruptions in ways that impact distribution. AI is able to fundamentally change the way software is priced because it can do the work, not just supplement or augment the work. Sarah Tavel’s blog post on this topic is a great read. Applications like Midjourney, Writer, and Runway are all data points to support this argument.
If we had to pick one of these camps, we lean towards the sustaining one, especially in the medium-term. We’ve seen the clear incorporation of AI into enterprise products, and although startups will surely emerge, we think more of the value will accrue to incumbents.
And yet, we actually don’t think that “choosing” per se is even important. Instead, we think that the whole sustaining vs. disruptive question could be a big red herring. If AI creates $10T worth of incremental value (see above graph; it has already created 25% of that) and if startups capture just 1% of that, $100B of startups will emerge. So even if AI ends up being an extreme version of a sustaining innovation and you’re bearish on its value, this is still a huge opportunity for VCs, and that’s really all that matters.
Coatue recently published a 2023 report on AI. If you haven’t read it yet, it’s really worth your time. There’s a lot of great content in there, but one slide stands out:
Incumbents clearly won the mobile platform shift, capturing $6.5T out of $7T of value (93%!!!). And yet, startups that rode this wave created $500B of enterprise value and included some of the most iconic startups in history – Airbnb, Uber, Doordash, Snap. If you were an investor who missed out on mobile, you made a massive and maybe career-ending error, even though incumbents captured the lion’s share of value.
So, the question investors should be asking isn’t whether AI is a sustaining or disruptive innovation. Rather, if they think AI will create a lot of value, the question is: where to invest?
Multiple times this year, we’ve returned to the frameworks of Christensen to hone our lens on where to invest in AI. Specifically, we have turned to the Innovator’s Dilemma, from which all innovation theory (including the idea of sustaining vs. disruptive innovations) is spawned. The core idea of the innovator’s dilemma is that incumbents tend to avoid initiatives that cannibalize, undercut, or otherwise threaten existing revenue streams. It's a Catch-22. They're well positioned to capitalize on a new innovation, but they would need to sacrifice existing business in the process. In these situations, startups have a chance to beat better-resourced incumbents because they move quickly towards an opportunity incumbents are disincentivized from pursuing.
This concept animates our view on AI. It is hard to make the argument that incumbents couldn’t build what new AI startups are building if they wanted to. However, what is important isn’t whether incumbents can copy a startup, but whether incumbents will and are incentivized to. This is the teaching of the Innovator’s Dilemma.
Although true defensibility is hard to come by in this new AI world, we believe there are opportunities to exploit the Innovator’s Dilemma, and these are the most fertile places for us to invest. Google is the most commonly cited example regarding AI. GenAI has the potential to make search as we know it irrelevant, and, lo and behold, Google has been chronically slow with GenAI, despite having all the assets to be the market leader. It’s hard to jeopardize one of the greatest cash cows the world has ever seen.
Where does the Innovator’s Dilemma exist for startups? A few ideas:
- Our portfolio company Etched – creating specialized chips for LLM inference that NVIDIA is not incentivized to produce because doing so would cannibalize the GPU and bypass the CUDA moat it has spent years and tremendous capital building.
- GPU clouds – one argument Coreweave makes about its solution is that its architecture is fundamentally different from the hyperscalers (dependent on hypervisors), meaning hyperscalers would need to re-architect their solution to truly optimize for AI. If this is true, it would mean GCP, Azure, and AWS would need to disrupt themselves to optimize for AI workloads.
- New app layer modalities – companies like Salesforce are heavily invested into some core organizing modality, like the “CRM.” Startups in the future could build apps with a totally different interface that just requires chat and voice. For Salesforce to unwind its app ecosystem and CRM interface would be almost unthinkable. This one is abstract, but we’re sure we’ll see new ways of app consumption that the incumbents probably won’t replicate because they’d be disrupting themselves in the process.
- LLM-native AIOps – data infrastructure businesses like Databricks and Snowflake built their architecture and business models prior to AIOps. While they’re both adapting quickly it is possible that hidden requirements of their business model will blind them to specific opportunities at the expense of others.
- Cloud security – much of cloud security is “dashboard and routing” centric. The products amount to dashboards that summarize lots of information and routing engines that send tasks to engineers. This architecture may not make sense if you build from the ground-up using GenAI interfaces and capabilities.
The concept of the Innovator’s Dilemma is connected to the data unlocks we wrote about last month, and also another concept we’re calling “compute unlocks” – solutions that help with speed, latency, and cost issues related to AI compute. In the examples above, an incumbent’s core architecture is the very bottleneck that necessitates an unlock. Because GPUs are not built for LLMs, costs are exorbitant. Because hyperscaler clouds are not optimized for AI workloads, speed is compromised. The list goes on. We think that these unlocks are actually examples of the Innovator’s Dilemma in disguise – an incumbent architecture lags, so there needs to be an unlock from a startup. These are two sides of the same coin.
Last but not least, one core component of the Innovator’s Dilemma is that disruption takes root in parts of the market that are overlooked by incumbents. These are the best places to look for investment opportunities. In the words of Clayton Christensen:
“First, disruptive products are simpler and cheaper; they generally promise lower margins, not greater profits. Second, disruptive technologies typically are first commercialized in emerging or insignificant markets. And third, leading firms’ most profitable customers generally don’t want, and indeed initially can’t use, products based on disruptive technologies.”
If there’s a core motto for our AI investing activity, it is: invest in AI startups building products that the incumbents aren’t incentivized to replicate, where the solutions enable a compute or data unlock and target an overlooked customer segment.
As always, please let us know if you have thoughts and feedback.
Until next time,
Brian & Tobias