Don't hate on GPT wrappers
We found ourselves at another venture event a few weeks back and ended up in a debate with a few other VCs about a potential investment. The details of the company aren’t super relevant, but the core question was on “moat” — or put another way, how defensible the company’s position would be and for how long.
This startup was on the radar of a bunch of firms because it is an AI startup and you can’t get fired anymore for skipping marketplace and SaaS pitches, but you can definitely get fired for skipping AI pitches. ESPECIALLY ones like this, where the founding team is focused on a vertical that they know super well and have gotten into a top accelerator and have real dollar investment from a really big tech company.
Sounds amazing! But then in the midst of the discussion, someone dared share one of the great poison pills of the AI era: “It’s a GPT wrapper, though… right?”
(Gasps and chuckles from the crowd, who has moved this unbeatable deal off into the deal dustbin — “it feels like somebody else could easily catch up to them…”)
But… why hate on GPT wrappers? We say that we want to invest in AI and we say that we want to work with AI native founders, but then we expect that narrowly funded startups AT THE SEED STAGE will have not only secured access to proprietary and exclusive data into perpetuity, but then ALSO built their own LLM? Are we expecting they will also have somehow built their own servers and chip sets, too?
I’m not saying that defensibility questions for this specific company and many others in the AI or “AI” sector aren’t valid. But I do think we have to remember where we are in the adoption of a new platform and how platform components tend to evolve.
PSA -> Before we dive in — we don’t see an investment opportunity for Front Porch at the AI “system layer” below… this will be won by the big guys. But we are focused on the application layer especially as the bundling battle outlined below plays out.
When I say platform components, here’s what I mean:
Infrastructure. Some would calls this “picks and shovels” — basically the foundational hardware and related software. Think compute, network and storage in the case of the Cloud cycle (won early by AWS and their S3 and EC2 innovations), or the processor, battery, antenna and display in the case of the Mobile cycle (won early by Nokia and Moto, then RIM, then Samsung and Apple, who each extended further and further into other platform components below).
Operating system. Think some guiding logic at the intersection of hardware and software. These protocols can be open (Linux, Android) or closed (Windows, Apple iOS). You can characterize things like Tim Berners-Lee’s link-based World Wide Web in the “operating system” category too, or even Microsoft Windows Azure and Linux, Ubuntu, etc facilitating efficient use of cloud infrastructure.
Infrastructure and a guiding model/logic (i.e., operating system) on-top are often referred to as the “system layer” of a platform — the things that have to work well and together in order for the applications to succeed. This is why so often in platform transitions, we have seen big companies play hard for the infrastructure and the operating system and a handful of powerful applications to be bundled together (resulting in years of monopoly claims against Microsoft and most recently Google and Apple in the forced use of their app stores and their associated economics).
That said, the fun for most early stage investors starts on top of the system layer:
Applications. Think of any front-end for users either built on top of system layer resources or meant to provide access to manage them. Mobile ushered in an explosion of apps. Many apps need to be found by users (through advertising or an app store), but many others (browsers, utilities, even app stores until the EU EMA and US OAMA are finally real) are embedded into the OS. We also have “applications” in things like Google Search, which leverages the web’s link-based system, or most everything SaaS that sits on top of cloud based infrastructure and can remain updated on central servers but be accessed from anywhere.
In other platform transitions, we see over time that super blurry lines start to emerge between operating systems and applications, where things like browsers that started off as apps become a system layer component for things like SaaS apps.
Data (the new gold!). The relatively open nature of the majority of the Internet means that anybody with access to Wikipedia and a few other sites is sitting on a ton of existing data (1,200+ petabytes and counting). And then users and apps create more and more data every day. Historically most of this data sat on a server somewhere, and was mostly used to target advertising, but now the closed nature of platforms like Meta and Amazon and TikTok and others means that the most lively and real-time source of data on the Internet sits within a walled garden with super high walls, though a few smaller proprietary datasets increasingly up for sale for AI purposes (i.e., the Reddit deal with Google).
Content (new! kind of). (I actually want to take this separately at a later date).
All that said, over time, a few things tend to happen —
The system layer fights for relevance
Platform components get bundled
New operating systems emerge (often apps themselves)
Competition results in consolidation
Governments get involved
If you’re not asleep yet… what I find most interesting about AI right now is how the competition and bundling of these components playing out — especially at the “operating system” (does OpenAI just become the standard open operating system of the AI era?) and data level (are publishers getting some power back?).
Which brings me back to my point on not hating on GPT wrappers.
One (sad?) version of the future is that the big tech entrenches all the way from infrastructure (except for a bunch of NVIDIA GPUs) to application, including all the data in between. Meta uses its investment in light-weight Llama 3 and massive set of proprietary first party data to launch an AI-first phone (or some other new form factor) and resurrect its hardware aspirations. Apple uses its ATT-veiled ad tech aspiration to extend its closed ecosystem into ads, which is unified with a growing set of first-party data that can then be leveraged with an AI module to create Super Siri. Microsoft leverages its relationship with OpenAI to resurrect its own mobile ambitions plus propel itself ahead of Google Cloud and AWS in Cloud.
(I could go on and on).
BUT… we have recent evidence of another version of the future, too.
In this version, LLMs get to some steady state of equivalent performance in terms of quality/speed/cost… and they get better/faster/cheaper every year on the back of a few big tech balance sheets and technical advantages. Then these LLMs become the “operating system” for AI — with some making their name as closed and private, and others as open source. Then, as long as (important!) it is not IMPOSSIBLE to shift between models as needed, then defensibility can be earned by applications through access to unique data sets and/or network effects, intellectual property or copyrights, purpose-built vertical service models, pricing, or speed to market.
What I like about this version of the future is that the application layer will be up for the taking and it may again be possible to leverage someone else’s system layer assets (including a common AI “operating system”) and built a great business.
Though of course one key here is that somehow, somewhere an underlying “system layer” of AI tech remains open, as it did in a lot of government-sponsored R&D as well as through open source innovators like Andy Rubin and Linus Torvalds — and doesn’t die in the current fear narrative that closed would be better and safer than open in AI.
So let’s not throw all the GPT Wrappers into the deal dustbin quite yet — but instead keep up hope for another open source hero who can give new entrants a chance to win as the AI platform transition frenetically paces ahead.
Because REMEMBER! Twilio was built on AWS and the PSTN… and CRISPR leveraged the system layer provided by the Human Genome Project… and isn’t Safari search is just a (pricey) Google search wrapper?… and Android begot Samsung’s One IU and Fire OS and a bunch of consoles, cameras and media players… and the first 20 years of Amazon was built on UPS… and Netflix didn’t always have their own content…