Where Does Value Accrue In The Edge AI Tech Stack?
Rebranding, and focusing on investing in AI
Happy Sunday. I’ve re-named this substack Investing In AI for a few reasons. First, I’ve spent 5 years writing about the emerging AI ecosystem and trying to educate people on what is happening in AI. But now the most interesting thing to me is how the presence of AI impacts business models, ecosystems, companies, and markets. That’s what this newsletter will be about.
Secondly, I’m starting a podcast about Investing In AI, and already have some awesome interviews lined up.
The newsletter and podcast won’t just focus on startups, but the impact of AI on investing in public and private markets. It will focus less on new technology, and more on what the new technology means for markets and companies. As AI grows, I feel like this is the right focus for me going forward. I hope you still enjoy it. The format will still stay the same with some interesting links to read, and some commentary.
— Interesting Links —
AI is Both Miraculous and Dangerous. EE Times.
Chinese Lab Aims for Big AI Breakthroughs. Wired.
The First Bots Get Fined For Ticket Scalping. The Verge.
Gaze Estimation and GANs. Synthesis Blog.
Plasmonics. IEEE Spectrum.
Thoughts On Personalization Algorithms. Alve Mine Blog.
— Research Links —
A Multilingual Benchmark for Navigation Instruction Following. Link.
Fidelity and Privacy of Synthetic Medical Data. Link.
— Commentary —
Edge AI is going to be a big industry - bigger than cloud computing even, as everything everywhere gets some intelligence in it. The question to ask then, as an investor, is where will the most value accrue? Today I want to attempt to answer that by looking at 3 key areas of edge ai: chips, tools, and applications.
CHIPS - While you might think there are a lot of computer chips, there really aren’t. Particularly in the processor market, the core architectures are all pretty similar and the variations between products are small compared to the space of possible computer architectures. AI is changing that.
GPUs don’t solve all the problems because if you double the size of a model, it typically quadruples the time and power it takes to run on a GPU. As a result of the proliferation of AI workloads and different architectures (CNNs, DNNs, RNNs, etc), and different technical approaches to solving them (neuromorphic, analog crossbar, in-memory compute, spiking neuron, variations on CPU/GPU optimization) the result is fragmentation in the chip market when it comes to AI at the edge.
I think long term as tools standardize and workloads consolidate and architectures settle on best-of-breed that AI chips will consolidate quite a bit, but I think that’s 15-20 years away. If you think of investing in both public and private markets on tech cycles, the next two, maybe three tech cycles will still have major fragmentation at the chip layer.
That doesn’t mean they are bad investments. Chip companies are awesome when they win. While they take longer to get to market, and more capital to get to revenue, they are massively scalable, and can grow revenue even faster than software companies once they go. My take on investing in the chip market is that you can make money but you have to go deep and be really smart and make bets on which chips will win because there are A LOT of options.
TOOLS - The tools space for edge can be broken down into a few parts. The NN development tools like Tensorflow and Pytorch, which aren’t edge specific but do have features to help target edge devices. The MLOps space that includes companies like Verta doing model management. Although I don’t think any of them do edge models right now. There are companies that help design and compress models for the edge like Deeplite and OctoML.
These levels will, in my opinion, be oligopolies because they need to work with so many different ecosystem partners that once the first two or three companies hit sufficient support for many tools and chips, it won’t make sense for others to do it. The overall value accrued to this part of the stack versus the chip part of the stack may be higher because the fragmentation at the chip layer may cause price competition and these tools will be frequently used parts of key workflows.
APPLICATIONS - The applications layer will have a few new solutions, and many old solutions that have been upgraded with AI. Think of a security camera with built-in face recognition. At the application layer of the edge AI tech stack, the value accrue and the defensibility will most likely mirror that of the underlying market. For example, if automotive is competitive then AI automotive will be competitive. If security cameras are commoditized then AI security cameras will become commoditized.
There can be big investment wins at this layer, but, you need to find the winners of the pre-AI era that are moving rapidly into AI. Or, you need to find use cases where AI builds a flywheel that somehow takes a commodity market and brings some uniqueness to the first company to win the AI version of it. I don’t know what that would be off the top of my head, but I am keeping my eye out.
While there will be winning bets in all three layers, I think the tools layer is the most likely to accrue high value per company with the lowest overall investment risk. The chip layer will have big winners but will require a lot of research about use cases and workloads and what is getting early traction. And the application layer will mirror what happened in the pre-AI versions of those use cases.
That’s all for this week. Thanks for reading.