Why Amazon Will Become An AI Chip Company

Capital intensive business with high fixed costs and the opportunity for vertical integration means Amazon should be interested.

Happy Sunday and welcome to Investing in AI. I’m Rob May, a Partner at PJC investing in AI and Robotics companies. If you are new, this newsletter is about how AI impacts products, business models, and markets both public and private. The newsletter has 3 parts. The first part has some useful AI news links. The second part has a couple of research papers I found interesting. The third part is some commentary on something AI related. This week I’m suggesting Amazon will get into the chip market. Read on to understand why.

— News Links —

AI to Replace Humans In Cybersecurity. ZDNet.

GPT-3 Is Ready For Business. IEEE Spectrum.

AI Generated Shareholder Letter. Mine Safety Disclosures.

New DRAM Could Accelerate AI. IEEE Spectrum.

Spying Eyes Everywhere Now Share An AI Brain. Wired.

Learning To Reason Over Tables. Google AI Blog.

— Commentary —

Chip fabrication is a hot space right now, after a couple of decades of being not so hot. I know a lot about hardware because I started my career at Harris in June of 2000, as a FPGA and ASIC designer. I loved the work but, it was a time when most everything was moving into software, and hardware was an afterthought for most products.

At the moment, it seems like everyone is worried about our ability to produce chips, but it wasn’t always that way. In 1987, when TSMC was formed in partnership with Phillips and the Taiwanese government, chip production didn’t matter that much. The technology stage of a chip fabrication facility is called a “process node” and is usually denoted by a size. So when you hear people say 32nm or 15nm or 7nm, that represents the process node of that fabrication facility.

Some chips can use slightly older process nodes, but part of the reason the chip fabrication business is so capital intensive is that the leaders have to keep pushing the process node ahead to smaller and more power efficient chips. TSMC started two process nodes behind the market leaders, and took a decade to catch up. Now, 35 years later, TSMC leads the pack.

But for most of recent history, the software has mattered more than the hardware, so chip fabrication was not considered a strategic business as it has been for the last year or two. So, what changed?

Workloads changed.

For most of the history of AI, AI tools were programmed and run on CPUs like everything else. But sometime around 2010, as neural networks came into vogue, some smart people realized that a GPU, which is designed to process graphics in computers, worked much the way a neural network does, and they hacked the GPU to train their neural net. It was a huge success, and NVIDIA, the world’s top GPU manufacturer, embraced neural nets.

But AI workloads continue to change, and this has led to an explosion of interesting chips. You have neuromorphic chips, spiking neural net chips, analog crossbar arrays, chips using memristors, and much more. AI workloads run a bunch of small computations in parallel instead of the serial nature of most processing workloads, so all of these chips are figuring out how to run AI workloads most efficiently from a time, power, or performance perspective.

For 60 years, most of our chips have centered on the “von neumann” architecture. Even though there are other ways to do compute, we had standardized and pushed performance forward primarily via power and speed enhancements of the same architecture. Now AI is blowing up that paradigm.

This brings me back to Amazon. The reason to get into the chip business now is that AI chips, because they are functionally different, can perform well on AI workloads without using cutting edge process node technology. So if TSMC is pushing 3nm process node tech, you don’t need a new AI chip to be fabricated with that. You can use something older and more stable and it still performs very well. In fact, it will outpeform a CPU built at 3nm process node.

What this means is that there is older equipment available, that most chip companies can’t really use, that is very valuable to the AI chip companies. That means it is a unique time to get into the chip industry.

On top of this workloads-changing-equipment moment, there is the fact that chip fabrication facilities are seen as strategic national assets, which means you can probably get matching funds or tax credits or some kind of economic benefit from putting capital to work in this area, if you ask the government for it.

I think Amazon is well positioned to do this. Chip fabrication is a capital intensive business, which Amazon likes. There are probably financial incentives to enter the market. And it’s a unique time to be able to do this without being on the cutting edge of process node tech. Plus, it’s a way for Amazon to engage in more vertical integration for AWS.

It makes sense to me, so, I expect to see Amazon enter the chip business within the next 24 months. Stay tuned to see if it really happens.

That’s all for today. Thanks for reading.

@robmay