The alchemy of transforming hardware into software

October 20, 2024

There is a class of company that I find extremely fascinating- those that transform hardware inputs to deliver software outputs. This class of company is the focus of this post.

The example that really captures today’s zeitgeist is of course AI. It’s no coincidence that talk about AGI & the intelligence explosion often coincides with discussions of some new industrial revolution. Plenty has already been written about how: “AI is a massive industrial process: each new model requires a giant new cluster, soon giant new power plants, and eventually giant new chip fabs.”

In other words, building an AI model is a massive industrial process in the input only for the output to be a large file with numbers [1], aka: hardware input, software output.

In some sense, it isn’t natural- a massive industrial process feels like it should deliver a massive industrial product, that’s certainly how it’s worked in the past. Big industrial factories ought to make cars, robots, and airplanes; now they are producing… files with numbers in them?

We’ll end up seeing it in the cost structure too, Sam Altman has said: “The most important thing to me is to drive techno-abundance from those two key inputs [AI & energy]… What I didn’t realise and got totally lucky with is how much they are the same problem. Eventually, the cost of intellgience should approximate the cost of energy… Energy will remain the constraint. I expect if we fast-forward many years, energy is the biggest constraint on the cost of AI and the ability to continue progress.” [2]

You have energy going in one end of a factory, and a bunch of transformations happen along the way, and you get artifical intelligence coming out the other end.

In other words, challenges in AI (the world of bits) asymptote towards challenges in the hardware (the world of atoms). Now, I would be remiss not to mention that OpenAI isn’t completely vertically integrated (as I understand it), they focus on the software side of the equation and offload the hardware side to Microsoft: OpenAI uses Azure as their primary cloud provider, and Azure of course relies on Nvidia’s GPUs. [3] Yet, it wouldn’t be a surprise to me if OpenAI’s biggest headaches often involves the hardware side of the equation.

What’s exciting is that the idea of hardware inputs, software outputs isn’t only limited to AI (I use the term mixware to describe such businesses). Cloud providers are a fantastic example- they require massive amounts of hardware (data centers full of servers and networking hardware) in the input to ultimately deliver software services (storage, computing, networking) in the output. Indeed, cloud is the business of taking IT infrastructure (hardware) and providing it’s software analogue, virtualised IT infrastructure (software) -as-a-service. In fact, this to me is what makes it the canonical example of mixware, because the transformation from hardware to software here is so one-to-one, as opposed to the AI example which is more of a drastic transformation.

Another example of a great mixware business is Google Maps. Google Maps is a massive database of roads, buildings, and other geographic features collected from quite an ‘industrial’ process. It required sending sattelites up to space to collect satellite imagery and humans driving all around the world across every street on the planet to take photographs of local infrastructure. All this hardware input is transformed into a marvel of a software product that I (and probably you) use every day.

When mixware is done right, it seems to benefit from the best of hardware and software. Hardware makes the business a lot more defensible and software makes it scalable and drops distribution costs to near zero. That isn’t to say it’s low hanging fruit- as the saying goes, ‘hardware is hard’ so it’s not easy to do. It doesn’t help that mixware opportunities are also probably quite rare as it requires the perfect alignment of some unique capability enabled by hardware that specifically manifests in the realm of software. This also means that I think mixware businesses often starts with some innovation in the hardware-side.

It’s important to note- mixware is not hardware + software in the output, like the iPhone or Tesla cars, it’s specifically the transformation of hardware inputs into software outputs.

One example of a mixware business that is vertically integrated is SpaceX’s Starlink, which is a vast satellite constellation that provides internet services. It’s just such an insanely good business and deserves to be mentioned as one of the best kinds of mixware businesses out there. Starlink is only made possible by a sequence of incredibly great hardware innovations, from increasingly large rockets that can carry larger and larger loads, to constellations of satellites that can always see one another at any given point in time, all to result in a great software output: abundant internet access. I should reiterate, my intuition is that the great mixware businesses all do something spectacular in the hardware-side that enable some specific capability that can only can be expressed in the world of software.

So, what are the mixware opportunities out there today? One that comes to mind is Northwood Space. Northwood Space is working on building ground stations which connect to satellites in space, with what seems to be a very well-reasoned approach, to match the explosion in satellites being added to the Earth’s orbit. My cursory understanding is that their innovation is to build many small groundstations as opposed to fewer large ones to improve satellite connectivity with what will possibly grow to be an exponentially increasing data load over time. So, what is the mixware opportunity here? Well, they could take on Amazon head-on, who themselves build ground stations as part of their ground station-as-a-service offering in AWS, and virtualise access to their ground stations.

I would love to hear other examples of mixware businesses out there- the alchemy of transorming an industrial process of hardware inputs into software is just so counterintuitive and fun!

[1] Of course, it’s a very useful file of large numbers.
[2] I must of course point out that just because, in the future, the biggest bottleneck to AI is energy, that doesn’t mean it’s the biggest bottleneck now. Even OpenAI has limited manpower of extremely smart researchers, there’s also the bottleneck of compute, and capital and so on.

Here are my thoughts:

• I think the mixware opportunity used

back