Railway Barons
In 1849, the California Gold Rush inspired Cornelius Vanderbilt to exit his dominant steamboat business (which handled much of New York’s commerce at the time) and enter into the railroad industry. As he took control of the New York Central and Hudson River Railroad, he created one of the earliest megacorporations in the US, providing infrastructure and support to the wide-eyed miners heading out west. Carnegie and Rockefeller followed suit, and the era of the robber barons came about through industrialists taking control of core infrastructure (Steel, Rail, and Oil) just as they rose to prominence.
Many have attempted to characterize the industry changes coming about from AI as analogous to the dot-com craze from two decades prior, but the industries couldn’t be more different. The dot-com boom came about from the proliferation of the internet and the world wide web, and the possibilities that could be realized. The most successful companies were those that mapped products from human life onto the new information superhighway: Facebook bringing human interaction online, Uber as a way to lease your car and time, AirBNB as the noveau house rental, or Google as the new encyclopedia and oracle. These companies were fundamentally product driven–nobody was making any money off of HTTPS and TCP. The value came about from the products that could be built.
Most people seem to think that the next decade will be similarly defined by a wave of brilliant AI applications–agents that plan vacations, copilots for every profession, and robots that complete everday tasks. And while I’m confident that these will exist, I believe that the real value will accrue to the infrastructure layers beneath them, not to the applications themselves.
This isn’t the consensus view. Venture capital still optimizes for product velocity and front-end innovation. But the world is shiting. When intelligence becomes a comoddity, the frontier moves downward to the primitives that enable scale, reliability, and new forms of behavior. Applications may be the visible surface area, but infrastructure is where the bottlenecks–and therefore the value–sit.
We’ve seen a microcosm of this dynamic through Crypto’s fat protocol thesis: that most value would be captured by foundational layers (Ethereum, Solana) rather than the apps built on top (eg. Uniswap). Even today with prediction markets, the value comes from the infrastructure that enables individuals to treat events as comoddities, and not necessarily the products and tools that leverage this underlying infrastructure.
The same paradigm is emerging in AI. One can point to ChatGPT as the killer AI application, but OpenAI is not an application company–it’s a research infrastructure one. Its most important product isn’t the chatbot; its a platform that lets any developer plug intelligence into their system, via primitives like function calling, memory, agents, and of course, the models themselves. And in an AI age of infinite minds, models and training are the new steel and oil.
Commodity Tennis
When looking for current value gaps, it is easy to feel like everything is priced in. AI Research companies already seem to be valued as if they are at max scale, and to some extent, they do operate at a volume that touches most of humanity. But just as with any infrastructure, the value gap lies in the commodity bottlenecks, and that is also the space with the most latent value for entrepreneurs.
Commodity bottlenecks are best thought of as analogous to the limiting reactant in a chemical reaction. For example, the combustion of Propane involves propane and oxygen mixing to release carbon dioxide, water, and energy. If my gas tank is running low, my energy output is limited by my propane, while if I try to run my car in space I’m limited by the lack of oxygen. The idea is the same for technology. For example, take perhaps the most important supply chain of the past two decades: the global semiconductor space. Software companies (eg. Google, Meta, Anthropic) need to train and deploy software, so they need quality chips (Apple, NVIDIA, Cerebras). Chip designers are dependent on fabs (TSMC, Intel), who in turn depend on EUV lithography machines (ASML), which depends on a single special type of mirror made by Zeiss that can reflect the precise wavelegnth of extreme ultraviolet light necessary for lithography. These chains have been the source of remarkable tension and value creation, as new market players with the ability to simplify one element in the line are able to benefit from the tailwinds and market pull that each player wants.
In general, infrastructure-style firms are more subject to the effects of this commodity tennis than product based firms. The reason for this is simply proximity–products generally are high level enough that they can abstract themselves from the lower elements in the stack, while infrastructure is subject both to a demand pull from above and also to the constraints of players below them in the chain. And because infrastructure generally has high barriers to entry, it is far more privy to monopolization: one need only look at NVIDIA’s dominant grip on the market, between its ownership of CUDA, the GPU business, and even of the research stack to some degree.
An interesting case study of the past few years has been Groq. Groq–founded by Johnathan Ross, the father of the TPU at Google–built something called a language processing unit (LPU), a chipset specifically optimized to run inference workloads. All the way back in 2016, the team had conceived of the need for ASICS that could handle the workload of an industry that so far was just playing board games and powering home assistants. LPUs are particularly exciting today because they help address some of the cyclical money problems in AI–by optimizing inference architecture, they enable labs to serve model access at 10x cheaper, which makes these companies offerings far more sustainable without the need for venture subsidy. Of course, as of two days ago, NVIDIA made an interesting not-acquisition of Groq for $20B, the largest in the company’s history.1 NVIDIA’s acquisition could either be a moat-widening strategy in which they offer Groq’s inference chips alongside NVIDIA products, or it may just be a means of preemptively shuttering out the competition. Considering that Google is disinentivized to ever sell their TPUs 2, it seems that the semi space supporting traditional AI research and training will remain as is for now, unless companies find spades of growth in edge compute and inference.
Considering that for much of its life technology has been focused on products first and foremost, and the new infrastructure focus is a heavy paradigm shift, we will have to find new ways to identify value. For example, bottlenecks in RAM and EV motors both trace back to specific rare earth magnets, for which the US currently has no sustainable manufacturing. There are plenty of real problems (materials, energy, etc.) that can be given a renewed focus thanks to the shift in industry, and a new way of thinking is necessary to successfully navigate and solve them.
Further Reading
https://groq.com/blog/the-groq-lpu-explained
As an aisde, the recent rise in “License & Acquihire” deals (Windsurf, Groq, ScaleAI, CharacterAI, Covariant, Adept) seem to be eroding the value structure that made Silicon Valley so historically successful. What is the incentive for early employees to take on so much risk and join a rocketship firm if there’s a significant chance they get ejected in order to help the founders and acquiring company skirt antitrust regulation? ↩︎
Google is not incentivized to sell TPUs because they are what enable Deepmind to train and offer their models at a competitive rate. Selling TPUs turns them into a commodity that equalizes the playing field to some extent between Deepmind and competing labs, and as we’ve seen with Samsung refusing to sell RAM to its own other divisions, it is quite possible for a company to merely be the sum of its parts if it is too open. Google is better positioned to keep its own training cheap and bet on the success of their models long term, with the option to raise prices later, at least strategically. ↩︎