To a large extent, the huge cost of compute infrastructure for artificial intelligence is upending the economics and strategy of “open source,” especially at Meta.
For the past three decades, the prevailing narrative of the technology industry has been one of "commoditizing the complement." This strategy, articulated by Joel Spolsky, founder of Stack Overflow, Trello, HASH, and Fog Creek Software (now Glitch), suggests that companies should strive to make the infrastructure surrounding their product as cheap and ubiquitous as possible to drive value toward their own unique layer.
By making the infrastructure a commodity, the industry enabled an explosion of innovation "above" the stack, for software as a service, mobile apps and digital services, for example.
That worked for Linux, Apache, and MySQL, which could be developed in the software realm.
In the AI era, the "infrastructure" is no longer just code that can be shared; it is massive compute power that is quite physical and therefore much more difficult to share, much less afford without clear and direct monetization mechanisms.
In the traditional software era, the barrier to entry was human ingenuity, not hardware access. A developer with a laptop could leverage a "LAMP" stack (Linux, Apache, MySQL, PHP/Python) to build a global platform.
This leveled the playing field, moving the competitive battleground to the application layer includinguser experience, network effects, and business model innovation.
That doesn’t work in an era where compute hardware is essential. Not only are investors demanding a financial return on those investments, but the infrastructure itself becomes a competitive moat, essentially changing the value of the “open” strategy.
In the past, a flourishing ecosystem where the "value" was found in what you built with the tools, rather than the tools themselves.
The rise of large language models fundamentally alters the math. For AI, the "infrastructure" is the model itself, and building that model requires a "factory" built on graphics processing units or other accelerators.
Unlike a software kernel that can be written once and distributed at zero marginal cost, a frontier AI model is a physical and linear manifestation of energy and silicon.
The difference is profound, a shift from "bits" to "atoms," virtual to physical. And though the outcome is as yet unsettled, it has been argued that only a handful of entities such as Microsoft, Google, Amazon, and Meta possess the balance sheets required to compete at the frontier because they can afford the physical infrastructure.
That is less the case now that “neocloud” providers offering high-performance computing as a service now are available.
Some might argue that, while the internet era was defined by who had the best idea, the AI era may be defined by who has the most power (electrical and financial). Again, the caveat is whether sufficient, affordable high-performance compute facilities are commercially available.
The rise of data centers focused on high-performance computing, includingCoreWeave, Nebius, Lambda Labs and others, will tend to shift the business models for some providers from capital investment to operational expenses.
So the infrastructure "compute moat" changes from an absolute barrier to entry to a variable cost. The point is that the ability to “own” a high-performance computing infrastructure is not the barrier it once seemed.
And there already are signs the terrain is shifting.
