Over the past three years, venture capital firms and Wall Street hyperscalers poured hundreds of billions of dollars into a singular thesis: the conviction that proprietary frontier Large Language Models (LLMs) represent the greatest natural monopoly in the history of commercial software. Corporate boardrooms repeated the same playbook: because pre-training a state-of-the-art foundation model demands tens of thousands of top-tier GPUs, industrial-scale megawatts, and petabytes of curated training tokens, only a small cartel of global tech giants could ever afford to compete at the technological frontier. Whoever raised that multi-billion-dollar capital moat would extract monopolistic pricing rents through closed API subscription gates forever.
As 2025 unfolds, that commercial house of cards is disintegrating in real time. The emergence of open-weight, distilled reasoning models —matching or exceeding the coding and mathematical benchmarks of the most expensive proprietary models while consuming a microscopic fraction of compute— has triggered an institutional repricing across equity markets.
Those who view this disruption as an unforeseen shock simply misunderstand the economic history of software. What we are witnessing in artificial intelligence is nothing more than the inexorable operation of an economic law formulated twenty-three years ago: Joel Spolsky's Law of Commoditizing Complements1 Joel Spolsky articulated this microeconomic principle in his seminal 2002 essay Strategy Letter V: Commoditize Your Complement. Spolsky demonstrated that commercial technology companies have an enduring economic incentive to commoditize and cheapen products that complement their primary revenue driver, driving competitor margins to zero to expand demand for their own offering..
Spolsky's Axiom: Destroy Your Complement's Margins
Spolsky's insight is disarmingly elegant: in any technical value chain, a firm's product relies on specific complements. A rational company will systematically work to commoditize those complements down to their marginal cost of production, obliterating supplier moats to maximize the economic value captured by its own core platform.
[ DEMAND FOR YOUR CORE PRODUCT ] ◄── (Expands when complements become free)
│
▼
[ COMPLEMENTARY COMPONENT ] ──► (Its margin compresses to marginal cost)
The history of computing is governed by this operational dynamic: - IBM standardized PC architecture so that third-party hardware could be cloned cheaply to accelerate unit sales; Microsoft capitalized on that silicon commoditization to monopolize the underlying operating system (MS-DOS and Windows). - Years later, Google acquired and distributed Android as an open-source operating system not out of philanthropic altruism, but to commoditize mobile software, guaranteeing that no hardware vendor could erect a tollbooth between mobile users and its search advertising index. - Meta open-sources frontier weights like Llama for the exact same strategic reason: it must compress the profit margins of OpenAI, Google, and Microsoft to zero, ensuring that foundation models never become a private tax on its social media graph and consumer hardware ecosystem.
The Illusion of the Weight-Based Moat
Executives at closed API labs committed a foundational category error: they believed that the weights of a neural network constituted an impregnable proprietary moat, comparable to the secret recipe of Coca-Cola.
They overlooked the reality that software and mathematics are pure information. The instant a proprietary frontier model exposes its generation API to the public internet, it implicitly leaks its probability distribution. Through synthetic distillation, targeted supervised fine-tuning, and algorithmic reinforcement learning centered around test-time compute2 The architectural pivot from massive pre-training toward deliberative test-time compute proved that model capability does not depend solely on static parameter scale, but on the algorithmic ability to explore reasoning search trees and self-correct intermediate logic prior to emitting final tokens., independent engineering teams worldwide can train comparable architectures at 1% of the original training capital expenditure.
A leaked internal Google memorandum in 2023 diagnosed this vulnerability with prophetic accuracy: «We have no moat, and neither does OpenAI. The open-source community is running circles around us because their iteration cycle is orders of magnitude faster and cheaper»3 Internal memorandum written by senior software engineer Luke Sernau (We Have No Moat, And Neither Does OpenAI), leaked in May 2023. Sernau analyzed how grassroots innovations (4-bit quantization, LoRA parameter adaptation, consumer-GPU memory optimizations) consistently undermined the scale advantages of proprietary clusters..
The Sovereignty of Local Silicon
For the systems architect and the rational investor, the collapse of this artificial capital barrier inaugurates the era of truly sovereign computing. So long as mission-critical workloads depend on an opaque API hosted within a commercial vendor's data center, an organization's infrastructure remains hostage to unilateral price hikes, terms-of-service deprecations, and corporate content filtering.
When frontier-class models run natively on local hardware —whether on a unified-memory workstation or a homelab server cluster— artificial intelligence ceases to be an unbacked operational subscription and transforms into a permanent capital asset. The commoditization of foundation models is not merely an economic victory over software monopolies; it is the restoration of technical autonomy into the hands of the sovereign individual.