AI Architecture Of Exclusion: The Need For New Anti-Trust Jurisprudence
The AI architecture of economic activities linked to it — which is almost everything now in the tertiary sector and would soon be in the primary and secondary sectors too through AI enabled robotics — is anything but democratic. A technology that was supposed to lower barriers, compress hierarchies, and distribute capability more broadly is being betrayed by the AI architecture that has grown around it.
The AI industry is controlled at every distinguishable layer by a handful of entities, who invest in one another, supply one another, and collectively preclude any meaningful competition. Anti-trust law is incapable of addressing this form of power that operates across layers, through contractual arrangements and the sheer cost of entry.
This piece examines the AI architecture, the mechanisms through which equal opportunity is being lost, and the reasons why existing anti-trust jurisprudence is inept to check it.
The Five-Layer AI Architecture of Control
The AI architecture is best conceptualized as a stack of five interdependent layers: silicon, chips, cloud infrastructure, models, and applications. Each layer has distinct economics and barriers to entry, but the dominant entities integrate across layers foreclosing opportunities across the spectrum.
At the very bottom of the stack sits TSMC (Taiwan Semiconductor Manufacturing Company) , which fabricates roughly 90 percent of the world’s most advanced chips. TSMC is not a chip designer. It takes blueprints from companies like Nvidia, Apple, and Google and turns them into physical silicon.
TSMC is the only company capable of fabricating the 3nm and 2nm chips at scale. Samsung and Intel are nominally competitors, but their capacity lags significantly. SIMC, the Chinese company, has made inroads in 7nm chips, but it is not significant enough. TSMC’s limited capacity to manufacture is the biggest chokepoint in the industry, leading to long waiting periods — TSMC is booked till 2028.
Chip design layer sits above the manufacturing layer, which is dominated by Nvidia, holding 80 percent plus share of the AI accelerator market. It depends entirely on TSMC to turn its designs into physical silicon. Other players include AMD, Intel and Google. Google sells Tensor Processing Units (TPUs), whereas others sell Graphic Processing Units (GPUs).
Nvidia’s dominance is not a factor of the quality of silicon. It sells GPUs, which are in fact less efficient than TPUs in strictly neural network operations used in AI. Nvidia’s dominance is rather a result of its software platform CUDA, which has become the de facto standard for AI development. Nvidia exploits it by restraining reverse engineering in its contractual terms, thus closing the opportunities to translate CUDA output for integration in other platforms.
Google has developed TPUs, which are designed specifically to perform tensor operations very fast and efficiently. Neural networks do enormous amounts of tensor arithmetic — mainly matrix multiplications and additions. But when a company builds its AI pipelines on CUDA, switching to a different platform like Google’s JAX is prohibitively expensive, requiring re-writing of millions of codes. Also Google doesn’t sell its chips per se except to gigawatt-scale buyers like Meta and Anthropic. It only rents them to others through cloud. TPUs are primarily used captive by Google only.
Huawei does design Neural Processing Units (NPUs), which are also specialized AI chips like TPUs, but they are mostly consumed domestically in China only, especially after the US export controls on China.
Above chip design sit the hyperscalers-Microsoft Azure, Amazon Web Services, and Google Cloud-who rent compute capacity. These entities are basically the gatekeepers of access to AI capability. The evidence is now apparent that cloud providers have been prioritizing GPU inventory for internal teams and major clients over smaller customers.
It is not discrimination per se but a structural feature of the market. The hyperscalers have incentive to allocate scarce GPU capacity to themselves and their strategic partners, or to those customers who generate the most revenue. Microsoft Azure prefers OpenAI, Amazon Web Service ties up strategically with Anthropic, and Google, of course, has Gemini.
The Chinese cloud market works under chip supply constraints. Prioritization is rather a norm there. Huawei Cloud has state-aligned priorities. Tencent Cloud, Alibaba Cloud, and Volcano Engine of ByteDance, have heavy internal consumption and strategic customer tiering systems. Baidu AI Cloud is an exception as its own consumer app traffic is relatively small.
Frontier model developers control the most capable AI models in the world. This layer sits on the top of the cloud infrastructure layer. Though there are many players in this layer, OpenAI, Anthropic, and Google together capture the vast majority of frontier model revenue and usage.
DeepSeek and other Chinese labs have made significant inroads, particularly in open-weight models, but they operate under severe constraints imposed by export controls and limited access to advanced manufacturing.
This layer is dominated by an apparent vertical integration. Microsoft financially backs OpenAI, and then bills it for Azure compute. Google and Amazon both fund Anthropic and make a show of competing for its workloads. These strategic partnerships create dependencies that make it difficult for new entrants to compete.
At the top are application companies; i.e., the startups and enterprises that build products on top of frontier models. This is the most visible layer to the public, and it gives the illusion of high competition and fragmented AI market. But the players in this layer have the least power.
The entities in this layer fully depend on the layers below them for survival. They cannot train their own models without incurring huge, prohibitive costs. They cannot build their own data centres. They cannot manufacture their own chips. They operate under a dependency hierarchy competing on thin margins and constant pressure from the platform owners, who can replicate their products at will.
The Mechanisms of Exclusion
The AI architecture is a system of structural exclusion that operates through four distinct mechanisms.
Get Virtus Legis’s stories in your inbox
Join Medium for free to get updates from this writer.
The most direct mechanism is access to compute. Training a frontier model requires huge investment and access to resources — even fine-tuning a smaller model requires significant compute resources. Cloud providers allocate the resources preferentially to their own teams, strategic partners, and largest customers. Thus the players at the bottom — startups, researchers, small enterprises — just can’t develop new models. All they can do is employ strategies like model distillation to cater to specific production tasks (eg. Shopify).
This is not a level playing field. It is a compounding structural advantage for vertically integrated players like Google, Microsoft and Amazon. The entities who align with them get the best resources to train the best models and attract the most users. The entities without partnerships are left to compete on scraps.
Nvidia’s CUDA platform is not merely a tool; it is an ecosystem that shapes how developers think, write code, and build products. The CUDA-specific libraries, the frameworks, the debugging tools, the community knowledge makes it next to impossible to switching to a competitor’s platform-whether AMD’s ROCm, Intel’s OneAPI, or Google’s JAX. Relying on one vendor’s stack makes players vulnerable to pricing changes, contract renegotiations, and strategic shifts. This is not a market mechanism but an inherent feature of the AI architecture, leading to unequal opportunities.
AI models require vast amounts of high-quality training data. The big players like Google enjoy advantageous access to data through its search index, Meta’s social graph and Amazon’s e-commerce data also enjoy a structural advantage that new entrants cannot replicate.
Furthermore, the entities with initial advantage continue to enjoy the strategic advantage as they generate operational data in the course of their operation that can be used to fine-tune models. The result is a feedback loop. The entities with the most users generate the most data. With the most data, they train the best models. With the best models, they attract the most users. And so on, it continues indefinitely.
In recent years, the AI players have developed a novel technique for eliminating emerging competitors without triggering merger review: the acqui-hire or stealth merger. Instead of acquiring outright, a large player pays huge sums to hire away its founding team. Microsoft’s absorption of Inflection AI, Google’s re-hire of Character.AI’s founders, and Meta’s acquisition of Manus AI’s core personnel are such examples.
Regulators have started examining whether such acqui-hire deals trigger anti-trust regulations, but the enforcement remains nascent, and the deals continue.
The Failure of Existing Anti-Trust Jurisprudence
The existing anti-trust jurisprudence is designed for a world of discrete markets and identifiable harms. It is incapable of addressing the form of power that has emerged under the new AI architecture.
Anti-trust law requires defining a relevant market, but the AI industry is not a collection of discrete markets. It is a vertically integrated stack. Traditional market definition cannot identify the form of power that operates through the vertical integration under the new AI architecture.
The existing anti-trust jurisprudence targets the consumer welfare standard; i.e, whether a practice raises prices or reduces output for consumers. But the new AI architecture forecloses business opportunity, which in consequence affect the consumer welfare. The focus on consumer harm makes the existing anti-trust jurisprudence oblivious to the structural exclusion characterizing the new AI architecture.
Section 1 of the Sherman Act requires proof of an agreement between competitors to restrain trade. But under the new AI architecture, the unilateral decisions about resource allocation, contractual terms (Nvidia), and investment relationships (acrui-hires) create structural barriers to restrain trade. Then there is possibility of algorithmic tacit collusion; i.e., price alignment that emerges from the interaction of algorithms rather than from any human agreement. Section 1 of the Sherman Act, which requires proof of a meeting of the minds, fails to address such collusion.
Finally, there is political resistance tied to perceived reliability of existing jurisprudence. The argument goes that AI markets remain competitive, that concentration concerns are overstated, and that enforcement would do more harm than good. But it ignores the structural features of the new AI architecture.
The New Anti-Trust Jurisprudence
The new anti-trust jurisprudence is indeed a candidate for organic growth, but recognizing its features upfront makes the change less painful. I recognize the following as basic essentials of any new jurisprudence.
Recognizing AI compute as an essential facility is but necessary for the new anti-trust jurisprudence. Those who control compute have the power to exclude. Thus obligations of non-discriminatory access and published pricing on cloud providers and chipmakers is necessary.
Nvidia’s CUDA dominance is a barrier to competition that has nothing to do with the quality of Nvidia’s hardware. Mandating interoperability, which will make Nvidia’s contractual terms, barring open-source translation, illegal, is necessary to reduce switching costs and thus open up competition.
The new AI arrangements like acqui-hires, minority shareholdings, and strategic partnerships should be assessed for their economic substance, not their legal form. If a transaction transfers control over a competitor’s technology and talent, it should be treated as a merger.
The AI architecture, as it has developed, will never provide equitable access to AI infrastructure because the returns are concentrated at the top. Imposing public duty on top layer companies irrespective of their legal structure, will establish a right to AI infrastructure as part of constitutional equality paradigm. Not treating AI infrastructure as a purely private good is not a choice but a necessity.
Consumer welfare standard is a useful tool, but it is not the only measure of competitive harm. The foreclosure of business opportunity, the concentration of innovative capacity, and the erosion of economic pluralism are harms that the new anti-trust jurisprudence should address.
The AI industry has produced extraordinary capabilities, but the entities that control the stack have built the AI architecture of exclusion. The existing anti-trust jurisprudence fails not only because it lacks tools but also because it was designed for a different economy. The power that the new AI architecture generates is not the power to raise prices; it is the power to determine the conditions of participation.
The choice has to be made between an AI economy that is open and pluralistic, and the one that is closed and concentrated. The new AI architecture makes it abundantly clear what kind of future is more likely under the old jurisprudence. The new anti-trust jurisprudence needs to change the AI architecture before it hardens into permanence.
ALSO CHECKOUT The Quick Artificial Intelligence Revolution and the Poor Publisher’s Paradox
Originally published at https://virtuslegis.com on September 28, 2026.
