The Layer Nobody Controls Yet Is the One Everyone Will Need
There is a pattern that repeats with enough consistency to take seriously: technologies do not concentrate where they are visible, but where they are supported. Social networks concentrated in distribution, not in content. The cloud concentrated in infrastructure, not in applications. Artificial intelligence is following the same geometry, but the point of control is one level lower than in any previous cycle.
In May 2026, David and Daniil Liberman, entrepreneurs with prior experience at Snap, published an argument in Fortune that deserves attention for what it describes structurally, not for who they are. Their thesis is precise: in artificial intelligence, whoever controls the compute controls access, and whoever controls access controls who can even exist in this economy. This is not a metaphor. It is an operational description of how the market functions today.
The numbers they cite are what give weight to the argument. NVIDIA concentrates 85% of the GPU market for data centers. Amazon, Microsoft, and Google control 63% of global cloud capacity. The United States manages approximately 75% of the world's high-performance computing capacity for artificial intelligence. China retains around 15%. The rest of the world shares the remaining 10%.
That does not describe a competitive market. It describes a geopolitical infrastructure with a market facade.
The Moment Snap Lost Without Losing the Product
The reference to the Snap episode of 2018 is not nostalgia. It is the analytical anchor of the article. The authors were in Santa Monica reviewing daily active user metrics when it became clear that the product, despite being technically superior in some dimensions, could not sustain growth against Instagram. Meta did not win because it had better design. It won because it controlled the layer beneath the design: the social graph, the distribution, the already-formed audience. Snap built on sand that Meta had already cemented.
That episode matters as an organizational diagnosis because it illustrates when metrics stop measuring what one believes they are measuring. User retention was not Snap's problem. Access to distribution was the problem. But if the dashboard only measures retention, the leadership team can arrive too late at the correct reading.
In artificial intelligence, the equivalent is more severe. A team building a language model may have better architecture, better data, better engineers. But if it lacks access to high-performance GPUs at an accessible price, if it depends on contracts with hyperscalers that can change rates or access policies without prior notice, then its technical advantage does not translate into competitive advantage. The layer it does not control neutralizes it before it can even demonstrate its worth.
This is exactly what the Libermans describe when they point out that artificial intelligence providers have withdrawn popular models despite user resistance, restricted access to APIs without warning, and adjusted developer capabilities under policies that no independent body can audit. This is not a moral critique. It is a description of how structural dependency operates when whoever concentrates the infrastructure decides to change the conditions.
Why This Concentration Is Qualitatively Different
When Meta bought Instagram in 2012 for one billion dollars, the market understood that the social distribution layer was consolidating. When Amazon Web Services scaled to become Amazon's primary source of profits, the market understood that the cloud was going to concentrate in few hands. In both cases, the concentration was visible from the application layer. Users, developers, and regulators could see it because they felt it directly.
What the article describes about artificial intelligence is different in a specific sense: the concentration occurs in a layer that most market participants do not monitor rigorously. Models are visible. Chatbots are visible. The AI products that users consume are visible. But GPUs, data centers, high-performance chip supply contracts, and preferential compute access agreements are the infrastructure behind what is visible, and it is there that the real concentration is forming.
The analogy the authors use with Bitcoin and Ethereum is interesting from a different angle than the one they emphasize. It is not only that decentralized protocols built a new layer beneath banking. It is that they did so because the existing financial architecture had frictions and control points that could not be removed from within. The relevant question for artificial intelligence infrastructure is not whether Gonka or any other decentralized project can displace AWS or Azure. The question is whether the market's incentive structure is sufficient to produce viable alternatives before the concentration becomes irreversible.
The historical evidence in infrastructure markets suggests that this window is narrow. Once hyperscalers reach certain thresholds of installed capacity, economies of scale and switching costs cause the structure to self-perpetuate. Not because it is illegal to switch, but because the operational cost of doing so exceeds the benefit for most actors.
What This Reveals About How Long-Term Bets Are — or Are Not — Designed
There is a dimension of the argument that the article does not develop completely but that proves analytically fertile: the problem of concentration in compute is not only a matter of public policy or market power. It is also a problem of how organizations distribute their attention between what works for them today and what might threaten them tomorrow.
The companies that built deep dependencies with hyperscalers over the last five years did so under a reasonable logic: the marginal cost of scaling in the cloud was lower than the cost of building proprietary infrastructure, and the speed of market access justified that dependency. That is the logic of exploiting what works. The problem is that this same logic, applied without counterweight, produces organizations that reach the point of lock-in without having anticipated it.
The pattern the Libermans identify in the compute market is exactly the same one that appears in organizations that over-exploited their core model and arrived too late to see that the ground had shifted beneath them. Snap, in their account, did not lose because it stopped innovating in product. It lost because it had no structural response to the dependency at the distribution layer. The relevant organizational learning is that dependencies that are not managed strategically become, over time, positions of vulnerability that cannot be negotiated when the provider decides to change the terms.
This applies to artificial intelligence startups that today operate on third-party model APIs. It applies to mid-sized companies that are building their data layer on a single cloud provider's infrastructure. It applies to countries that have no compute policy of their own and assume that the availability of American infrastructure is a permanent feature of the environment.
The advanced chip export controls mentioned in the article are not a hypothetical example of how compute can be used as a geopolitical instrument. They are evidence that it is already being used that way. When a power can decide which countries access a certain level of computational capacity, and that decision directly affects what artificial intelligence applications can be built in those territories, the conversation ceased being about market competition long ago. Two countries are setting the conditions of access for 191. That is the current design of the system.
The linguistic asymmetry the authors point out adds an additional layer that ordinarily does not appear in market concentration analyses. Language models trained predominantly in English do not only favor English-speaking users in terms of output quality. They make them financially more efficient: prompts in other languages consume more tokens to produce the same output, which translates into higher costs and more restrictive context limits for users who do not operate in English. A uniform tariff is not an equal price. It is a rate that discriminates by language with technical architecture as the mechanism.
Compute as a Prerequisite, Not a Competitive Advantage
There is a distinction the article establishes with precision and that deserves to be underscored because it changes the nature of strategic analysis. In social networks, you could build an alternative platform. TikTok proved it was possible. Social capital was not physically concentrated; it was distributed in the attention habits of users, and those habits could be redirected.
In artificial intelligence, compute is not a competitive advantage. It is the floor of participation. Without access to high-performance GPUs, a competitive model cannot be trained. Without cloud contracts, inference cannot be operated at scale. Without advanced chips, an entire country is excluded from certain capabilities. Concentration at this layer does not generate competitive disadvantage: it generates direct exclusion.
That makes the organizational implications more urgent than in previous cycles. A company that depended on Facebook for distribution could, with effort and resources, attempt to build an audience through other channels. A company that depends on compute infrastructure concentrated in three actors does not have, today, a structurally equivalent alternative to turn to if those actors change the conditions.
The promise of projects like Gonka, the decentralized protocol that the authors themselves are building, is to create that alternative before the window closes. They do not need to be better than AWS on AWS's own terms. They need to be sufficiently functional so that the dependency ceases to be total. That is a more modest and more realistic threshold than winning market share from the hyperscalers.
What the market has not yet resolved is whether that threshold can be reached with sufficient speed to have an effect before the concentration solidifies at a point from which it no longer generates pressure for change. Previous cycles suggest that infrastructure that arrives late rarely changes the structure of the market. Infrastructure that arrives before the moment of closure can set the rules of the next game.
---
The Libermans' article describes with precision a structural dynamic that is already underway. But the problem they identify is not only one of market or regulation: it is one of dependency design that most organizations are building today without managing it as a strategic risk. When compute becomes the prerequisite for existence and that prerequisite is in the hands of three actors who can change their terms unilaterally, companies that have no explicit policy toward that dependency are not delegating a technical decision. They are ceding a position that cannot afterwards be recovered with speed.
