The Case for Big A.I. Is Beginning to Fracture

The A.I. boom was built on the assumption that a few laboratories would own machine intelligence and capture extraordinary profits. The market is starting to signal a different future, one in which intelligence becomes abundant, interchangeable and controlled closer to the organizations that use it.

Category: Strategy. Written by Jaime Garcia, Founder, SnowRock. Published July 9, 2026. 18 min read.

In short

• The monopoly thesis is weakening. Capability leads have proved temporary, and low-cost and open-weight models stay close enough to satisfy most buyers.
• Value is migrating from owning the most powerful model to distribution, customization and implementation. The model is becoming infrastructure while the businesses that deploy it capture the productivity.
• For operators the question shifts from which model is best to which system gives the most control, usefulness and return, favoring sovereignty, multi-model deployment and internal integration.

The artificial-intelligence boom was built on a single assumption: that a small number of frontier laboratories would achieve durable technological dominance and capture extraordinary profits. The market is beginning to signal a different future, one in which intelligence becomes abundant, interchangeable and controlled closer to the organizations that use it.

Is this an AI bubble? The thesis the boom was built on

For much of the boom the industry rested on a remarkably concentrated thesis. A small number of laboratories would build increasingly capable models. One might eventually establish a decisive lead, and that advantage would compound as the system helped design stronger successors, attracting more customers, capital, computing power and talent.

The winner would not merely become another successful software company. It could become the dominant supplier of machine intelligence to the global economy, and that possibility justified extraordinary investment. Technology companies committed hundreds of billions of dollars to chips, data centers, energy infrastructure and model development. Investors accepted immense operating losses and ambitious valuations because the eventual prize appeared almost unlimited. If intelligence became a scarce proprietary service, the laboratory controlling the strongest system could charge premium prices, establish deep customer dependence and capture an unusually large share of the value that artificial intelligence created. The model may become infrastructure. The greater value may be created by the businesses that decide how to use it.

That assumption now looks less secure. Frontier models continue to improve, but technological leadership has proved temporary. Lower-cost systems remain close enough in capability to satisfy many customers, and open-weight models can be adapted, operated privately and integrated into an organization without permanent dependence on a single provider. The emerging market may reward distribution, customization and implementation more than ownership of the most powerful general-purpose model.

The sovereignty argument

Palantir has become one of the most forceful advocates of an alternative structure. Its leadership argues that companies and governments should not surrender their data, workflows and operational intelligence to a small group of frontier-model providers, and should instead retain control over the systems they deploy, the information those systems use and the environments in which they operate. This position is often described as artificial-intelligence sovereignty.

The principle extends beyond national autonomy. For a corporation, sovereignty means retaining control over its models, data, permissions, workflows and intellectual property. For a government, it can mean ensuring that sensitive operations do not depend on a foreign or privately controlled provider. For a military or intelligence service, it can mean operating systems inside secure environments without exposing information to an outside platform.

Open-weight models support that objective. Unlike closed systems reached mainly through a commercial interface, they can often be downloaded, modified and operated inside an organization own infrastructure. The organization can fine-tune the system on proprietary information, set its own restrictions and decide where data is processed. That control is attractive even when the underlying model is slightly less capable. A company may prefer a system that performs 95 percent as well but can be customized extensively, operated privately and bought at a fraction of the price.

The strategic question is therefore changing. It is no longer simply which model is best. It is which system gives the organization the greatest control, usefulness and economic return.

A critique that is commercial as well as philosophical

Palantir is not a neutral observer. It sells software designed to integrate data, models and operational systems inside businesses and governments, and a future in which customers build customized environments rather than relying entirely on one frontier laboratory aligns directly with its commercial interests. Its partnership with Nvidia strengthens that position, letting the two offer computing infrastructure, model access and operational software that keep the frontier laboratory out of the center of the relationship.

That does not make the criticism wrong. Companies routinely identify genuine market weaknesses while positioning their own products as the remedy. The relevant question is whether the underlying argument is supported by the market, and increasingly it is. Customers have grown more sensitive to model cost, data control and dependence on external providers. Governments are examining national-security and privacy exposure. Open alternatives are improving quickly. Many enterprises have found that the most capable available model is not always the most economically useful one. The market is separating intelligence from implementation. Frontier labs produce the capability. Other companies decide whether it becomes valuable inside the organization.

The original monopoly thesis

The strongest economic case for enormous frontier investment was built around technological takeoff. Under this theory a sufficiently capable system would help researchers design a stronger successor, that successor would accelerate the next, and improvements would compound until one laboratory achieved a lead competitors could no longer close. Even a small initial advantage could become decisive, and the result would resemble a natural monopoly on advanced intelligence.

The leading system might outperform rivals by such a margin that customers would have little practical alternative, and the provider could charge exceptionally high prices because access would produce extraordinary value. That possibility helped justify unprecedented capital expenditure. If the race could produce a dominant machine intelligence, spending tens or hundreds of billions of dollars to win it might look rational. The return would not resemble an ordinary technology product. It could resemble ownership of a foundational economic resource.

Much of the current investment cycle still depends implicitly on this belief, even as industry leaders speak less often about imminent superintelligence or the elimination of most human work. The language has softened. The capital commitments have not.

No durable leader has emerged

The actual competitive pattern has been more fluid. One laboratory releases the strongest model, a rival closes the gap, a new system leads in coding, mathematics or reasoning, and within months another provider matches or exceeds it. No company has held a decisive lead for long.

Open-weight models have also stayed closer to the frontier than many expected. They may lag the best proprietary systems on the most demanding evaluations, but the difference is often too small to decide a purchase. Most organizations do not need the strongest model in the world. They need a system that is reliable enough for a specific workflow, affordable at high volume and compatible with their data and security requirements. A model that performs modestly better can be economically inferior when it costs several times as much.

This is especially true for repetitive enterprise tasks. Customer support, document classification, internal search, routine analysis and content generation may not require frontier-level reasoning. A lower-cost system can complete the work adequately while offering greater control and lower operating expense. The frontier laboratories may keep producing the strongest models. That does not guarantee they capture the greatest profits.

Intelligence is showing signs of commoditization

Commoditization occurs when products become similar enough that customers weigh price, availability and integration more heavily than brand or technical distinction. Artificial intelligence has not become a full commodity, but it is moving in that direction. Performance keeps improving across the industry, techniques spread quickly, researchers change employers, papers circulate globally, and customers can switch providers or split workloads across several models. Each factor limits durable pricing power.

Capable, inexpensive systems from China accelerated the shift. DeepSeek showed that a lower-cost provider could approach frontier performance without matching the spending of the largest American laboratories. The broader lesson was not that every model would become identical, but that the performance gap might stay too narrow to support monopoly economics. A slightly inferior model sold at a dramatically lower price can capture substantial share, especially when the customer can operate it independently, customize it and avoid sending sensitive data outside.

The market may come to resemble cloud computing, databases or electricity. Providers can build large and profitable businesses, the underlying service becomes increasingly standardized, and much of the value shifts to the applications, processes and institutions built on top of it.

The price-resistance signal

The market has already shown resistance to premium pricing. As frontier laboratories tried to charge more for their most advanced models, some enterprise customers questioned whether the improvement justified the expense. The calculation is simple. A model must generate new revenue, eliminate substantial labor or reduce risk by enough to offset its operating cost. A stronger benchmark score has little value if the company cannot convert it into a measurable business result.

Organizations experimenting with AI often find that implementation costs exceed the price of the model itself. Data must be cleaned and connected, permissions established, workflows changed, systems monitored, outputs reviewed, and legal and compliance requirements met. In many cases the model is the easiest component and the bottleneck lies inside the organization. That reduces the advantage of paying heavily for the strongest general-purpose system, because a less expensive model integrated well may create more value than a superior model deployed poorly. It is one reason corporate demand for premium frontier access can stall even as overall adoption rises. The market is growing. The spending is moving toward cheaper models and application layers.

Distillation and the erosion of scarcity

Frontier laboratories increasingly describe model distillation as a serious threat. Distillation lets one system learn from the outputs of another. A smaller or weaker model is trained on examples generated by a stronger system, absorbing part of its behavior and performance. The frontier provider views this as appropriation. Its competitor views it as a method of development. The dispute reveals how hard it may be to preserve scarcity in artificial intelligence.

A model behavior can be observed at scale. Users can generate millions of outputs, study response patterns and use the result to improve other systems. Even when a competitor cannot reproduce the strongest model exactly, it may capture enough useful behavior to narrow the gap. The largest laboratories are therefore seeking legal and political protection alongside technical leadership, arguing that weaker providers, particularly foreign ones, are extracting proprietary value unfairly. The concern is understandable, and it also exposes the weakness of the original thesis. If the frontier advantage can be copied, approximated or compressed quickly, it may not support long-term monopoly pricing, and the leading laboratory must keep spending heavily simply to stay temporarily ahead. That is a far less attractive business than an asset whose advantage compounds permanently.

The value may reside in diffusion

The ultimate economic effect of artificial intelligence will depend on diffusion, the process by which a technology spreads across companies, institutions, industries and ordinary work. A powerful model sitting inside a data center creates limited value on its own. The value appears when the system is integrated into medical practices, factories, banks, schools, logistics networks, government agencies and software products.

That process is slow and difficult. Organizations contain legacy systems, regulatory constraints, internal politics, fragmented data and employees who may resist changing established routines. Artificial intelligence must be adapted to those conditions. It must understand proprietary information, respect permissions and produce work in a form people can trust. Those requirements favor companies close to the customer. Consultancies, software providers, systems integrators and internal technology teams may capture much of the value by translating general intelligence into specific operational improvement. The laboratory supplies capability. The implementing organization creates productivity. If diffusion determines impact, the largest model developers may be less central than investors assumed, essential infrastructure providers without becoming dominant profit engines.

The utility scenario

A utility provides an indispensable service under constrained economics. Electricity is essential, yet the producer of electricity does not capture all the value created by the factories, offices and homes that use it. Artificial intelligence may develop similarly. Model providers could sell access to machine reasoning as a metered service, and customers would buy computing capacity and intelligence the way they buy cloud storage, bandwidth or electricity.

The service could generate enormous revenue while competition and standardization limit margins. Customers would choose among providers on price, reliability, latency, security and compatibility, workloads could move between systems, and open models could be run internally when privacy or cost justified it. The largest returns would then accrue to companies that use AI to create superior products, cut labor expense or redesign industries. This does not make frontier laboratories unimportant, because utilities are foundational. It means their market value should reflect infrastructure economics rather than assumptions of permanent monopoly control. The distinction is enormous. A company valued as the future owner of machine intelligence may be radically overpriced if it becomes one supplier among many.

Open models change the distribution of power

Open-weight artificial intelligence changes who controls the technology. A closed model centralizes authority with its provider, which sets pricing, access, restrictions, updates and the terms under which a user can keep operating the system. An open-weight model distributes more of that authority. Organizations can host it themselves, modify it and preserve a version even if the original developer changes direction.

That flexibility matters most for governments, defense organizations, hospitals and regulated industries, which may not want critical operations dependent on a commercial service whose rules can change without notice, and which may need to inspect, test or modify the system more deeply than a closed provider allows. Open models can therefore become strategically preferable even when they are not technically superior, because control is itself a feature. The market may gradually divide between organizations that prioritize convenience and those that prioritize sovereignty, with consumer applications and small businesses favoring managed frontier services and large enterprises and governments increasingly demanding systems they can govern directly.

The national-security dimension

Sovereignty also has geopolitical implications. Governments do not want critical administrative, military or intelligence functions dependent on a foreign model provider, since reliance on an external laboratory can expose sensitive information, invite service restrictions or cut off access during a political dispute. National governments may therefore back domestic models, computing infrastructure and data ecosystems even when superior foreign alternatives exist.

This can fragment the global market. The United States may favor American providers, China will keep developing independent systems, European governments may seek regional alternatives consistent with local regulation, and Middle Eastern states are investing in their own computing and model capabilities. Artificial intelligence may not converge toward one dominant global system. It may divide into national and institutional stacks. That fragmentation weakens the monopoly thesis further, because a frontier laboratory can lead technologically while being excluded from major portions of the global market.

The risks of centralized dependence

Reliance on a small group of frontier companies creates several systemic risks. The first is financial. The largest laboratories and their partners are committing extraordinary capital on uncertain returns, and if anticipated revenue fails to arrive, the contraction could reach technology stocks, credit markets, utilities, energy infrastructure and semiconductor suppliers. The second is operational. Thousands of organizations could depend on a few providers, so an outage, cyberattack, policy change or price increase could disrupt many industries at once.

The third is political. A privately controlled system may become embedded in government, defense or public services without clear democratic accountability. The fourth is informational. Organizations may transfer proprietary data and decision processes to external providers, widening the provider visibility into sensitive operations. The fifth is competitive. A few laboratories controlling access to advanced intelligence could decide which companies, countries and individuals receive the strongest capabilities. These risks strengthen the case for a more distributed architecture, and they also create a hard policy problem, because decentralization can reduce dependence while making dangerous capabilities harder to contain.

The centralization argument still has strength

The case for large centralized laboratories has not disappeared. Frontier development requires enormous computing resources, advanced chips, specialized talent and sophisticated research infrastructure, and smaller organizations may be unable to fund or coordinate work at that scale. Centralized providers can also invest more heavily in safety testing, security and reliability, while a widely distributed open model can be harder to update or control when vulnerabilities appear, and bad actors can modify it, remove safeguards and operate it anonymously. Some advanced capabilities may stay concentrated for technical and security reasons.

The largest laboratories also hold substantial distribution advantages. OpenAI has a globally recognized consumer product. Google, Microsoft and Meta can integrate models into platforms used by billions. These companies may keep powerful market positions even when their models are not uniquely superior. The question is not whether Big A.I. disappears. It is whether frontier leadership produces the extraordinary economic dominance implied by current investment and valuations.

The application layer may win

The greatest value in technology does not always accrue to the company producing the foundational component. Semiconductor makers enabled personal computing while software companies captured much of the commercial value. Telecommunications networks enabled the mobile internet while application and platform companies became more valuable than many network operators. Cloud infrastructure enabled thousands of software businesses, and the providers grew enormous but did not capture the full value created on top of their systems.

Artificial intelligence may follow the same pattern. Model laboratories provide capability. Application companies turn it into a product customers understand and will pay for. The strongest businesses may be those controlling distribution, proprietary data, specialized workflows or trusted customer relationships. A healthcare company using AI to improve clinical operations may create more defensible value than the laboratory supplying the model. A defense platform integrating several models into mission systems may matter more to the customer than any one provider. A financial institution running a customized system on its own data may keep the productivity gain internally. The model becomes replaceable. The workflow does not.

Integration is harder than intelligence

Frontier laboratories have focused mainly on raising model capability. Enterprise adoption depends on a different set of problems. The system must connect to existing databases, understand internal terminology, follow permissions, operate inside legal and regulatory boundaries, produce auditable results, know when to involve a human and perform reliably despite incomplete or inconsistent data. These are not benchmark problems. They are implementation problems.

A model can achieve extraordinary test results and still fail inside a company because the company itself is fragmented. This is why diffusion may stay human-directed far longer than the most aggressive automation narratives suggest. Artificial intelligence can accelerate work. It cannot instantly resolve organizational dysfunction. Human bottlenecks, incentives and institutional constraints remain decisive, and the value may accrue to whoever understands the organization well enough to redesign it.

Big A.I. and its customer conflict

Frontier providers face an added tension. To improve their products they benefit from access to customer interactions, feedback and operational data, yet customers increasingly want to protect that information. A company may fear that its proprietary workflows, research or intellectual property could help improve a general model that is ultimately sold to competitors, and even when providers promise contractually not to train on customer data, organizations may stay cautious about sending sensitive information outside their own environments.

This concern grows more serious in defense, intelligence, healthcare and finance. The more valuable the application, the more sensitive the data. The more sensitive the data, the stronger the case for internal or sovereign deployment. That creates a ceiling on the closed-platform model. The frontier laboratory can offer convenience and high capability, but it may struggle to become the sole intelligence layer for institutions unwilling to surrender control.

The capital-expenditure assumption

The boom is still being financed through historic investment. Data centers, chips, power systems and transmission infrastructure are being built on the assumption that demand for model computation will rise dramatically. That forecast can hold even if frontier laboratories never achieve monopoly economics, because artificial intelligence may become ubiquitous while the price of intelligence falls.

The distinction matters. A technology can transform the economy without producing exceptional returns for every company building it. Railroads changed commerce while many railroad investors lost money. The internet created enormous value while large numbers of internet companies failed. Telecommunications infrastructure enabled the digital economy while repeated investment cycles destroyed capital. Artificial intelligence could follow the same path, the social impact immense, the investment returns concentrated elsewhere or arriving far below current expectations.

What would prove the big-lab thesis correct

The centralized frontier model would regain strength if one laboratory achieved a sustained, commercially meaningful capability gap, one that persisted long enough that customers could not substitute a cheaper system without significant loss. The leading model might autonomously conduct scientific research, operate companies, write production-grade software or perform professional work at a level rivals could not approach. The provider would also need to prevent that capability from being replicated through open research, employee movement, distillation or alternative architectures, and to hold secure control over computing resources, proprietary data and distribution. Most of all, the system would need to generate customer value far exceeding its cost. If all of those conditions emerged, monopoly-like economics could become plausible. The evidence so far points toward continued competition.

What would prove the decentralized thesis correct

The alternative would be validated if several conditions continue. Open-weight systems stay close to the frontier, model prices keep falling, enterprises deploy multiple systems, customization and proprietary data outweigh raw benchmark leadership, governments demand sovereign infrastructure, application companies capture most of the customer relationship, and model providers struggle to hold differentiated margins. Under that structure artificial intelligence becomes an abundant input rather than a scarce product. The frontier laboratories remain important, but no single one controls the market. Intelligence becomes modular, and organizations combine several models, switching among them by cost, capability and risk. This increasingly resembles the direction of the market.

A more competitive future

The AI future may be less centralized than both its supporters and critics once assumed. The dominant early narrative imagined convergence, one system growing powerful enough to absorb or supersede the others, with economic and perhaps political authority concentrating around whoever controlled it. The emerging reality looks more unsettled. Models compete, open alternatives spread, governments seek control, companies build proprietary systems, customers resist premium prices, and technological advantages narrow quickly.

This structure can be called decentralized, democratic, fragmented or simply competitive, and each term captures part of the picture. A competitive market would reduce the risk of one laboratory becoming an unaccountable gatekeeper to machine intelligence. It could also make advanced capabilities harder to govern, because the same openness that empowers companies and countries can empower malicious actors. Decentralization changes the risk. It does not eliminate it.

The wrong bet may not be on A.I.

The mistake may not have been betting that artificial intelligence would be economically transformative, which remains highly plausible. The mistake may have been assuming that transformation would naturally produce a single dominant owner of intelligence. Markets often create value differently from how early investors expect. The technology becomes cheaper, capability spreads, and competition shifts toward distribution, integration and customer control, until the foundational provider becomes one participant in a much larger ecosystem.

Artificial intelligence may still reshape work, science, defense and economic production. The principal beneficiaries may not be the companies building the largest general-purpose models. They may be the organizations that understand how to combine intelligence with proprietary data, operational authority and real-world implementation. The frontier laboratories have asked investors to believe the race can be won. The market is beginning to suggest the race may never end, and if that is true, the future of artificial intelligence will not belong to a single machine or laboratory. It will belong to the institutions best able to direct many forms of intelligence toward specific human purposes.

The SnowRock take

Strip out the market-structure debate and a practical instruction remains for anyone running a mid-market company. You do not have to predict which laboratory wins. You have to make sure your business is not fragile to the answer.

In practice that means three habits. Treat models as interchangeable parts rather than partners, so you can move a workload from one system to another without re-architecting your business. Keep your proprietary data, your permissions and your workflows on your side of the line, because that is the asset that compounds, not the model behind it. And spend where the value actually sits, on integration, on the judgment to pick the few tasks worth automating, and on the controls that let you deploy without fear, rather than on the highest benchmark score. Rent intelligence. Own the workflow. However the market settles, the company that owns its own operations keeps its leverage.

That is our whole position in one line. The frontier will keep moving, prices will keep falling, and the leaders will keep trading places. None of it should decide whether your business works. Build so that it does not.