Part I: The Economics of Intelligence
Why the AI industry is about to optimize for intelligence per dollar rather than intelligence itself
I have become increasingly convinced that the artificial intelligence industry is approaching one of the most important economic inflection points since the launch of ChatGPT, yet the overwhelming majority of investors remain focused on a variable that, while certainly important today, may ultimately prove far less valuable than the market currently assumes. The dominant conversation continues to revolve around which company possesses the smartest model, which frontier laboratory tops the latest benchmark, which reasoning model scores highest on increasingly obscure academic evaluations, and which AI company has temporarily claimed leadership on the industry’s ever-changing leaderboard. While those discussions undoubtedly attract headlines, I believe they risk missing the much larger economic transition quietly unfolding beneath the surface, because the variable that ultimately determines where profits accrue throughout the AI ecosystem is unlikely to be intelligence itself, but rather the amount of intelligence delivered for every dollar spent.
Perhaps the simplest way to understand this transition is through an analogy that repeatedly comes to mind whenever I think about enterprise AI adoption. When a company needs someone to reconcile its books, prepare financial statements, or process invoices, it does not recruit a PhD in pure mathematics, not because that individual lacks the ability to perform the work, but precisely because they possess far more capability than the task economically justifies. The value created by accurate bookkeeping is fundamentally capped. Once the accounts are correct, there is little incremental benefit from employing substantially more intelligence, regardless of how exceptional that intelligence may be. Rational organizations therefore optimize around economics rather than capability, hiring the least expensive person capable of consistently producing work that meets the required quality standard while deploying scarce intellectual talent toward problems where additional intelligence genuinely creates incremental value.
The economics change entirely, however, when the problem itself changes. If the objective is discovering a breakthrough Alzheimer’s treatment, developing a revolutionary semiconductor architecture, or solving one of the most difficult scientific questions facing humanity, then suddenly the cost of hiring the world’s brightest researchers becomes almost irrelevant relative to the potential economic value generated by success. A single breakthrough may create tens or even hundreds of billions of dollars in value, making the salaries of a handful of elite scientists almost immaterial within the overall economics of the project. In these situations, intelligence itself becomes the scarce resource, and maximizing capability rather than minimizing cost becomes the rational economic decision.
I believe artificial intelligence is now approaching exactly this bifurcation. During the past two years, enterprises have overwhelmingly defaulted to using frontier models for virtually every conceivable task, whether summarizing emails, extracting information from invoices, classifying customer support tickets, translating documents, drafting meeting notes, generating routine software code, or searching internal knowledge bases, largely because the industry had only recently crossed the threshold where large language models became broadly useful for knowledge work. When a technology first becomes commercially viable, organizations naturally reach for the highest-performing solution available because they are still trying to answer a much more fundamental question, namely whether the technology works at all, and cost optimization remains secondary while businesses validate capability.
That behavior, however, is unlikely to represent a stable equilibrium because pilot programs inevitably become production systems, and production economics eventually dominate technology economics. We are already beginning to observe that transition occurring across enterprise AI deployments as organizations discover that the challenge is no longer determining whether AI creates value, but rather determining whether it creates sufficient value relative to the rapidly expanding costs associated with deploying it at scale. Once finance departments begin scrutinizing AI budgets with the same discipline applied to every other enterprise technology investment, the optimization process inevitably shifts away from maximizing intelligence and toward maximizing intelligence per dollar spent.
The remarkable speed at which inference costs continue collapsing only accelerates that transition. Stanford’s AI Index estimates that the cost of delivering GPT-3.5-level performance declined by more than 280-fold between late 2022 and late 2024, while Andreessen Horowitz concludes that maintaining a fixed level of model capability becomes approximately ten times cheaper every year, and Epoch AI arrives at similar conclusions across multiple reasoning benchmarks, suggesting that capabilities commanding premium prices today rapidly become commodities tomorrow. Even OpenAI’s Chief Financial Officer, Sarah Friar, recently observed that inference costs between GPT-4 and the company’s latest generation of models had declined by approximately 97% within only two years, illustrating just how extraordinary the industry’s deflationary dynamics have become. Although each organization measures these trends somewhat differently, they all converge upon the same underlying conclusion, namely that intelligence is becoming dramatically less expensive at a pace rarely witnessed anywhere else in modern technology.
At precisely the same time that inference costs continue collapsing, enterprises are beginning to encounter an entirely different challenge, one that may ultimately prove even more important. Companies are increasingly discovering that the AI budgets they originally expected to last an entire fiscal year are being exhausted within only a few months as usage expands far more rapidly than initially anticipated. Sam Altman recently remarked that enterprise customers increasingly tell OpenAI they have effectively consumed their planned annual AI spending within the first quarter and are now asking not for smarter models, but for more efficient ones. That observation should not be dismissed as merely another anecdote because it signals that AI has entered a fundamentally different stage of commercialization. Organizations have already concluded that artificial intelligence works. Their attention is now shifting toward ensuring that it works economically.
The technological evolution occurring beneath the surface reinforces exactly the same conclusion. During the earliest years of the large language model revolution, industry participants largely assumed that better models simply required more parameters, larger architectures, and exponentially greater amounts of compute. Increasingly, however, frontier laboratories are discovering that carefully trained smaller models, enhanced through better datasets, improved reasoning techniques, synthetic training data, and sophisticated distillation methods, can approach the performance of substantially larger systems while requiring only a fraction of the inference cost. Meta has already demonstrated this philosophy internally by using its largest frontier models primarily as teachers while deploying significantly smaller distilled models across its advertising and recommendation infrastructure, thereby reserving maximum intelligence for learning while optimizing production around economics rather than benchmark scores.
The consequence is that artificial intelligence increasingly begins resembling human labor markets rather than scientific competitions. No rational organization staffs every position with Nobel Prize winners, just as no enterprise will ultimately route every inference request toward the world’s most expensive frontier model. Tasks involving frontier scientific research, advanced mathematics, complex engineering, autonomous reasoning, or pharmaceutical discovery will almost certainly continue relying upon the most capable AI systems because the economic upside remains effectively uncapped. The overwhelming majority of enterprise workloads, however, involve document classification, customer support, workflow automation, information extraction, software maintenance, compliance monitoring, enterprise search, contract review, and countless other tasks where reliability, consistency, governance, and economics matter substantially more than squeezing another one or two percentage points out of a benchmark leaderboard.
This is why I believe the AI industry may be approaching its own Linux moment. Open-weight models such as Llama, DeepSeek, Qwen, GLM, Kimi, and others do not need to surpass every proprietary frontier model across every benchmark to fundamentally reshape industry economics. They simply need to become sufficiently capable for the overwhelming majority of enterprise workloads, because once that threshold is crossed, purchasing decisions become increasingly driven by return on investment rather than raw capability. Enterprise CIOs have never selected critical infrastructure solely because it ranked first on a benchmark. They optimize around security, governance, reliability, compliance, integration, vendor support, operational simplicity, and total cost of ownership. Artificial intelligence is unlikely to prove any different.
If that view is correct, then the defining metric of the AI industry gradually shifts away from intelligence itself and toward intelligence per dollar, while the frontier models become increasingly concentrated around problems whose economic value genuinely justifies paying for the very highest levels of capability. That distinction forms the foundation of the investment thesis that follows, because once intelligence itself becomes increasingly abundant, investors should begin asking a different question. Rather than debating who builds the smartest model, we should increasingly ask who captures the economic value when intelligence becomes inexpensive enough to be embedded into virtually every workflow throughout the global economy. That, in our view, is where the next phase of the AI investment story truly begins.
Part II: The Great Value Migration
Why the owners of installed compute may ultimately capture more value than the sellers of new compute
If the central argument of the first part is correct, namely that artificial intelligence is steadily evolving toward maximizing intelligence per dollar rather than intelligence itself, then the natural question for investors becomes where the economic value ultimately migrates as that transition unfolds. The market’s answer today appears remarkably straightforward. Buy the companies supplying the picks and shovels. Buy Nvidia, Broadcom, the ASIC designers, the networking vendors, the memory manufacturers, and anyone else selling the hardware required to build the next generation of AI infrastructure. That strategy has unquestionably been the correct one over the past several years, as hyperscalers embarked on one of the largest capital expenditure cycles in the history of technology, deploying hundreds of billions of dollars into GPUs, networking equipment, power infrastructure, cooling systems, and entirely new AI campuses designed to support what everyone expects will become an explosion in AI inference demand.
Yet I increasingly believe the market is asking the wrong question. Investors remain almost entirely focused on who sells the next GPU, when the more important question may ultimately become who owns the last one. That distinction may appear subtle today, but it fundamentally changes the economics of the entire AI ecosystem because it shifts attention away from one-time hardware sales toward the recurring cash flows generated by infrastructure that has already been deployed, much of which will continue processing AI workloads for years after the initial capital expenditure has been incurred.
From our perspective, there are only two broad scenarios under which the industry can evolve over the coming years, and what makes this investment debate particularly interesting is that both outcomes appear considerably more favorable for the hyperscalers than current market pricing implies.
The first possibility is that AI models continue becoming dramatically more efficient through a combination of better architectures, distillation, quantization, speculative decoding, routing algorithms, compiler optimization, and increasingly sophisticated inference techniques, allowing cloud providers to extract substantially more useful work from the hardware they already own. Instead of requiring another US$100 billion of annual capital expenditure simply to remain competitive, existing GPU clusters gradually become more productive with each successive model generation, enabling hyperscalers to support exponentially larger inference workloads without matching that growth dollar for dollar through new hardware purchases. In this world, the AI infrastructure already sitting inside Microsoft’s Azure, Amazon’s AWS, and Google’s Cloud becomes significantly more valuable than investors currently appreciate because every software breakthrough effectively extends the economic productivity of previously deployed hardware.
Should that scenario materialize, the implications for cloud economics become extraordinarily attractive. Capital expenditure naturally begins to stabilize, depreciation gradually declines as earlier investments mature, while revenue continues compounding because enterprise token consumption keeps expanding. The result is that free cash flow inflects sharply upward as hyperscalers transition from businesses that absorb enormous amounts of capital into businesses increasingly monetizing infrastructure already sitting on their balance sheets. What the market currently views as one of the largest expenses in technology may eventually prove to be one of the largest productive asset bases ever assembled, generating attractive returns long after investors have stopped worrying about the initial investment required to build it.
The alternative scenario is, in many respects, even more constructive. This is the Jevons Paradox scenario, where improvements in efficiency do not reduce demand but instead accelerate it because lower costs make entirely new applications economically viable. As inference becomes dramatically cheaper, companies stop rationing AI usage and begin embedding intelligence into virtually every workflow throughout their organizations. Agents execute continuously rather than occasionally, software increasingly calls models automatically instead of waiting for humans to initiate requests, coding assistants repeatedly evaluate their own work before producing a final answer, customer support systems consult multiple models simultaneously, and enterprise software begins treating inference as an always-on utility rather than an expensive premium feature. Every individual token becomes cheaper, yet the total number of tokens processed expands exponentially because businesses suddenly discover thousands of new use cases that were previously uneconomical.
History suggests this is exactly how technological progress usually unfolds. When storage became dramatically cheaper, humanity did not store the same amount of information for less money. We stored vastly more information. When bandwidth became dramatically cheaper, we did not simply reduce internet bills. We transformed text-based websites into streaming video platforms. When cloud computing reduced the cost of deploying software infrastructure, businesses did not purchase fewer servers. They built entirely new categories of software that would never have existed under the economics of on-premise computing. Artificial intelligence appears likely to follow exactly the same trajectory, where falling inference costs accelerate demand sufficiently to overwhelm any reduction in revenue generated per individual token.
The remarkable feature of these two scenarios is that both appear highly constructive for the owners of cloud infrastructure. If model efficiency improves faster than demand, hyperscaler capital expenditure slows while free cash flow accelerates. If demand grows faster than efficiency, hyperscalers continue expanding infrastructure while simultaneously generating substantially larger revenues from AI services running across increasingly productive hardware. In neither scenario do we arrive at an outcome that appears structurally negative for the cloud platforms themselves. Instead, the debate becomes one of relative beneficiaries rather than absolute winners and losers.
This is why I believe the market continues misunderstanding what hyperscaler capital expenditure actually represents. Many investors continue treating AI infrastructure spending as though it were simply another operating expense suppressing near-term profitability, when in reality it increasingly resembles productive capital formation. Throughout economic history, transformative infrastructure investments have almost always appeared financially unattractive during the construction phase because they consumed enormous quantities of capital before generating meaningful cash flows. Railroads, electricity grids, telecommunications networks, fiber-optic cables, and cloud computing all followed precisely the same pattern. The upfront investment appeared excessive until utilization reached a level where operating leverage became overwhelming, at which point those same assets began generating extraordinary returns on invested capital.
The debate therefore should not revolve around whether hyperscalers are spending too much on AI infrastructure. The more important question is whether those assets will ultimately produce sufficient economic output to justify the investment, and increasingly the evidence suggests the answer is yes. The market, however, appears to be pricing a rather peculiar middle ground where semiconductor companies continue benefiting from assumptions that capital expenditure remains elevated indefinitely, while hyperscalers simultaneously trade as though that same spending permanently suppresses returns on capital. I find that combination increasingly difficult to reconcile because either AI infrastructure becomes dramatically more productive over time, allowing free cash flow to inflect sharply higher, or AI demand grows rapidly enough to justify continued investment. Neither outcome appears fundamentally bearish for the cloud providers.
Perhaps the most important development underpinning this thesis is that inference itself increasingly resembles a utility rather than a premium technology product. Electricity provides a useful analogy because consumers rarely know, or particularly care, which power station generated the electricity arriving at their homes. They simply expect power to arrive reliably, securely, and at the lowest possible cost. Artificial intelligence appears to be evolving toward exactly the same equilibrium. Very few enterprises ultimately care whether routine document classification is performed by GPT-7, Claude 8, DeepSeek, Llama, Qwen, or another open-weight model. They care that the answer satisfies the required quality threshold, integrates seamlessly into existing workflows, complies with security and regulatory requirements, and does so at the lowest possible total cost. Once inference begins resembling a utility rather than a luxury service, the industry’s economics naturally shift away from rewarding intelligence itself and toward rewarding the infrastructure responsible for delivering that intelligence at scale.
Another analogy may be even more appropriate. Frontier AI laboratories increasingly resemble airlines, while hyperscalers increasingly resemble airports. Airlines compete relentlessly on service quality, customer experience, route networks, fleet modernization, and operational efficiency, yet they also face continuous pressure to improve because yesterday’s premium offering rapidly becomes today’s industry standard. Airports operate under an entirely different economic model because they benefit regardless of which airline ultimately wins market share. Every aircraft still lands, every passenger still walks through the terminal, every airline still pays landing fees, and every additional flight simply increases the utilization of infrastructure that already exists.
The same economic logic may increasingly apply to artificial intelligence. OpenAI, Anthropic, Google DeepMind, xAI, Meta, DeepSeek, and future frontier laboratories will undoubtedly continue competing aggressively to build the world’s smartest models, with benchmark leadership changing hands multiple times over the coming decade. Yet every enterprise inference still runs inside somebody’s data center, consumes somebody’s GPUs, utilizes somebody’s networking infrastructure, and ultimately depends upon somebody’s cloud platform. The airlines compete vigorously for passengers. The airports quietly collect rent regardless of which airline wins.
This distinction becomes even more powerful as model competition intensifies. Open-weight models continue improving, proprietary models become increasingly interchangeable across routine enterprise workloads, and pricing pressure gradually emerges at the model layer as customers optimize around intelligence per dollar rather than absolute capability. Yet every token still consumes compute, every inference still traverses cloud infrastructure, and every enterprise workload still depends upon secure, scalable, globally distributed computing resources. Per-token economics may compress at the model layer, but infrastructure margins remain remarkably resilient because the physical act of serving inference continues regardless of which model ultimately performs the computation.
In our view, this represents one of the largest migrations of economic value currently taking place within artificial intelligence. The value itself does not disappear. It simply changes ownership. Rather than concentrating primarily within the companies building frontier models, an increasing proportion of that value migrates toward the platforms orchestrating trillions of AI requests every day. As intelligence becomes increasingly abundant and model capabilities continue converging, the long-term winners may not necessarily be those producing the smartest models, but rather those owning the infrastructure through which the world’s intelligence flows. If the first phase of the AI revolution rewarded the creators of intelligence, the second phase may increasingly reward those who distribute it. That, in our view, is where the real investment opportunity begins to emerge.
Part III: The Orchestration Layer
Why the company that owns the routing layer may ultimately own enterprise AI
If the first phase of the AI revolution was defined by building the most intelligent models, and the second phase by extracting more economic value from installed infrastructure, then I believe the third phase will revolve around something far less glamorous but potentially far more valuable: orchestration. In other words, the question gradually shifts away from who builds the smartest model and toward who decides which model should perform each individual task, because once enterprises begin deploying AI at production scale rather than pilot scale, managing intelligence becomes considerably more important than merely accessing it.
One of the biggest misconceptions surrounding enterprise AI today is the assumption that organizations will ultimately standardize on a single frontier model across every conceivable workload. That may appear logical while frontier capabilities remain meaningfully differentiated, but history suggests enterprise technology almost never evolves in that direction. Companies do not purchase one database, one programming language, one cybersecurity product, or one cloud service simply because it ranks highest on a benchmark. Instead, they build technology stacks optimized around cost, reliability, governance, performance, and business requirements, with different tools performing different functions depending upon the economics of each workload.
Artificial intelligence is likely to evolve in exactly the same way. The overwhelming majority of enterprise inference does not require frontier intelligence. Document classification, contract extraction, invoice processing, customer support, software testing, enterprise search, meeting summarization, translation, compliance monitoring, and countless other routine business processes simply require a model that consistently clears the required quality threshold at the lowest possible cost. More complex reasoning tasks, strategic planning, scientific research, advanced software engineering, and autonomous agentic workflows may continue relying upon the most capable frontier systems, but those represent only a relatively small proportion of overall enterprise token consumption. The result is that organizations increasingly begin routing workloads dynamically, assigning simpler tasks to smaller, cheaper models while escalating only the most demanding requests toward frontier systems where the additional intelligence genuinely creates incremental economic value.
This transition fundamentally changes where enterprise value resides. Once routing becomes the dominant deployment model, individual AI models gradually become interchangeable components sitting behind a much larger orchestration platform. Enterprises stop purchasing intelligence directly and instead purchase a system capable of deciding which intelligence should be deployed at any particular moment. Rather than asking employees to choose between GPT, Claude, Gemini, Llama, DeepSeek, Qwen, GLM, Kimi, or future models, the orchestration layer quietly makes that decision automatically based upon latency, cost, accuracy, governance requirements, security policies, regulatory restrictions, customer preferences, and workload complexity. The remarkable consequence is that the model itself gradually moves into the background while the orchestration platform becomes the primary customer relationship.
This reminds me strongly of another important technological transition. For years, investors believed operating systems represented the economic center of enterprise computing because every application ultimately depended upon them. Linux fundamentally changed that assumption by making the operating system itself increasingly abundant while shifting value toward cloud infrastructure, enterprise software, cybersecurity, managed services, and higher-level applications built on top of it. The operating system never disappeared. It simply became less economically important than the surrounding ecosystem.
I believe open-weight AI models have the potential to produce a remarkably similar outcome. DeepSeek, Qwen, Llama, GLM, Kimi, MiniMax, and many other open-weight models are steadily improving at a pace that few investors fully appreciate. Importantly, they do not need to surpass every proprietary frontier model across every benchmark to reshape industry economics. They simply need to become sufficiently capable for perhaps eighty percent of enterprise workloads, because once that threshold is crossed, purchasing decisions become driven primarily by economics rather than leaderboard rankings. The model itself increasingly resembles Linux: freely available, highly capable, continuously improving, and ultimately valuable not because it captures the economic rent itself, but because it enables an entirely new ecosystem to emerge around it.
This is why I increasingly believe the AI industry is approaching its own Linux moment. Open-weight models commoditize intelligence. Orchestration monetizes it. That distinction is critical because it fundamentally alters where the industry’s competitive moats reside.
Today, much of the discussion focuses on benchmark leadership, parameter counts, reasoning scores, and scientific evaluations. Those metrics undoubtedly matter at the frontier, but they matter considerably less inside a Fortune 500 procurement committee. Enterprise CIOs rarely purchase technology because it tops an academic leaderboard. They purchase technology because it integrates seamlessly into existing identity systems, satisfies governance requirements, complies with regulatory standards, provides audit trails, supports security policies, delivers predictable service-level agreements, simplifies procurement, reduces operational complexity, and minimizes total cost of ownership. In practice, reliability almost always defeats theoretical superiority once technologies mature.
This distinction may prove enormously important for investors. Winning another benchmark generates headlines. Winning procurement decisions generates recurring revenue. The orchestration layer sits precisely where those procurement decisions occur.
Amazon’s Bedrock illustrates this transition particularly well. Although many investors continue viewing Bedrock primarily as a marketplace through which customers access Claude or other frontier models, the platform has evolved into something considerably more important. Bedrock increasingly functions as the enterprise operating system for artificial intelligence, allowing organizations to access well over one hundred model variants from numerous providers while automatically routing workloads toward whichever model best satisfies the desired combination of cost, latency, capability, and governance. Intelligent Prompt Routing, AgentCore, memory management, observability, security, identity integration, browser automation, tool calling, and production monitoring all transform Bedrock into far more than simply another API endpoint. It becomes the decision-making layer governing enterprise AI itself.
Microsoft’s Azure AI Foundry follows an almost identical philosophy. Rather than requiring customers to commit to a single frontier laboratory, Foundry allows enterprises to orchestrate increasingly diverse AI ecosystems while embedding those workflows directly inside Azure’s existing security architecture, compliance frameworks, developer tools, and enterprise software stack. Google Vertex AI pursues a similar objective, although naturally with greater emphasis on Gemini. Regardless of their individual implementation strategies, all three hyperscalers appear to be converging toward the same destination, namely becoming the operating system through which enterprise AI workloads are orchestrated rather than merely supplying the infrastructure upon which they execute. This transition also creates switching costs that I believe the market significantly underestimates.
Today, investors often discuss switching costs as though they reside primarily at the model layer, asking whether enterprises will remain loyal to OpenAI, Anthropic, Google, or another frontier laboratory. I increasingly believe the opposite may ultimately prove true. Once organizations begin building retrieval pipelines, fine-tuned models, evaluation frameworks, memory architectures, security policies, compliance controls, observability dashboards, identity integrations, and autonomous agent workflows inside Azure AI Foundry, AWS Bedrock, or Google Vertex, changing the underlying model becomes relatively straightforward while migrating the orchestration platform itself becomes substantially more difficult. Models gradually become replaceable. The orchestration layer becomes deeply embedded inside enterprise operations.
One of my readers recently made an observation that perfectly captures this transition by suggesting that, over time, the orchestration harness may become just as valuable as the model itself because it ultimately governs how intelligence is deployed throughout the organization. I believe that insight deserves considerably more attention than it currently receives. The competitive moat increasingly shifts away from possessing the smartest model and toward owning the workflow through which every model is accessed.
Another implication follows naturally from this framework. Many investors assume that cheaper models inevitably reduce compute demand because each individual inference consumes fewer resources. I believe the opposite is considerably more likely. Routing does not reduce inference. It expands it. Once organizations realize they can solve routine tasks at a fraction of today’s cost, they stop rationing AI usage altogether. Agents begin operating continuously rather than intermittently. Models repeatedly verify their own outputs, consult multiple reasoning chains, retrieve larger context windows, evaluate competing responses, and execute increasingly sophisticated autonomous workflows. Individual tokens become dramatically cheaper, but total token consumption accelerates because intelligence becomes economically viable for an ever-expanding range of applications.
This is Jevons Paradox expressed through enterprise software. Efficiency does not reduce demand. Efficiency creates entirely new demand. Every additional workflow, regardless of which model ultimately performs the inference, still traverses Microsoft’s Azure, Amazon’s AWS, or Google’s Cloud. Every routed request still consumes networking capacity, storage, GPUs, memory, security services, logging infrastructure, compliance systems, monitoring tools, and orchestration software. The cloud platform captures value regardless of whether the customer ultimately selects GPT, Claude, Gemini, Llama, DeepSeek, or another open-weight model.
In many respects, this is precisely why I believe the orchestration layer may become one of the most valuable positions in the entire AI stack. Frontier laboratories will undoubtedly continue competing aggressively for benchmark leadership because intelligence remains essential for solving humanity’s most difficult problems. Yet as artificial intelligence expands from a niche technology into the operating system of the global economy, enterprises will increasingly care less about which individual model produced an answer and considerably more about whether the entire system remains secure, reliable, compliant, cost-efficient, and seamlessly integrated into existing business processes.
History repeatedly demonstrates that as technologies mature, value migrates away from the invention itself and toward the infrastructure coordinating its widespread adoption. Artificial intelligence appears increasingly likely to follow exactly the same path. The companies that ultimately control enterprise AI may not necessarily be those producing the smartest models, but rather those quietly orchestrating billions of decisions every day while making the complexity of artificial intelligence almost entirely invisible to the customer. In our view, that is where some of the industry’s widest and most durable competitive moats are likely to emerge.
Part IV: Government, Geopolitics, and the New AI Order
Why regulation may accelerate the rise of the hyperscalers rather than slow it.
Up until this point, the thesis has been built almost entirely around economics. Falling inference costs, rapidly improving open-weight models, enterprise token optimization, and the migration of economic value toward infrastructure all point toward the same conclusion, namely that the hyperscalers are becoming increasingly central to the future AI ecosystem. However, there is another force quietly reinforcing exactly the same outcome, and unlike technological progress, this force is unlikely to follow Moore’s Law or any predictable engineering roadmap. It is geopolitics.
For much of the internet era, technology companies operated under the assumption that software could move freely across borders, allowing innovations developed in one country to become globally available almost instantaneously. Artificial intelligence is proving fundamentally different because frontier models are increasingly viewed not merely as commercial products, but as strategic national assets whose capabilities extend into cybersecurity, intelligence gathering, military applications, scientific research, and critical infrastructure. Once governments begin viewing AI through the lens of national security rather than purely commercial competition, entirely new economic dynamics begin to emerge.
Recent developments in the United States illustrate this shift remarkably well. The Trump administration’s Executive Order promoting advanced artificial intelligence innovation and security establishes a framework under which certain frontier models may undergo government evaluation before broad commercial release, particularly where advanced cyber capabilities are involved. Although the framework remains voluntary rather than mandatory licensing, it introduces something that previously did not exist: a structured relationship between frontier AI laboratories and the federal government regarding the deployment of the most capable models.
That development may appear incremental today, but I believe its long-term implications are considerably larger than markets currently appreciate. The debate is no longer simply about who builds the smartest model. It is increasingly about who gets access to that model, under what conditions, and through which infrastructure.
The Anthropic episode illustrates this changing landscape. Earlier this year, access to Anthropic’s most advanced models became subject to export-control considerations, creating a situation where commercial deployment was no longer determined solely by technical readiness or customer demand, but increasingly by geopolitical considerations. Regardless of one’s view regarding the merits of such policies, the broader direction appears unmistakable. Frontier AI models are gradually becoming strategic technologies subject to government oversight in much the same way that advanced semiconductor manufacturing equipment, cryptography, aerospace technology, and certain defense capabilities have been for decades.
This creates a problem that many enterprises have yet to fully appreciate. If different models become available in different jurisdictions, at different times, under different regulatory frameworks, then enterprises can no longer build AI strategies around a single model provider. Global companies operating across dozens or even hundreds of countries require flexibility because regulatory requirements, data sovereignty rules, export controls, and model availability may differ substantially from one jurisdiction to another. A multinational bank, pharmaceutical company, or industrial manufacturer cannot simply pause operations because one frontier model becomes temporarily unavailable within a specific region. That reality dramatically increases the value of orchestration. The orchestration layer no longer optimizes merely for cost and performance. It increasingly optimizes for compliance.
An enterprise operating across Europe, the United States, the Middle East, and Asia may ultimately require different routing decisions depending not only on workload complexity, but also on geography, data residency, cybersecurity requirements, export restrictions, customer agreements, and local regulation. Suddenly, choosing the optimal AI model becomes an extraordinarily complex optimization problem that extends far beyond benchmark performance.
This is precisely where the hyperscalers possess structural advantages that become increasingly difficult to replicate. Microsoft, Amazon, and Google already operate some of the world’s largest globally distributed cloud infrastructures, with decades of experience managing identity systems, encryption, cybersecurity, compliance, sovereign cloud deployments, regulatory certifications, audit requirements, and government relationships across virtually every major jurisdiction. They have spent years building trust with enterprise CIOs, financial regulators, healthcare providers, defense contractors, and governments because cloud computing required solving many of these governance challenges long before artificial intelligence arrived.
In many respects, AI simply inherits those advantages. The enterprise customer no longer asks only whether GPT performs slightly better than Claude or Gemini on a reasoning benchmark. The enterprise increasingly asks a different series of questions. Can this workload legally run in Germany? Can customer data remain inside Japan? Will this model satisfy financial regulators? What happens if one provider becomes temporarily unavailable?Can workloads automatically reroute without disrupting operations? Can every inference be audited months later? Can we demonstrate compliance during a regulatory review? These are not machine learning questions. They are enterprise infrastructure questions.
History consistently suggests that enterprise technology markets reward reliability at least as much as technical superiority. CIOs rarely purchase infrastructure based solely upon benchmark rankings because downtime, compliance failures, or security breaches often cost organizations far more than marginal differences in technical performance. Artificial intelligence appears unlikely to behave differently. The smartest model may attract headlines, but the most reliable platform often wins procurement decisions.
Another consequence of this evolving geopolitical environment is that model providers themselves become increasingly dependent upon hyperscalers. As frontier development grows more computationally intensive, regulatory scrutiny increases, and global deployment becomes more complicated, independent laboratories increasingly require partners capable of supplying infrastructure, compliance, cybersecurity, sovereign cloud capabilities, enterprise distribution, and global customer relationships. The hyperscalers therefore become more than infrastructure providers. They increasingly become strategic distribution partners through which frontier models reach enterprise customers.
This creates an interesting asymmetry. Every additional model strengthens the orchestration platform. Every additional frontier laboratory makes multi-model routing more valuable. Every additional regulatory framework increases the complexity of enterprise deployment. Each of these trends reinforces the position of the cloud platforms rather than weakening it. Ironically, the more competitive the model ecosystem becomes, the more valuable the orchestration layer becomes because enterprises require a neutral platform capable of managing that complexity. This ultimately brings us back to the central investment question.
Over the past two years, investors have understandably focused on identifying which company possesses the smartest AI model because intelligence itself represented the industry’s primary bottleneck. I increasingly believe that bottleneck is beginning to shift. Intelligence continues improving rapidly across both proprietary and open-weight models, inference costs continue collapsing, and enterprise customers are becoming increasingly focused on economics, governance, and deployment rather than benchmark leadership alone.
Throughout the history of technology, investors have repeatedly overestimated the value of the invention itself while underestimating the value of the infrastructure enabling its widespread adoption. Railroads transformed commerce, yet freight networks captured recurring economic rents. The internet transformed communication, yet cloud computing became one of the greatest businesses ever created. Smartphones transformed daily life, yet operating systems and app stores ultimately became trillion-dollar distribution platforms.
Artificial intelligence may follow exactly the same pattern. The market continues debating who builds the smartest model. I increasingly believe that is becoming the wrong question. The more important question is who owns the infrastructure through which trillions of AI decisions will ultimately flow every single day. Intelligence is becoming increasingly abundant. Inference is becoming increasingly inexpensive. Models are becoming increasingly interchangeable across a growing proportion of enterprise workloads. What remains scarce are global cloud infrastructure, enterprise trust, orchestration, governance, security, compliance, distribution, and the ability to integrate all of those capabilities into a seamless platform that allows organizations to deploy artificial intelligence at global scale. History suggests that scarce assets consistently capture the largest share of long-term economic rents.
That is why our conviction extends beyond the companies creating intelligence itself and increasingly toward the companies building the operating system of the AI economy. The first chapter of artificial intelligence was about inventing intelligence. The next chapter may prove to be about distributing it efficiently, securely, and economically across every enterprise, every industry, and eventually every corner of the global economy. In our view, that is where one of the largest investment opportunities of the coming decade is likely to emerge.
Conclusion: Long these following names: NVDA, TSMC, Sk Hynix, Micron, Samsung Electronics, Microsoft. Alphabet, Amazon, Meta.
