AI is not just a product trend. It is a shift in who holds leverage along three stacked markets: chips, cloud, and models. Each layer already had its own incumbents, pricing logic, and political economy. The current wave of large-scale AI rearranges that stack rather than simply adding one more line item to it. Viewed from the balance sheet of a serious AI deployment, three questions quietly determine who captures value: who controls advanced accelerators, who owns the datacenters that feed them power and cooling, and who sits closest to the end user through models and workflows. The rest of the industry orbits those constraints. ## Legacy structure before the current wave Before large-scale deep learning models without centralizing data and generative models became central, the stack looked roughly like this. At the bottom sat chip designers and fabs. Nvidia, AMD, and Intel designed GPUs and CPUs. TSMC, Samsung, and a handful of others transformed designs into silicon. Bargaining power at this layer already mattered, but demand was diversified across gaming, HPC, enterprise, and mobile. Above that sat the hyperscale clouds. AWS, Azure, and Google Cloud aggregated demand, turned it into virtual instances, and monetized utilization. They bought hardware in bulk, amortized datacenter build-out across many workloads, and sold compute plus storage plus managed services. At the top sat SaaS and enterprise software. CRM, ERP, ticketing, collaboration tools, marketing automation. Some used ML internally, but the models were invisible to most buyers. Few customers cared which GPU rendered their dashboard or scored their lead. AI models existed, but they did not distort the structure. They fit inside existing budgets as one technical component among many. ## The AI shock to demand Large-scale training runs and inference workloads changed the proportion of everything. Training a frontier model consumes hundreds of millions of dollars of compute. Scaling up inference for consumer and enterprise assistants pulls in continuous GPU time. These are not occasional batch jobs. They are always-on demands with strong latency constraints and spiky traffic. That demand spike had two immediate effects. First, it turned advanced accelerators into a constrained resource. The availability of high-end GPUs or bespoke AI chips became the binding factor for many companies. Product roadmaps and fundraising stories began to pivot around "access to compute." Second, it pulled model providers into the center of the stack. Instead of being an internal module owned by a SaaS vendor, the model itself became a standalone product, often exposed via API from a separate company. That changed who sits between the user and the underlying cloud. ## Chip vendors and the new bottleneck Nvidia sits at the center of this story. Its combination of hardware, software stack, and ecosystem makes it a quasi-utility for modern AI. The hardware side is visible: successive generations of GPUs tuned for tensor workloads, with careful attention to memory, interconnect, and power efficiency. Less visible but equally important is CUDA, the software stack that became the default environment for many ML frameworks and tools. This creates several layers of leverage. A company that controls the dominant accelerator platform can charge premium margins without losing volume, because alternatives impose migration costs and performance risks. It can bundle hardware with networking, systems reliability engineering, and software. It can influence which kinds of workloads are easy or hard to run efficiently. Competitors exist. AMD pushes its ROCm stack and GPUs. Large players invest in their own AI ASICs. Startups explore specialized accelerators. All face the same obstacle: inertia. The installed base of models, tooling, and operational knowledge ai how teams actually repartition tasks between humans and models around the incumbent's platform is enormous. The result is a chip layer with unusual pricing power and an unusually strong say in the pace and shape of AI expansion. Hyperscalers and model labs negotiate hard, but they negotiate from a position downstream of that bottleneck. ## Clouds as capital allocators and gatekeepers Hyperscale cloud providers transform capex into rentable capacity. In AI, that role becomes more central and less interchangeable. Building domain specific assistants for law finance and medicine datacenters that can host dense GPU clusters entails specific costs: high-voltage power, cooling, networking, physical footprint, long-term contracts with utilities and governments. These are not easily replicated by smaller players. Cloud providers that secure long-term hardware supply, build or retrofit datacenters for AI, and finance multi-year build-outs move into a position similar to that of pipeline operators in the energy world. They allocate scarce capacity between internal AI projects, external customers, and strategic partners. This changes their leverage in three ways. - They can bundle compute with model access, storage, security, and higher-level AI services.

• They can offer better unit economics to customers who adopt their particular AI stack and managed services.
• They can decide which model vendors or tool providers receive favorable placement and commercial support. Clouds already had power as the default infrastructure choice. AI intensifies it by making access to certain types of compute non-fungible. A generic CPU workload can move across providers with some pain. A large-scale GPU workload pinned to a specific networking topology, managed service, or proprietary accelerator stack moves far less easily. ## Model providers in the middle Model providers insert a new layer between clouds and end users. Some are independent labs that rent compute and offer models via API. Others are internal units of the clouds themselves. Some are hybrid: a lab partly backed or owned by a single cloud, with joint go-to-market and revenue sharing. Their leverage rests on three assets. - Weights and training pipelines asset. The models themselves, plus the know-how and data that went into them.
• Developer mindshare. Tooling, documentation, and community that makes integration easier and creates a default choice.
• Product adjacency. Assistants, agents, and domain solutions that sit directly in front of users and capture behavioral data. Model providers that control large user-facing surfaces accumulate detailed traces of how people use their systems. That feedback loop can feed back into fine-tuning, alignment, and new features. It also builds switching costs: moving away means losing not just the model, but a growing history of interactions. At the same time, these providers depend heavily on cloud and chip players. Their cost base is dominated by compute. Their ability to serve traffic depends on datacenters they do not own. Their bargaining power reflects how unique their models are, how easy it is for customers to swap them out, and how deeply they are embedded in workflows. Open-weight models and self-hosting shift this dynamic slightly. When enterprises can run competitive models on their own infrastructure or through specialized hosting providers, the bargaining power of model vendors shifts toward those who can offer better weights plus better tooling, rather than monopoly access. ## Vertical integration and realignment The natural talking to computers still hard response to this three-way dependency is vertical integration. Cloud providers launch their own model families and assistants. They want to sell not just raw compute, but a full AI stack: from training through inference to application frameworks. That captures higher margins and reduces dependence on external labs. Model providers seek closer ties with clouds or build their own infrastructure footprints. Joint ventures, exclusivity agreements, and long-term capacity deals appear. In some cases, model labs effectively become anchor tenants in cloud datacenters, securing favorable pricing in exchange for traffic and strategic alignment. Chip vendors experiment with going up-stack: reference systems, software platforms, and partnerships with both clouds and model labs. Sometimes they even back model projects directly to stimulate demand and showcase their hardware. The result is less a simple hierarchy and more an ecosystem of alliances ai communities govern themselves, often with exclusivity clauses and complex incentives. A model provider deeply tied to one cloud may gain privileged access to hardware and distribution there, but lose bargaining power elsewhere. A cloud that backs its own model family may face tension between promoting it and remaining friendly to independent model vendors. ## Scenarios that shape who captures value Several trajectories emerge over the medium term. In one, clouds dominate. Their control over capital-intensive datacenters and integration into enterprise workflows allows them to absorb much of the model layer. Independent labs become specialized suppliers or get acquired. Model differentiation matters less than the completeness and integration of the cloud's own AI platform. In another, chip vendors maintain outsized influence. Supply constraints persist, alternatives lag, and the cost structure of AI remains heavily skewed toward the hardware layer. Clouds and model providers compete aggressively downstream, but margins compress, while chip margins remain robust. In a third, models themselves retain strong differentiation. Transferable skills, safety reputations, and ecosystem gravity around certain model families create a layer that enterprises insist on, independent of cloud choice. Multi-cloud and on-prem hosting options give buyers enough flexibility to avoid total lock-in with any one infrastructure provider. Reality is likely to mix elements of all three, with differences by region and sector. Public policy adds another variable: export controls on advanced chips, national ambitions for "sovereign AI," and regulation of cloud infrastructure all skew incentives. ## Implications for everyone else in the stack For SaaS vendors and startups that build on top, these shifts are not abstract. A company that depends primarily on a single model API is exposed to both pricing and strategic choices made far upstream. If that model provider tightens terms, prioritizes its own applications, or changes behavior for safety or branding reasons, downstream products absorb the shock. A company that depends primarily on a single cloud's AI stack inherits that provider's strengths and weaknesses. Integration and performance are excellent, but strategic options narrow. Negotiating power later in the life of the product shrinks. On the other hand, fragmentation in the model layer and gradual diversification in hardware create room for specialized hosts, optimization platforms, and open-weight ecosystems. Those actors derive their leverage from helping buyers navigate and hedge against the power of the big three layers. ## Conclusion The headline is simple. AI changes who holds market power because it makes certain resources and capabilities non-substitutable for extended periods: cutting-edge accelerators, dense AI-ready datacenters, and a small number of model families with real adoption and trust. The detailed shape of that power will keep evolving, but the core constraints are already visible on every serious company's planning sheet.