GPU Dependency Is Resource Dependency Theory in a $200B Market

I kept running into the same structure reading supply chain coverage of the GPU shortage. A hundred articles about NVIDIA backlog, TSMC capacity, export controls, and the only question was whether supply could catch up. Nobody asked whether the problem was a structural property of the market, not a temporary blip. I think it is a resource dependency crisis hiding inside a supply chain story.

Pfeffer and Salancik published The External Control of Organizations in 1978. Their central argument was that organizational survival depends on securing critical resources from the environment, and that when resource control is concentrated, power asymmetries emerge. They showed that under conditions of uncertainty, power becomes a stronger predictor of who gets resources (Sutton and Staw, 1995, discuss this specific claim in their theory quality paper, which I have in my library). That sentence, written almost fifty years ago, is the most concise explanation I have found for what is happening in AI computing right now.

Consider the concentration. Industry reports suggest that NVIDIA controls roughly eighty percent of the AI GPU market. TSMC handles approximately ninety percent of advanced chip fabrication worldwide. A single company designs the chips that nearly every large model trains on. A single company builds them. If you need the latest GPU to train a competitive model, you have no real alternative at scale. AMD is emerging. Intel is trying. Custom ASICs exist, but they are years behind the CUDA software ecosystem. The five conditions that Pfeffer and Salancik identified as producing dependency are all present. Resource criticality: models cannot train without these GPUs. Concentration: one dominant supplier. Discretion: NVIDIA decides allocation, timing, and priority. Alternatives: emerging but not yet viable for large-scale training. Uncertainty: every cycle of export controls, new architecture releases, and allocation changes reshuffles who gets what.

When you map the industry response against resource dependency theory, the pattern is textbook. Microsoft invested billions in custom AI chips through its Maia project and committed to OpenAI infrastructure. That is not a technical decision. It is a dependency-reduction strategy. Microsoft is diversifying its resource base so that it is not entirely at NVIDIA's discretion for its most strategic initiative. Amazon built Trainium. Google has TPUs. Every hyperscaler is doing the same thing. From a resource dependency lens, this is exactly what the theory predicts: firms that face high resource concentration will seek vertical integration or find substitutes to reduce the power imbalance.

China's response is the same logic under tighter constraints. Stockpiling GPUs before export controls tightened is resource accumulation under anticipated scarcity. Pushing domestic fabrication through SMIC and developing alternative architectures like Cambricon is search for substitute resources. The Chips Act in the US, Europe, and Japan is collective action to reduce dependency on a single fabrication geography. Pfeffer and Salancik would have recognized every move.

The situation I find most underexamined is what happens to the thousands of SaaS companies racing to add AI features. They cannot build custom chips. They cannot negotiate allocation priority with NVIDIA. They enter a GPU allocation lottery where cloud providers decide who gets capacity and when. Their dependency is double: on the cloud provider that holds the GPUs and on NVIDIA that supplies the cloud provider. The firms that survive this period will not be the ones with the best AI strategy. They will be the ones that found a way to train and serve models without depending on a supply chain that a single company controls.

I wrote earlier about how vendor lock-in is a strategy, not an accident, and the GPU market is the most extreme version of that dynamic I have ever seen. The difference is that usually a firm can eventually leave a bad vendor. In this market, if NVIDIA decides your workload is not a priority, there is nowhere to go at scale. Widespread reports of fifty-two week lead times in 2023 and 2024 were not operational failures. They were the signature of a concentrated market exercising discretion.

Tushman and Anderson (1986) argued that environmental uncertainty is about how well future states can be predicted, and that munificence or scarcity of resources constrains organizational options. They were citing Pfeffer and Salancik when they made that argument, and I think it holds directly. The AI GPU market is not munificent for anyone outside the top few firms. It is constrained, concentrated, and uncertain. Wade and Hulland (2004) built on the same environmental dimensions from Pfeffer and Salancik to argue that resource characteristics shape competitive dynamics. In the GPU market, the resource itself is scarce, the supplier base is concentrated, and the switching costs are astronomical because the entire software stack runs on CUDA.

My reading of the next three years is this. The concentration will loosen slightly but will not break. Alternative architectures will reach viability for inference and for some training workloads, and the hyperscalers will succeed in reducing their direct NVIDIA dependency through custom silicon. But the broader market of firms that need compute will remain dependent. The firms that survive in AI will be those that follow the Pfeffer and Salancik prescription: reduce dependency through vertical integration into the supply chain, or find genuine substitutes. Everyone else will be waiting in line.

I need to double-check this next claim, but my recollection is that TSMC production capacity is already fully committed well into 2027. Even if NVIDIA wanted to increase supply dramatically, the fabrication constraint is a separate dependency that even NVIDIA cannot control. That is how deep the resource concentration runs in this market. The chip designer depends on the fabricator. The cloud provider depends on the chip designer. The SaaS company depends on the cloud provider. At every layer, resource dependency theory predicts the same outcome: the actor at the narrowest point in the chain holds power over everyone above and below. And they will use it.