AI Supercomputing Is Not a Moat

Every major cloud provider and chip vendor now offers AI supercomputing. Google has TPU pods. AWS has Trainium clusters. Microsoft and OpenAI pour billions into supercomputing infrastructure. NVIDIA sells GPUs to anyone with a budget. Recent industry reports suggest that around 40% of enterprises will adopt hybrid computing architectures by 2028, up from roughly 8% today, and that more than twenty vendors will offer unified developer platforms leveraging supercomputing by then.

I kept reading this as another version of a debate I already knew. The numbers change. The logic does not.

Carr published "IT Doesn't Matter" in 2003. His argument was straightforward. As information technology becomes ubiquitous and standardized, it becomes commodity infrastructure. It is necessary for participation in the modern economy, like electricity or railroads, but it cannot generate sustained competitive advantage because every competitor can access the same technology. The argument turns on scarcity. Sustained advantage requires that some firms possess a resource others do not. When everyone can rent the same GPU, the same TPU, the same cloud API, no one holds a scarce resource.

Carr was right about infrastructure. He was wrong about the slide from infrastructure to capability. That slide is exactly what AI supercomputing reveals again, twenty years later.

Barney (1991) established the VRIN framework. Sustained competitive advantage requires resources that are valuable, rare, inimitable, and non-substitutable. I stress that VRIN are characteristics of resources, not types of resources, because the wrong question is "is AI supercomputing VRIN?" The right question is whether a specific resource bundle that includes AI supercomputing has VRIN characteristics in a specific competitive context. GPU clusters alone are valuable, certainly. But they are not rare. AWS, Google, Microsoft, Oracle, and a dozen startups all sell access. They are not inimitable. NVIDIA sells the same H100 to every buyer. They are not non-substitutable. AMD MI300X and custom ASICs already compete. The hardware fails every VRIN test.

The mistake is stopping at the hardware. Bharadwaj (2000) demonstrated that firms with high IT capability, defined as the organizational ability to mobilize IT infrastructure, human IT resources, and IT-enabled intangibles, achieve significantly higher profit ratios and lower cost ratios than matched controls. The mechanism is social complexity, path dependence, and causal ambiguity. Two firms rent identical GPU clusters. One embeds the output into organizational decision routines, retrains its analysts, and restructures its workflows. The other treats the cluster as a faster way to run the same scripts. The hardware is identical. The capability is not.

I think this is the same pattern Carr missed in 2003. He looked at IT infrastructure and concluded that ubiquity destroys advantage. He did not follow the chain from infrastructure to organizational capability to differential performance. The same blind spot is repeating now. Every AI vendor sells compute as a commodity. Every consulting firm sells AI adoption frameworks. None of that is the source of advantage. What is rare is the organizational ability to reconfigure workflows, routines, and decision structures around what the compute produces.

This is where dynamic capabilities theory does real work that Carr's argument cannot dissolve. Teece et al. (1997) defined dynamic capabilities as the organizational ability to sense opportunities and threats, seize them by mobilizing resources, and transform by reconfiguring routines and resource portfolios. Eisenhardt and Martin (2000) sharpened the construct by insisting that dynamic capabilities are not vague abilities but specific, identifiable processes like product development, strategic decision-making, and alliancing. They also clarified that dynamic capabilities are necessary but not sufficient for advantage. The advantage resides in the resource configurations these capabilities produce.

Connecting this to the AI supercomputing question is the part that made the whole thing click for me. Renting GPU time is an ordinary capability. Embedding AI into how your organization makes decisions, detects fraud, prices products, or manages supply chains is a potentially dynamic capability. The first is available to every firm with a credit card. The second is heterogeneously distributed because it is embedded in routines that are socially complex, causally ambiguous, and path-dependent. Two hospitals can buy the same radiology AI platform. One integrates it into clinical workflows, retrains radiologists, and changes diagnosis protocols. The other installs the software and hopes. The platform is the same. The outcome will not be.

The productivity paradox that Brynjolfsson (1993) named had four explanations: mismeasurement, time lags, redistribution, and mismanagement. I keep coming back to mismanagement because it is the one that keeps recycling. Organizations invest in AI supercomputing without aligning processes, skills, or governance. They measure success by how many GPUs they deployed or how many models they trained, not by whether decisions improved or workflows changed. Torres, Sidorova, and Jones (2018) showed this pattern empirically for business intelligence and analytics. BI and analytics reach firm performance through business process change capabilities and functional performance, not through analytical output by itself. Dashboards do not create value. Acting on what they show does. The same holds for AI supercomputing. The compute output creates zero value until the organization reconfigures something around it.

Melville et al. (2004) gave us the integrative BVIT model that refuses the direct shortcut. IT resources interact with complementary organizational resources, which improve business process performance, which then drives organizational performance, all moderated by the competitive environment. AI supercomputing follows the exact same mediated path. The hardware is an IT resource. The complementary resources are the human skills, the redesigned workflows, the governance structures, and the decision routines that turn raw compute into action. Without those complements, the hardware is a cost center with a fast processor.

I wrote about the productivity paradox in a previous post and made the point that the mismanagement explanation from Brynjolfsson's 1993 paper keeps repeating. AI supercomputing is the latest proof. The companies that treat it as a moat because they spent more on GPUs than their competitors are making the same mistake as the companies that bought more servers in 2003 and expected competitive advantage by default.

The VRIN test applies to the bundle, not the component. The bundle that includes AI supercomputing, organizational routines designed around its output, human expertise to interpret and act on models, and governance structures that manage risk and accountability, that bundle can be valuable, rare, hard to imitate, and hard to substitute. The GPU alone cannot. The difference is organizational capability, not compute budget. That was true in 2003. It is true in 2026. And I suspect it will still be true when the next generation of hardware arrives.
internal_links:
- /notes/the-productivity-paradox-is-still-alive
---