Which AI infrastructure and middleware layers are most at risk of being absorbed by hyperscalers and foundation model providers in the next 18 months?

The AI application layer, particularly companies building simple applications or thin wrappers around foundation models, is at the most significant risk of absorption or displacement within the next 18 months [2, 7, 11]. Foundation model (FM) providers are aggressively pursuing vertical integration, moving up the stack to build their own agentic capabilities and applications that compete directly with their customers [4, 17, 19]. OpenAI and Google are already offering agent-building tools, while Anthropic is developing applications that rival its own ecosystem partners [2, 4]. This trend is driven by the challenging economics of the AI stack, where a large portion of venture capital invested in application startups passes through to FM providers and hardware companies, resulting in poor margins for the application layer [3, 5]. The most defensible application businesses are vertical-specific players that act as an "outsourced AI team" for enterprises, leveraging deep domain expertise and proprietary, real-time data to solve the complex "last 20%" of a workflow that general-purpose models cannot address [6, 13, 15].

The middleware and AI infrastructure layer presents a more contested outlook. Some investors view this "picks and shovels" space as a defensible investment, betting that FM providers will prefer to operate as an infrastructure layer akin to AWS rather than competing in tooling [14, 16, 18, 30]. However, evidence points to increasing risk, with some experts arguing that midsize AI infrastructure startups face significant pressure from FMs, leading to market consolidation . OpenAI's acquisition of a company like Weights and Biases signals a clear intent to build a more complete, vertically integrated stack that includes MLOps and developer tooling . The AI coding market has become the **primary battleground** for major labs, indicating that core developer middleware is a strategic priority for them, not a layer they will cede to third parties . This suggests that while some infrastructure may remain independent, tools closely tied to model interaction and development are vulnerable.

The core compute layer, dominated by hyperscalers, appears least at risk of disruption by foundation model providers in the near term. A symbiotic relationship exists where hyperscalers like AWS and Google provide the immense capital and infrastructure required by FM providers, capturing significant value in the process [10, 21]. This dynamic solidifies the hyperscalers' central role, with the frontier model market expected to consolidate into an oligopoly of these large, well-capitalized players and labs [22, 29]. While there is a long-term risk that large AI companies may insource their compute over a 5-10 year horizon , the immediate threat is minimal. In fact, the commoditization of the FM API layer, which suffers from **near-zero switching costs**, incentivizes model providers to expand into applications and middleware to capture value, thereby increasing pressure on the layers above them rather than the core compute providers below [25, 26].