The relentless buzz around AI has sparked a crucial debate: Are we in an AI bubble poised to burst? But framing it as a single, monolithic bubble is a dangerous oversimplification. As someone who's spent years building and studying these systems, I see a far more nuanced reality: a series of distinct bubbles, each inflated by different forces and set to deflate at different times. Understanding these layers – and their inherent risks – is now paramount.

The Perilous Position of 'Wrapper' Companies

The first, and most vulnerable layer, comprises the companies that essentially repackage existing AI models. These firms often take an API like OpenAI's, slap on a user-friendly interface, sprinkle in some prompt engineering, and then charge a premium subscription. While some, like Jasper.ai, saw early success, the cracks are already forming. Their fundamental weakness? They don't own anything.

These wrapper companies are threatened by larger platforms that can easily absorb their functionality. Microsoft could bundle an AI writing tool into Office 365; Google could integrate an AI email assistant into Gmail. According to The Verge, the biggest risk for these companies is that their product becomes a feature of a larger platform. Moreover, these businesses face commoditization: as foundation models improve and costs decrease, their value proposition erodes. Customers face zero switching costs. The white-label AI market, as TechCrunch reports, exemplifies this fragility. The timeline? Expect significant failures in this segment by late 2025 and throughout 2026.

There are exceptions. Consider Cursor, a wrapper-layer company that has achieved real defensibility by deeply integrating into developer workflows and creating proprietary features. But these are outliers, not the norm.

Foundation Models: A More Defensible, But Still Precarious, Position

The companies building the Large Language Models (LLMs) themselves – OpenAI, Anthropic, Mistral – occupy a more defensible position, but they aren't immune. Economic researcher Richard Bernstein points to OpenAI as an example of potential bubble dynamics, noting the vast AI deals against current revenue.

These companies possess significant technological moats: model training expertise, access to compute, and performance advantages. The key question is whether these advantages are sustainable. As baseline capabilities converge, the competitive edge will increasingly come from inference optimization and systems engineering. As I've seen firsthand, those who can scale the memory wall and deliver faster time-to-first-token will command premium pricing. The winners won’t just be those with the largest training runs, but those who can make AI inference economically viable at scale.

Another concern is the circular nature of some investments, where companies like Nvidia essentially subsidize their own customers, potentially inflating demand. Still, these companies have massive capital backing and strategic partnerships. Consolidation is inevitable, with perhaps 2-3 dominant players emerging between 2026 and 2028.

The Solid Foundation: Infrastructure

Here's where my contrarian viewpoint emerges: the infrastructure layer – Nvidia, data centers, cloud providers, memory systems – is the least bubbly. Global AI capital expenditures already exceed $600 billion in 2025. That sounds bubbly, right? But infrastructure retains value regardless of which specific AI applications succeed. Think of the fiber optic cables laid during the dot-com bubble; they weren't wasted. They enabled YouTube, Netflix, and cloud computing.

Nvidia's Q3 fiscal year 2025 revenue hit approximately $57 billion, a significant increase year-over-year. These aren't vanity metrics; they represent genuine infrastructure investments. The chips, data centers, and memory systems being built today will power whatever AI applications ultimately succeed. This integrated approach to memory and storage represents a fundamental architectural innovation.

Short-term overbuilding is possible in 2026, but long-term value retention is expected as AI workloads expand over the next decade. This isn't just about commoditized storage; it's about the entire memory hierarchy, from GPU HBM to DRAM to high-performance storage systems.

The AI boom won't end with a single dramatic crash. Instead, expect a cascade of failures starting with the most vulnerable wrapper companies, followed by consolidation in the foundation model space, and finally, a normalization of infrastructure spending. For builders, the risk isn't being a wrapper; it's staying one. The key is to move upstack, own the workflow, and create proprietary data and deep integrations. The winners won't just be software companies; they'll be distribution companies. The AI revolution is undeniably real, and understanding these layers, understanding the nuances of each level, is paramount to ensuring a company's survival as the landscape continues to evolve and mature.