AI Bubbles: Multiple Peaks & Timelines, Not One Burst

The Three Layers of the AI Ecosystem

Layer 3: The Wrapper Companies – Primed for a Fall

The most precarious position isn’t building AI; it’s repackaging it. These companies take readily available large language models (LLMs) like those from OpenAI, add a user-friendly interface, and often some basic prompt engineering, then charge a subscription fee – sometimes as high as $49 per month – for what essentially amounts to a polished ChatGPT front-end. Jasper.ai, for example, rapidly achieved approximately $42 million in annual recurring revenue by offering a marketing-focused wrapper around GPT models.

But cracks are appearing. These businesses face existential threats on multiple fronts:

• Feature Absorption: Major tech platforms like Microsoft, Google, and Salesforce can easily integrate similar functionality directly into their existing suites – Office 365, Gmail, and CRM, respectively – effectively rendering standalone wrappers obsolete.
• Commoditization: As foundation models improve and become more accessible, the value proposition of simple wrappers diminishes. Falling prices and increasing similarity between models compress margins to unsustainable levels.
• Zero Switching Costs: Most wrapper companies lack proprietary data, deep integrations, or embedded workflows. Customers can effortlessly switch to competitors or directly utilize the underlying LLM with minimal disruption.

The white-label AI market exemplifies this fragility, with businesses facing vendor lock-in and API limitations. They are, in essence, building on leased land, subject to the whims of the property owner.

Cursor represents a rare exception, demonstrating defensibility through deep integration into developer workflows, unique features beyond simple API calls, and the cultivation of strong network effects. However, Cursor remains an outlier; most wrappers lack this level of integration and user lock-in.

Timeline: Expect significant failures in this segment between late 2025 and 2026 as large platforms absorb functionality and users recognize the limited value proposition.

Layer 2: Foundation Models – Navigating a Precarious Middle Ground

Companies building LLMs – OpenAI, Anthropic, and Mistral – occupy a more defensible, yet still vulnerable, position. Economic researcher Richard Bernstein highlighted OpenAI’s $1 trillion in AI deals, including a $500 billion data center buildout, juxtaposed against projected revenue of only $13 billion, as a clear indicator of bubble-like dynamics.

These companies possess genuine technological moats: expertise in model training, access to substantial compute resources, and performance advantages. The critical question is whether these advantages are sustainable, or if models will ultimately commoditize into low-margin infrastructure utilities.

Engineering Excellence is Paramount: As foundation models converge in baseline capabilities, the competitive edge will increasingly hinge on inference optimization and systems engineering. Companies that can overcome the “memory wall” through innovations like extended KV cache architectures, achieve superior token throughput, and deliver faster response times will command premium pricing and market share. The winners won’t simply be those with the largest training runs, but those who can make AI inference economically viable at scale.

Furthermore, the circular nature of investment – Nvidia investing heavily in OpenAI’s data centers, which then purchase Nvidia’s chips – raises concerns about artificially inflated demand. Despite these concerns, these companies benefit from substantial capital backing, genuine technical capabilities, and strategic partnerships.

Timeline: Consolidation is anticipated between 2026 and 2028, resulting in 2 to 3 dominant players and the acquisition or closure of smaller model providers.

Layer 1: Infrastructure – The Most Resilient Layer

Here’s a contrarian perspective: the infrastructure layer – encompassing Nvidia, data centers, cloud providers, memory systems, and AI-optimized storage – represents the least bubbly segment of the AI boom. While global AI capital expenditures and venture capital investments are projected to exceed $600 billion in 2025, potentially reaching $1.5 trillion according to Gartner, infrastructure retains inherent value regardless of which specific applications succeed.

The fiber optic cables laid during the dot-com bubble weren’t rendered useless when that bubble burst; they enabled the rise of YouTube, Netflix, and cloud computing. Similarly, the chips, data centers, and storage infrastructure being built today will power whatever AI applications ultimately prevail.

Nvidia’s Q3 fiscal year 2025 revenue reached approximately $57 billion, a 22% increase quarter-over-quarter and a 62% increase year-over-year, with the data center division alone generating roughly $51.2 billion. These figures demonstrate genuine demand and substantial infrastructure investment. Nvidia’s financial results underscore the ongoing demand for AI infrastructure.

Modern AI infrastructure encompasses the entire memory hierarchy – from GPU HBM to DRAM to high-performance storage – representing a fundamental architectural innovation, not merely a commodity play.

Timeline: Short-term overbuilding and inefficiencies are possible in 2026, but long-term value retention is expected as AI workloads expand over the next decade.

Frequently Asked Questions About the AI Bubble

• What is the biggest risk for AI wrapper companies?

  The biggest risk is feature absorption by larger tech companies. When giants like Microsoft or Google integrate similar functionality directly into their existing products, the value proposition of standalone wrappers evaporates.

• Are foundation model companies also at risk of a bubble bursting?

  Yes, but to a lesser extent than wrapper companies. While they possess technological moats, the commoditization of models and the need for significant infrastructure investment pose substantial challenges.

• Which layer of the AI ecosystem is considered the most resilient?

  The infrastructure layer – including Nvidia, data centers, and cloud providers – is considered the most resilient because it provides essential resources regardless of which specific AI applications succeed.

• What should AI application builders do to increase their chances of survival?

  They should move up the stack by owning the entire workflow surrounding the AI interaction, building execution layers, and focusing on distribution to create a strong user base.

• What is the predicted timeline for the AI bubble to burst in different layers?

  Wrapper companies are expected to face significant failures between late 2025 and 2026, foundation models will consolidate between 2026 and 2028, and infrastructure spending will normalize after 2028.