Not One AI Bubble, But Three: How Wrappers, Models, and Infrastructure Will Deflate on Different Timelines

The question “Are we in an AI bubble?” badly undershoots what’s actually happening. Treating AI as a single economic unit, destined either for glorious transformation or spectacular collapse, ignores how unevenly risk is distributed across the stack.

The reality, drawn from current spending patterns, company strategies and early market shakeouts, is more nuanced: we don’t have one AI bubble; we have at least three. Each sits at a different layer of the ecosystem, and each has its own economics, competitive dynamics and likely timeline for deflation.

For AI founders, product leaders and technical decision-makers, that distinction is not academic. It determines where margins will evaporate first, where consolidation is most likely, and where today’s aggressive capital flows are most likely to leave durable value behind.

The three-layer AI stack: Why “one bubble” is the wrong frame

Arguments about an “AI bubble” tend to revolve around two extremes. On one side are those who see current valuations and capex commitments as an economic time bomb. On the other are those who insist that because AI is transformative, today’s froth will be justified in hindsight.

Even major industry figures are split but increasingly acknowledge bubble-like dynamics. Meta CEO Mark Zuckerberg has pointed to signs of financial instability around AI. OpenAI’s Sam Altman and Microsoft co-founder Bill Gates both recognize classic bubble conditions—overexcited investment, inflated valuations and a long tail of doomed projects—while still arguing that AI will ultimately reshape the economy.

Both camps share a flawed assumption: that “AI” is a monolithic market that will rise and fall together. In practice, the ecosystem breaks down into three distinct layers with very different risk profiles:

• Layer 3 – Wrappers and light applications: Tools that sit on top of existing APIs and models, often adding UX, workflow glue and prompt engineering rather than core technical innovation.
• Layer 2 – Foundation models: Frontier labs and model providers building large language models and related systems that power everything above them.
• Layer 1 – Infrastructure: GPUs, data centers, cloud platforms, memory systems and AI-optimized storage that underpins training and inference workloads.

These layers are correlated but not synchronized. They are likely to deflate on different timelines, for different reasons, and with very different long-term outcomes. Understanding where your company sits in this stack—and how to move to a more defensible position—is now a strategic necessity, not a theoretical debate.

Layer 3: Wrapper companies are closest to the edge

The most fragile part of the current boom is also the most crowded: “wrapper” companies and white-label AI offerings that repackage existing models rather than build or own them.

These products commonly take the form of a specialized user interface on top of an API like OpenAI’s, plus some prompt engineering and workflow scaffolding. A customer pays $49 per seat for what is, at its core, a more user-friendly way to talk to the same underlying model available directly from its provider.

Some of these companies have shown impressive early traction. Jasper.ai, for example, wrapped GPT models in a marketer-friendly interface and reportedly reached about $42 million in annual recurring revenue in its first year. That kind of growth validated the intuition that there was space between raw model access and end-user workflows.

But structural weaknesses are now harder to ignore:

1. Feature absorption by platforms
Large incumbents can fold the core value of many of these tools directly into their existing suites. A $50/month AI writing assistant can become a tab in Office 365. An email copilot can be bundled into Gmail. CRM-native AI can replicate standalone sales assistants.

Once this happens, the standalone product’s pricing power collapses. What looked like a product is reclassified as a feature, and platforms with distribution and bundling leverage can give that feature away to drive stickiness elsewhere.

2. The commoditization trap
Wrappers mostly pass model inputs and outputs through a thin value layer. As frontier models converge in capability and pricing falls, differentiation erodes. If OpenAI or another model provider improves prompting, instruction-following or domain tools, a wrapper’s main “secret sauce” can vanish overnight.

Because wrappers rarely own the model, they are structurally downstream of their suppliers’ roadmap. Their margins, and often their core value, can be squeezed without warning.

3. Zero switching costs
Most wrapper tools lack proprietary data, deep workflow integration or hard-to-replicate network effects. A customer can move from one AI writing tool to another—or back to the model provider’s own interface—in minutes. There is little lock-in beyond habit.

White-label vendors face a related but inverted risk: they build platforms that others brand as their own, but they themselves are constrained by proprietary systems and API limitations. They are building businesses on “rented land” where the landlord controls pricing, access and capabilities.

4. Rare but real exceptions
Cursor stands out as an exception that illustrates a possible escape path. While it still depends on underlying models, it has built defensibility by deeply embedding into developer workflows, adding proprietary functionality beyond basic API calls and cultivating user-specific configurations and habits that raise switching costs.

That doesn’t disprove the fragility of the wrapper layer; it reinforces it. Most wrappers do not have this degree of workflow ownership, and without it, they are exposed to fast-moving suppliers above and consolidating platforms below.

Timeline for deflation
Based on current dynamics, the application and wrapper layer is likely to see visible failures first, with meaningful shakeouts from late 2025 through 2026. As enterprise buyers standardize on embedded platform features and cost pressure intensifies, many thinly differentiated tools will collapse or be acquired for modest sums.

Layer 2: Foundation models face consolidation, not extinction

The middle layer—foundation model providers such as OpenAI, Anthropic and Mistral—is more robust but still exhibits bubble characteristics in capital intensity and expectations.

Economic analyst Richard Bernstein points to OpenAI as a stark illustration: the company is tied to roughly $1 trillion in AI-related deals, including around $500 billion in data center buildout, while reportedly generating only about $13 billion in revenue. The gap between current and expected earnings “certainly looks bubbly,” as he puts it.

Yet, unlike most wrappers, model companies do have meaningful moats today: specialized training expertise, privileged access to compute, hard-won systems experience and performance advantages at the frontier.

The strategic question for this layer is not whether models are useful—they clearly are—but whether:

• Capabilities will converge to the point that most models are functionally interchangeable for mainstream use cases; and
• Price competition and open-source alternatives will compress margins, turning many providers into quasi-utility infrastructure with limited differentiation.

Engineering, not just scale, will pick the winners
As baseline capabilities stabilize and become table stakes, the battleground is shifting away from “who trained the biggest model” toward “who can run powerful models economically, reliably and fast at scale.”

Three areas stand out:

• Memory and inference efficiency: Techniques such as extended KV cache architectures and improved memory management matter because they mitigate the “memory wall” that constrains inference throughput.
• Token throughput and latency: Providers that deliver higher tokens-per-second and faster time-to-first-token create measurably better user experiences and can support more demanding workloads per unit of compute.
• Systems-level optimization: From scheduling and caching to model routing and hardware utilization, systems engineering increasingly defines unit economics and scalability.

The labs that combine strong research with world-class systems engineering are best positioned to survive as the market matures. Those that focus solely on ever-larger training runs without equal attention to inference economics are far more exposed.

Capital loops and distorted demand signals
There is also concern about circular investment patterns in this layer. For example, Nvidia is committing around $100 billion to fund OpenAI’s data centers, which OpenAI in turn fills with Nvidia’s own chips. That effectively means a key supplier is subsidizing one of its largest customers, potentially amplifying demand signals beyond what is sustainable.

Even if some of that demand proves overstated, however, the foundation model category is unlikely to disappear. These companies have major strategic partnerships with hyperscalers and enterprises, and their models are embedded in workflows up and down the stack.

Timeline for deflation and consolidation
Between 2026 and 2028, the model layer is likely to undergo consolidation rather than collapse. Expect a small number—perhaps two or three—dominant global providers to emerge, with several others acquired by large tech companies. Smaller, undercapitalized labs without clear technical or go-to-market differentiation are the most likely casualties.

Layer 1: Infrastructure looks bubbly on paper, but durable in practice

At the bottom of the stack sits what may be the least appreciated—and least fragile—part of the AI economy: infrastructure. This includes GPU vendors such as Nvidia, data center operators, major cloud providers, and the memory and storage systems that support AI training and inference.

By the numbers, this layer does look like a classic bubble. Global AI-related capital expenditures and venture investments are already estimated to exceed $600 billion in 2025. Gartner projects that worldwide AI spending could reach about $1.5 trillion. Those are dot-com-era levels of enthusiasm.

Yet there is a crucial distinction: infrastructure can retain utility even if specific AI applications and model providers fail.

During the dot-com boom, enormous sums were poured into fiber-optic networks that went underutilized for years after the crash. But that capacity was not wasted—it later enabled YouTube, streaming media, large-scale SaaS and cloud computing. The future eventually caught up to the infrastructure.

Real revenue vs. pure speculation
Unlike many speculative booms, parts of today’s AI infrastructure buildout are already translating into large, realized revenues. Nvidia, for example, reported about $57 billion in revenue for its Q3 fiscal 2025, up 22% quarter-over-quarter and 62% year-over-year. Around $51.2 billion of that came from its data center division alone, reflecting concrete demand from organizations building or expanding AI infrastructure.

These numbers do not guarantee that every new data center or cluster will be fully utilized in the short term, but they indicate that buyers are making substantial, real capital commitments—often as part of long-term strategies to enable AI workloads across multiple potential applications.

From commodity boxes to integrated memory hierarchies
Modern AI infrastructure is more than just racks of generic servers and storage. It encompasses a tightly integrated memory hierarchy—from GPU high-bandwidth memory (HBM) and DRAM through to high-performance storage that acts as a “token warehouse” for inference.

This integration is itself an architectural shift. Designing infrastructure that can efficiently serve large models at scale, keep utilization high and manage data movement intelligently is a nontrivial technical challenge. That makes this layer less of a pure commodity market and more of an evolving systems landscape.

Timeline for overbuild and normalization
In the near term, some overbuilding and “lazy engineering” are likely, especially around 2026. Certain facilities may sit partially idle, similar to dark fiber in the early 2000s. But as AI workloads continue to mature and expand over the next decade, much of this capacity is likely to be absorbed.

For founders and technical leaders, the implication is that while specific infrastructure bets can be mistimed, the layer as a whole is structurally more resilient than the application and model tiers above it.

The cascade: How the unwind is likely to unfold

If this is not a single AI bubble, then its deflation will not be a single event. Instead, we are likely to see a cascade—starting at the top of the stack and propagating downward over several years.

Phase 1: Wrapper and white-label shakeout
Application-layer companies with thin differentiation face the earliest and sharpest pressure. As platforms absorb popular features and model providers improve their own UX and tools, hundreds of venture-backed startups will struggle to justify their valuations.

There are already more than 1,300 AI startups reportedly valued above $100 million, including nearly 500 “unicorns” at $1 billion or more. Many of these will not generate the growth or margins implied by those valuations once competition, commoditization and platform bundling are fully felt.

Phase 2: Foundation model consolidation
As performance converges and costs remain high, only the best-capitalized model companies with strong technical and commercial moats are likely to remain independent. A few large acquisitions—on the order of three to five—by hyperscalers and major tech companies are plausible as they seek to internalize capabilities and reduce dependence on external labs.

Phase 3: Infrastructure normalization
At the bottom of the stack, infrastructure spending is more likely to normalize than collapse. Some data centers may operate below capacity for a period, but as AI use cases expand and mature, that capacity is poised to be put to work, much as unused fiber eventually was.

The net effect is a multi-year redistribution of value and power across the stack, not a single catastrophic implosion.

Implications for AI founders and product builders

For teams building in and around AI, the biggest risk is not merely sitting in the wrapper layer—it’s staying there.

Move from wrapper to true application
If your product primarily passes prompts to a model and returns responses, your defensibility is weak. To improve it, you need to own more of the workflow before and after the AI interaction. That means:

• Embedding into critical daily processes (e.g., drafting, review, approval, execution)
• Managing state, history, context and collaboration—not just single queries
• Integrating with systems where work actually gets done (CRMs, IDEs, productivity suites, line-of-business tools)

From application to vertical SaaS
Robust AI products increasingly look like vertical SaaS platforms with AI at the core, not AI features in search of a problem. That often requires:

• Execution layers that let users act on AI output without leaving your product
• Proprietary data creation and enrichment as a byproduct of use
• Deep, opinionated workflows that make switching genuinely painful

Prioritize distribution as a moat
In a world where underlying models are accessible and improving for everyone, your long-term edge is rarely the LLM itself. It is how you acquire, retain and expand users:

• Channels and partnerships that are hard to replicate
• Community and ecosystem effects around your product
• Strong tenant isolation and data guarantees that enable enterprise trust

Winning AI companies will look at least as much like distribution and workflow businesses as they do pure technology plays.

How to read the timelines if you’re allocating capital

For investors and technical decision-makers, the staggered timelines across the three layers suggest different risk and return profiles.

Short term (next ~18 months): Expect heavy volatility in wrappers and application-layer startups. Many will fail, but some will successfully move “upstack” into owning workflows, data and vertical SaaS positions.

Medium term (2–4 years): Foundation model providers will likely consolidate, with a small number of dominant players and a handful of acquired specialists. Pricing, performance and partnership structures may look very different by 2028 than they do today.

Long term (5+ years): Infrastructure investments are more likely to be validated by growing workloads, even if some assets are mis-timed or overbuilt in the short run. Architectural innovations in memory, storage and data movement will matter more than any single model release.

The key for capital allocators is to distinguish between:

• Froth that depends on fleeting feature gaps or distribution arbitrage; and
• Assets and capabilities that remain useful no matter which specific applications or providers win.

The bottom line: A roadmap, not a verdict on AI

Asking whether we are in “an AI bubble” misses the point. We are in multiple bubbles layered on top of each other, each with different characteristics and deflation paths.

Wrapper companies at the top of the stack are most exposed and likely to pop first. Foundation model providers will not vanish but will consolidate under pressure from economics and competition. Infrastructure, despite headline-grabbing capex numbers, is the most likely to retain long-term value, even if the path there includes a period of overbuilding and underutilization.

For builders, this is less a warning than a map. Knowing which layer you occupy—and which bubble dynamics apply to you—can be the difference between being swept away in the correction and emerging as one of the durable platforms that define the post-bubble AI landscape.