• AI value has polarized into a barbell: foundation model companies on one side, deep vertical specialists on the other.
• The middle — API aggregators, generic chatbots, prompt wrappers, “AI platforms” — is getting crushed from both directions.
• The structural implication is blunt: either own the intelligence layer ($5–10B) or own the domain layer ($1–5B). Nothing in between scales.

For weekly coverage of how these patterns evolve across markets, funding, and product strategy, follow This Week in Business AI:
https://businessengineer.ai/p/this-week-in-business-ai-the-2025

THE PATTERN: VALUE CONCENTRATES AT THE EXTREMES

This is the clearest structural pattern in the 2025 unicorn data:

Value clusters at the deepest layers — intelligence (foundation)
and domain expertise (verticals).

Everything in the middle gets squeezed.

The barbell shape tells you exactly where value accrues.

LEFT EXTREME — FOUNDATION PLAYERS ($5–10B)

Companies:

• Thinking Machines Lab — $10B
• Reflection AI — $8B
• Reka — $1B
• Anthropic / OpenAI ecosystems (large but not shown for scale reasons)

Why this pole captures value:

• Frontier model performance
• Data feedback loops
• Capital intensity that deters entrants
• Distribution via platform embeddings
• National-level and hyperscaler-level support

Owning the intelligence layer means owning the raw resource that powers the rest of the stack.

Winner-take-most.

RIGHT EXTREME — VERTICAL SPECIALISTS ($1–5B)

Companies:

• Harvey — Legal AI
• Cursor — Developer AI
• Hippocratic — Healthcare AI
• Decagon — Customer Service AI

Why verticals thrive:

• Domain-specific data moats
• Clear ROI stories
• Embedded workflows
• Industry trust and compliance advantages
• Early enterprise adoption

Vertical AI avoids commoditization by going deep, not broad.

Winner-take-many.

THE SQUEEZED MIDDLE — WHERE VALUE DOES NOT FORM

This is the graveyard of 2023–2025 startup formation:

1. Generic Chatbot Wrappers

No defensibility.
Zero differentiation.
Easy to copy.

2. API Aggregators

Middlemen with no lock-in.
Destroyed by platform pricing power.

3. Basic AI Workflows

Not proprietary.
Absorbed by larger tools.

4. Prompt Engineering Tools

Features, not companies.
Eaten by model providers and infra layers.

5. Horizontal AI Platforms

“All-in-one” = “AI for nothing.”
No target customer.
No domain moat.

These categories get crushed by:

• upstream pressure from foundation models
• downstream pressure from vertical specialists

This is the structural force that defines the barbell.

THE STRUCTURAL IMPLICATION

LEFT EXTREME: Own the Intelligence Layer ($5–10B)

Foundation models → capital-heavy, compute-heavy, defensibility-heavy.

RIGHT EXTREME: Own the Domain Layer ($1–5B)

Vertical apps → workflow-heavy, outcome-heavy, ROI-heavy.

THE MIDDLE: Gets Squeezed from Both Sides

There is no room for horizontal AI platforms.
No room for shallow tooling.
No room for generic wrappers.
The economics crush these categories before they scale.

WHY THE BARBELL EMERGED

Three underlying forces shape the barbell:

1. Commoditization of the Middle

Platform providers absorb anything generic or horizontal.

2. Specialization at the Edges

Models differentiate through scale.
Vertical AI differentiates through domain expertise.

3. Capital Concentration

Investors fund the extremes because they have:

• deeper moats
• clearer outcomes
• higher margins
• more defensible market positions

The middle lacks both defensibility and scale.

THE STRATEGIC INSIGHT: PICK A SIDE — INTELLIGENCE OR DOMAIN

If you’re building in AI, the barbell distribution forces a decisive choice:

• Do you own the intelligence?
  or
• Do you own the domain?

Any attempt to sit between these poles — platform, wrapper, toolkit, aggregator — is crushed by the structural physics of the stack.

Depth wins.
Breadth dies.
The middle is suicide.

For ongoing analysis across all five structural patterns, tune into:
https://businessengineer.ai/p/this-week-in-business-ai-the-2025