The Power Hierarchy in AI

The Infrastructure–Application Sandwich • Value accrues to the floor and the ceiling • Models get squeezed

  1. Power in AI does not sit in the model layer. It sits above (the application ceiling) and below (the infrastructure floor) (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
  2. As model capabilities converge, they become interchangeable — a cost center, not a profit center.
  3. Whoever controls customer interaction or compute supply controls the economics.

WHERE VALUE ACCRUES IN THE AI STACK

Control the floor (infrastructure) or the ceiling (applications) — not the middle.

The AI stack has hardened into a three-layer power structure:

  • The Ceiling: Applications
  • The Middle: Models
  • The Floor: Infrastructure

Only the top and bottom layers compound value.
The middle layer compresses.

This is the infrastructure–application sandwich (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).


THE INFRASTRUCTURE–APPLICATION SANDWICH

1. THE CEILING: APPLICATIONS

Controls customer relationship → Captures margin

Applications determine:

  • who owns the user
  • who owns the workflow
  • who controls the transaction
  • who captures the value

Commerce Layer

ACP (OpenAI) • A2P (Google)

Enterprise Layer

Microsoft Copilot • Google Vertex

Vertical Layer

Harvey (legal) • Hippocratic (healthcare)

Applications are the monetization engines of the agentic economy.


POWER CONCENTRATION AT THE CEILING

11.4% → ChatGPT conversion

vs 10.2% direct visits
AI-mediated commerce outperforms human browsing.

1M+ Shopify merchants onboarding via ACP

Agent-mediated transactions at scale.

Whoever owns the agent interface owns demand (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).


▼▼▼ SQUEEZE FROM ABOVE ▼▼▼

Application platforms compress model profits by:

  • multisourcing models
  • treating them as interchangeable
  • prioritizing integration over innovation
  • routing traffic based on cost-performance

The ceiling pushes down.


2. THE MIDDLE: MODELS

Converging capabilities → Squeezed margins → Cost centers

Model capabilities are rapidly converging:

  • Opus: 80.9%
  • GPT-5.1: ~75%
  • Kimi K2: 71.3%
  • Gemini 3: closing the gap

The differences are shrinking month by month.
This eliminates model-level differentiation and collapses margins.

Models become utilities — not monopolies.

(as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new)


MODEL COMMODITIZATION

$0.15 vs $10

The pricing gulf between Kimi-style efficiency models and premium US frontier models is now 10×.

This is the biggest evidence of commoditization in the industry.

As long as open-weight and INT4-native models keep improving, model vendors will operate under constant margin pressure.


▲▲▲ SQUEEZE FROM BELOW ▲▲▲

Infrastructure providers also push models downward:

  • NVIDIA raises GPU costs
  • hyperscalers standardize APIs
  • custom silicon (TPU, Trainium) captures more value
  • cloud providers bundle models at low cost

The floor rises; the ceiling presses down.
The model layer gets trapped between two monopolistic forces.

This is structural — not temporary.


3. THE FLOOR: INFRASTRUCTURE

Table stakes to compete → The floor keeps rising

The floor layer includes:

  • silicon
  • data centers
  • energy
  • distributed networks
  • global GPU supply

This layer controls:

  • unit economics
  • deployment speed
  • training feasibility
  • platform power

Infrastructure is the scarce layer.
Models and applications are abundant.

(as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new)


INFRASTRUCTURE LAYER SCALE

$57B — NVIDIA Q3 revenue

Record-breaking GPU demand.

$500B — Stargate commitment

Infrastructure-first strategy for OpenAI + SoftBank + Oracle + MGX.

$5T — NVIDIA market cap

Company valuation driven entirely by infrastructure dominance.

Infrastructure is no longer “support.”
It is the moat.


THE STRUCTURE OF POWER

Applications pull value up • Infrastructure pulls value down • Models get crushed in the middle

The Ceiling (Apps)

  • owns identity
  • owns workflows
  • owns purchase events
  • owns enterprise lock-in
  • owns agentic commerce

The Middle (Models)

  • interchangeable
  • price-pressured
  • ML research no longer defensible
  • open-weight alternatives spreading
  • cost centers, not profit centers

The Floor (Infra)

  • controls silicon
  • sets capacity
  • raises floor pricing
  • dictates who can train frontier models

This is the new AI hierarchy.


THE ZERO-SUM SHIFT

Value migrates up and down — never stays in the middle.

Three truths now define the AI economy:

  1. Models will never recapture margin.
    They have become the commodity layer.
  2. Applications will monopolize customer-facing value.
    Whoever owns the agentic interface owns the money.
  3. Infrastructure will monopolize compute-facing value.
    Whoever owns the chips and data centers owns the bottleneck (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).

THE STRATEGIC INSIGHT

The middle layer dies first.

For all the hype about “model superiority,” the truth is simple:

  • Infrastructure prints free cash flow
  • Applications capture margin
  • Models generate cost

This is why:

  • OpenAI moved into infrastructure
  • Google doubled down on TPU
  • AWS vertically integrated silicon
  • China optimized for INT4 efficiency
  • Enterprises shifted to agent-level integrations

Everyone is escaping the middle.


THE BOTTOM LINE

Power in AI is governed by the floor and the ceiling — not the middle.

  • The floor determines who can train
  • The ceiling determines who gets paid
  • The middle becomes a low-margin utility

This is the power hierarchy of the AI era (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).

Whoever controls the lowest level of compute or the highest level of customer interaction will dominate the next decade of AI economics.

Scroll to Top

Discover more from FourWeekMBA

Subscribe now to keep reading and get access to the full archive.

Continue reading

FourWeekMBA