The defining framework of AI value distribution • Models commoditize • Value flows to the edges
- The AI stack behaves like every prior platform era: value moves to the edges — infrastructure and applications (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
- Models sit in the middle and get squeezed from both directions as benchmarks converge and pricing collapses.
- Strategic winners either control the floor (compute) or the ceiling (applications). Everyone else becomes a commodity.
THE PLATFORM ERA PATTERN REPEATS
PC era: Intel/Microsoft won. Mobile era: Apple/Google won. AI era: Infrastructure + Applications will win.
Every platform cycle ends the same way:
- The layer that controls distribution wins.
- The layer that controls bottlenecks wins.
- The middle gets crushed.
The AI era is no different.
Models are not the place where enduring power accumulates (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
THE SANDWICH
The AI economy is now a three-layer value hierarchy:
- The Ceiling → Applications
- The Filling → Models
- The Floor → Infrastructure
Value flows upward into applications and downward into infrastructure.
The middle collapses.
1. APPLICATIONS (THE CEILING)
Controls the customer relationship → Captures the margin → Sets the price
This is the most powerful layer in the AI economy.
Why Ceiling Wins
- Owns customer, data, workflows
- Controls willingness to pay
- Can switch underlying models at will
- Treats models as interchangeable suppliers
- Converts better than human browsing
Example: AI-mediated commerce converts at 11.4 percent vs 10.2 percent direct.
Applications are the consumer-facing aggregation layer (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
Where the Ceiling Lives
- Commerce: ACP (OpenAI), A2P (Google)
- Enterprise: Agent 365, Vertex
- Vertical: Harvey (legal), Hippocratic (healthcare)
The ceiling sets the economic rules for every other layer.
▼ PRESSURE (FROM ABOVE)
Applications push down on the model layer:
- Demanding lower price per token
- Rotating models based on cost-performance
- Treating them like pluggable components
- Capturing the full customer lifetime value
The ceiling extracts margin from the middle.
2. MODELS (THE FILLING)
Converging benchmarks → Commodity economics → Cost center, not profit center
The model layer is structurally weak because differentiation is collapsing.
Evidence of Convergence
- Opus ≈ GPT-5.1 ≈ Gemini 3 ≈ Kimi K2
- INT4 inference narrowing compute gaps
- Open-source matching proprietary performance
- Compression architectures equalizing capability
Benchmarks converge → pricing collapses → margins evaporate (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
Pricing Reality
Kimi K2: $0.15
GPT-5: $10
A 10× price gap for nearly overlapping capability.
This is the definition of commoditization.
▲ PRESSURE (FROM BELOW)
Infrastructure squeezes the model layer:
- GPU scarcity raises model training costs
- Custom silicon (TPU/Trainium) forces dependency
- Cloud platforms bundle models at super-low margin
- Fine-tuning becomes the only viable differentiator
The floor pushes upward.
The ceiling pushes downward.
Models get trapped.
3. INFRASTRUCTURE (THE FLOOR)
Table stakes to compete → Floor keeps rising → Barriers compound
Infrastructure dictates:
And unlike models, infrastructure does not commoditize (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
Why Floor Wins
- Capital requirements eliminate most competitors
- Technical advantages compound over years
- Training lock-in creates moat
- Supply-chain control becomes geopolitical leverage
Example: TPU offers 30–40 percent cost advantage for Google.
Infrastructure Examples
- NVIDIA: $57B Q3 revenue
- Trainium: 1M chips
- TPU: 2027 external customers
- Stargate: $500B buildout
- xAI: 1M GPUs target
Infrastructure is becoming the global bottleneck.
VALUE CAPTURE DYNAMICS
Ceiling captures the margin → Floor captures the moat → Middle gets squeezed
Why Ceiling Wins
- Controls pricing
- Owns interface
- Owns payments
- Owns retention
- Switches models freely
Why Middle Loses
- Benchmarks converge
- Open-source closes gaps
- Differentiation collapses
- Cost-center economics
Why Floor Wins
- Massive CapEx
- Distributed compute advantages
- Silicon + datacenter vertically integrated
- Few competitors can enter
(as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new)
STRATEGIC POSITIONING
A company must choose a layer — or it will be crushed.
✔ CONTROLS BOTH LAYERS (WINNERS)
Google: TPU + Gemini + A2P
Apple: Silicon + Applications + Device moat
Amazon: Trainium + Bedrock + Commerce rails
These companies have maximum leverage.
⚡ CONTROLS ONE LAYER (STRONG BUT FRAGILE)
NVIDIA: Infrastructure dominance
Microsoft: Application distribution
OpenAI: Transitioning to infrastructure with $500B Stargate
These players hold power — but incomplete protection.
✘ STUCK IN THE MIDDLE (DOOMED)
- Pure-play model companies
- API-only providers
- Undifferentiated fine-tuning shops
These firms face two-sided pressure and cannot keep margin.
🚀 ESCAPE ROUTES
If you are stuck in the middle, you must escape upward or downward.
Upward (Ceiling)
- Build agentic verticals (Harvey, Hippocratic)
- Own customer workflow
- Create distribution leverage
Downward (Floor)
- Acquire silicon access
- Build specialized infrastructure
- Secure compute sovereignty
Move to the floor or ceiling — fast (as per analysis by the Business Engineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
THE BOTTOM LINE
The sandwich is the framework: infrastructure and applications dominate; models are the commodity layer.
The AI economy rewards:
- control of compute (floor)
- control of customer (ceiling)
The middle layer compresses into a low-margin utility.
This is the core power structure of the AI era — and it will define the next decade of winners and losers (as per analysis by the BusinessEngineer on https://businessengineer.ai/p/this-week-in-business-ai-the-new).
More Resources
