The AI stack is often described as a hierarchy, but economically it behaves more like a curve. Specifically, a smile curve: value pools concentrate at the ends (equipment and applications) and at the bottleneck (HBM), while the middle layers operate with thinner margins. Once you understand this curve — and the physics behind it — you see the underlying logic of where capital flows and why certain layers accumulate disproportionate power.

Crucially, as explained in The AI Memory Chokepoint (https://businessengineer.ai/p/the-ai-memory-chokepoint), memory bandwidth has become the limiting factor of AI performance. This fundamentally reshapes where value is captured along the chain.

1. The Value Chain — From Sand to Software

The chain moves from commodity inputs to strategic bottlenecks:

• Raw Materials (~5% margin)
  Silicon, copper, rare earths — essential but undifferentiated.
• Equipment (~45% margin)
  ASML’s EUV monopoly, Applied Materials, Lam — the highest-margin part of the manufacturing upstream.
• HBM Manufacturing (~8–10% margin today, rising)
  SK Hynix, Samsung, Micron — the new chokepoint of AI scaling.
  HBM is ~$30–50/GB and is 50–60% of GPU BOM.
• GPU/Accelerators (~65% margin)
  NVIDIA, AMD, Google TPU — the platform layer, capturing massive value through ecosystem lock-in.
• Cloud Infrastructure (~30% margin)
  AWS, Azure, GCP — massive capex, modest margins, high switching costs.
• AI Applications (~70%+ potential margin)
  Software that scales indefinitely. But can’t exist without hardware.

This chain makes a fundamental point explicit:

AI software’s upside exists only because hardware provides the floor.
And that floor is constrained by HBM.

2. Where Value Accumulates — Understanding the Smile Curve

Plot the gross margins of each layer and you get the “smile”:

• Left end: Equipment has huge pricing power due to monopoly conditions.
• Middle: GPU vendors capture enormous platform value.
• Right end: Applications scale economically.

But the most interesting point is not the ends — it’s the inflection point in the middle:

HBM captures outsize value not because of margins, but because of scarcity.

Value pools are not just about margin — they are about control of constraints.

And HBM is the constraint.
Every model, every GPU, every inference pass routes through this layer.

3. Key Insights — Who Actually Captures the Most Value

1. HBM = Chokepoint Value

The pricing power doesn’t come from brand or product differentiation; it comes from:

• scarce supply
• critical function
• zero substitutes

This is why HBM commands 50–60% of GPU cost even though it is not the main die.

As detailed in The AI Memory Chokepoint, performance now scales with memory bandwidth, not compute — flipping the entire economic hierarchy of the AI stack.

2. NVIDIA’s Platform Value Capture

NVIDIA doesn’t just sell hardware; it sells an ecosystem:

• CUDA
• cuDNN
• TensorRT
• Software libraries
• Enterprise tooling
• Model optimization stacks

This produces:

• 65%+ gross margins
• $3T+ market cap
• The strongest economic flywheel in AI

NVIDIA captures more value than cloud providers because cloud depends on NVIDIA hardware to differentiate.

3. Equipment Moats Are Impenetrable

ASML’s EUV systems cost $200M+ each and carry extraordinary margins.
No EUV → no advanced nodes → no AI accelerators → no models.

Equipment vendors are the most secure layer in the entire chain — politically, technologically, and economically.

4. AI Apps Have Infinite Upside — But They Rest on the Chokepoint

Applications enjoy:

• zero marginal cost
• zero distribution friction
• extreme scalability

But without reliable, abundant memory supply, AI applications don’t exist.
Value creation in apps is infinite; value-dependence is absolute.

4. The Value Chain Insight — The Strategic Map

The entire chain follows three structural truths:

1. Bottlenecks Capture Value

Scarcity → constraint
Constraint → pricing power
Pricing power → value capture

HBM is now the most strategic bottleneck in AI — more than compute, more than cloud.

2. Platforms Beat Products

NVIDIA’s dominance isn’t about chips; it’s about switching costs.
Cloud providers are forced to buy NVIDIA to stay competitive.
Startups are forced to optimize for NVIDIA to survive.

This creates compound value capture.

3. Software Needs Hardware

Every high-margin AI application depends on an upstream bottleneck that is:

• geographically concentrated
• slow to scale
• controlled by three firms
• fundamentally constrained by physics

This is the structural vulnerability of the entire AI economy.

The Bottom Line

When you trace dollars through the semiconductor stack, a truth becomes obvious:

AI profitability is downstream of hardware scarcity.
And the scarcest resource is HBM.

This is why cloud capex is exploding, why hyperscalers are signing multi-year supply agreements, why governments are subsidizing memory fabs, and why The AI Memory Chokepoint (https://businessengineer.ai/p/the-ai-memory-chokepoint) is the single most important structural reality in the entire market.