Skip to content

Amazon and Tesla: Modular Evolution in Action

Note: This is the extended case study. See Chapter 12, Section 5 for the summary.

"But we're not Amazon or Tesla."

Half right. You don't have their scale. But the architectural patterns that enable their evolution aren't exclusive to trillion-dollar companies. They're design decisions anyone can make.

Amazon: Addition Without Replacement

Amazon's AI services didn't evolve through rewrites. They evolved through addition.

2017: SageMaker launched with three modular components: notebooks, built-in algorithms, one-click deployment. Customers could use only build and train capabilities while hosting elsewhere1. Modularity baked in before generative AI existed.

2023: Bedrock arrived as a parallel service, not a replacement. SageMaker for deep customization. Bedrock for serverless foundation model access2. AWS added in parallel, didn't deprecate.

2024: Amazon Q and SageMaker Unified Studio created integration points, not forced migrations. JumpStart models register with Bedrock while maintaining existing infrastructure3.

The pattern: independent services solving distinct problems. Customers adopt what they need. Nobody forced onto migration treadmills.

flowchart LR
    subgraph Y2017["2017"]
        SM["SageMaker<br/>ML Platform"]
    end

    subgraph Y2023["2023"]
        BR["Bedrock<br/>Foundation Models"]
    end

    subgraph Y2024["2024"]
        Q["Amazon Q<br/>Unified Studio"]
    end

    SM --> |"Not replaced"| BR
    BR --> |"Integrated"| Q
    SM --> |"Still works"| Q

    INTUIT["Intuit:<br/>Uses Both"] -.-> SM & BR
    SALES["Salesforce:<br/>Uses Both"] -.-> SM & BR

    style SM fill:#1e6fa5,stroke:#155a85
    style BR fill:#7345b0,stroke:#5b3590
    style Q fill:#1a8a52,stroke:#14693e

The Integration That Didn't Break

Intuit's CTO at re:Invent 2023: they moved to AWS, built AI/ML with SageMaker, then added Bedrock in the generative AI era. Extended, not replaced4. Salesforce achieved 30% faster deployments and 40% cost savings while preserving backward compatibility5.

When you design services with well-defined interfaces, customers evolve on their own timeline.

Tesla: Every Vehicle Becomes a Training Machine

Tesla proves modular evolution works for physical products.

Their firmware separates battery management, autopilot, infotainment, thermal systems into independent modules—each updated independently. Traditional automakers use single-block firmware where one issue requires reflashing everything6.

The FSD stack operates as a separate layer with its own versioning. Hardware abstraction distinguishes HW3 from HW4 vehicles, with automatic binary selection7. Customers with 2019 vehicles receive updates alongside 2024 vehicles.

Shadow Mode: The Data Flywheel

FSD runs silently in background on every Tesla even when not enabled. The system makes hypothetical decisions and compares them to driver choices. Disagreements become training data8.

Two million vehicles become a distributed data collection network. The entire fleet feeds the training pipeline.

flowchart TB
    subgraph FLEET["2M+ Tesla Fleet"]
        direction LR
        V1["Vehicle 1"]
        V2["Vehicle 2"]
        VN["Vehicle N"]
    end

    subgraph SHADOW["Shadow Mode (Always Running)"]
        direction TB
        S1["FSD Makes<br/>Hypothetical Decision"]
        S2["Compare to<br/>Driver's Actual Choice"]
        S3{"Disagreement?"}
    end

    subgraph TRAINING["Training Pipeline"]
        direction TB
        T1["Edge Cases<br/>Collected"]
        T2["Model Retrained"]
        T3["OTA Update"]
    end

    FLEET --> SHADOW
    S1 --> S2 --> S3
    S3 -->|"Yes"| T1
    S3 -->|"No"| DISCARD["Discard"]
    T1 --> T2 --> T3
    T3 -->|"Improved Model"| FLEET

    style FLEET fill:#1e6fa5,stroke:#155a85
    style TRAINING fill:#1a8a52,stroke:#14693e

FSD v11 to v12 shows what this enables. V11: neural networks for perception, 300,000 lines of C++ for decision-making. V12: end-to-end neural networks, those 300,000 lines reduced to 2,000-3,0009.

Musk in 2023: "There's no line of code that says there is a roundabout... There are over 300,000 lines of C++ in version 11, and there's basically none of that in version 12"10. A fundamental transformation delivered via OTA update.

What Went Wrong (And How They Recovered)

Modular architecture doesn't prevent failures. It enables recovery.

FSD Beta v10.3 pushed Saturday night October 2021. By Sunday, vehicles swerved into adjacent lanes while FSD wasn't even enabled. Tesla's rollback accidentally removed beta access entirely11. Modular architecture allowed rapid (if botched) recovery.

Version 13: 489 miles between disengagement versus 183 for v12.5—a 2.7x improvement12. Each iteration compounds because the architecture permits rapid iteration.

The Cost Structure Reality

Tesla's approach: $10 billion in AI training infrastructure 2023-2025. Over $1 billion on custom Dojo supercomputers. $2 billion on H100 clusters13.

You don't need $10 billion. But evolutionary architecture requires upfront modularity investment. Payoff: capability improvements deploy in weeks, not quarters.

Andrej Karpathy: "Progressing from prototypes that work 90% of the time to 99.999% requires exponentially more effort"14. Shadow mode and modular updates close that gap incrementally.

Patterns That Transfer

Design services that solve distinct problems. SageMaker for customization. Bedrock for simplicity. Each usable independently.

Create integration points, not forced migrations. Salesforce uses both because AWS built the connection. Your services should work the same way.

Separate platform from capability. Tesla's vehicle platform stays stable while FSD evolves. Your business logic should stay stable while AI evolves.

Turn usage into improvement. Shadow mode makes every driver a training source. Your usage should feed your improvement cycle.

Enable rollback and dual code paths. Recovery options when things break. Version management when hardware varies.

The question isn't whether you have their resources. It's whether you're making the same architectural choices they made before they had those advantages.

References

Chapter Overview | Next: HubSpot and Gong: Platform and Flywheel Examples →

  1. AWS. Introducing Amazon SageMaker ↩

  2. Leanware. Amazon SageMaker vs Amazon Bedrock: What's the Difference ↩

  3. AWS Documentation. JumpStart Foundation Models Registration with Bedrock ↩

  4. YouTube. Intuit at re:Invent 2023 ↩

  5. AWS Machine Learning Blog. How Amazon Bedrock Custom Model Import Streamlined LLM Deployment for Salesforce ↩

  6. LinkedIn. Tesla Software Architecture ↩

  7. Shop4Tesla. Tesla FSD v12.6 Update 2025 ↩

  8. Not a Tesla App. Tesla's FSD Shadow Mode ↩

  9. Fred Pope. Tesla FSD 12: Machine Learning ↩

  10. Fred Pope. Tesla FSD 12: Machine Learning ↩

  11. Dev.to. Tesla Pushes Out Dangerous Software Update ↩

  12. Creative Strategies. Tesla AI Autonomy FSD v13 Update ↩

  13. Tom's Hardware. Elon Musk Spent Roughly $10 Billion on AI Training Hardware in 2024 ↩

  14. Global Advisors. Quote: Andrej Karpathy ↩