Acquired · Pattern · P1
Vertical integration and value chain
Control over upstream or downstream steps that improves quality, timing, or margin.
Control over upstream or downstream steps that improves quality, timing, or margin.
The mechanism
Owning the steps above and below your core stops margin from leaking and turns each layer into a reason the others are hard to leave. NVIDIA could have sold chips and stopped; instead it stacked silicon, the Mellanox interconnect it bought in 2019, CUDA, the libraries, and the cloud on top of each other, so a customer who adopts one layer inherits switching costs in all the rest, and a DGX box leaves the building at $150,000 to $300,000. Costco ran the same logic into its own shelves with Kirkland Signature, a private label that now does more than $50 billion a year, larger than Nike, capturing the margin a national brand would have taken and deepening the member trust at the same time. Ferrari keeps engines, chassis, and most of the supply chain inside Maranello so a design change does not have to wait on a supplier and the racing team's learnings cross the street into road cars within weeks.
The tension
Integration buys control by spending flexibility, and the bill comes due in capital and in the parts of the chain you still do not own. Every layer you bring in-house is a fixed cost and a commitment: Ferrari makes its cars by hand in one town, which is the provenance and also a hard ceiling on how fast it can ever grow. And no stack is ever really complete, so the leverage leaks at the link you skipped. NVIDIA designs the whole system but does not fab it; the actual silicon is made by TSMC on machines only ASML knows how to build, which means the company with the most admired full-stack moat in technology still depends on two suppliers it cannot replace, and the link you do not own is exactly where a competitor or a government can apply pressure.
Vertical integration in Maranello1947-present
Engines, chassis, and most of the supply chain stay in-house at Maranello. Quality control + brand provenance + employment of the same families across generations.
- Ben:Ferrari makes all of their cars mostly by hand, inefficiently and one-off in a single town, Maranello, Italy.[Acquired Ferrari, ch. Intro]
- Ben:Ferrari can operate on a much shorter timescale between design and actually getting a car out to customers since there's less overhead in getting started. They could change their engine casting spec without involving a supplier.[Acquired Ferrari, ch. Post-IPO Ferrari]
- Ben:They can also get learnings from their F1 team quickly into cars since it's right across the street.[Acquired Ferrari, ch. Post-IPO Ferrari]
Kirkland Signature as private label flywheel1995-present
Costco's private label outsells every national brand competitor in category after category (vitamins, batteries, coffee, water, vodka). Margin captured directly, supplier leverage extended, member trust deepened — Kirkland is the trust collateral made into product.
- Ben:Kirkland Signature does more revenue alone than all of Nike. $52 billion a year, doesn't even include the Kirkland gas.[Acquired Costco, ch. Intro]
- David:I think Kirkland Signature as a unified brand might be the largest brand in the world by revenue.[Acquired Costco, ch. Intro]
Centralised infrastructure, independent brand identity1990s-present
LVMH centralises real estate (Place Vendôme, Champs-Élysées, Bond Street), media buying, supply chain, leather tanneries, watch ateliers. But each maison runs its own creative direction. The shared back-office is the cost moat; the independent fronts are the brand moats.
Vertical integration of hardware + software + IP1985-present
Nintendo designs the console, owns the silicon path (with NVIDIA on Switch), publishes the killer-app software, owns the IP. Each layer reinforces the others. Competitors can match any one layer; matching all four simultaneously is the moat.
Full-stack moat: silicon + system + software2006-present
Don't just sell chips. Sell chips + boards + interconnect (Mellanox) + reference designs + CUDA + libraries (cuDNN, TensorRT) + frameworks (PyTorch native support) + cloud (DGX Cloud). Each layer adds switching costs to every other layer.
- David:These three things that NVIDIA has been building, the dedicated Hopper data center GPU architecture, the Grace CPU platform, the Mellanox-powered networking stack, they now have a full suite solution for generative AI data centers.[Acquired NVIDIA Part III, ch. Full-stack solution]
- David:NVIDIA sells these DGX systems for $150,000–$300,000 a box. With all these three new legs of the stool — Hopper, Grace, and Mellanox — these systems are just getting way more integrated, way more proprietary, and way better.[Acquired NVIDIA Part III, ch. DGX economics]
- TSMC2021
- Rolex2026
- Costco2026
- Lockheed Martin2023
- CAA (with Michael Ovitz)2021
- DoorDash2020
- Convoy (with CEO Dan Lewis)2019
- Episode 45: HTC, Google and the Future of Mobile2017
- Episode 42: Opsware (with special guest Michel Feaster)2017
- Acquired Episode 39: Whole Foods Market2017
- Trader Joe’s2025
- Hermès2024
Keyword density across title, chapters, description, and source list.