ExpertsFMNvidia Part II: The Machine Learning Company (2006-2022)
Acquired · April 20, 2022
5 Selected Lessons:
case · 3 reasoning_pattern · 1 framework · 1
The 'stew of powers' analysis for assessing a platform's competitive moat
Source
Acquired — Nvidia Part II: The Machine Learning Company (2006-2022) (April 20, 2022) · at 1:48:21 ▸ · at 1:51:26 ▸
When this applies
when assessing whether any company's competitive position is durable
The takeaway
Evaluate a company's moat by explicitly checking it against each named category of competitive power, rather than relying on a vague sense of 'this company seems strong.'
Why it works
Enumerating against a fixed checklist of power types (counter-positioning, scale economies, switching costs, network economies, process power, branding, cornered resource) forces a systematic rather than anecdotal assessment of what is actually defensible.
One level up
A real moat is usually a combination ('stew') of several of these powers reinforcing each other, not any single one alone.
Where it stops holding
This stops working when the categories are applied superficially without evidence for each, producing a checklist illusion of rigor rather than real analysis.
Detail
procedure:
- List the standard power categories: counter-positioning, scale economies, switching costs, network economies, process power, branding, cornered resource.
- For each, ask whether and how strongly it applies to the company's current position.
- Identify where competitors (direct or disruptive) could erode specific powers (e.g., process power eroded once TSMC serves everyone).
rationale: NVIDIA's largest identified powers today are scale economies (amortizing CUDA's cost over 3M+ developers) and switching costs (developer lock-in to CUDA); process power has weakened since TSMC serves all comers.
In their words
To the extent that you have lock-in to developing on CUDA, which I think a lot of people really have lock-in on CUDA, then that's major switching costs
Recognizing and doubling down the instant a dormant bet gets validated
Source
Acquired — Nvidia Part II: The Machine Learning Company (2006-2022) (April 20, 2022) · at 47:51 ▸ · at 57:41 ▸
When this applies
when early, ambiguous signals suggest your long-dormant bet has suddenly found its real market
The takeaway
Maintain enough attention on adjacent, even unrelated fields that when your core technology unexpectedly fits a new need, you can recognize it and mobilize instantly rather than after competitors do.
Why it works
Because NVIDIA was paying close attention to all emerging compute-intensive workloads (not just its original target market), it could recognize the matrix-math overlap between graphics and deep learning the moment it appeared and move its whole organization behind it quickly.
Where it stops holding
This stops working when an organization is too narrowly focused on its original target market to notice or credit an adjacent breakthrough.
Detail
approach: Watch broadly across fields for where your core technical capability (e.g., massively parallel matrix math) might unexpectedly be the missing piece, and move fast once you see it.
signals:
- an unrelated field's breakthrough running on your infrastructure
- researchers independently gravitating to your platform
- your own engineers' insight that two 'different' workloads reduce to the same math
rationale: Speed of recognition and mobilization converts a lucky convergence into durable first-mover advantage (e.g., CUDNN library, enterprise AI ecosystem).
In their words
And so once NVIDIA saw that, it was basically instant. The whole company just latched onto it.
CUDA: betting the company on a market that didn't exist yet
Source
Acquired — Nvidia Part II: The Machine Learning Company (2006-2022) (April 20, 2022) · at 13:09 ▸ · at 16:55 ▸ · at 19:45 ▸ · at 23:03 ▸
When this applies
when leadership must decide whether to make a company-defining bet on a platform for a market that is currently small or nonexistent
The takeaway
Commit resources to build the enabling infrastructure for a plausible future market before you can prove it exists, because 'if you don't build it, they can't come.'
Why it works
By building the full developer stack (compiler, SDKs, libraries, evangelism) exclusive to your hardware, you create durable differentiation and switching costs that a commodity hardware business cannot get — but only once/if a market materializes that needs it.
One level up
Durable competitive moats often require accepting years of unprofitable, unproven investment that looks irrational in the moment.
Where it stops holding
This stops working when the anticipated market never materializes, leaving a stranded multi-year, multi-hundred-employee investment with no return — a real risk the hosts note could have happened.
Detail
situation: NVIDIA dominated gaming GPUs but Jensen wanted durable differentiation beyond commoditized hardware; a researcher's anecdote suggested GPUs could serve scientific computing.
decision: Commit the company to building CUDA, a full programming language/platform/library stack, starting 2006, despite the addressable market being estimated at only a couple billion dollars and six-plus years from usability.
outcome: CUDA took over a decade to find its true market (deep learning via AlexNet in 2012) and became the foundation of NVIDIA's current platform dominance and highest valuation.
lesson: Jensen's own phrase: 'if you don't build it, they can't come' — build infrastructure ahead of unproven demand if you believe durable differentiation requires owning the platform.
In their words
The phrase he uses is, if you don't build it, they can't come. So it's not even like, yeah, I'm pretty sure if we build it, they will come. It's one step removed from that. It's like, well, if we don't build it, they can't even possibly come.
AlexNet: an unplanned breakthrough validates a decade-long bet
Source
Acquired — Nvidia Part II: The Machine Learning Company (2006-2022) (April 20, 2022) · at 42:53 ▸ · at 45:24 ▸ · at 47:51 ▸ · at 57:41 ▸
When this applies
when a long-shot infrastructure investment is vindicated by an external breakthrough nobody could have predicted
The takeaway
A platform built for one speculative purpose can become indispensable when an unrelated field independently discovers it needs exactly that infrastructure.
Why it works
Deep learning had existed for decades but was computationally infeasible on serial CPU architecture; NVIDIA's embarrassingly-parallel GPU/CUDA architecture happened to be exactly the missing ingredient, turning an unrelated, unproven investment into the foundational technology for an entire industry overnight.
One level up
Long-horizon infrastructure bets can pay off through convergence with external breakthroughs you did not cause and could not have predicted — but you must have already built the infrastructure to capture it.
Where it stops holding
This stops working when no external breakthrough ever arrives to validate the infrastructure, leaving the investment stranded — exactly the risk NVIDIA faced for 1986-2011.
Detail
situation: Decades-old deep learning algorithms were computationally infeasible on CPUs; NVIDIA had spent years building CUDA and massively parallel GPU architecture for an unrelated scientific-computing market.
decision: A University of Toronto team (Krizhevsky, Sutskever, Hinton) implemented deep learning in CUDA on NVIDIA GPUs for the 2012 ImageNet competition.
outcome: AlexNet cut the error rate from ~25% to 15%, a landmark breakthrough that is credited as the 'Big Bang moment' for modern AI, and instantly made NVIDIA's unrelated 15-year CUDA investment the central infrastructure of the field.
lesson: Jensen did not plan or foresee AlexNet; the payoff came from recognizing and immediately doubling down once the convergence became visible.
In their words
This was the Bing Bang moment for artificial intelligence, and NVIDIA and CUDA right there.
Public market slow to recognize the 'picks and shovels' play
Source
Acquired — Nvidia Part II: The Machine Learning Company (2006-2022) (April 20, 2022) · at 52:58 ▸ · at 55:14 ▸ · at 1:51:26 ▸
When this applies
when an infrastructure/supplier company quietly underlies an entire emerging industry but the market hasn't yet connected the dots
The takeaway
When a new industry wave emerges and a single infrastructure provider underlies nearly every participant, betting on that provider carries less company-specific risk than betting on which application wins, and is often underpriced during the early hype phase.
Why it works
In an emerging ecosystem, it's easier and safer to identify the common infrastructure provider that every player depends on than to pick which individual application-layer company will win.
One level up
Position for certainty (the layer everyone needs) over prediction (who wins the application layer).
Where it stops holding
This stops working when the infrastructure layer itself becomes commoditized or is replaced by in-house/alternative solutions, or when the application wave itself fails to materialize.
Detail
situation: By 2016 nearly every deep-learning startup Andreessen Horowitz saw was built on NVIDIA's platform, yet NVIDIA traded under $5/share for years through 2012-2016.
decision: Andreessen noted internally his firm would hypothetically put all capital into NVIDIA stock if they ran it like a hedge fund; the hosts admit they, as venture capitalists who should have known, also failed to act.
outcome: NVIDIA stock rose from under $5 to over $300 at its high within several years.
lesson: Investors who should have recognized the infrastructure dependency missed it for years even with the information in plain sight — recognizing the picks-and-shovels layer early is valuable precisely because it's often overlooked.
In their words
We've been investing in a lot of companies applying deep learning to many areas. And every single one effectively comes in building on NVIDIA's platforms