ExpertsFMNvidia Part I: The GPU Company (1993-2006)
Acquired · March 28, 2022
5 Selected Lessons:
case · 2 reasoning_pattern · 2 priority · 1
Grading a company's historical performance against both its direct rivals and the full value chain
Source
Acquired — Nvidia Part I: The GPU Company (1993-2006) (March 28, 2022) · at 1:48:02 ▸ · at 1:51:17 ▸
When this applies
when evaluating how well any company performed in a historical period, not just against named competitors but against everyone in its value chain who might have captured more value
The takeaway
Grade historical business performance against the full value chain's alternative uses of capital and risk, not merely against the nearest competitors.
Why it works
A company's performance should be judged not only by whether it beat its direct competitors, but by comparing its value capture to every other actor in the value chain (suppliers, platform owners, customers) who might have captured more value with less risk.
One level up
Fair retrospective evaluation requires an explicit counterfactual comparison set, not just a narrow 'did they win their market' question.
Where it stops holding
This stops working when the value chain comparison is too speculative to support (e.g., insufficient data on what alternative value capture would have looked like).
Detail
hierarchy:
- First: did the company survive and win against its direct competitors? (NVIDIA: yes, beat ~90 rivals)
- Second: did another player in the value chain capture more value with less risk? (Microsoft arguably did)
- Final grade accounts for both: strong but not perfect
rationale: I think I give NVIDIA for this period of time an A because they're basically the only company that survived... But it's not an A-plus because Microsoft.
In their words
I think I give NVIDIA for this period of time an A because they're basically the only company that survived. ATI did for sure, of course, but in a very different fashion. And they created this whole industry, almost inarguably, created and shepherded this whole industry. But it's not an A-plus because Microsoft.
A venture bet can be 'intellectually wrong' about the market yet still be one of the best returns ever, because of the founder
Source
Acquired — Nvidia Part I: The GPU Company (1993-2006) (March 28, 2022) · at 1:39:12 ▸
When this applies
when evaluating, in hindsight, whether backing a specific team in a crowded, undifferentiated proliferation-phase market was the 'right' call
The takeaway
In markets with many undifferentiated entrants, back the founder's demonstrated capacity for reinvention and intellectual honesty rather than the specific market thesis, since the thesis is likely to be wrong for everyone but the survivor's execution quality determines who wins.
Why it works
In a proliferation-phase market with no visible differentiation, the deciding variable for which company survives consolidation is often the founder's capacity for repeated self-reinvention under existential pressure, a quality that's hard to assess at the time of initial investment but determines the outcome more than the original market thesis.
One level up
Outcome quality in high-variance, crowded markets is often decoupled from the accuracy of the original thesis and instead tracks the resilience/adaptability of the specific team chosen.
Where it stops holding
This stops working when the founder's reinvention ability can't actually be assessed ex ante, or when the market consolidates on factors (capital, distribution) entirely outside founder control.
Detail
approach: Judge the historical bet by asking whether the investors backed a reasonable-enough market with a founder who had the capacity to repeatedly reinvent the business.
signals:
- Market thesis later proven technically wrong in its specifics (quadrilaterals vs. triangles)
- Company nonetheless survives and dominates
- Founder shows repeated willingness to scrap sunk strategic bets and re-bet the company
rationale: They were wrong intellectually, and yet they were right... frankly, of Jensen.
In their words
The magic was that Jensen really figured it out early that they were in a business that was totally at the mercy of Moore's Law. And so like in having that initial realization as early as they did
A near-bankruptcy crisis can seed the exact capability that defines the next decade's opportunity
Source
Acquired — Nvidia Part I: The GPU Company (1993-2006) (March 28, 2022) · at 1:35:49 ▸ · at 1:57:01 ▸
When this applies
when a company is forced into an emergency capability under existential pressure that seems narrowly tactical at the time
The takeaway
Pay attention to capabilities built under crisis pressure — they may be more strategically valuable long-term than the crisis that forced them.
Why it works
A capability built purely to survive a crisis (software simulation of hardware) can, once institutionalized and generalized, become the foundational competency for an entirely new and much larger market years later.
One level up
The true strategic value of a capability is often not visible until long after the emergency that produced it has passed.
Where it stops holding
This stops working when the crisis-born capability is too narrowly tied to the specific emergency to generalize to any future market.
Detail
approach: Trace how a tactical crisis response became a core company competency years later.
signals:
- Capability was adopted only because of existential necessity
- Capability later generalizes far beyond its original narrow use
- Company continues investing in and expanding the capability after the crisis passes
rationale: It became very clear to the company very early on the benefits of being able to simulate something rather than having to do it in the real world.
In their words
The reason I'm talking about it in this episode is that DNA comes from the fact that in order to survive, when they had nine months left, the way that they saved themselves was with simulation.
Shipping a chip without a physical prototype, using a slow software emulator, to beat a cash-out deadline
Source
Acquired — Nvidia Part I: The GPU Company (1993-2006) (March 28, 2022) · at 43:22 ▸ · at 46:54 ▸ · at 49:23 ▸ · at 1:35:49 ▸
When this applies
when your standard process timeline is longer than your actual survival runway and you must compress it radically
The takeaway
When physical iteration cycles are too slow to survive, substitute simulation for physical testing even at steep per-unit speed cost, because it collapses the overall calendar timeline.
Why it works
Simulating the product in software, even at a huge speed penalty (one frame per 30 seconds), let the team find bugs and validate design correctness without waiting for slow physical fab iteration cycles, trading painstaking manual verification time for calendar time.
One level up
Simulation converts a hard physical-cycle-time constraint into a labor/patience constraint, which is often the only constraint you can actually compress under pressure.
Where it stops holding
This stops working when no adequate simulation/emulation technology exists for the domain, or when the gap between simulated and real-world behavior is too large to trust the result.
Detail
situation: NVIDIA needed to design and ship a new chip in nine months instead of the normal two years, with no time for a prototype-and-iterate cycle.
decision: They spent roughly a third of their remaining cash ($1M) on slow chip-emulation technology from an unproven startup, manually verifying output frame-by-frame at one frame per 30 seconds, then taped out directly to mass production without a physical prototype.
outcome: The Riva 128 shipped on time, sold 1 million units in four months, and the emulation vendor (NVIDIA's only customer) later went out of business.
lesson: Betting the company on an unproven tool can be rational when it's the only path that fits the actual survival timeline.
In their words
They had to spend a million dollars just to get the emulator hardware and software to do this. Which I think they had generated some revenue, but it was still like a third of the cash that they had in the entire bank account.
NVIDIA out-cadenced the industry by compressing the design-ship cycle to six months
Source
Acquired — Nvidia Part I: The GPU Company (1993-2006) (March 28, 2022) · at 49:23 ▸ · at 53:33 ▸ · at 1:00:50 ▸
When this applies
when a company discovers, under existential pressure, a repeatable process that lets it outpace the rate of change competitors assume is fixed
The takeaway
A compressed cycle time born of crisis can become a durable competitive cadence if you institutionalize the process rather than reverting to 'normal' once the crisis passes.
Why it works
Once a crisis forced NVIDIA to compress its cycle time once, the company turned the one-off emergency fix into a standing operating cadence, letting it deliver generational performance gains at roughly 3-4x the frequency the rest of the industry assumed was the physical limit (tied to Moore's Law's 18-24 month doubling).
One level up
Constraints that force extreme speed sometimes reveal a faster permanent operating rhythm that was never actually necessary to run at the old pace.
Where it stops holding
This stops working when the compressed cycle was a one-time emergency measure that sacrificed quality/sustainability and can't be repeated without burning out the team or the balance sheet.
Detail
situation: Forced by near-bankruptcy to ship the Riva 128 in nine months instead of two years.
decision: NVIDIA kept the compressed, simulation-driven design process as its ongoing operating model.
outcome: NVIDIA shipped new generations doubling performance every six months, versus competitors' 18-24 month cycles, effectively outpacing the price-performance rate implied by Moore's Law.
lesson: An emergency process improvement can become a lasting structural advantage if institutionalized.
In their words
NVIDIA is now on a cycle starting in 1997, 1998, where they are doubling the performance that they are delivering at a given price point to the market every six months.