Apr 20, 2022 · 2h 15m · acquired
Nvidia: The Machine Learning Company (2006-2022) · Acquired
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this deep-dive episode of Acquired, hosts Ben Gilbert and David Rosenthal trace NVIDIA's journey from a 1990s 3D gaming graphics pioneer to the dominant full-stack platform powering modern artificial intelligence. They analyze Jensen Huang's visionary gamble on CUDA, the deep learning revolution, strategic data center expansions, and NVIDIA's durable economic moats.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Ben and David hold 98.6% of the talking time here. How this is scored →
Ben and David as informed peer 6.5 Guest teaching 0.5 Guest disagreement 0.1 Ben and David pushing back 0.1
speaking balance: gold is Ben and David, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 33:15 Challenging the Wintel open-ecosystem comparison
Ben gently challenges David's characterization of Wintel as an open hardware standard by pointing out that it was restricted to the IBM-compatible architecture rather than truly open hardware.
Hardest push from Ben and David ▶ 33:15 Pushing back on the Apple vs. Microsoft analogyBen pushes back on David's assertion that Windows software ran on any hardware by noting it was limited strictly to Intel/IBM architecture and incompatible with PowerPC or Apple machines.
Biggest teaching moment ▶ 38:40 The Adreno anagram origin revealedBen educates David on the origin of Qualcomm's Adreno GPU brand name, revealing that it was created as an exact anagram of ATI's Radeon after Qualcomm acquired the division.
Ben and David hold their own ▶ 1:31:20 Deep technical breakdown of DLSS inferencingBen demonstrates deep domain expertise by clearly explaining the exact machine learning mechanisms behind DLSS temporal and spatial pixel inferencing in real-time rendering.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Ben and David as informed peer | Guest teaching | Guest disagreement | Ben and David pushing back | Why |
|---|---|---|---|---|---|---|
| Acquired Community and LP Show Update | 6 | 1 | 0 | 0 | David and Ben walk through the transition from graphics pipeline to proprietary driver writing, highlighting the six-month release cadence that outpaced Intel. The hosts operate in complete collaborative alignment with zero friction. | |
| The Genesis and Audacious Bet of CUDA (2006) | 7 | 1 | 0 | 0 | David details Jensen Huang's massive gamble starting CUDA in 2006 to address general-purpose computing before a clear commercial market existed. Ben adds depth by comparing CUDA's full abstraction stack to Apple's developer frameworks. | |
| The AMD Rivalry, Earnings Shock, and Financial Crisis Crash | 6 | 0 | 0 | 0 | The hosts recount AMD's acquisition of ATI, NVIDIA's subsequent 2008 earnings miss, and the catastrophic 80% stock crash during the financial crisis. Both hosts emphasize the rarity of Jensen doubling down rather than pursuing a fire sale. | |
| Deconstructing CUDA: Architecture, Strategy, and the Closed Platform | 8 | 1 | 1 | 1 | David and Ben explore CUDA's proprietary full-stack architecture, contrasting its closed ecosystem against Wintel and Android. Ben politely refines David's analogy about open vs. closed models by pointing out the historical nuances of the Microsoft-Intel alliance. | |
| Mobile Ambitions: Tegra, Icera, and Strategic Pivots | 7 | 2 | 0 | 0 | David explores the Tegra mobile chip era and the Icera acquisition, noting how repurposed chips ended up in the Tesla Model S and Nintendo Switch. Ben brings in trivia regarding Qualcomm's Adreno GPU anagram origin from Radeon. | |
| The 2012 Deep Learning Big Bang: ImageNet, AlexNet, and cuDNN | 8 | 1 | 0 | 0 | David narrates the landmark 2012 ImageNet breakthrough with AlexNet running on CUDA GPUs, calling it AI's big bang. Ben complements this with technical context on Brian Catanzaro and Andrew Ng's work establishing cuDNN. | |
| Commercializing AI: Advertising Giants and Investor Skepticism | 7 | 0 | 0 | 0 | The hosts analyze how algorithmic recommendations and ad targeting provided the initial multi-trillion dollar monetization engine for deep learning. They review Wall Street's delayed realization of NVIDIA's emerging AI monopoly. | |
| Crypto Mining Waves, GPU Volatility, and Supply Dynamics | 7 | 1 | 0 | 0 | Ben explains the mathematical mechanics behind why parallel GPU architecture excelled at cryptocurrency mining and why crypto crashes caused secondary market supply shocks. David connects this volatility to ongoing Wall Street skepticism. | |
| Data Center Transformation and the Mellanox Acquisition | 8 | 1 | 0 | 0 | The hosts review the rapid rise of data center revenue and the strategic acquisition of Mellanox for high-speed interconnects and DPUs. Ben highlights the transformation from selling single PCIe cards to high-margin bundled data center solutions. | |
| Sponsor Spotlight: Vouch and Employment Practices Liability | 1 | 0 | 0 | 0 | Mid-roll sponsorship segment discussing Employment Practices Liability insurance with sponsor Vouch. Scores reflect standard commercial read dynamics. | |
| The Attempted ARM Acquisition and Strategic CPU Ambitions | 7 | 0 | 0 | 0 | David and Ben discuss the strategic motivations behind NVIDIA's attempted $40B acquisition of ARM and its eventual collapse under antitrust scrutiny. Ben breaks down the distinct business models between ARM instruction set licensing and core designs. | |
| GTC 2022, Trillion-Dollar TAM, and Software Monetization | 8 | 0 | 0 | 0 | The hosts examine GTC 2022 announcements, including the Hopper GPU and software licensing strategy, while critically assessing Jensen's trillion-dollar TAM claims. Ben highlights the astronomical price-to-sales multiple compared to Big Tech peers. | |
| Gaming Engineering: Real-Time Ray Tracing and DLSS | 8 | 0 | 0 | 0 | Ben explains the technical mechanics of real-time ray tracing and DLSS, illustrating how AI neural networks upscale gaming resolutions in real time. They also review NVIDIA's LHR software limits on crypto hash rates. | |
| Future Growth Engines: Automotive Drive and the Omniverse | 7 | 0 | 0 | 0 | The hosts evaluate the DRIVE automotive platform and the enterprise Omniverse simulation framework. Ben distinguishes Omniverse from consumer metaverse concepts, framing it as an industrial digital twin platform. | |
| Competitive Landscape: Startups, Hyperscalers, and Bear Cases | 8 | 0 | 0 | 0 | Ben and David debate bear cases, contrasting custom hyperscaler silicon like Google TPUs and radical wafer-scale architectures like Cerebras against NVIDIA's CUDA moat. They conclude that building a competing software ecosystem remains the ultimate hurdle. | |
| Hamilton Helmer's Seven Powers Applied to Modern NVIDIA | 8 | 0 | 0 | 0 | The hosts map Hamilton Helmer's 7 Powers to modern NVIDIA, identifying Scale Economies and Switching Costs derived from CUDA lock-in as the primary moats. Ben mentions recent Lapsus group leaks targeting driver proprietary code. | |
| Acquired Playbook: Mission Expansion, Fabless Economics, and Margins | 8 | 1 | 0 | 0 | Ben breaks down NVIDIA's capital efficiency, highlighting its 37% operating margin and modest $1B CapEx enabled by TSMC's fabless manufacturing model. David adds context on Morris Chang's role in pioneering the pure-play foundry model. | |
| Sponsor Spotlight: SoftBank Latin America Fund and VTEX | 1 | 0 | 0 | 0 | Sponsorship segment highlighting the SoftBank Latin America Fund and its investment in VTEX. Scores reflect an informational advertising read. | |
| Grading NVIDIA: High Expectations, Physical AI, and Valuation | 7 | 0 | 0 | 0 | David and Ben discuss the expectations-investing framework for grading NVIDIA stock, weighing the requirement for real-world physical AI expansion against potential pandemic pull-forward risks. | |
| Carve Outs: The Expanse's Memory's Legion and Sony RX100 | 3 | 0 | 0 | 0 | Casual outro segment where David recommends James S.A. Corey's 'Memory's Legion' short story collection and Ben shares his experience using the Sony RX100 point-and-shoot camera. |
Statements from this episode (40)
Training a single speech model requires more math operations than sand grains
“If you are training one single speech recognition machine learning model these days, one, just one model, the number of math operations like adds or multiplies to accomplish it It's actually greater than the number of grains of sand on the earth.”
Rosenthal: NVIDIA was first graphics card company to write in-house drivers
“All the other graphics cards companies at the time and most peripheral companies, they let the further downstream partners rate the drivers for what they were doing. NVIDIA was the first one that said, no, no, we want to control this. We want to make sure cons…”
Rosenthal: NVIDIA reached nearly a $20 billion market cap by mid-2007
“They go from sub a billion dollar market cap company after the tech crash up to five to six billion dollars kind of by 2004, 2005. Stock keeps going on a tear. By mid 2007, the stock reaches just under twenty billion dollar market cap.”
Gilbert: CUDA was not a useful platform for six-plus years after 2006
“Because while CUDA development began in 2006, That was not a useful usable platform for six plus years at NVIDIA.”
Gilbert: NVIDIA has 1,100 employees with 'CUDA' in their LinkedIn title
“I searched LinkedIn for people who work at NVIDIA today and have the word CUDA in their title. There are 1100 employees dedicated specifically to the CUDA platform.”
Rosenthal: NVIDIA market cap dropped 80% from $20B high in 2008
“So, you know, we said the high was around a twenty billion dollar market cap.
It drops 80%.
Eight zero.”
Gilbert: NVIDIA's investment in CUDA was an iPhone-sized bet
“Those were big bets relative to the company's size at the time, but this bet is like an iPhone-sized bet.”
Gilbert: NVIDIA's latest consumer GPU has over 10,000 cores
“Today there's over 10,000 cores on their most recent consumer graphics card.”
Rosenthal: NVIDIA has never charged a dollar for CUDA
“NVIDIA to this day, now this may be changing, we'll talk about this at the end of the episode, has never charged a dollar for CUDA”
Rosenthal: CUDA software cannot be deployed on non-NVIDIA chips
“You do any of this work, you cannot deploy it on anything but NVIDIA chips.”
Rosenthal: NVIDIA Tegra powered the original Tesla Model S infotainment
“It powered the original Tesla Model S touchscreen. So, like, before any of the autopilot autonomous driving stuff, they were the processor powering just the infotainment, the touchscreen infotainment in the Model S.”
Rosenthal: NVIDIA Tegra powers the Nintendo Switch
“The Tegra platform still to this day is the main processor of the Nintendo Switch.”
Rosenthal: Android hardware value chain yields no profits by Google's design
“My impression of the whole Android Value chain ecosystem is that there's no profits to be made anywhere, and Google keeps it that way on purpose.”
Rosenthal: Jensen Huang did not anticipate AlexNet's breakthrough
“Jensen did not plan Alex net or see it coming cause nobody saw Alex net coming.”
Rosenthal: AlexNet on CUDA was the 'big bang' moment for AI
“We cannot overstate the importance of this moment, not just for NVIDIA, but for, like, computer science, for technology, for business, for the world, for us staring at the screens of our phones all day every day. This was the big bang moment for artificial int…”
Gilbert: NVIDIA and Stanford matched Google Brain's 1,000 nodes with three
“And so, in 2013, Brian Catanzaro, who's a research scientist at NVIDIA, published a paper with some other researchers at Stanford, which included Andrew Ng, where they were able to take this unsupervised learning approach that had been done inside the Google B…”
Rosenthal: Google and Meta depend fundamentally on Nvidia
“This is Google and Facebook and a whole list of other companies on the consumer side, and they're all dependent on NVIDIA.”
Andreessen: Every deep learning startup in 2016 built on Nvidia
“Marc Andreessen gave an interview. He said, 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. It's like when people were all building on Windows in the nin…”
Rosenthal: Nvidia stock was under $5 per share across 2012-2015
“So literally NVIDIA stock in 20, like recent, like this is now known, 2012, 1314, 15. It doesn't trade above like five bucks a share. And NVIDIA today, as we record this, is I think about two 20 a share. The high in the past year has been well over 300.”
NVIDIA Took Until 2016 to Reclaim 2007's $20B Market Cap
“It's not until 2016 that Nvidia gets back to the twenty billion dollar market cap peak that they were in 2007 when they were just a gaming company.”
Rosenthal: 2018 Crypto Mining Bust Caused NVIDIA Revenue to Decline
“And then when the crypto winter hit in 2018 and the end of the ICO craze and all that, the mining rig demand It fell off and this had become so big for Nvidia that their revenue actually declined.”
Gilbert: NVIDIA data center revenue tripled to over $10.5B annually
“So then insanely over the last two years, it three X the data center, the segment three X it is now doing over 10 and a half billion a year in revenue. And it's basically the same size as the gaming segment.”
Rosenthal: NVIDIA barred consumer GeForce cards from data centers in 2018
“In 2018 they actually do change the terms of the user agreements of their consumer cards, of GeForce cards, that you cannot put them in data centers anymore.”
Rosenthal: NVIDIA Hopper will be TSMC's first 4nm chip
“I think Hopper is going to be the world's first four nanometer process chip using TSMC's new four nanometer.”
Gilbert: NVIDIA trades at 3x Apple's P/S and 2x Microsoft's
“Nvidia has literally three times the price to sales ratio of Apple, or price to revenue as Apple, and nearly two x Microsoft.”
Gilbert: NVIDIA's 66% gross margins prove their durable moat
“They now have a 66% gross margin, so that illustrates to me how seriously differentiated they are and how much of a moat they have versus competitors in order to price with that kind of margin.”
Gilbert: Unbundling software from hardware creates strategic conflict without segmentation
“Oftentimes when someone has hardware that is differentiated by software and services, and then they decide to start selling those software and services a la carte, it's a strategy conflict. It's your classic vertical versus horizontal problem, unless you are g…”
Gilbert: NVIDIA Grew Revenue 60% in Previous Fiscal Year
“They're growing at 60% a year. This is a thirty-year-old company that grew 60% in revenue last year.”
Gilbert: NVIDIA Artificially Restricted Consumer GPU Crypto Mining
“What they did was they artificially crippled the card to make it worse at crypto mining.”
Gilbert: NVIDIA Will Likely Shift Away From Board Partners to Direct Sales
“So I suspect that shifts more over time. I can't imagine a company that wants as much control as NVIDIA does loves the ad and board partner thing, but they've built a business on it, and so they're not really willing to cannibalize and alienate. But I bet if t…”
Gilbert: Every tech company failed at in-car differentiation except Tesla
“Every technology company has had their fanciful attempt to try to create a meaningfully differentiated experience in the car, all have failed. You think about Microsoft and the Ford Sync. You think about CarPlay, kind of, maybe a little bit works. And the only…”
Gilbert: Enterprises will model everything in Omniverse before real-world rollout
“Because I think now that we have the compute, the ability to gather the data, and the ability to actually, you know, run these simulations in a way that has a efficient way of running it and a good user interface to understand the data, people are going to sto…”
Rosenthal: The enterprise metaverse will run mostly without human presence
“This is what an enterprise metaverse is going to be. This is not designed for humans. Humans may interact with this There will be UI. You'll be able to be part of it. The purpose of this is for simulating Applications. And most of it, I think, is going to run …”
Rosenthal: Replicating NVIDIA's 15-year CUDA stack is an enormous challenge
“This is 15 years of CUDA and the hardware underneath it and the libraries on top of it that Nvidia has built. To go recreate that and surpass it on your own is such an enormous, enormous bite to bite.”
Ben Gilbert: NVIDIA's Scale Moat Relies on Amortizing 1,000+ CUDA Engineers Across Millions
“The whole CUDA investment. Not at first, but definitely now is predicated on being able to amortize that a thousand plus employees spend over the base of the three million developers and all the people who are buying the hardware to use what those developers c…”
NVIDIA and CUDA constitute a platform on par with Apple and Microsoft
“NVIDIA is a platform, in my mind, no doubt about it. CUDA and NVIDIA and general purpose computing on GPUs is a platform, so whatever, you know, all of the stew of powers that go into making that, that go into making Apple, Microsoft, you know, and the like, G…”
Gilbert: A modern GPU contains dozens of miles of microscopic wiring
“The stat that I was going to bring up, which was just mind-bending to think about, is that there are dozens of miles of wiring on a GPU.”
Gilbert: NVIDIA spends only $1 billion annually on CapEx
“Nvidia only has a billion dollars of capex per year.”
Gilbert: NVIDIA generates a 37% operating margin, beating Apple
“Another one that I wanted to point out, it's a freaking hardware company, I know we didn't, they're not a hardware company, but they're a hardware company, with 37% operating margins, so this is even better than Apple”
Gilbert: Nvidia needs physical AI to sustain its historical growth
“The A plus is definitely that it does exist in the physical world, and they are the dominant provider of everything you need to be able to accomplish that. And if the real world stuff, you know, these little robots that run around factory floors, and the auton…”