NVIDIA – The Full-Stack Maestro

Summary

---

Podcast Summary: AI to ROI – "NVIDIA – The Full-Stack Maestro"

Host: Ray Rike
Co-Host: Peter Buchanan
Date: May 13, 2026

---

Episode Overview

In this Big Story Edition, Ray Rike and Peter Buchanan explore NVIDIA's astonishing evolution from a GPU manufacturer into what they now call "the full-stack AI maestro." The discussion charts the company’s six-month transformation into a dominant AI platform provider, encompassing hardware, software, networking, and strategic investments that have positioned NVIDIA at the center of the global AI economy. The hosts break down the company's financial meteoric rise, major new products, and its increasingly sprawling ecosystem, while offering insights for enterprise buyers navigating this fast-moving landscape.

---

Key Discussion Points & Insights

1. From GPU Maker to AI Maestro

• Transformation Narrative: NVIDIA’s role has expanded from orchestrating AI hardware (the “maestro”) to rewriting the music—and even "building the entire AI concert hall."
  • “It’s going beyond just orchestrating and conducting that orchestra. It’s rewriting the music. … I think they may be building the entire AI concert hall.” – Ray Rike [02:33]
• Historic Comparison: Ray likens NVIDIA’s rise to IBM’s dominance in the mid-20th century—an analogy underscoring scale and influence.
  • “The only company I think was ever close … was what IBM was in the 1960s through 80s.” – Ray Rike [05:04]

2. Revisiting the AI Economy Thesis

• Dependency on NVIDIA: Since October 2025, the thesis that NVIDIA is indispensable to the high-growth AI market has only intensified, given continued scarcity of high-performance chips and the lack of alternatives.
  • “The AI market in its high growth form couldn’t exist without Nvidia … everyone had to go to them.” – Peter [02:53]
• Ecosystem & Flywheel: Strategic investments convert free cash flow into long-term, sticky customer relationships, compounding their market velocity.
  • “Convert cash … into long-term customer relationships. … Compound the velocity of that flywheel.” – Ray [04:07]

3. GTC Announcements: Vera Rubin and Groq3 LPU

• Vera Rubin Supercomputer:
  • A fully integrated AI supercomputer, using seven NVIDIA chips and fast liquid cooling. Can be installed in two hours vs. two days for legacy systems.
  • “This Vera Rubin … is fully liquid cooled and they can install this in two hours versus two days.” – Ray [06:18]
• Licensing the Groq3 LPU:
  • A $20B deal, acquiring leading inference chip IP, talent, and assets from Groq (founded by Jonathan Ross, ex-Google TPU team), enabling up to 35x token throughput over Blackwell systems.
  • “…completed this $20 billion licensing transaction with Grok… to deliver 35x token throughput over the current the Blackwell systems…” – Peter [07:19]
• Inference Bottleneck Breakthrough:
  • Moving from serial (GPU) to parallel inference (LPU) architecture addresses a major bandwidth bottleneck for real-time, agentic AI workloads.
  • “With … LPU architecture … it becomes more of a parallel process. So it eliminates that bottleneck of token by token processing.” – Ray [08:22]

4. Software & Networking: Rounding Out the Stack

• Networking Growth:
  • Surprising cost driver for enterprises; $11B in quarterly revenue (263% YoY growth), with Optical partnerships (e.g., Coherent, Lumentum) to attack network capacity bottlenecks.
  • “One of the number one unexpected costs was the networking capacity…. That’s exactly why Nvidia’s networking segment … generated over $11 billion just in the last quarter…” – Ray [09:31]
• GTC Software Launches:
  • • Nemo Claw: Agent toolkit for enterprise AI; partners include Adobe, Salesforce, SAP.
  • • Dynamo: Open-source inference OS
  • • Open Shell: Open-source runtime, launch partners include Box, Cisco, Atlassian, Salesforce, etc.
  • “Nemo Claw was the first serious attempt by a major hardware vendor to make autonomous agents safe for production enterprise deployment.” – VentureBeat quote via Ray [11:54]
  • • Nemotron: Open-source “frontier models” for on-prem deployment, countering rise of China’s open models.

5. Strategic Partnerships & Customer Base Diversification

• Verticals & Industry Penetration:
  • Key customers and pilots include Mercedes-Benz (self-driving systems), Eli Lilly (drug discovery), Uber (autonomous vehicles), and retail innovation.
  • “Mercedes-Benz will be shipping vehicles equipped with the Nvidia's self-driving system.... Eli Lilly actually partnering... using their infrastructure and... models to accelerate drug discovery.” – Ray [14:31]
• Sovereign AI & New Cloud Players:
  • $30B+ in national government contracts post-October 2025 (France, Netherlands, Canada, Singapore, Middle East).
  • Increasing deals with NEO cloud providers and industry-specific leaders to diversify away from hyperscalers.
  • “They have basically venture investments in market leaders … in all of them. … They’re also very big in sovereign AI.” – Peter [16:00]

6. NVIDIA’s Financial Dominance

• Staggering Financials:
  • FY2026 revenue: $215.9B (65% YoY growth, 8x since FY2022), with $190B (90%) from data center revenues.
  • Free cash flow: $97B for FY2026 (47% margins).
  • “That is a 47% free cash flow margin … That is a hundred and nine percent rule of 40. … I've never seen anything like this.” – Ray [19:12]
  • Gross margins reach 80%—rare for hardware companies (“They have 80% gross margins. ... benefit of being the designer and not the manufacturer.” – Ray [20:23])
  • Consistently beating guidance; Q1 FY2027 guidance at $78B revenue, $5B above consensus [21:44]
• CFO’s Confidence:
  • “Right now we are the king of inference. … With the Vera and Rubin platform and their Grok LPUs, that position will only get stronger.” – Colette Kress, CFO, relayed by Ray [21:44]
  • “That’s downright chesty from a CFO, frankly.” – Peter [21:44]

7. Building Moats: Investments, Licensing, and Ecosystem Lock-in

• Strategic Investments:
  • $50B across 170 VC deals since 2022; from 12 corporate deals in 2022 to 67 in 2025.
  • Major stakes in key players: $10B+ in Anthropic, $20B Groq licensing, investments in OpenAI, X.ai, Cursor, Synopsys, Synthesis, Nibias, and optical networking startups.
  • “They're investing everywhere ... have their fingers and tentacles everywhere. But is there a strategic pattern here, Peter, or am I missing something?” – Ray [24:23]
  • “Largest share … by deal count, because that's where the chips actually run … Then they put money into foundation model developers.” – Peter [24:32]

8. Competitive & Macro Risks

• Custom Silicon/ASICs:
  • Hyperscalers’ in-house chip development (TPUs, custom ASICs) at a lower cost threatens NVIDIA’s future share.
  • “TPUs cost between 10 to 50% of what an equivalent Nvidia chip. So that's one [risk].” – Ray [25:39]
• Counterarguments:
  • Integration, multi-cloud needs, and NVIDIA’s full-stack approach offset some of these risks.
  • “Nvidia is an accelerated computing company. It's not a GPU company. … Buying the stack together is a better deal for most enterprise customers.” – Peter [26:50]
• Inference Economics Pressure:
  • SaaS and AI-native companies’ margins squeezed by high inference costs, feeding margin pressure back to NVIDIA.
  • “There's going to be a real push as we mature this AI industry to reducing inference costs.” – Ray [27:41]
  • “There is no CUDA moat in inference.” – MIT Researcher Paul Krajowski, relayed by Ray [28:27]
• Geopolitical & Structural Threats:
  • Export controls (esp. to China), TSMC dependency (now NVIDIA’s #1 customer, surpassing Apple), and potential capital market tightening among hyperscaler customers.

---

Notable Quotes & Memorable Moments

• “NVIDIA’s transformed into that full stack player … everything’s bonded basically as an AI factory for enterprises and governments alike.” – Peter [33:09]
• “I think they're investing everywhere to have their fingers and tentacles everywhere.” – Ray [24:23]
• "No one can even come close to that. Not even Palantir." – Ray on NVIDIA's metrics [19:12]
• “That’s downright chesty from a CFO, frankly.” – Peter [21:44]

---

Enterprise Buyer Takeaways [31:07]

1. Full-Stack Evaluation:
   • Don’t just compare GPU costs; analyze the whole stack (infrastructure, networking, inference, software, security).
2. Model & Inference Economics:
   • Optimize model choices and procurement to better manage inference costs.
3. Strategic Partnerships:
   • Gain better terms or access by engaging with NVIDIA’s investment/partnership programs—these offer advantages not available through standard procurement.
4. Custom Silicon Caution:
   • Realistically assess fit before assuming custom in-house silicon will solve all enterprise AI infrastructure needs; NVIDIA is likely to remain part of the solution.

---

Conclusion & Final Thoughts

• Full-Stack Leadership Confirmed: NVIDIA’s rapid execution, financial advantage, and broad bets across the ecosystem cement it as the foundational company for enterprise and government AI infrastructure for at least the next 3-5 years.
  • “Think about the long term impact of AI in your company. … Keep an eye on Nvidia and what they're doing.” – Ray [33:42]
  • “…the speed with which they've done it and their ability to just basically grab the brass ring is really impressive.” – Peter [33:09]

---

For ongoing coverage:
Find the newsletter at ai2roi.substack.com

---

Transcript

[A]

Foreign. Welcome back to another episode of AI to roi, the Big Story Edition. I'm Ray Reich, founder and CEO of Benchmarket. And joining me as always, is my co host, Peter Buchanan.

[B]

Yep, I'm Peter Buchanan. I'm the managing partner of New Plan. And Ray, we're diving into something really extraordinary today. Nvidia's complete transformation from a GPU maker to a full stack AI platform company in six months.

[A]

Okay. Hey, Peter, as you know, I've always been in software. I'm a business and financial metrics guy. So I may need you to help bail me out when I don't understand exactly some of their chips and technology.

[B]

Okay, I'll do that.

[A]

The good news is this isn't a technology story per se, it's not just another earnings story. But since we published our deep dive on Nvidia in the newsletter back in October, the company has moved faster and further than I think anyone could anticipated. And by the way, we're going to talk about the financials later on. Remember when cloud costs first started surprising people? Finance would open the AWS invoice, no one could explain what drove it, and Engineering would spend a week building a spreadsheet to explain it. And it was still wrong. AI is doing that again, except faster, more dynamic, and spread across more systems and departments. Today, a single Enterprise has multiple AI costs running across multiple vendors such as AWS, inference costs on Anthropic and OpenAI, GitHub, Co Pilot, Cursor, and a handful of AI agents that no one billing system sees the entire spend picture. So Finance asks, what did we spend on AI this month? And the answer takes three days to find out and it's still probably wrong. And what did we get for it? Most don't even try to answer that one. That is a problem Maverick was built for. Maverick gives finance, IT and engineering a single source of truth for AI spend so you can allocate costs, enforce budgets, and connect investments to business outcomes. Learn more at Maverick AI that's M A V V R I K AI now onto the show. Yeah, they're bonkers, in fact, and a spoiler alert that we always do. You know, we said that Nvidia was the kind of AI maestro, right?

[B]

Yeah.

[A]

Well, it's going beyond just orchestrating and conducting that orchestra. It's rewriting the music. And I'll tell you what, I think they may be building the entire AI concert hall. But hey, let's start with what we talked about back in October, man, five months ago. Tell us, Peter.

[B]

So we started. Let's start. Yeah, we use that as the baseline. So on October 4th of 2025 we wrote an article that concluded that Nvidia was the maestro of the AI economy. And we had three elements to our argument. The first is the AI market in its high growth form couldn't exist without Nvidia because there's no other place to get the high performance chips that the products also were capacity constrained and there were no alternatives, so everybody had to go to them. And then they were actively building an ecosystem by investing in companies that would then use their platform. So there was a big debate at the time about circular deals with hyperscalers and some of the other up the stack application companies that sort of guaranteed that Nvidia was going to make money, but frankly they were going to make it anyway because there was nowhere else for people to go.

[A]

Yeah, but you know, back then the poster child of that was the $100 billion Nvidia and OpenAI partnership. Now there were milestones and it was time based, but to me that was the playbook in action. Right. Convert cash, which they have a lot of cash and we'll find out They've got almost 45% free cash flow margins. So they invested that free cash flow into long term customer relationships. They wanted to really broaden and cement their place in the entire ecosystem and basically compound the velocity of that flywheel. How's that held up?

[B]

Yeah, so that thesis held up. I mean, what do you think?

[A]

You tell me. Well, maybe I should tell you. Okay, here, I'll start it.

[B]

Okay. Yeah.

[A]

You know, the Economist is always a good publication to quote and they describe Jensen's vision perfectly. Transform Nvidia into a foundational company on which the rest of the AI economy rests. This means not just selling chips, but to me the most interesting move is them bundling software into this complete AI stack and embedding Nvidia's technology across the industry. In fact, Nvidia is so much more than a chip company today. And the I was trying to think of parallels in the history of technology, Peter. And the only company I think was ever close to what Nvidia is here in 2026 was what IBM was in the 1960s through 80s. And most people listening to this podcast, if you're not a historian of technology industry, you may not know, but IBM dominated the technology industry for over 20 years.

[B]

Well, I lived it because I'm from a big blue family. My dad worked across IBM and we bled blue there in Bethesda, Maryland. So yeah, I remember it really deeply. So We've got new products coming out Ray from Nvidia because they just had their gtc. So what's the summary of that?

[A]

Well, so they unveiled this, the Vera Rubin platform. And this is a great example of an integrated AI supercomputer with all the components coming from Nvidia because that platform has several. In fact, I think it's seven chips including Vera Rubin, NVLink, I'm not going to name them all, but the fact that they're integrating all this together and then they extended that with the, I call it the reverse Aqua hire by licensing the GROQ3 LPU. This is the key differentiator and in fact this Vera Rubin, and this is important in the hardware space I'm learning, is fully liquid cooled and they can install this in two hours versus two days. In fact, when they announced this at their GTC just a week ago, Microsoft Azure even went live on this before the keynote even finished. But hey, let's talk more about the Grok licensing deal, if possible.

[B]

Sure. So on Christmas Eve, in the dark of night, right before Santa left to make his rounds, Nvidia completed this $20 billion licensing transaction with Grok. And Grok makes these specialized LPU chips for Inference, which is the new big thing in AI. It uses a lot more power and electricity and computing cycles than even training does. What that licensing deal said is a non exclusive license. Their ip, their physical assets for where they were making these things. And key engineering staff is in most of the engineering staff plus the founder Jonathan Ross, who started Grok after working on Google's TPU team. So. So the deal was Fast tracked after OpenAI signed a $10 billion pact with a competing inference chip company called Cerebras. And so they very quickly combined the Rubin GPU chip and the Grock3 LPU chip to deliver 35x token throughput over the current the Blackwell systems that everybody's installed now. So it's a big rapid accomplishment.

[A]

Yeah. And I'm just going to bring this up to a level that even I can understand. Basically with today's GPUs Nvidia's core product that it started from, it's a very serial process. Each token that's generated and that response is done one at a time. So it becomes a memory bandwidth problem. So now with the LPU architecture, my understanding is, is it becomes more of a parallel process. So it eliminates that bottleneck of token by token processing. Is that correct?

[B]

It is. When you look at there's a great analyst with exponential view Azim Mazar he described Nvidia's GROK acquisition as a direct read on how Nvidia views the inference market, what it's about to become. So Nvidia believes it's going to be a million fold expansion in compute demand driven by agentic AI. And as great as their GPUs were, they're not going to catch up with that. And it would take a couple years for them to develop the types of chip and the software integration and all that sort of stuff. But Grox solves this problem for them. Really great technology. Now the company GROK still exists, they can sell this technology to other people. But of course Nvidia is going to fork it in this integrated environment and it's going to be hard to beat. So beyond this though, this supercomputer, there's software and networking that are becoming really important.

[A]

Yeah, of course I want to spend more time on software because that's what I'm really familiar with. But even on the networking side, we did some research on unexpected cost of AI for enterprises and one of the number one unexpected costs was the networking capacity that was required. So we needed to get more throughput in the network. And that's exactly why Nvidia's networking segment it generated over $11 billion just in the last quarter of their fiscal 26. It was a 263% year over year increase. They also have a lot of partnerships including with companies like Coherent and Lumentum which are co packaged optics. So bottom line they're saying hey we, we know that network capacity is going to be a bottleneck, so let's build and partner there to reduce that bottleneck.

[B]

Yeah, and I think that that's the next place, I bet that's the next place you see a really big Nvidia acquisition because there are a lot of companies coming up to challenge them in that particular space and I bet they take the best one off the board because nobody can outbid them. So that basically another quote from the Economist basically says that moving data between processors is as important as the processors themselves because it's a bottleneck. Much like using a GPU for massive amounts of inference would be a bottleneck. This is sort of the same thing. So what about that software layer though, Ray? We'll get to your favorite thing now.

[A]

Okay, so of course at their gtc, their big conference, annual conference, they announced something called Nemo Claw. Now OpenClaw was an AI agent framework that went from zero to a billion dollar acquisition by OpenAI in about five months. And what Nvidia has done is they've created this agent toolkit called Nemo Claw. It's built on top of the Open Claw framework. They've already got 14 kind of classic enterprise partners and Adobe Salesforce and SAP actually partnering with them. And VentureBeat actually said, hey, Nemo Claw was the first serious attempt by a major hardware vendor to make autonomous agents safe for production enterprise deployment. They also introduced Dynamo, it's an open source inference operating system. And then they also introduced Open Shell which is an open source sandbox runtime. And for that the launch partners included Box, Cisco, Atlassian, Salesforce, ServiceNow, CrowdStrike. So a lot of exciting software. But then they decided not to stop and they even introduced something about I think developing their own generative AI model called Nemotron. Can you tell me a little bit about that?

[B]

It's an open source, a Nematron family of open source frontier models. They run locally on Nvidia software and they're honestly the best US counterweight to the onslaught of these excellent Chinese open source models. And so Nvidia is not fooling around here because one of the things they said at GTC and well, Jensen said in his three hour keynote presentation is that Nvidia is going to spend at least $26 billion over the next five years out of free cash flow with no outside funding required to create best in class frontier models based on open source. Nvidia really now has the whole stack to the point where now they're taking that whole stack and they're trying to really push distribution with some incredible advantages. For example, they've spent the last six months really, really diversifying their customer and partner base. So they were selling a lot to hyperscalers. AWS committed to a million Nvidia GPUs by the end of 2027. Azure, you mentioned before, deployed the first Vera Rubin rack conveniently timed to GTC. But they've also got anthropic Meta Mistral AI, huge AI OpenAI commitments and they are starting to get integrated into the mission critical AI strategies of vertical industry market leaders. So who are they, Ray?

[A]

Yeah, in fact this is one of the other really interesting things. In fact, a former person who worked for me, she was my head of marketing a couple times. She's now ahead of their retail industry marketing and they're doing a lot in retail and she's been there for five years, so it's been a great journey. But then Jensen recently announced that they've been working with Mercedes Benz and Mercedes Benz will be Shipping vehicles equipped with the Nvidia's self driving system. So Elon Musk said 10 years ago, hey, someday Tesla probably won't be a automotive manufacturing company will provide the self driving platform for other OEMs. Nvidia is now in that market. And then we have Eli Lilly actually partnering with Nvidia and using their infrastructure and their open source models to accelerate drug discovery. And they have a lot of other specialized open source AI models for the automotive industry. I think it's called Apple Myo for robotics and even biomedical research. And then they also were just associated by Bloomberg with Uber to help Uber launch their own autonomous vehicles. So I will tell everyone listening to this podcast, go to the Nvidia website and go to the industries menu and you'll see so much amazing research. What's happening? Use cases not only they're a great technology company, they're a pretty good research company. Peter.

[B]

They are a pretty good research company. We like to use that research from time to time. And I think one of the things to say before we move to the next section here is in these categories that where Mercedes is deploying self driving technology. Well, Uber Nvidia has invested about $5 billion in two self driving software companies that advance these technologies. And if you go through these vertical industries, they have basically venture investments in market leaders or contenders in all of them. So they're also very big in sovereign AI. So between when we wrote this article in October and Today, they've signed $30 billion in contracts with governments that are building national AI infrastructure. They've done it in France, they've done it in the Netherlands, they've done it in Canada, they've done it in Singapore, they've done it in the Middle East. So Hyperscalers accounted for 50% of their revenue in the fourth quarter of their fiscal year 26 which ended in the end of January. But the faster growing part are these sovereign entities, companies like GM and Eli Lilly, Neo cloud partners like N Scale and Nebius. So they are also diversifying away from the really big names, the companies that also have AI foundation models.

[A]

That's really smart because the level of CapEx going on with the hyperscalers, it's not sustainable long term. And long term I'm talking about looking out at 2030 to go into the sovereign entities, the enterprises including the software stack, the new cloud operators, really smart. But hey Peter, it's the financial performance that makes Nvidia the maestro of the AI ecosystem.

[B]

Can I talk a little bit about that gobsmacking? It's just gobsmacking. First of all, if you look at the end of fiscal year 2022, Nvidia's revenues are pretty impressive. About $27 billion more than anyone ever thought they'd get to. But if you look at their fiscal end of fiscal 2026 ending on January 21st, they were at $215.9 billion.

[A]

Hey Peter, I was going to say that's up what, 65% from just the prior year and 8x since 22.

[B]

Yes, absolutely.

[A]

What about data center? How big is data center revenue as

[B]

a portion of that data center revenue is over $190 billion. It's about 90%. But of course you have to break that data center revenue down into GPUs, networking software, etc. Etc. But the big number, the one that really gets your attention, is free cash flow for the year, $97 billion. And it's not staying in the bank, it's being spread across the market.

[A]

Peter, I'm sorry to interrupt you, but I just. I'm the SaaS metrics guy.

[B]

Right, you are.

[A]

We talk about the rule of 40 all the time. I just did some quick math. If that is a 47% free cash flow margin, which I'm pretty sure it is, they have a growth rate of 65% year over year. That is a hundred and nine percent rule of 40. No one can even come close to that. Not even Palantir.

[B]

No. So have you ever seen that in your career? No one ever. You've never seen that in all your metric ness?

[A]

No. Well now let's be fair. When companies are very small, like 1 million and they're growing 300%, but for a company at scale, and that's at least to me, 100 million. I've never seen anything like this and not even in. And Peter, I don't want to get on too big of a rabbit hole here with you, but when you think about a chip or hardware company, you don't even think about gross margins in that 70. 80%. But they're a hardware company. Yes. Now they're becoming a full stack platform company. They have 80% gross margins. And I guess that's the benefit of being the designer and not the manufacturer.

[B]

Yeah, that's exactly right. It's just amazing. So Ray, how did, how did Jensen summarize this last earnings call? It was just a few days ago. What did he actually say?

[A]

Yeah, well, you want a CEO and I'm going to also share something. Their CFO said something like, hey, Blackwell cells are off the charts and cloud GPUs are sold out. He also said that compute demand keeps accelerating. Man, that's a really good market to be in. And I think he even said we've entered the virtuous cycle of AI, right? And I believe he said that after Q3 and then it was Q4, which was even better results. I mean, revenue climb is, I think you mentioned 68.4 billion in Q4, which ended on January 25, 2026. That was up 73%. But I have to tell you what their CFO said. I've never heard a CFO say this. And the CFO is Colette Kress. And she said right now we are the king of inference. And with the Vera and Rubin platform and their Grok LPUs, that position will only get stronger. That's pretty positive optimistic commentary from a cfo, Peter.

[B]

That's downright chesty from a cfo, frankly. You know, to be. It's, it's putting it out there, it's very confident. So they provided Q1 2027 guidance of $78 billion. That would be $10 billion increase quarter over quarter. And that analysts were saying $72. 73 billion. So that's $5 billion over. That would be the fourth straight quarter above consensus guidance. Plus our friends at the Economist, we're loving them today noted that their cash flow advantage versus its most formidable competitors, the AMDs of the world. All these hyperscaler chip efforts, Broadcom, their cash flow advantage is insurmountable. And then we go to the next moat, Ray, and that moat is ventures, licensing and ecosystem lock in. So let's talk about Nvidia building that competitive moat just beyond hardware. So the investment approach is pretty simple. Nvidia backed firms are more likely to buy Nvidia backed chips. They would get really upset if they didn't, even though it's not technically in any of these contracts. Since 2022, Nvidia has committed over $50 billion across 170 venture deals. So corporate deal volume went from 12 deals in 2022 to 67 deals in 2025. And their internal venture capital fund grew from one deal in 2022 to 30 deals in 2025. So again, they're just spreading that money around and creating basically guaranteed customers, but also doing a lot of things for those customers to make their businesses better.

[A]

Yeah, and by the way, they're really hedging their bets across the entire AI ecosystem. I mean, I think they invested close to $10 billion in anthropic, right? Yeah, they did that $20 billion license deal, reverse aqua hire with Grok. They participated in Xai's $20 billion fundraise. They participated in Bat Cursor's Series D, which was 2.3 billion. They took a $2 billion equity stake in Synopsys, which is an EDA kind of software design software. Then they participated in Synthesis Series E. Another $2 billion investment in Nibias, $4 billion in companies developing optical interconnects that we talked about earlier for that networking capacity. So to me it seems like they're investing everywhere to have their fingers and tentacles everywhere. But is there a strategic pattern here, Peter, that I'm just missing?

[B]

There is. And so they, they've invested. They have cloud and infrastructure providers at the bottom of the stack. So the up and comers, the core weaves together AI Lambda. That's the largest share of their portfolio by deal count because that's where the chips actually run that people are buying from them. And then they put money into foundation model developers. Obviously OpenAI and Anthropic are the biggest two examples. But they've also put money in Nostril here, Xai, that's the second most active area. Jensen Huang just basically says he doesn't see investment risk in these NEO cloud providers that journalists and analysts always fret over because Nvidia has probably better vision into the pipelines of basically every hyperscaler and every NEO cloud. And he sees the demand flowing really heavily into these NEO cloud providers. So the real question Ray is what risks could derail Nvidia?

[A]

Well, one of my favorite leaders early in my career, and he wrote a great book called Only the Paranoid Survived, was Andy Grove from Intel. Now this was back when cpu, because of the PC revolution, I mean they thought they were going to grow forever. So same thing with Nvidia. They got to really think about where is my competition going to come from. So two directions that I kind of see first are the hyperscalers and their own custom ASICs, things like Google's TPUs. They now support Gemini 3. And Gemini 3 was trained entirely on GPUs, not Nvidia's GPUs. Then Google started to sell those TPU access externally. Anthropic is a deal to use like 1 million of them. I think Bernstein said that Google's TPUs cost between 10 to 50% of what an equivalent Nvidia chip. So that's one. And then some of these hyperscalers, Amazon, Meta and even now Microsoft, their own developing their own processors for production So I think that's one core threat.

[B]

Right. But you can argue that these custom Asics, they're internal assets and if you want those, you have to buy from that specific hyperscaler. Multi cloud systems are difficult to do. You've got sovereign AI buyers that want kind of a kit, or you've got model developers that can't build their own silicon. So what Jensen Huang told Stratecheri in a recent interview with Ben Thompson is Nvidia is an accelerated computing company. It's not a GPU company. And they use software in this new stack as a moat because buying the stack together is a better deal for most enterprise customers. Ray, what about the inference economics?

[A]

Well, I spend a lot of time with CFOs and one of the biggest concerns and issues for CFOs in SaaS and AI native software companies is the impact of those inference costs that they're paying to OpenAI or Anthropic to their gross margins. Where they were used to 75 to 80% gross margins, they're now having to deal with 30, 40, 50% gross margins. So there's going to be a real push as we mature this AI industry to reducing inference costs. And when the inference providers, the model companies, et cetera, are pushed on price, that's going to push down to the infrastructure providers like Nvidia. And I think it was an MIT research fellow, his name is Paul Krajowski. He actually said that there is no cuda mote in inference.

[B]

Right. But Nvidia is arguing that by pairing the Groq LPU chips with Vera Rubin, they get the low latency, the high throughput that pure GPU architectures don't and that they've basically solved that problem. So they feel pretty good about that. So I think we ought to move to some other sort of macro risks that are important for Nvidia. The first is export control escalation. So Nvidia is restarting production of the H200 chips which they're now, right now they're allowed to sell them to China. Jensen Huang has noted that if China's domestic chip ecosystem develops unchallenged, those chip companies combined with the open source models are going to be really a handful for American, really any Western technologies to compete with. Getting Nvidia chips into places like China actually undermines the Chinese chip companies. The second thing for them that's a problem is TSMC concentration. Nvidia is a fabulous company. TSMC makes their chips. They are now TSMC's largest customer, larger than Apple. They Passed them last year. Apple actually has to compete for capacity for the first time. And you have the geopolitical risk of China rattling sabers at Taiwan. If China took over Taiwan, what would happen to Nvidia or a lot of Western technology companies, businesses? And the third is the circular capital risk. So we've talked in our writings and we've talked in today's podcast that basically three hyperscalers account for over half of Nvidia is receivable, and those same companies are building competing chips. So if the capital markets tighten or neo cloud players don't quite make it and have liquidity problems, then demand for Nvidia's chips could drop. And so these are real risks.

[A]

Yeah, they're risks, but I wouldn't mind being sitting in their seats right now for the next couple years. I think they're fairly manageable. But, you know, here we are. God, our time's already up. Let's bring this home with a few takeaways for part of our core audience that listens to the AI to ROI podcast, and that's enterprise buyers. I'd like to kind of maybe give two or three ideas of what enterprise buyers should do based upon this dominance of Nvidia. Is that okay?

[B]

Sure, go ahead.

[A]

So first of all, you know, people are making make versus buy decisions. So if you're making that make decision, you might want to evaluate more than just the gpu. You want to look at the GPU infrastructure, the inference cost, and networking the software. How do you optimize that? How do you ensure you have security of your AI agent AI? So just don't look at that infrastructure GPU cost. And then the second thing is the model inference economics, you really need to look at that because you may make some decisions of what models you use. Maybe you bring in some small language models, maybe you bring them in house versus using third party. And then that's going to impact a procurement approach also, correct, Peter?

[B]

It is. So you don't want to be a regular customer with Nvidia. So Nvidia's investment and partnership program creates commercial arrangements where you can maybe jump the line a little bit to procure supply. You could agree to infrastructure commitments. You could do joint go to market. Those aren't available through standard procurement. So figuring out a way where they're encouraging you to go deep. If you think you're going to use Nvidia, you totally want to do that. And fourth, you shouldn't assume that custom silicon solves your problem, because these custom chips are optimized for their particular platforms and there are a lot of applications because data resides in different places or you want to have something on premise where you can't run those chips. And so Nvidia needs to be at least part of your solution, possibly all of it.

[A]

Coming back to why become a strategic partner if you're working with Nvidia and you're thinking about building your own infrastructure, not just a customer. So let's wrap this up. Peter, what do you want to say here at the end to wrap it up?

[B]

I'd say Nvidia's transformed into that full stack player, compute, networking, inference software, agent, runtime, stack, the racket, total scale. Everything's bonded basically as an AI factory for enterprises and governments alike. You can see this coming going into the fall, but the speed with which they've done it and their ability to just basically grab the brass ring is really impressive.

[A]

And last thing I would say is think about the long term impact of AI in your company. Which means you need to keep an eye on Nvidia and what they're doing because they're going to be strategic to what you try to accomplish, at least for the next three to five years. Which means you should probably keep tuning in to AI to roi, but also subscribe to our newsletter because we every week are keeping up to speed on everything happening in the AIO ecosystem, especially with Nvidia. And you can subscribe at ai2roi. That's AI, the number2roi.substack.com hey Peter, thank you so much for all the great research you've done here.

[B]

Well, it's great to do it and it's great to be here with you today. So we'll see you next week.

[A]

Okay, Bye bye everyone. Sam.