NVIDIA (NVDA) has been making frequent investment moves recently, first announcing a $5 billion investment in Intel, followed by committing up to $100 billion to invest in OpenAI. The market has responded positively to these developments, particularly following OpenAI's collaboration with Oracle.

However, skeptical voices have emerged in the market, claiming that NVIDIA, OpenAI, and Oracle are engaged in "circular revenue" practices, suggesting that financial manipulation outweighs actual revenue generation.

On September 25, in the latest episode of the BG2 podcast, BG2 host and Altimeter Capital founder Brad Gerstner, along with Altimeter Capital partner Clark Tang, engaged in a conversation with NVIDIA CEO Jensen Huang. During this dialogue, Huang addressed various market concerns.

Huang believes that investing in OpenAI represents an excellent opportunity, viewing OpenAI as the next trillion-dollar Hyperscaler company.

Additionally, Huang specifically explained why ASIC chips are not entirely competitive with NVIDIA GPUs. He emphasized that NVIDIA serves as an AI infrastructure provider, offering capabilities that extend beyond hardware and software layers to include continuous iteration speed, reliability from scale advantages, and overall energy efficiency.

Therefore, Huang believes NVIDIA's moat is now "wider" than it was three years ago, while Brad Gerstner even suggested that NVIDIA could become the first company to reach ten trillion dollars in value.

The following is a compiled transcript of the interview (edited for length):

**Brad Gerstner**: Jensen, welcome back. Your red glasses look great on you. It's been over a year since your last podcast appearance. You now derive over 40% of your revenue from inference, and inference is about to take off due to chain of reasoning.

**Jensen Huang**: Most people haven't truly internalized this yet - this is essentially an industrial revolution.

**I. Investing in OpenAI Isn't a Prerequisite for Cooperation - It's an Investment Opportunity**

**Brad Gerstner**: Honestly, since then, it feels like we're validating that podcast every day. In AI timescales, this year feels like a hundred years. I recently rewatched that episode, and many observations struck me.

What impressed me most was when you pounded the table saying - when everyone thought pre-training was entering a trough, many said pre-training was finished, hardware construction was excessive. That was about a year and a half ago. You said inference wouldn't just be 100x or 1000x - it would be a billion times. This brings us to today. You just announced a massive collaboration, and we should start there.

**Jensen Huang**: I want to formally state that I believe we now have three Scaling Laws. First is the pre-training Scaling Law. Second is the post-training Scaling Law. Post-training essentially involves AI practicing a skill until it gets it right, trying many different approaches. To achieve this, inference must be performed. So training and inference are now integrated through reinforcement learning in a very complex way - this is post-training.

Third is inference. Past inference was "one-shot," but the new inference we understand now is "think before answering." Think first, then respond - the longer you think, the higher the answer quality. During thinking, you retrieve, verify facts, learn things, continue thinking, continue learning, then output the answer rather than generating immediately. So thinking, post-training, pre-training - we now have three Scaling Laws, not one.

**Brad Gerstner**: You mentioned these last year, but are you more confident this year that "inference will improve a billion-fold, bringing higher-level intelligence"?

**Jensen Huang**: I'm more confident this year. The reason is looking at today's agent systems. AI is no longer a single language model but a system composed of multiple language models running concurrently. Some are using tools, some are doing retrieval - many things happening, and it's multimodal. Look at generated videos - absolutely incredible.

**Brad Gerstner**: This leads to this week's key moment - everyone's talking about your major Stargate collaboration with OpenAI. You'll become the preferred partner and invest $100 billion in the company over time. They'll build 10 "gigs" (Gigawatts). If all 10 gigs use NVIDIA, the revenue contribution could reach $400 billion. Help us understand this collaboration - what does it mean to you, and why is this investment reasonable?

**Jensen Huang**: I'll answer the latter question first, then return to my narrative. I believe OpenAI will very likely become the next multi-trillion-dollar hyperscale company.

Just like Meta is hyperscale, and Google is too - they'll have both consumer and enterprise services simultaneously. They're very likely to become the next multi-trillion-dollar hyperscale company. If so, being able to invest in them before they reach that scale would be one of the smartest investments we could imagine. You must invest in what you understand, and we happen to understand this field. So this investment's returns will be outstanding.

We're happy to invest, but it's not necessary, nor a prerequisite for cooperation - they offered the investment opportunity, which is wonderful.

We collaborate with OpenAI on multiple projects. First, Microsoft Azure construction, which we'll continue advancing - this collaboration is progressing very well with years of construction ahead. Second, OCI (Oracle Cloud Infrastructure) construction - I think about 5-7 GW to build. We're advancing this with OCI, OpenAI, and SoftBank. These projects are all contracted and being implemented with substantial workloads. Third is CoreWeave. Everything CoreWeave-related... I'm still talking about OpenAI, yes, everything's in the OpenAI context.

So the question is: what is this new partnership? It's helping OpenAI build AI infrastructure for the first time. That is, we're collaborating directly with OpenAI on chips, software, systems, and AI factories to help them become a fully self-operating hyperscale company. This will continue for quite some time as a supplement to their existing construction.

They're experiencing two exponential curves: The first exponential is customer numbers growing exponentially because AI is improving, use cases are improving - almost every application now connects to OpenAI, so they're experiencing a usage exponential.

The second exponential is computational demand's exponential growth. Computing power for each use case is exploding. Previously it was one-click inference; now you must think before answering. These two exponentials combined dramatically elevate computational needs. We'll advance all this construction. Therefore, this new collaboration is "incremental" to all existing collaborations, supporting this amazing exponential growth.

**Brad Gerstner**: You mentioned an interesting point - you believe they'll very likely become a multi-trillion-dollar company and represent a good investment. Simultaneously, you're helping them build their own data centers. Previously, they outsourced data centers to Microsoft; now they want to build "full-stack factories," like Elon and X, right?

**Brad Gerstner**: Think about Colossus's advantages - they build full-stack as hyperscalers. Even if they don't use all capacity themselves, they can sell to others. Similarly, Stargate is building massive capacity - they think they'll use most of it but can also sell excess. This sounds very much like AWS, GCP (Google Cloud), or Azure - is that the idea?

**Jensen Huang**: I think they'll likely use it themselves, like X mostly does. But they want to establish direct relationships with us - direct engineering collaboration and direct procurement relationships. Like the direct relationship between Zuck, Meta and us. Our direct relationship with Sundar and Google, our direct partnership with Satya and Azure. They're large enough now and believe it's time to establish these direct relationships. I'm happy to support this, and Satya (Microsoft CEO) knows, Larry (Google co-founder) knows - everyone knows.

**II. Expectation Divergence Between Wall Street and NVIDIA: Understanding AI Demand Scale**

**Brad Gerstner**: There's something I find rather mysterious here. You just mentioned Oracle's $300 billion, Colossus construction - we know some sovereign nations are building (AI infrastructure), hyperscalers are building too. Sam speaks in trillion-dollar terms about all this. Yet the consensus among 25 sell-side analysts covering your stock shows growth slowing from 2027, predicting 8% annual growth from 2027-2030. These people's only job is making growth predictions for NVIDIA. Obviously...

**Jensen Huang**: We're comfortable with this. Look, we often easily exceed expectations.

**Brad Gerstner**: I understand. But this remains an interesting "divergence." I hear this daily on CNBC and Bloomberg. Many worry shortages will turn to oversupply - they don't believe sustained high growth. They say, fine, we'll believe your 2026, but 2027 might see oversupply and reduced demand. But consensus forecasts actually show this won't happen. We've also made our own forecasts incorporating all these numbers.

The result is, even two and a half years into the "AI era," beliefs remain vastly divergent. What Sam Altman, you, Sundar, and Satya believe differs greatly from what Wall Street still believes. Do you remain comfortable with this?

**Jensen Huang**: I don't think this is inconsistent either. First, we "builders" should build for "opportunity." We are builders. Let me give you three thinking points to help you feel more confident about NVIDIA's future.

The first point, at the physical laws level - the most critical point - general-purpose computing has reached its end; the future belongs to accelerated computing and AI computing. This is the first point.

You can think of it this way: How many trillions of dollars of global computing infrastructure needs to be updated and replaced? So first you must recognize the end of general-purpose computing. No one disputes this. Moore's Law is dead, people often say. So what does this mean? General-purpose computing will shift to accelerated computing. Our collaboration with Intel acknowledges that general-purpose computing needs to merge with accelerated computing, creating new opportunities for them. General-purpose computing is migrating to accelerated computing and AI.

The second point: AI's primary use cases are already everywhere - in search, in recommendation engines, etc. Basic hyperscale computing infrastructure previously executed recommendations via CPU; now it must execute AI via GPU. You replace traditional computing with accelerated computing and AI; you replace hyperscale computing from CPU to accelerated computing and AI. This is the second point. Just satisfying Meta, Google, ByteDance, Amazon - moving their traditional hyperscale methods to AI represents hundreds of billions in market opportunity.

So even without discussing new AI opportunities, just AI changing old methods to new methods. Then discuss the future. Yes, so far I've actually only discussed "mundane" things. The old way is wrong. You won't use kerosene lamps anymore - you want electricity. That's enough.

Then there's the more incredible part: when you move toward AI, toward accelerated computing, what new applications emerge? That's all the AI-related opportunities we're discussing - massive opportunities. How to understand? Simply think: previously, electric motors replaced physical labor; now we have AI - I call them AI supercomputers, AI factories - they'll generate tokens to enhance human intelligence. And human intelligence accounts for 55%-65% of global GDP... let's say $50 trillion. This $50 trillion will be enhanced by something.

Back to the personal level: suppose I hire an employee earning $100,000 and equip them with a $10,000 AI. If this AI doubles or triples that employee's productivity, would I do it? Absolutely. I'm already equipping everyone in the company - yes, every co-agent, every software engineer, every chip designer already has AI collaborating with them - 100% coverage.

The result is we make better chips, increasing quantities, advancing faster. The company therefore grows faster, hires more, has higher productivity, higher revenue, higher profits. Now apply NVIDIA's story to global GDP - what's likely to happen is this $50 trillion will be... let's take a number, enhanced by $10 trillion. And this $10 trillion needs to run on a machine.

AI differs from past IT in that previously software was pre-written, ran on CPUs, didn't really "move" much. In the future, AI must generate tokens, and machines must generate these tokens - they're "thinking," so software is continuously running. Previously it was write-once, now it's continuous writing, continuous thinking. To make AI think requires factories. Assume this $10 trillion in tokens has a 50% gross margin - $5 trillion of that needs factories, needs AI infrastructure.

So if you tell me global annual capex is roughly $5 trillion, I'd say that number makes sense. This is roughly the future outline - from general-purpose computing to accelerated computing; replace all hyperscale with AI; then use AI to enhance human intelligence, covering the global economy.

**Clark Tang**: As of today, we estimate the annual market size at approximately $400 billion, so TAM from now to target is a four to five-fold increase.

**Jensen Huang**: Exactly. Last night (Beijing time September 24), Alibaba's Eddie Wu said from now until the end of the 2020s, they'll increase data center power (consumption) ten-fold. Right? You just said four-fold? That aligns. They'll increase power ten-fold, and our revenue is almost positively correlated with power. (Note: Eddie Wu stated that compared to 2022's GenAI inaugural year, by 2032 Alibaba Cloud's global data center energy consumption will increase 10-fold) He also said token generation doubles every few months.

What does this mean? Performance per watt must improve exponentially. This is why NVIDIA must frantically advance performance per watt improvements, and revenue per watt is basically revenue.

**Brad Gerstner**: In this future, there's an assumption that I find fascinating from a historical perspective. For two thousand years, global GDP basically didn't grow. Then the industrial revolution came, GDP accelerated; then the digital revolution, GDP accelerated again. You're now saying the same thing as Scott Bessent (current U.S. Treasury Secretary) - he believes next year global GDP growth will reach 4%. You're essentially saying global GDP growth will accelerate because we're providing the world with "billions of colleagues" to work for us. And if GDP is output under given labor and capital, then it must...

**Jensen Huang**: Definitely will grow. Look at what's happening with AI - AI's technical forms, availability, such as large language models and AI agents, are all driving a new "agent industry." This is unquestionable. OpenAI is the fastest revenue-growing company in history - they're growing exponentially. So AI itself is a high-growth industry. Because AI needs underlying factories and infrastructure, this industry is growing, my industry is also growing. And because my industry is growing, industries beneath us are also growing - energy is growing, power and facilities are growing. This is simply an energy industry renaissance. Nuclear, gas turbines... look at those infrastructure companies in our ecosystem - they're doing great, everyone's growing.

**Brad Gerstner**: These numbers have everyone talking about potential "oversupply" or "bubbles." Zuckerberg said on a podcast last week there might be short-term "air pockets," Meta might spend an extra $10 billion or so. But he said it doesn't matter. For his business's future, it's too critical - it's a necessary risk. But from a game theory perspective, it's somewhat like a "prisoner's dilemma."

**Jensen Huang**: Happy prisoners.

**Brad Gerstner**: Let me go through this again. Today we estimate by 2026, there'll be $100 billion in AI revenue, excluding Meta, and excluding GPUs running recommendation engines plus search and other workloads - let's just calculate $100 billion.

**Jensen Huang**: But how big is the hyperscale industry itself? What's this industry's current base?

**Brad Gerstner**: In the trillions of dollars.

**Jensen Huang**: Right. This industry will deploy AI first - it's not starting from zero, you must start from here.

**Brad Gerstner**: However, skeptics would say we must grow from 2026's $100 billion to at least $1 trillion in AI revenue by 2030. You just mentioned $5 trillion. From a bottom-up global GDP derivation perspective, can you see achieving $100 billion to $1 trillion in the next five years?

**Jensen Huang**: Yes, and I'd say we're actually already there. Because Hyperscalers have already migrated CPU to AI - their entire revenue base is now AI-driven.

**Brad Gerstner**: Yes.

**Jensen Huang**: No AI, no TikTok, right? No AI, no YouTube Shorts. The customized, personalized content Meta creates can't be done without AI. Previously, those things relied on human pre-creation, providing several options, then recommendation engines selecting. Now AI generates unlimited options.

**Brad Gerstner**: These transitions have already occurred: we've migrated from CPU to GPU, mainly for those recommendation engines.

**Jensen Huang**: Right. Zuck would tell you, and he said this when I was at SIGGRAPH, they actually arrived somewhat late. Meta's GPU usage is only about a year and a half, two years old. Using GPUs for search is even newer - just beginning.

**Brad Gerstner**: So the argument is, the probability of having $1 trillion in AI revenue by 2030 is almost certain because we've almost already reached it.

Then we only discuss "incremental." Whether you do bottom-up or top-down, I just heard your top-down analysis by global GDP percentage. So what do you think is the probability of "oversupply (glut)" appearing in the next three to five years?

**Jensen Huang**: Before we completely convert all general-purpose computing to accelerated computing and AI, I believe the probability of oversupply is extremely low.

**Brad Gerstner**: How many years will that take?

**Jensen Huang**: Until all recommendation engines are AI-based, until all content generation is AI-based. Because consumer-facing content generation is largely built on recommendation systems, so all of this will shift to AI generation. Until traditional hyperscale completely migrates to AI - from shopping to e-commerce, everything migrates over.

**Brad Gerstner**: But all these new construction projects we're discussing are "trillion-level," always investing ahead. So if you see slowdowns or oversupply, do you still "have to" put money in? Or once you see slowdown signals, can you contract at any time?

**Jensen Huang**: Actually it's the opposite, because we're at the end of the supply chain - we respond to demand. Now, all VCs will tell you - you know too - globally what's in shortage is "computing," not GPU quantities. Just give me orders, I'll manufacture. Over the past two years we've opened up the entire supply chain - from wafer starts, to packaging, to HBM memory, etc. - we've ramped everything up. Need to double, we double - the supply chain is ready. We're now waiting for demand signals. When cloud service providers, hyperscalers, and customers make annual plans and give us forecasts, we respond and build according to those forecasts.

The problem is, every forecast they give us is wrong because predictions are consistently low. So we're always in "emergency catch-up" mode - this has continued for several years, with each round of forecasts significantly higher than the previous year.

**Brad Gerstner**: But still not enough. For example, last year Satya seemed to moderate slightly - some called him the "more steady adult in the room," tempering expectations. But a few weeks ago he said again, we built two "gigs" this year too and will accelerate in the future. Are you seeing those traditional hyperscalers - compared to CoreWeave or Elon's X, or compared to StarGate - previously somewhat slower ones now doubling down, and...

**Jensen Huang**: Because the second exponential has arrived. We already have one exponential growing - AI applications and penetration growing exponentially. The second exponential is "inference and thinking" - this is what we discussed a year ago. I said then, once you advance AI from "one-shot, memorize and generalize" to "reasoning, retrieval and using tools," AI is thinking, it will use more computing power.

**Clark Tang**: Returning to your earlier point, hyperscale customers need to migrate internal workloads from general-purpose computing to accelerated computing regardless - they'll build continuously through cycles. I think different hyperscalers have different load structures, uncertain digestion speeds, but now everyone believes they're seriously under-configured.

**Jensen Huang**: One of my favorite applications is traditional data processing - structured and unstructured data processing. Soon we'll announce a major plan for "accelerated data processing."

Data processing occupies the vast majority of today's world's CPUs - it still runs entirely on CPUs. Go to Databricks, mostly CPUs; go to Snowflake, mostly CPUs; Oracle's SQL processing, mostly CPUs.

Everyone's using CPUs for SQL/structured data. In the future, all this will migrate to AI data. This is an extremely huge market that we'll advance into. But you need NVIDIA's full capabilities - the acceleration layer and domain-specific "recipes." Data processing layer "recipes" need us to build, but they're coming.

**III. "Circular Revenue" Skepticism: Investment Opportunities Aren't Tied to Any Conditions**

**Brad Gerstner**: There's another skeptical point. Yesterday I opened CNBC - they talked about "oversupply, bubbles." Switch to Bloomberg - "round-tripping and circular revenues." For viewers at home, this refers to companies forming seemingly transactional but actually lacking real economic substance arrangements, artificially inflating revenue.

In other words, growth doesn't come from real customer demand but from financial numbers. So when you, Microsoft, or Amazon invest in companies that are simultaneously your major customers, like you investing in OpenAI while OpenAI purchases tens of billions in chips.

Please remind us and everyone: when Bloomberg and other analysts make a big deal about "circular revenue," what exactly are they misunderstanding?

**Jensen Huang**: Building 10GW scale is roughly $400 billion. That $400 billion is mainly supported by their offtake (absorption capacity/downstream demand), which is growing exponentially.

(Expenditure) This must be supported by their own capital, equity financing, and available debt - these are three tools. How much equity and debt they can raise depends on their confidence in future revenue. Savvy investors and lenders will comprehensively weigh these factors. This is their company's business, not mine.

We certainly need to work closely with them to ensure our construction can support their continued growth, but the revenue side is unrelated to the investment side. Investment opportunities aren't tied to any conditions - it's a pure investment opportunity. As mentioned earlier, this company will very likely become the next multi-trillion-dollar hyperscale company. Who wouldn't want to hold its equity? My only regret is they invited us to invest in their early years when we were too "poor" - we didn't invest enough, really should have put all our money in.

**Brad Gerstner**: And the reality is, if you don't do your core job well, say Vera Rubin ultimately isn't a good chip, they can buy from others, right? They have no obligation to use your chips. As you said, you view this as an opportunistic equity investment.

**Jensen Huang**: We invested in xAI, invested in CoreWeave - these are all great investments.

**Brad Gerstner**: Returning to the "circular revenue" discussion, there's also a fundamental point - you put everything on the table, telling everyone what you're doing. And what's the economic substance behind it? It's not both parties shuffling revenues back and forth. We see users paying monthly for ChatGPT, 1.5 billion monthly actives using this product. You just said every company in the world must either embrace all this or be eliminated. Every sovereign nation views this as "life-or-death" for their national defense and economic security, just like nuclear power.

**Jensen Huang**: Ask anyone - which person, company, or nation would say "intelligence" is optional for us? None. That's the foundation. The key is "automation of intelligence."

**IV. Moore's Law Is Dead - Now We Need Extreme Software-Hardware Co-design**

**Brad Gerstner**: I've asked enough about demand - let's discuss system design. I'll pass this to Clark next. In 2024 you switched to annual release cadence, right?

After Hopper, 2025's Grace Blackwell is a massive upgrade requiring major data center retrofits. 2026's second half will have Vera Rubin, 2027 has Rubin Ultra, 2028 has Feynman. How's the annual release cadence going? Why switch to annual releases? Is NVIDIA's internal AI enabling you to execute annual releases?

**Jensen Huang**: Yes, the answer is definitely yes. Without it, NVIDIA's speed, pace, and scale would all be limited. Now without AI, it would be impossible to build the products we have today. Why do this? Remember Eddie (Wu Yongming) said in earnings or conferences, Satya said, Sam also said... token generation rates are rising exponentially, user usage is rising exponentially. I remember OpenAI saying weekly active users are around 800 million, right? Just two years since ChatGPT's launch.

**Brad Gerstner**: And these users' each usage generates more tokens because they're using "inference-time reasoning."

**Jensen Huang**: Exactly. So the first point: under two stacked exponentials, unless we improve performance at incredible speed, token generation costs will continue rising.

Because Moore's Law is dead, transistor unit costs remain almost unchanged annually, and power is roughly unchanged. Under these two "law" constraints, unless we invent new technologies to reduce costs, even giving the other party a few percentage points discount cannot offset the pressure from two exponential growths. Therefore, we must improve performance annually at a pace keeping up with this exponential.

For example, from Kepler (note: released April 2012) all the way to Hopper (note: released March 2022), we achieved roughly 100,000x improvement. That was the beginning of NVIDIA's AI journey - 100,000x in ten years. Hopper to Blackwell, due to NVLink and others, we achieved 30x system-level improvement in one year. Next, Rubin will bring another wave of "x" (multiples), Feynman another wave of "x"...

The reason we can do this is because transistors themselves can't help much anymore. Moore's Law basically only has density increasing, performance isn't improving correspondingly. So we must completely disassemble the problem at the system level - all chips upgrade simultaneously, software stack and system upgrade simultaneously - this is extreme "co-design."

Previously no one achieved this level. We simultaneously change CPU, reshape CPU, with GPU, network chips, NVLink vertical scaling, Spectrum-X horizontal scaling. Of course, we also build larger systems, doing cross-domain interconnects across multiple AI factories. And advancing at annual pace. So we're also technically forming "exponential on exponential." This lets customers continuously reduce token costs while making tokens smarter through pre-training, post-training, and "thinking." AI gets smarter, usage increases, more usage means exponential growth.

**Brad Gerstner**: What is extreme co-design?

**Jensen Huang**: Extreme co-design means you must simultaneously optimize models, algorithms, systems, and chips.

When Moore's Law could still drive things, just making CPUs faster made everything faster. That was innovation "inside the box" - just make that chip fast. But if chips no longer get faster, what do you do? You must jump out of the original framework to innovate.

NVIDIA changed the industry because we did two things - invented CUDA, invented GPU, and implemented the concept of large-scale co-design.

This is why we cover so many industries. We're building massive libraries and co-design. First, full-stack extreme isn't just software and GPU but extends to data center-level switching and networking, and all software inside them: switches, network interfaces, vertical and horizontal scaling, optimization across all layers. The result is Blackwell's 30x improvement over Hopper. Moore's Law simply can't do this - this is the result of extreme co-design.

**Brad Gerstner**: All this stems from extreme co-design.

**Jensen Huang**: Yes, this is why we do networking, switching, vertical/horizontal/cross-domain scaling, CPU, GPU, NIC. Also why NVIDIA's software is so rich. Our open-source software contribution submissions globally - few can compare. And this is just the AI field. Don't forget we're in computer graphics, digital biology, autonomous driving, etc. The scale of software we produce is extremely considerable, allowing us to do deep and extreme co-design.

**Brad Gerstner**: I heard from one of your competitors that you doing this can reduce token generation costs. But simultaneously, your annual release cadence makes it almost impossible for competitors to keep up. Because you give the supply chain three years of visibility, supply chain locks deeper, knowing what scale to supply to.

**Jensen Huang**: You might think of it this way: to enable us to do hundreds of billions of dollars of AI infrastructure construction in one year - think how much capacity we must prepare in advance a year ago. We're talking hundreds of billions of dollars in wafer starts, DRAM procurement. At this scale, almost no company can handle it.

**V. Why NVIDIA's Moat Is Wider Now: How to View ASIC Competitiveness**

**Brad Gerstner**: Is your moat today wider than three years ago?

**Jensen Huang**: Yes.

First, there are more competitors than ever, but difficulty is also greater than ever. Because wafer costs are rising. Unless you do co-design at extreme scale, you can't deliver that "multiple" level growth - that's the first point. So unless you do 6-8 chips annually simultaneously, it won't work. The point isn't making one ASIC but building an AI factory system. This system has many chips - they're all co-designed. Together they deliver the "10x" we can get at almost fixed intervals.

So first, co-design must be extreme. Second, scale must be extreme. When your customers deploy 1 GW, that's 400,000-500,000 GPUs - making 500,000 GPUs work together is a miracle.

So customers are taking huge risks purchasing these. Think - which customer would place a $50 billion procurement order on one architecture? A completely new, unverified-at-scale architecture?

However excited you are, however much everyone cheers for you, when you just showed the first silicon, who would give you a $50 billion order? Why would you dare start $50 billion in wafers on a chip that just taped out? But for NVIDIA, we dare, because our architecture is highly mature with accumulated credibility. Second, our customers' scale is extremely amazing. Moreover, our supply chain's scale is also extremely amazing. Who would start these for a company in advance, pre-build so much, unless they're convinced NVIDIA can deliver to the end? Right? They believe we can deliver to all customers globally, willing to start billions of dollars in supply at once.

This is the "scale" story.

**Clark Tang**: Following this point, globally one of the biggest debates is "GPU vs ASIC," like Google's TPU, Amazon's Trainium. Arm to OpenAI, Anthropic are reportedly developing in-house... You said last year we build "systems," not "chips," and you drive performance improvements at every stack layer. You also said many of these projects might never reach production scale - actually... most won't reach it. Given TPU's apparent success, how do you view this evolving landscape today?

**Jensen Huang**: Google's advantage is "foresight."

Remember they made TPU v1 before everything began. This is no different from startups. You should start when the market hasn't grown big yet, not wait until the market rises to trillion-level. All VCs understand one fallacy: the market is big, just getting a few percentage points can build a big company. This is wrong. You should get "almost all share" in a very small field - this is what NVIDIA did back then, and what TPU did back then.

So today's challenge for ASIC makers is: the market looks "fat," but don't forget, this "fat market" has evolved from one chip called GPU to the "AI factory" I just described.

You just saw me announce CPX (note: Rubin CPX GPU, NVIDIA's chip specifically designed for long-context reasoning) - this is a chip for "context processing and diffusion video generation," very specialized but very important data center workloads. I also hinted earlier there might be an "AI data processing" processor. Because you need "long-term memory" and "short-term memory." KVCache processing is very heavy - AI memory is crucial. You want your AI to have good memory. Processing KVCache around the entire system is very complex - maybe it also needs a dedicated processor.

You can see, NVIDIA's perspective today overlooks all AI infrastructure - how will these amazing companies make diverse and changing workloads flow through systems? Look at Transformers - this architecture is rapidly evolving. Without CUDA being so useful and easy to iterate, how would they conduct such massive experiments to decide which Transformer variants, which attention algorithms to adopt? How to do "disaggregate/refactor"? CUDA can help you do all this because it's "highly programmable."

So looking at our business, you must go back three to five years when those ASIC projects started - the industry then was "lovely and simple," just GPUs. A year or two later, it's already huge and complex; two more years, its scale will be enormous. So as latecomers trying to kill into a massive market, this battle is very hard to fight.

**Clark Tang**: Even if those customers succeed with ASICs, their compute fleets should have an "optimized ratio," right? I think investors like black-and-white, but actually even if ASICs succeed, there needs to be balance. There will be many different chips or components joining NVIDIA's accelerated computing ecosystem to adapt to emerging workloads.

**Brad Gerstner**: In other words, Google is also your big customer.

**Jensen Huang**: Google is our big GPU customer. Google is special - we must give respect. TPU has iterated to v7, right? This is also an extreme challenge for them. What they're doing is extremely difficult.

So I want to first sort out chip categories. One category is "architectural" chips: x86 CPU, Arm CPU, NVIDIA GPU - architectural level, with rich IP and ecosystems, very complex technology, built by architecture owners.

Another is ASIC. I once worked at LSI Logic, the company that invented the ASIC concept. As you know, LSI is long gone. The reason is, when market scale isn't too big, ASIC is great - find a foundry/design service company to help you package, integrate, and manufacture, they'll charge you 50-60 points gross margin.

But when ASIC's target market becomes bigger, a new approach appears called COT (Customer-Owned Tooling) - who does this? Like Apple's phone chips, the volume is too big - they would never pay others 50%-60% gross margin for ASIC, they'll master the tools themselves.

So when TPU becomes a big business, it will also move toward COT, no doubt. That said, ASIC has its place - video transcoder markets will never be too big; Smart NIC markets won't be too big either.

So when you see an ASIC company with ten or even fifteen ASIC projects, I'm not surprised, because maybe five are Smart NICs, four are transcoders. Are they all AI chips? Certainly not. If someone makes a processor customized for a specific recommendation system as ASIC, of course that's possible. But would you use ASIC for that "basic computing engine"? Know that AI workloads change extremely fast. There are low-latency loads, high-throughput loads; chat token generation, "thinking" loads, AI video generation loads, now you're talking about...

**Clark Tang**: The compute fleet's "main backbone."

**Jensen Huang**: That's NVIDIA's positioning.

**Brad Gerstner**: More plainly, it's like "chess vs checkers." Those making ASICs today, whether Trainium or other accelerators, are essentially building "one component in a larger machine."

While you're building a "very complex system, platform, factory," now you're also starting to do a degree of "opening." You mentioned CPX GPU - in a sense, you're "splitting" workloads to hardware slices most suitable for them.

**Jensen Huang**: Exactly. We released something called "Dynamo" - disaggregated orchestration for AI loads, and open-sourced it, because future AI factories are disaggregated.

**Brad Gerstner**: You also released NVLink Fusion, even opening to competitors, including Intel which you just invested in - this lets them also access the factory you're building. No one is crazy enough to want to build complete factories alone. But if they have good enough products, attractive enough, end customers say we want to use this to replace some Arm GPU, or replace your inference accelerator, etc., they can plug in.

**Jensen Huang**: We're very happy to connect these. NV Fusion is a great idea, and we're also happy to work with Intel - it brings Intel's ecosystem in. Most global enterprise workloads still run on Intel. It fuses Intel's ecosystem with NVIDIA's AI ecosystem and accelerated computing. We'll do the same fusion with Arm. Later we'll do it with more people. This opens opportunities for both sides - it's win-win, very big win-win. I'll become their big customer, they'll also bring us to bigger market opportunities.

**Brad Gerstner**: Closely related to this, you made a shocking point: even if competitors' ASIC chips are already cheaper today, even if they drop prices to zero, people would still buy NVIDIA's systems. Because a system's total operating costs - power, data centers, land, etc. - and "intelligence output," choosing you is still more cost-effective, even if opponents give chips away free.

**Jensen Huang**: Because just land, power, facilities cost $15 billion.

**Brad Gerstner**: We tried doing the math behind this. For many unfamiliar people, this sounds illogical - you price competitor chips at zero, considering your chips aren't cheap, how could it still be more cost-effective?

**Jensen Huang**: There are two perspectives. One is from revenue angle. Everyone is "power" constrained. Suppose you get an additional 2GW of power, then you want 2GW converted to revenue. If your "tokens per watt" is double others', because you did deep and extreme co-design, your performance per watt is better, then your customers can generate double revenue from their data centers. Who doesn't want double revenue? Even if someone gives them a 15% discount - say our 75% gross margin vs others' 50%-65% gross margin - this difference absolutely cannot make up for the 30x gap between Blackwell and Hopper.

Even if we view Hopper and others' ASIC as same level, Blackwell still has 30x room. So on the same GW, you'd give up 30x revenue. This cost is too high. Even if opponents give chips free, you only have 2GW power available - your opportunity cost is prohibitively high. You'd always choose the strongest "per watt" system.

**Brad Gerstner**: I heard from a hyperscaler CFO that given your chips' performance improvements, especially with power per watt (tokens/gigawatt) and "power supply" as hard constraints, they had to upgrade to new cycles. Looking ahead to Rubin, Rubin Ultra, Feynman, will this curve continue?

**Jensen Huang**: We now do six to seven chips annually, each part of a system. System software is everywhere. To achieve Blackwell's 30x requires tuning and optimization across these six to seven chips. Imagine I do this annually - bang, bang, bang consecutively. If you only make one ASIC in this "chip stew" while we optimize everywhere in the pot, this becomes a very difficult problem.

**Brad Gerstner**: This brings me back to the initial moat question. We've been investing for a long time, investing across the ecosystem, also invested in your competitors like Google, Broadcom.

But when I start from first principles - you switched to annual release cadence, co-research with supply chain, scale far exceeding everyone's expectations, this requires dual scale requirements for balance sheets and R&D. You advanced NVFusion, CPX, etc., through acquisition and self-development. Therefore, your moat is widening, at least in "building factories or creating systems."

But interestingly, your valuation multiples are lower than those people. I think part stems from "law of large numbers" - a $4.5 trillion company can't get bigger. But a year and a half ago I also asked you, if the market will increase AI loads 10x or 5x, we also know Capex trends. In your view, combining the advantages just discussed, what's the probability that revenue won't be "much higher"?

**Jensen Huang**: I'd answer this way - our opportunities are far greater than market consensus.

**Brad Gerstner**: I believe NVIDIA will very likely become the first $10 trillion company. I've been in this business long enough. Ten years ago, people said there couldn't be $1 trillion companies in the world. Now we have ten. Today's world is bigger, right?

**Jensen Huang**: The world has gotten bigger. And people misunderstand what we're doing. People remember us as a "chip company" - yes, we make chips, the world's most amazing chips. But NVIDIA is actually an AI infrastructure company.

We are your "AI infrastructure partner." Our partnership with OpenAI is the best proof. We are their AI infrastructure partner. We cooperate with customers in many ways. We don't require anyone to buy everything from us. We don't require you to buy full racks - you can buy one chip, one component, our network, or just our CPU. Some people only buy our GPUs, pairing with others' CPUs and networks. We basically sell however you like. My only request is, buy some of our stuff.

**Brad Gerstner**: You said, not just better models, but "world-class builders" are needed. You said maybe the nation's strongest builder is Elon Musk. We've discussed Colossus One, where he assembled 200,000+ H100/H200 into one "coherent" large cluster. Now he's doing Colossus Two, possibly 500,000 GPUs, equivalent to millions of H100s in "effective" coherent clusters.

**Jensen Huang**: If he's first to reach 1GW, I wouldn't be surprised.

**Brad Gerstner**: What advantages do "builders" who can both do software and models and understand how to build these clusters have?

**Jensen Huang**: These AI supercomputers are extremely complex. Technically complex, procurement complex (financing), getting land, power, and facilities complex, construction complex, lighting up complex. This is probably one of humanity's most complex systems engineering projects. Elon's advantage is: in his mind, these systems are integrated, all interdependencies are in one person's brain, including financing. Yes, and...

**Brad Gerstner**: He himself is a "big GPT," a "big supercomputer."

**Jensen Huang**: Right, the ultimate "GPU." He has strong urgency - he really wants to build it. When "will" meets "capability," incredible things happen.

**VI. Sovereign AI: AI Is Becoming Every Country's Infrastructure**

**Brad Gerstner**: Another area you're deeply involved in is sovereign AI... Looking back 30 years, you probably couldn't imagine you'd frequently visit the White House today. The President says you and NVIDIA are crucial to U.S. national security. Facing these, first give me background - if nations didn't view this as "life-or-death," at least equivalent to how we viewed "nuclear" in the 1940s, you wouldn't appear in those places. Now if there isn't a government-funded "Manhattan Project," then it's funded by NVIDIA, OpenAI, Meta, Google.

**Jensen Huang**: No one needs atomic bombs, but everyone needs AI. That's the huge difference. AI is modern software. This is what I said from the beginning: from general-purpose computing to accelerated computing, from humans writing code to AI writing code - this foundation can't be forgotten. We've reshaped computing. It needs to be popularized, which is why all nations realize they must enter the AI world, because every country must stay modern in computing. No one would say: you know, I used computers yesterday, tomorrow I'll rely on sticks and fire. So everyone must move forward - it's just that computing has been modernized.

Second, to participate in AI, you must write your own history, culture, values into AI. As AI gets smarter, core AI learns these things quickly, no need to start from zero. So I think every country needs certain sovereign capabilities. I suggest everyone use OpenAI, use Gemini, use Grok, use Anthropic... use various open models. But they should also invest resources to learn how to "build" AI - not just for language models, but for industrial models, manufacturing models, national security models. They need to cultivate a complete set of "their own intelligence." Therefore, every country should have sovereign capabilities.

**Brad Gerstner**: Is this also what you're hearing and seeing globally?

**Jensen Huang**: Yes. They'll all become customers of OpenAI, Anthropic, Grok, Gemini, but simultaneously need to build their own infrastructure. This is the big idea NVIDIA is pursuing - we're building "infrastructure." Just like every country needs energy infrastructure, communications and internet infrastructure, now every country needs AI infrastructure.