So, next up in our interview series, well, it's Pat Gelsinger. Do I need to say more? Good to see you, Pat.
And great to see you. Thanks for having me on.
We're here at Playground Global, which is Pat's new home. So, I just have to ask, how's it going?
Well, you know, as I'd say at this phase of my career, just turned 65, just had my 65th birthday. You know, I want to do things that matter, right? You know, if they succeed, make a difference with people I enjoy. Pretty simple at this point in life. And Playground's a place I can do that. I have about 10 companies that I'm deeply involved in now, on the board of most of those. And I think some of these are going to be industry-shaping companies. And if I get to help build great companies with great leadership teams at this phase of my career, okay, this is good. And so far, you know, I'm a rookie in venture capital, but the team's embraced me and we're making a difference.
Did you ever see yourself eventually migrating into this sort of role? It's a bit hard to go from a sustained, big company sort of position into VC.
Well, remember 45 years, right? Not a day was I, you know, when I left Intel the first time, to EMC, literally I retired at night, started the next morning. Anyway, you know, officially I retired. And when I went from EMC to VMware, literally I did both jobs for two weeks. When I left VMware to come back to Intel, I did both jobs for four weeks. This has been 45 incredible years. And then all of a sudden, my wife walks into my study literally the next day after the announcement departing Intel and she said, 'You're not done yet.' I think that meant don't be home too much. There was a little bit too much energy there. So, please. And that began a 100-day journey for me, of figuring out, okay, what do I want to do next? And I interviewed a lot of venture firms, private equity firms, a bunch of CEO roles, government roles. It was when the administration was coming in, you know. University roles. And at the end of that, it was sort of like, okay, you know, I want to make a difference. And Playground came to the top of that list. And then also my Glue company, the faith technology company. So I'm splitting my time between the two of those, making a difference eternally and making a difference in the future of science and bringing those two together. Okay, that's a pretty good position of life.
I can't imagine how many people, you know, came out of the woodwork to offer you advice on the next stage.
Of course, and a number of those, I wanted their advice. 100 meetings in 100 days to figure out what was next. And here we are a year later and it's going pretty well. And still semis, a lot of semis, but you know, I'm doing nuclear operating. I didn't know anything about nuclear, now I'm doing that. I'm doing superconducting, Josephson junctions. It's sort of in computing, but wow, those are like really different looking transistors from what we thought about. Worrying about cryogenics, stretching my mind into quantum computing. And for a digital Boolean guy my entire life, you know, it's like you got to turn your brain to a 90-degree angle to think about qubits and how they're functioning and entangled and how you use them to compute. So, a lot of learning for me personally. But as I say, the only disappointment I have at this phase of my career is I'm not 35 years younger. This is the greatest time to be a technologist in human history.
So in my mind I'm seeing you at a desk pouring through papers of potential startups trying to find the one that's got something a bit unique. Is that how it works?
Well, you know, I'm reading a lot, so lots of papers. That's part of it. A lot of that work gets done by our associates. So, you know, 'hm, that one sounds pretty interesting' and they'll seek my advice on how it would fit into the industry. They'll go do work and we'll meet. I like reading, but even more so I like interacting. To me, if I could read for 10 hours or meet with the company for one hour, the one hour is more valuable.
You'd rather have the college lecture than the reading?
Well, it's not the college lecture. It's the questioning, the interaction. I want a few smart people at the table. I want to ask questions. I want to hear their questions, how they respond. That's my best learning modality. But yeah, my brain is being stretched in so many dimensions. I'm learning about bioengineering. How semiconductors are being applied to bioscience as well. What I call the trinity of computing, the fusion of classical, AI, and quantum effects, coming together. I think that will be, literally, the problems that we will now be able to compute will be almost everything that we've been able to express mathematically will now become computable. So science, biology, and of course areas like...
Some not NP-hard, I guess.
Yeah. But even so, you just have this powerful new tool coming over the horizon with quantum computing where many of those quadratic functions that were uncomputable before, okay, now they're sort of like, yeah, great, let's go attack those portions of the problem as well.
So how do you marry the fact that you're not completely green and brand new, but these are new degrees of freedom for you to explore, but then also evaluating whether a startup is actually worth the investment? There is something to be said for the person who spent 40 years in a space and their add-in to that discussion.
Yeah. And you sort of come down to sort of three questions, right? One is, okay, can we make the technology work? Is there real proof? And we do hard tech things, so there really is this question, does the tech work? And then you're sort of saying, okay, how does it get to market? How do we bring it into an at-scale deployment? So, combining the business side. And then the third is, do we have the team to do it? The leadership team. And obviously, for each one of those, the tech, I mean, you're at Playground, right? We have more brilliant technologists to ask those hard questions. What are going to be the milestones in the way to show that we get there? The market insertion, and by the way, that's part of my unique value. I'm able to call up CEOs throughout the industry, 'Hey, don't let this one get stuck down here as some entry-level person. As you evaluate, this one matters. Let's work together on how to differentiate your offerings, your GPUs, your networking components, your business using these.' And so that connectivity helps quite a lot with number two, driving the business development and accelerating it into the marketplace. And then obviously with 45 years of major leadership roles, my job is to shape leaders into great CEOs, great CXO teams as well. And I give them some pretty tough feedback, sometimes. Some of them are great leaders, couldn't present to save their life. Okay, we're going to make you at least a mediocre presenter. In other cases, they're great presenters, but they're not great leaders. How do we build a leadership team around you? In other cases, well, we got to tell you how to make a sale. There's all these things associated with building a great company. And since I've been in these big leadership roles, I really enjoy bringing that into the CEOs. But the hardest thing, Ian, I don't have my hands on the wheel. Every once in a while it's sort of like...
Well, you can lead a horse to water, right?
Yeah. For some of those, but it also does give me a lot more flexibility as well. So, tell me how you solved this one. By Monday morning, I'm taking the weekend and taking my grandkids skiing.
You got that veteran status that you can take the weekend off. You're a startup, you still got to work.
If you like this content, there are multiple ways to support the channel. You can like and subscribe this video and many thanks for doing so. There's also Patreon which gives you access to our Discord. There's a merchandise store and a newsletter. Links in the description. For all of you who do contribute, thank you. You are keeping me well-fed.
I've been seeing some of the magic you guys have done with our good friend Mark Wade over at Ayar Labs. They just did their Series E, mighty impressive, and they seem to have funding from everyone now.
Yeah. Well, you know, we also, I'll say there's this characteristic where you get a term sheet, somebody ready to lead the round, price the round, and then okay, people are ready to jump in. And there's this sort of interesting thing, what's going to cause that dam to break? Right. For obviously that was...
Did you mean that in the sense of people don't want to fund or it just may end up being oversubscribed? Everybody is sort of around the table. Who's going to put the lead term sheet down for this round? Who's going to price the round? Who's going to say, 'I'm going to put the biggest check into this round.'
And as soon as, you know, them or maybe it's two who are sort of co-leading the round, then others are ready to come into it. Okay, the round is going to get formed. We're going to see it. Now, we want to be part of that as well. And that was very much the case with Mark and the Series E. Once we had that lead term sheet, sort of broke the dyke a bit, then the energy unleashed of many other people wanting to come into the round and quite a few strategics as well. So to me, as I'm learning the venture game, if we could call it that, having that network of people that I am confident that they're going to be leading with me to form rounds in timely fashions, because if it takes us three months to form a round, okay, that's three months later in the market as well. And I don't want to let some of the hottest companies languish in competitiveness by not moving them forward effectively.
Speaking to a few people in the last couple of months since Nvidia's acquisition of Groq, a lot of people who are raising have suddenly found it's getting easier in '26. Are you finding similar?
Clearly that was catalytic, yes. Nvidia is not going to acquire 10 companies in that space. They were clearly deficient in having optimized inference. And as I said at GTC last year in the pre-show, we need to make inferencing 10,000 times better. Not 10x, 10,000 times.
That's a great headline. The question is how.
Yeah, right. Which, when you think about it, it's replacing search. We're going to, now with OpenClaw, something none of us quite predicted even though everybody was predicting agentic. It's just demonstrating that, okay, we got to make inferencing a lot better. And my 10,000x was a number I pulled out based on some math of where search was and so on in terms of energy, compute, cost. But that led Nvidia, as proud as they are and should be around the incredible progression of the GPU, saying okay, the GPU is great for training, it's great for some of the waterfall training into inferencing, but it's not an optimized inference chip. And that led them to do Groq, obviously. But now, there's 20 companies pursuing that assignment to say, how can we be 10x better or 100x better than where Nvidia just described the LPU with Groq.
I had to tell you, but I'm tracking 150 these days.
It's anything can precede, all the way up to...
I got to see your list here. So, yeah. The variety of SRAM-based and data flow machines and people talking about HBM and high flash and...
Well, so that brings me on to the fact that when I look at the companies that often put you in their press releases when hanging around, you know, Snowcap for example, this sub-zero, almost absolute zero type of computing. It seems like a lot of bets you're making are non-von Neumann these days. Despite, you know, you literally just saying you've spent so long in digital logic and traditional compute, for lack of a better term. x86 is still high and mighty, and I know you're very proud of that. So, to turn around and say, well, hey, let's look at the analog side of the business. Let's look at different ways of computing. Is that where you see the future going or is this just a we have to optimize for every workload sort of play?
Yeah. Well, I do think, part of that is the language I've used called the trinity of computing where we've always believed in accelerators. I've always believed in accelerators and whether they were SIMD accelerators, SIMT accelerators, MIMD, there's other forms of how you attack workloads. To me, architecture in that sense, trying to fit the world through a von Neumann lens means that all workloads look von Neumann. Is the workload defining the architecture or is the architecture defining the workload? And in this sense, what we're now, and I think clearly, the surge in AI has clearly said, hey, different architectures enable different workloads to emerge at scale. And that's what's made this period exciting, that all of a sudden, data-centric workloads, with matrix functions against them, are able to do really interesting things. But it's the workload that matters. And in this sense, I think I was saying that before even Jensen was saying that in some of our debates. So it's the workload that matters. Which workload is interesting, solving unique and wonderful problems, and then what's the best architecture to run it on? There's nothing that you saw yesterday at GTC that can't run on the CPU.
Right, yeah, right. That's the beauty of von Neumann type machines.
Can you run Jensen on the CPU? Maybe he's got a little bit of magic there, but I know what you mean.
You know, because it may not run efficiently. It may take a long time. And as we said, there's algorithms that only can run on the quantum machine. No, I can run them on a classical machine. They just take a billion years and run out of memory. I can't get there from here using that computing architecture. So I do think in that sense we're going to see this multiplicity or heterogeneity of what the compute architectures are. Most of the interesting problems will combine them. Yesterday Jensen announced the CPU, what a backward thing for him to do. Why are we going back to the CPU?
It was funny because at the beginning of the presentation, I'm not sure if you caught it, he said, you know, the CPU is old. It's the legacy. And then as you're right, he then announced the CPU.
Yeah. Right. So sort of, of course I noticed that. I'm a CPU guy. Of course I noticed it. It's like, so first we're going to spend half the keynote throwing it under the bus, we don't need it anymore. And then we're going to announce that we now have the best one because we need it. Right. Which is it? Well, the answer is the workloads need both. There are things that run like crap on a GPU. If I have control flow, I need my code. If I have control flow related functions, they're terrible in a GPU.
Anything predicated and...
Yeah. Right. Your basic if-then-else is a terrible thing to do on a long pipeline GPU. You may be able to run six different dimensions of parallelism, but if-then-else is not a parallel function. So in that sense, I really do see it as a heterogeneous view. That's why I call it the trinity of computing. Classical, that's going to be control flow, tool-based, analytics, operating systems, all of those things, more in the CPU. I'm going to have this raft of things that are data-centric, algorithmic, that are going to be great on the AI. And then I'm going to have these things that only work because of a quantum machine, being able to apply entangled qubits against those problems. And with that, okay, now we can open up a whole lot of workloads that are not computable today.
It's, I know people are going to be shouting at the camera because again we discussed this back in the day and your good friend Greg Lavender was on top of this, it's the software strategy, right? You can't have a heterogeneous installation without having the software that attacks it. And I've always been a vocal critic of the write-once-compile-to-target because it's never as efficient as you need it. And we know working with hyperscalers, they're going to extract every percentage point of power and efficiency out of code and you can't do that at a high level where you might have this heterogeneous. So what, it sounds mad to ask, but what can we do here? We've had standards in machine learning, for example, on Windows ML and Intel tried to do it with oneAPI and everything else, but there just doesn't seem to be any consolidation right now like we had in GPUs.
Yeah. And you know, I do think we sort of abstract, abstract, abstract, and then all of a sudden we need to collapse. Right. When there is that, and I think everybody was, a little bit over a year ago, was the DeepSeek moment. Right. Well, what did they do? I'll say they collapsed. They went to really understanding what the machine was doing and getting very targeted on how to use the machines to be much more efficient.
Stick people in a box and they'll engineer their way out.
Yeah. Right. And what does an engineer do? He produces great results in the constraints that he has. And to me, that was part of what made DeepSeek such a defining moment because they sort of tunneled through the stack, said we really know what's going on in hardware. We're going to align our algorithmic functions against what hardware we have available and is capable. And I think we're going to see some of that happening again. I also think, one of our portfolio companies, NextSilicon, building a programmable data flow machine, saying, hey, this has all gotten too complicated. We have to create an abstraction sitting beneath a more programmable hardware embodiment that is able to statically and dynamically reconfigure itself both in network topology and in compute resource to the most efficient way of the current running workload. And as we've all seen, in AI workloads, the characteristics of the workload change dramatically over, am I in the prefill phase? And as you go through the different phases of computation, big GPUs are turning on and then they're turning off and power rails are bouncing across the scale. Networks are being overloaded and then they're being underutilized. So for that, I think there has to be this layer of programmability. Part of the reason we're quite excited about NextSilicon. But clearly, we have DRAM shortages today. Imagine that, four years of great DRAM business. This ain't ever happened in industry's history. Now you have extraordinary logic and DRAM demand. That never happens in the industry.
Everybody says we're so used to these bubbles in these commodity markets. And the question I get asked a lot by investors, and I'm sure you do, is when is it going to end?
Yeah. And I'm not sure about you, but I struggle to predict that end because realistically the only way it's going to end from my perspective is if somehow the bottom comes out of the market.
Yeah. Well, you know, to some degree, Moore's Law, Koomey's Law, whichever way you view it, we're in a Promethean period of compute expansion, right?
Right. And in this sense, and I think OpenClaw clearly has been sort of that next accelerant, that all of a sudden, how many tokens per day are you going to use as an engineer? Because, hey, I'm going to spend $100,000 on a token. I'm going to give you a $200,000 salary. At what point does that become...
That's the argument I've made. It's the how much do you want to put into an engineer.
Yeah. And there doesn't seem to be any end in that value proposition right now. Now obviously as I said we have to make inference 10,000 times better. I do think some of these compute-memory architectures are going to have meaningful breakthroughs that are going to make inferencing dramatically cheaper and not 10x like we saw from Groq yesterday, but thousand-x or 10,000x better.
Did you see the Talos announcement earlier in the year?
No, I didn't follow that one.
So they're doing what looks like a structured ASIC design. So you bake in the communication metal layers your model. And they're seeing 10, 14,000 tokens per second.
Yeah. So, pure ASIC style, traditional ASIC by definition. I'm not sure how you feel about people calling things GPUs versus ASICs. You being the traditionalist you are.
Tomato, tomato, right? Some investors sometimes don't realize. But at the end of the day, market acceptance will be based on tokens per second, tokens per second per watt, aggregate throughput, capabilities, latency. It will deduce to real engineering measurable results over time.
So my argument against that has always been it doesn't matter how many tokens you produce if your tokens are useless, right? The value of a token. And I'm sure you're seeing the same thing I am that right now the biggest workload that has value in the token for the output is code. People are willing to pay if you accelerate your engineer. Or I often cite IBM because they have a consultancy business where they generate tokens for their clients and they can upsell and the client just sees the reduction in cost. Is there anything we can do to stop talking about just garbage tokens?
Well, let me disagree with you a little bit. I'm now starting to see very meaningful agentic business process workflows, where people are turning nominally low-end white-collar business process flows into highly agentic workflows, where I truly am putting agents to solve things for me that make me more productive, more scalable as well. And by the way, that's sort of what led to the whole SaaS implosion. As people started to see, oh, what's going to happen to my Salesforce, my Oracle, all of those start changing as well. And those are very real. And I think that probably becomes the AI pathway into most enterprise workloads as well. Once we solve security problems and people get comfortable with where they're running and whether that's on-prem or in the cloud or what the security model is, I think that's powerful. I do think there's always going to be a little bit of 'my token is different than your token'. So I do think there's going to be, okay, what really is the token? But good benchmarking is going to help us sort that out over time. And I've done more benchmarks in my career than most humans ever should consider. But there's lies, damned lies, and benchmarks. But you still are always seeking the best metrics that allows us to deduce differences in hardware. That's going to get harder as we've already talked about with regard to heterogeneous architectures as well. So the role of good benchmarking is going to get more important in this next phase. And by the way, it's going to be critical because people are going to be making billions or tens of billions of dollars of capital decisions based on that work. So, it's going to get to be more, not less important because otherwise I'm going to build data centers that produce crappy tokens.
Yeah. I think we started calling it benchmarketing.
Well, depends on the graph you get from the... Remember, my code is in the SPEC benchmark, right? One of the most venerable benchmarks of all time. My code is still in there. In SPEC92 or SPEC95.
Uh, yes, all right, from the very earliest days of SPEC, right? And my code was there. So, which test, which subtest?
Well, if you go into the SPEC benchmark, you look, the espresso flow, that was mine. We had the compiler flow, that was my code that we were using in the compiler when we initially did it. I think I have a third one in there as well.
So, would you trust yourself to do that again?
Oh my, at the time it was great. Of course, when people give me those numbers, I think they suck today. It's like everybody was trying to break cache sizes and so on with them at the time. And today, it's just not a valid view. So I would just yell at our engineers in my last four years at Intel when they would show me SPEC numbers or SPECrate numbers because it's just like, okay, that code is so old. It's been so tortured into cache footprints. There's no longer a meaningful view of system performance.
And there's a small argument here because I'll speak to ARM and they still value SPEC2006 because of the embedded market. The embedded market still relies on it. But no, I completely understand your point. Going back a little bit for a second and when we have these agentic setups, I often see that especially if you look into the wider industry, a lot of people are playing with it and it seems a very personal implementation on people improving their workflows. And I struggle to really see where it's going to offer it at scale. And the only sort of workload I'm seeing where it is actually being applied at scale, you buy your software program and it offers your agentic flow, is chip design. Because our good friends at Synopsys and Cadence are leaning on it heavily than almost anyone else. So how do we bridge that gap? Is that just a time thing or is that just a familiarity thing or...
Well, I do think, and that's a great place to start. So let's just say that's wonderful. There's a whole lot of other engineering workflows other than that in material science and CFD and all of the, now we're worried about hypersonics and so there's just all sorts of wonderful things that open up across the greater EDA space, not just chip design. And the explosions that are happening in bioengineering. Okay, now you've created orders of magnitude of additional complexity as you start looking at three-dimensional molecular models. You can't make nice 2D simplifications like you can in a lot of the chip design flows. So I do think in this sense, we're at the beginning of these processes, not anywhere close to the end. And I do think that every one of these as you open one up, somebody's going to be the one that sort of jumps ahead and then what happens is it's sort of like cockroaches. There's a little bit of food in that corner. A lot more industry innovation will run in that direction.
So given your background and it relates a little bit to my background on the HPC side, I speak a lot with the people over there almost on a weekly basis and they're getting frustrated by the whole march of machine learning. Especially as we look at 64-bit precision being reduced in traditional accelerators and I know that a lot of them are wary of 64-bit emulation in the 8-bit side. Do you have any good news for him?
Well, a couple of thoughts. I mean, when people started talking to me about BF4, right? I was like, what are you talking about? I was worried if 64 bits was good enough. We needed the 80-bit modes that we put into i486.
Yeah. Sort of, I was worried if 64 bits was good enough, right? We needed the 80-bit modes that we put into, right? i486.
Right, you know, for precision things, you know, and now we're saying that four bits is good enough, right? And I think even there, you know, four bits is, you know, people are doing interesting work, but what they find is models actually don't settle very well when you get to precision that low. So, you know, useful bits seem to be more in the 8 to 16 bits for most model work. You know, I'm not a model expert in that regard, but you know, they sort of run some of those benchmarks, right? At 4 bits, but the reality is most machines aren't actually operating there. And, you know, emulating 64 bits in a 16-bit environment is very different than saying, hey, we're going to do it in the 4-bit environment. But I do think as we move to the next phases of science, right, and not just doing large language model performance, as we move to the next phase of science where the LLM almost becomes a leaf node, not the core computational node, right? And you're going to be back to worrying very much about precision.
So I think there will be over the next couple of years the revenge of the HPC guys. Okay, because, you know, now when I want to start using LLMs in the context of a CFD, right, for my airplane or hypersonic design, I'm going to be fusing these together and I'm not going to be possible to give up precision for this portion of the workload even though this portion of the LLM may be perfectly happy running on a BF8 machine, right? I'm going to need them both. So I do think that the next phase of application of AI is going to be increasingly in the science domain, will be increasingly in three dimensions, not in flatter language representations. And as a result, all this pursuit of getting to 4-bit, 2-bit, etc., I think some of that's going to be for naught because the real machines are going to be combining those together in a much more central way in workloads. So now maybe I'm just trying to convince myself that all that hard work I did on 64 and 32-bit floating point is going to come back, you know, reality. But I think it's based on the view of workloads. The next phase is not doing more language modeling, it's much more science modeling. As you get to science, I need real precision to look at many of those algorithmic domains. Again, it comes back to the workload.
Yeah, it's speaking to a few people who run university supercomputers, they're saying because of the price of accelerators these days and the high demand, they're almost seeing cloud be more cost effective.
Right. And if there's one thing you don't want, it's a student accidentally clogging up a thousand GPUs in the cloud with bad code. So yeah, I think they're struggling. And you mentioned them before, our friends at Next Silicon are trying to solve part of that.
I think there's a community who really want more love.
I think so as well. And you know, remember, Nvidia began its journey into AI through the HPC door. So I do think there's going to be good science being done there. I think the national labs matter a lot. I think some of the government programs matter a lot here. And people like Next Silicon, I think, will be helpful. But I do think, hey, there needs to be more networking work, right? I joke that NVL72 is an engineering marvel and a manufacturing nightmare, right? It's just how do you build those things, scale them? You're hitting just the limits of copper. We saw a bunch of optical conversations. I think we need new models that become resilient networks, work like we're doing with Delos Data, right? So that we create more flexibility in the underlying network because we can't have failure rates that we're seeing today for these super large configurations. Then you're just spending all your time either restarting and redoing or checkpointing, right? And that's not a computing architecture suitable for a lot of the workloads that we want to get to. So I do think some of these things are going to move us more rapidly to optical, move us more radically to resilient networks, move us more radically back toward data flow machines with the full gamut of precision. And then, of course, we're going to do radical things like Snowcap, which are just, you know, thousand times better.
I'll go speak to people like IBM with the Z processors and then some of the automotive players and they speak about resilience as a very hardware-focused element. Yet with the lost data and others, we're seeing it more as a software application on top of the hardware. Is there a right balance?
I think the workload again is the determinant, right? In the sense that, hey, if the software has assumed faulty hardware, right, and has built resilience into the software layer, right? Okay, then software is the right answer, right? But if in fact a whole lot of workloads, right? And the sort of the difference between a virtual machine and a Kubernetes machine, the difference is okay, who is taking care of resilience? In the VM, I'm assuming the hardware is right and I'll run any software including containers, right? In the opposite view, I can only run containerized software because I'm not presuming resilience in the hardware. And for a lot of things, hardware is better for resilience, right? I can see memory errors, I can correct memory errors, I can see link failures.
You certainly could. And I think now as we're looking at machine learning and failure rates that we're seeing in machine learning, we need more in the hardware, right? It's not resilient enough anymore for some of these incredibly large cluster sizes that we're seeing emerge as well. So, I'd say, I don't think there's a one answer to that question, right? And I do like Jensen's code design perspective this way because, hey, there are certain things that I think are done better in software resilience, but there's a whole lot of things that, okay, don't let the hardware guys off the hook, right? You're going to build, you know, 59's hardware. You're going to be able to prove 10 to the 12th, 10 to the 14th error rates, right? And until your material structures have proven that, get back in the lab and finish your freaking work, right? To prove something that I can really build on into the next generation.
Last time we recorded one of these, it was at Intel Foundry Connect. I'm sure you remember. Back in the five nodes in four years days. How much do you spend these days thinking about next generation foundry and packaging technology?
Well, obviously, Xite, you know, I'm spending a lot of time on how we build next generation, as I call it, waking Moore's law up from its nap, right? So spending a lot.
