WEBVTT 00:00:00.000 --> 00:00:02.319 Welcome back. Today, we are doing a really deep 00:00:02.319 --> 00:00:05.620 dive into a company that I think defines resilience, 00:00:06.139 --> 00:00:09.720 advanced micro devices or, you know, AMD. It 00:00:09.720 --> 00:00:11.339 really is. This isn't just a corporate history. 00:00:11.380 --> 00:00:14.179 It's a story about being the perpetual underdog 00:00:14.179 --> 00:00:17.879 for, what, decades? For decades. Always the challenger, 00:00:18.019 --> 00:00:19.940 always fighting for a little bit of market share. 00:00:20.140 --> 00:00:22.359 And now when you look at them, they're not just 00:00:22.359 --> 00:00:24.899 participating in high performance computing and 00:00:24.899 --> 00:00:27.679 AI. They're actively shaping it. It's one of 00:00:27.679 --> 00:00:29.920 the most incredible turnarounds in tech. The 00:00:29.920 --> 00:00:32.679 hook here is just so compelling. How does a company 00:00:32.679 --> 00:00:34.420 go through multiple near -death experiences, 00:00:34.840 --> 00:00:38.259 lose almost every big fight for years, and then 00:00:38.259 --> 00:00:40.340 come back to completely reshape an industry? 00:00:40.640 --> 00:00:42.759 And the power shift is so recent. I mean, you 00:00:42.759 --> 00:00:45.880 look at 2022. For the first time ever, AMD's 00:00:45.880 --> 00:00:48.280 market CAG actually surpassed Intel's. Which 00:00:48.280 --> 00:00:50.340 is just staggering. That's not just a stock price 00:00:50.340 --> 00:00:52.119 thing. That's the market saying, we believe in 00:00:52.119 --> 00:00:54.740 your future more than the incumbents. And that 00:00:54.740 --> 00:00:57.149 future is all about the data center. Their numbers 00:00:57.149 --> 00:01:00.289 back it up. In 2024, they hit, what, over $25 00:01:00.289 --> 00:01:03.149 billion in revenue, around 28 ,000 employees. 00:01:03.509 --> 00:01:05.810 But the real story is where they're going, the 00:01:05.810 --> 00:01:09.730 AI space. Exactly. These massive deals, like 00:01:09.730 --> 00:01:12.370 the one with OpenAI announced in late... 2025. 00:01:12.950 --> 00:01:15.530 We're talking about a multi -billion dollar commitment 00:01:15.530 --> 00:01:19.230 for six gigawatts of AI processors. When you're 00:01:19.230 --> 00:01:21.849 making deals that big, you're not the underdog 00:01:21.849 --> 00:01:23.969 anymore. You're a pillar of the industry. Right. 00:01:24.030 --> 00:01:27.049 So our sources today give us that whole picture, 00:01:27.150 --> 00:01:29.890 the early days, the legal wars, the strategic 00:01:29.890 --> 00:01:32.390 moves. It all paints a picture of how they got 00:01:32.390 --> 00:01:34.030 here. Yeah. So let's start at the very beginning. 00:01:34.170 --> 00:01:35.930 So let's go right back to the beginning, the 00:01:35.930 --> 00:01:40.280 foundational years. Right. May 1st, 1969. And 00:01:40.280 --> 00:01:43.040 it's the classic Silicon Valley story. You've 00:01:43.040 --> 00:01:44.939 got Jerry Sanders and seven of his colleagues 00:01:44.939 --> 00:01:47.819 all frustrated leaving Fairchild Semiconductor. 00:01:47.920 --> 00:01:50.159 We hear about Fairchild all the time. Intel's 00:01:50.159 --> 00:01:51.920 founders, Noyce and Moore, they came from there, 00:01:51.920 --> 00:01:54.560 too, just a year earlier. What was going on at 00:01:54.560 --> 00:01:56.359 Fairchild that it was just bleeding all this 00:01:56.359 --> 00:01:59.560 top tier talent? It was a classic case of institutional 00:01:59.560 --> 00:02:02.760 friction. Fairchild was a giant, a pioneer, but 00:02:02.760 --> 00:02:05.260 it got slow. It got bureaucratic. And people 00:02:05.260 --> 00:02:07.640 like Jerry Sanders, who was famously charismatic 00:02:07.640 --> 00:02:10.439 and, well, flamboyant. He wasn't the type to 00:02:10.439 --> 00:02:12.979 wait for permission. He and his engineering team 00:02:12.979 --> 00:02:15.639 felt they could move faster, execute better, 00:02:15.819 --> 00:02:17.900 but the corporate structure was holding them 00:02:17.900 --> 00:02:21.639 back. They saw an opportunity and decided, like 00:02:21.639 --> 00:02:24.620 Intel, to just go build it themselves. And Sanders' 00:02:24.740 --> 00:02:27.259 whole philosophy was aggressive from day one, 00:02:27.419 --> 00:02:29.979 high performance, high risk. Absolutely. He had 00:02:29.979 --> 00:02:32.419 that thin quote about how you have to be willing 00:02:32.419 --> 00:02:35.180 to work hard and take calculated risks. And that 00:02:35.180 --> 00:02:37.360 really guided their first big strategic decision. 00:02:37.789 --> 00:02:40.569 Which wasn't about some revolutionary new chip. 00:02:40.729 --> 00:02:44.449 It was about quality control. Exactly. They decided 00:02:44.449 --> 00:02:46.509 to guarantee their products to the United States 00:02:46.509 --> 00:02:49.389 military standard, which for a startup was a 00:02:49.389 --> 00:02:51.669 huge and expensive promise to make. So they were 00:02:51.669 --> 00:02:53.530 basically eating their initial profit margins 00:02:53.530 --> 00:02:55.870 just to get a foot in the door. They were, but 00:02:55.870 --> 00:02:58.169 it was brilliant. You have to remember, in the 00:02:58.169 --> 00:03:01.090 late 60s, early 70s, microchips were just not 00:03:01.090 --> 00:03:03.240 reliable. That was the industry's... biggest 00:03:03.240 --> 00:03:05.259 problem. So if you're a big customer, especially 00:03:05.259 --> 00:03:07.800 the military, you can't have components failing. 00:03:08.020 --> 00:03:10.919 Right. So AMD comes in not with a brand new invention, 00:03:11.060 --> 00:03:14.259 but by taking existing logic chips and just promising 00:03:14.259 --> 00:03:16.139 they'll be rock solid. They guaranteed quality 00:03:16.139 --> 00:03:19.580 that built trust. It built a customer base and 00:03:19.580 --> 00:03:21.439 it gave them a revenue stream they could use 00:03:21.439 --> 00:03:23.400 to fund their own designs. So they established 00:03:23.400 --> 00:03:26.099 a beachhead on reliability and then they started 00:03:26.099 --> 00:03:28.509 to innovate. Precisely. Their very first product 00:03:28.509 --> 00:03:32.729 in 1970 was the AM9300, a shift register. But 00:03:32.729 --> 00:03:35.129 their first proprietary success came later that 00:03:35.129 --> 00:03:38.030 year with the AM2501 logic counter. And then 00:03:38.030 --> 00:03:41.810 into RAM in 71, the pace was pretty frantic. 00:03:42.009 --> 00:03:44.969 It was. By the end of 1971, they'd hit $4 .6 00:03:44.969 --> 00:03:47.129 million in sales. They were proving they could 00:03:47.129 --> 00:03:49.129 design and deliver. But the real game changer 00:03:49.129 --> 00:03:52.210 was 1975, when they decided to go head -to -head 00:03:52.210 --> 00:03:55.229 with Intel in the brand new microprocessor market. 00:03:55.469 --> 00:03:57.750 This was the big one. Intel had invented the 00:03:57.750 --> 00:04:00.330 microprocessor with the 4004 just a few years 00:04:00.330 --> 00:04:03.569 earlier. AMD jumped in with the AM9080. Which 00:04:03.569 --> 00:04:06.009 was a clone, right? A reverse -engineered Intel 00:04:06.009 --> 00:04:08.689 8080. It was. That was common practice for market 00:04:08.689 --> 00:04:10.889 entry. But it immediately set up this rivalry 00:04:10.889 --> 00:04:12.889 that would be defined by intellectual property. 00:04:13.169 --> 00:04:15.550 And that led to this incredibly important cross 00:04:15.550 --> 00:04:19.550 -licensing agreement in 1976. Why would Intel, 00:04:19.790 --> 00:04:22.149 the market leader, agree to license its tech 00:04:22.149 --> 00:04:24.829 to a competitor? They almost had to. It was because 00:04:24.829 --> 00:04:28.910 of a key technological change. Microcode. So 00:04:28.910 --> 00:04:31.629 microcode is basically software that lives on 00:04:31.629 --> 00:04:34.810 the chip itself. Exactly. It translates the complex 00:04:34.810 --> 00:04:37.730 instructions into signals the silicon can actually 00:04:37.730 --> 00:04:40.490 understand. And because it was software, it was 00:04:40.490 --> 00:04:43.029 protected by copyright. Ah, so for AMD to keep 00:04:43.029 --> 00:04:45.110 making compatible chips, they needed a legal 00:04:45.110 --> 00:04:47.910 license for that microcode. Precisely. The 1976 00:04:47.910 --> 00:04:50.740 agreement gave them that license. And in return, 00:04:50.980 --> 00:04:53.899 Intel got access to some of AMD's other designs. 00:04:54.079 --> 00:04:56.600 It was a detente, really, a necessary evil for 00:04:56.600 --> 00:04:58.560 both sides. An arrangement that somehow lasted 00:04:58.560 --> 00:05:01.540 all the way until 1995. It did. And it formed 00:05:01.540 --> 00:05:03.699 the legal foundation for the entire PC industry 00:05:03.699 --> 00:05:05.819 that was about to explode. And while that was 00:05:05.819 --> 00:05:07.759 going on, AMD kept growing. They had a joint 00:05:07.759 --> 00:05:09.939 venture with Siemens. They opened a new plant 00:05:09.939 --> 00:05:13.319 in Austin. And sales hit $100 million by 1978. 00:05:13.720 --> 00:05:15.220 They were getting ready. They were streamlined, 00:05:15.500 --> 00:05:17.839 capitalized, and positioned perfectly for the 00:05:17.839 --> 00:05:19.889 war that was coming. Okay, so if that first decade 00:05:19.889 --> 00:05:22.230 was them just, you know, getting their feet wet. 00:05:22.350 --> 00:05:24.430 This next part is where the real fight begins. 00:05:24.610 --> 00:05:27.470 It's basically survival. And it all comes down 00:05:27.470 --> 00:05:32.149 to one company, IBM. The IBM PC. When IBM decided 00:05:32.149 --> 00:05:36.069 to use Intel's x86 chips, they had one very strict 00:05:36.069 --> 00:05:38.189 rule. A non -negotiable rule. They would not 00:05:38.189 --> 00:05:40.810 bet their entire PC business on a single supplier. 00:05:41.230 --> 00:05:43.649 It was too risky. So they forced Intel to find 00:05:43.649 --> 00:05:46.970 a licensed second source manufacturer, and that 00:05:46.970 --> 00:05:50.350 manufacturer was AMD. It was. This led directly 00:05:50.350 --> 00:05:52.829 to that 10 -year technology exchange deal in 00:05:52.829 --> 00:05:56.889 1982. AMD was now officially licensed to make 00:05:56.889 --> 00:05:58.509 the chips that would power the PC revolution, 00:05:58.850 --> 00:06:01.589 the 8086, the 8088, and so on. They were indispensable, 00:06:01.930 --> 00:06:04.050 riding this massive wave. They were thriving. 00:06:04.459 --> 00:06:06.879 But inside Intel, the mood started to change 00:06:06.879 --> 00:06:10.319 by around 1984. They realized AMD wasn't just 00:06:10.319 --> 00:06:12.240 a backup supplier anymore. They were a competitor, 00:06:12.439 --> 00:06:14.519 a real one. And Intel decided they wanted out 00:06:14.519 --> 00:06:16.100 of the deal. And they showed it by their actions. 00:06:16.399 --> 00:06:18.319 When the next generation chip, the revolutionary 00:06:18.319 --> 00:06:22.160 8386, was ready, Intel stopped cooperating. They 00:06:22.160 --> 00:06:24.459 refused to share the technical details AMD needed 00:06:24.459 --> 00:06:27.100 to make it. Which was a direct breach of the 00:06:27.100 --> 00:06:29.439 contract. It was an existential threat to AMD. 00:06:29.860 --> 00:06:32.199 Totally. If they couldn't make the 386, they'd 00:06:32.199 --> 00:06:34.720 be stuck making obsolete hardware. They would 00:06:34.720 --> 00:06:37.860 become irrelevant overnight. So in 1987, they 00:06:37.860 --> 00:06:41.300 fought back. They invoked arbitration. And Intel's 00:06:41.300 --> 00:06:44.019 response was, well, the nuclear option. They 00:06:44.019 --> 00:06:46.860 just canceled the entire 1982 agreement. And 00:06:46.860 --> 00:06:49.199 that kicked off one of the longest, most bitter 00:06:49.199 --> 00:06:51.959 legal fights in tech history. The arbitration 00:06:51.959 --> 00:06:54.420 alone took five years. It finally came down in 00:06:54.420 --> 00:06:58.279 1992. And the ruling was in AMD's favor. The 00:06:58.279 --> 00:07:00.879 arbitrator agreed Intel had breached the contract. 00:07:01.180 --> 00:07:04.100 But Intel didn't stop there. They appealed. They 00:07:04.100 --> 00:07:05.800 fought it all the way to the Supreme Court of 00:07:05.800 --> 00:07:07.779 California, which finally upheld the decision 00:07:07.779 --> 00:07:12.139 in 1994. AMD was vindicated, but the damage was 00:07:12.139 --> 00:07:13.740 done. They couldn't just sit around and wait 00:07:13.740 --> 00:07:15.639 for the courts, though. That uncertainty forced 00:07:15.639 --> 00:07:18.120 them into a radical new strategy. This is the 00:07:18.120 --> 00:07:20.199 moment that saved the company. They couldn't 00:07:20.199 --> 00:07:21.779 wait for blueprints that might never arrive, 00:07:21.839 --> 00:07:23.420 so they had to start designing their own chips 00:07:23.420 --> 00:07:26.240 from scratch. That's a process called clean room 00:07:26.240 --> 00:07:28.850 design. We basically have engineers who have 00:07:28.850 --> 00:07:30.970 never seen the original design try to replicate 00:07:30.970 --> 00:07:33.449 its function just by observing how it works. 00:07:33.589 --> 00:07:35.230 It's incredibly difficult. You're building a 00:07:35.230 --> 00:07:39.410 compatible chip using your own totally new microcode 00:07:39.410 --> 00:07:41.709 and internal design. They were betting the entire 00:07:41.709 --> 00:07:44.209 company that their engineers were smart enough 00:07:44.209 --> 00:07:47.649 to pull it off. And they did. By 1991, they had 00:07:47.649 --> 00:07:52.209 the AM386. Their own clone of the Intel 386. 00:07:52.209 --> 00:07:54.810 And the market was starved for it. They sold 00:07:54.810 --> 00:07:57.709 a million units in just a few months. Then came 00:07:57.709 --> 00:08:01.589 the AM486 in 93, which was hugely popular with 00:08:01.589 --> 00:08:04.069 big companies like Compaq. This forced evolution 00:08:04.069 --> 00:08:05.990 turned them from a second source manufacturer 00:08:05.990 --> 00:08:09.110 into a true competitor. A homegrown competitor. 00:08:09.370 --> 00:08:11.910 And that led to their first... truly in -house 00:08:11.910 --> 00:08:15.470 CPU in 1996, the K5. And the name is a story 00:08:15.470 --> 00:08:18.329 in itself. The K stood for kryptonite. This is 00:08:18.329 --> 00:08:20.750 a direct shot at Intel, who everyone saw as Superman. 00:08:20.970 --> 00:08:22.970 It was AMD saying, we're the one thing that can 00:08:22.970 --> 00:08:25.050 hurt you. The competition was heating up. They 00:08:25.050 --> 00:08:27.350 bought a company called NextGen in 96 for almost 00:08:27.350 --> 00:08:30.589 a billion dollars. A huge acquisition. They basically 00:08:30.589 --> 00:08:33.149 bought the engineering talent they needed to 00:08:33.149 --> 00:08:35.789 create the K6 processor, which came out in 97 00:08:35.789 --> 00:08:39.220 and competed really well. against Intel's Pentium 00:08:39.220 --> 00:08:41.860 II Sec. But the big break came in 1999 with the 00:08:41.860 --> 00:08:44.700 Athlon, the K7. This is where they intentionally 00:08:44.700 --> 00:08:47.419 severed compatibility with Intel. Right. Because 00:08:47.419 --> 00:08:49.299 of all the legal issues around Intel's motherboard 00:08:49.299 --> 00:08:52.419 connectors, AMD just made their own, called Slot 00:08:52.419 --> 00:08:55.200 A. For the first time, you couldn't just swap 00:08:55.200 --> 00:08:57.379 an AMD chip for an Intel one. You had to choose 00:08:57.379 --> 00:09:02.080 a side. A huge strategic risk? It was. But while 00:09:02.080 --> 00:09:04.820 this CPU war was raging, they had another big 00:09:04.820 --> 00:09:07.309 business running in the background. Flash memory. 00:09:07.509 --> 00:09:09.269 And they decided to sell it off, which seems 00:09:09.269 --> 00:09:11.389 counterintuitive. Why sell a profitable business? 00:09:11.809 --> 00:09:13.929 Because the war with Intel was costing them a 00:09:13.929 --> 00:09:16.769 fortune. They needed cash and they needed focus. 00:09:17.529 --> 00:09:20.409 In 2003, they spun off their flash memory division 00:09:20.409 --> 00:09:22.929 into a joint venture called Spansion. So they 00:09:22.929 --> 00:09:25.450 cashed out a huge asset to fuel the main fight. 00:09:25.649 --> 00:09:27.970 Exactly. They sold off the rest of it by 2005. 00:09:28.389 --> 00:09:30.570 Yeah. It was a painful but necessary move to 00:09:30.570 --> 00:09:32.990 pour every single resource they had into winning 00:09:32.990 --> 00:09:35.269 the microprocessor war. They knew they couldn't 00:09:35.269 --> 00:09:37.490 fight on two fronts at once. So after all that, 00:09:37.610 --> 00:09:41.110 they're laser focused. And then in 2003, they 00:09:41.110 --> 00:09:43.490 pull off one of their biggest wins ever with 00:09:43.490 --> 00:09:46.230 the K8 architecture. K8 was, I mean, it wasn't 00:09:46.230 --> 00:09:48.409 just them catching up. This was a genuine leap 00:09:48.409 --> 00:09:51.450 forward. It introduced three really fundamental 00:09:51.450 --> 00:09:54.090 innovations. First was 64 -bit computing. The 00:09:54.090 --> 00:09:57.309 Bi -86 -64 instruction set, which they branded 00:09:57.309 --> 00:10:02.519 AMD 64. This was a huge deal. broke the 4GB memory 00:10:02.519 --> 00:10:05.100 barrier that 32 -bit chips had, which was becoming 00:10:05.100 --> 00:10:07.379 a serious bottleneck. And Intel didn't have it. 00:10:07.500 --> 00:10:10.340 They didn't. Intel was forced to adopt AMD's 00:10:10.340 --> 00:10:12.399 standard, which was a massive symbolic victory 00:10:12.399 --> 00:10:15.240 for AMD. Okay, so that's one. What was the second 00:10:15.240 --> 00:10:17.299 big innovation? They moved the memory controller 00:10:17.299 --> 00:10:19.600 from the motherboard directly onto the CPU die 00:10:19.600 --> 00:10:21.919 itself. Which sounds technical, but what does 00:10:21.919 --> 00:10:24.519 that actually do? It drastically reduces latency. 00:10:24.840 --> 00:10:27.539 Before, the CPU had to go off -chip to talk to 00:10:27.539 --> 00:10:30.320 the memory, which is slow. By putting the controller 00:10:30.320 --> 00:10:33.679 right on the CPU, data access became way faster. 00:10:33.799 --> 00:10:35.980 The whole system felt quicker. So why didn't 00:10:35.980 --> 00:10:38.279 Intel do that first? They were kind of locked 00:10:38.279 --> 00:10:40.580 into their existing architecture. Changing it 00:10:40.580 --> 00:10:42.299 would have been a huge expensive headache for 00:10:42.299 --> 00:10:45.500 them and all their partners. AMD, as the challenger, 00:10:45.620 --> 00:10:47.779 could afford to take a bigger risk on a new design. 00:10:48.059 --> 00:10:51.120 And the third piece was hyper transport. Right. 00:10:51.179 --> 00:10:53.700 A super fast connection that let all the components 00:10:53.700 --> 00:10:56.240 talk to each other. It was all part of this cohesive 00:10:56.240 --> 00:10:59.029 vision for a more efficient system. And the products 00:10:59.029 --> 00:11:01.649 they launched, the Opteron for servers and Athlon 00:11:01.649 --> 00:11:04.289 64 for desktops, were incredibly successful. 00:11:04.549 --> 00:11:08.669 For a while there, around 2005, AMD was the undisputed 00:11:08.669 --> 00:11:11.929 technology leader. They launched the first dual 00:11:11.929 --> 00:11:15.529 -core CPUs for both servers and desktops. They 00:11:15.529 --> 00:11:17.690 were on top of the world. But then they made 00:11:17.690 --> 00:11:21.590 another huge risky move. In 2006, they bought 00:11:21.590 --> 00:11:25.009 the graphics card company ATI for $5 .4 billion. 00:11:25.409 --> 00:11:28.429 A massive, massive gamble and one that nearly 00:11:28.429 --> 00:11:30.610 bankrupted them when the financial crisis hit 00:11:30.610 --> 00:11:32.590 a couple of years later. Was it worth it? All 00:11:32.590 --> 00:11:35.190 that debt? In the long run, absolutely. It was 00:11:35.190 --> 00:11:37.769 essential. ATI was a graphics giant with their 00:11:37.769 --> 00:11:41.200 Radeon line. By buying them, AMD suddenly had 00:11:41.200 --> 00:11:44.480 top -tier IP in both CPUs and GPUs. And that's 00:11:44.480 --> 00:11:47.200 where the Fusion idea came from, the APU. Exactly, 00:11:47.360 --> 00:11:49.840 the Accelerated Processing Unit. The vision was 00:11:49.840 --> 00:11:52.860 to put the CPU and the GPU on a single piece 00:11:52.860 --> 00:11:55.799 of silicon. This would be cheaper, more power 00:11:55.799 --> 00:11:58.539 efficient, and way faster because they could 00:11:58.539 --> 00:12:01.799 share data directly. A vision that would, as 00:12:01.799 --> 00:12:04.220 we'll see, eventually save the company. But before 00:12:04.220 --> 00:12:06.620 that, they made another huge change. They stopped 00:12:06.620 --> 00:12:09.490 making their own chips. They went fabless. It 00:12:09.490 --> 00:12:12.570 was a matter of survival. By 2008, building a 00:12:12.570 --> 00:12:16.149 new fab, a fabrication plant, cost a truly astronomical 00:12:16.149 --> 00:12:18.889 amount of money. AMD just couldn't compete with 00:12:18.889 --> 00:12:21.029 Intel's spending power. So they spun off their 00:12:21.029 --> 00:12:23.450 manufacturing arm into a new company, Global 00:12:23.450 --> 00:12:26.110 Foundries. They did. This let them focus all 00:12:26.110 --> 00:12:28.789 their money and brainpower on one thing, chip 00:12:28.789 --> 00:12:31.870 design. It was a huge philosophical shift, but 00:12:31.870 --> 00:12:33.769 a necessary one. But right after all these big 00:12:33.769 --> 00:12:37.129 moves, things went south, fast. The bulldozer 00:12:37.129 --> 00:12:39.690 architecture in 2011 was a disaster. It really 00:12:39.690 --> 00:12:41.710 was. It was a totally new design that was supposed 00:12:41.710 --> 00:12:43.669 to be revolutionary, but it just didn't perform. 00:12:43.909 --> 00:12:46.970 In real world tests, it fell way behind what 00:12:46.970 --> 00:12:49.730 Intel was offering. And a bad product leads to 00:12:49.730 --> 00:12:52.669 corporate chaos. It did. There were CEO changes, 00:12:52.889 --> 00:12:56.730 massive layoffs in 2011, 2012, 2014. The company 00:12:56.730 --> 00:12:58.889 was bleeding money and looked like it might not 00:12:58.889 --> 00:13:01.149 survive. And the sources say there was one thing 00:13:01.149 --> 00:13:03.289 that kept them afloat during this dark period. 00:13:03.549 --> 00:13:06.029 The semi -custom business. This is where that 00:13:06.029 --> 00:13:09.210 eye -wateringly expensive ATI acquisition paid 00:13:09.210 --> 00:13:12.389 off. They could combine their CPU and GPU technology 00:13:12.389 --> 00:13:16.429 into a single custom chip. A custom APU. And 00:13:16.429 --> 00:13:18.289 they sold those chips to Sony for the PlayStation 00:13:18.289 --> 00:13:21.730 4 and to Microsoft for the Xbox One. So the gaming 00:13:21.730 --> 00:13:24.789 consoles. Yes. Both consoles, launched in 2013, 00:13:25.190 --> 00:13:28.490 ran on custom AMD hardware. That business provided 00:13:28.490 --> 00:13:31.090 a steady, predictable, high -volume stream of 00:13:31.090 --> 00:13:34.429 cash that acted as a financial lifeboat. It kept 00:13:34.429 --> 00:13:37.090 them alive and, crucially, gave them the money 00:13:37.090 --> 00:13:39.129 they needed to fund their next big bet. So those 00:13:39.129 --> 00:13:41.049 console deals, they were basically a lifeline. 00:13:41.070 --> 00:13:43.049 They bought them time. And capital. The capital 00:13:43.049 --> 00:13:45.149 they needed for what was basically a bet -the 00:13:45.149 --> 00:13:48.110 -company project. Zen. An entirely new architecture, 00:13:48.350 --> 00:13:50.730 built from the ground up, led by the legendary 00:13:50.730 --> 00:13:53.230 engineer Jim Keller. The pressure must have been 00:13:53.230 --> 00:13:56.169 incredible. Immense. They set this staggering 00:13:56.169 --> 00:13:59.889 goal for themselves. A 40 % increase in IPC. 00:14:00.460 --> 00:14:03.019 instructions per cycle, over their last generation 00:14:03.019 --> 00:14:06.240 chip. An IPC is the real measure of a processor's 00:14:06.240 --> 00:14:08.460 efficiency, how much work it gets done per clock 00:14:08.460 --> 00:14:12.000 tick. Exactly. And 40 % is an almost unheard 00:14:12.000 --> 00:14:14.120 of generational leap. And they didn't hit it. 00:14:14.259 --> 00:14:17.220 No, they blew past it. In 2017, they announced 00:14:17.220 --> 00:14:21.340 they had achieved a 52 % IPC increase. 52! That's 00:14:21.340 --> 00:14:23.960 just, that's a revolution, not an evolution. 00:14:24.340 --> 00:14:26.100 It put them right back in the game. It was a 00:14:26.100 --> 00:14:28.620 fundamental technological triumph. It showed 00:14:28.620 --> 00:14:30.460 that all the pain and restructuring had paid 00:14:30.460 --> 00:14:33.340 off. So what was the magic? What made ZEN so 00:14:33.340 --> 00:14:35.840 good so fast? It was a combination of things. 00:14:35.960 --> 00:14:38.399 They moved to a much more efficient 14 nanometer 00:14:38.399 --> 00:14:40.840 manufacturing process. They introduced simultaneous 00:14:40.840 --> 00:14:43.879 multithreading, or SMT, which is their version 00:14:43.879 --> 00:14:46.379 of Intel's hyperthreading. So one physical core 00:14:46.379 --> 00:14:48.840 can act like two virtual ones. Much better for 00:14:48.840 --> 00:14:51.340 multitasking. Right. But the real secret sauce 00:14:51.340 --> 00:14:54.519 was the move to a modular, chiplet -based design. 00:14:54.799 --> 00:14:57.759 Okay, so instead of making one giant, monolithic 00:14:57.759 --> 00:14:59.799 chip, which is expensive and hard to do without 00:14:59.799 --> 00:15:02.580 defects... They started making smaller, modular 00:15:02.580 --> 00:15:05.259 chiplets, which they called core complex dyes, 00:15:05.299 --> 00:15:08.019 or CCDs. These are much cheaper and easier to 00:15:08.019 --> 00:15:09.740 manufacture with high yields. And they connect 00:15:09.740 --> 00:15:11.740 these little chiplets together to build a bigger 00:15:11.740 --> 00:15:14.460 processor. Exactly. Using a super -fast interconnect 00:15:14.460 --> 00:15:17.309 they called Infinity Fabric. This is the key. 00:15:17.429 --> 00:15:20.149 It allowed them to scale up core counts for their 00:15:20.149 --> 00:15:23.269 server chips called Epic, much more cheaply and 00:15:23.269 --> 00:15:25.649 efficiently than Intel could with its monolithic 00:15:25.649 --> 00:15:28.509 designs. A massive competitive advantage in the 00:15:28.509 --> 00:15:30.610 data center. A huge one. And it powered their 00:15:30.610 --> 00:15:33.169 new consumer lineup, too, the Ryzen chips. And 00:15:33.169 --> 00:15:35.889 the market responded immediately. It was a swift 00:15:35.889 --> 00:15:40.279 and brutal comeback. By 2019, AMD's Ryzen desktop 00:15:40.279 --> 00:15:42.779 processors were actually outselling Intel's. 00:15:42.899 --> 00:15:45.299 For the first time in years, they were winning 00:15:45.299 --> 00:15:47.059 the consumer market. And they kept the pressure 00:15:47.059 --> 00:15:50.860 on. Zen 2 in 2019, Zen 3 in 2020. They just kept 00:15:50.860 --> 00:15:53.139 executing. Flawlessly. And they weren't just 00:15:53.139 --> 00:15:55.940 winning on core count anymore. By 2020, with 00:15:55.940 --> 00:15:59.200 chips like the Ryzen 5 5600X, they were competing 00:15:59.200 --> 00:16:01.759 and sometimes even beating Intel on single -threaded 00:16:01.759 --> 00:16:04.240 performance, which had always been Intel's stronghold. 00:16:04.519 --> 00:16:07.000 And all the while, that semi -custom business, 00:16:07.259 --> 00:16:09.779 the console lifeline, was still going strong. 00:16:09.980 --> 00:16:12.360 Oh yeah. The new generation of consoles, the 00:16:12.360 --> 00:16:14.720 PlayStation 5, the Xbox Series X and S, even 00:16:14.720 --> 00:16:16.899 the Scheme Deck, they all run on custom chips 00:16:16.899 --> 00:16:19.940 based on the Zen 2 architecture. It's the perfect 00:16:19.940 --> 00:16:22.759 dual strategy. Predictable console money funding 00:16:22.759 --> 00:16:25.159 the high -margin, high -performance CPU war. 00:16:25.940 --> 00:16:29.100 And that Zen comeback, led by Lisa Su, it just 00:16:29.100 --> 00:16:30.899 completely changed the game. It gave them the 00:16:30.899 --> 00:16:33.610 foundation for, well... For what's happening 00:16:33.610 --> 00:16:36.289 right now. This massive, almost aggressive pivot 00:16:36.289 --> 00:16:39.330 into the data center, into AI. And to do it, 00:16:39.330 --> 00:16:42.129 they started making some truly monumental acquisitions, 00:16:42.149 --> 00:16:45.289 starting with Zilinx in 2022. The price tag on 00:16:45.289 --> 00:16:47.950 that was just mind -boggling. $50 billion. That's 00:16:47.950 --> 00:16:49.309 one of the biggest deals in the history of the 00:16:49.309 --> 00:16:51.230 industry. What on earth did they get for $50 00:16:51.230 --> 00:16:53.549 billion? They got market leadership in a very 00:16:53.549 --> 00:16:56.129 specific, very important type of chip, FPGAs, 00:16:56.230 --> 00:16:59.399 field programmable gateways. So for anyone listening, 00:16:59.620 --> 00:17:02.720 what makes an FPGA so valuable in the world of 00:17:02.720 --> 00:17:08.440 AI? Flexibility. A normal CPU or GPU, its function 00:17:08.440 --> 00:17:11.359 is fixed when it leaves the factory. An FPGA 00:17:11.359 --> 00:17:13.759 can be reconfigured, reprogrammed in the field 00:17:13.759 --> 00:17:17.079 after it's been made. So you can create custom 00:17:17.079 --> 00:17:20.880 hardware pipelines for new... AI models or specialized 00:17:20.880 --> 00:17:22.960 networking tasks that are changing all the time. 00:17:23.059 --> 00:17:26.299 Precisely. It gives data centers incredible adaptability. 00:17:26.559 --> 00:17:30.019 So AMD didn't just buy a product. They bought 00:17:30.019 --> 00:17:33.099 the ability to offer customized adaptable silicon 00:17:33.099 --> 00:17:36.359 for any future workload. It's a huge strategic 00:17:36.359 --> 00:17:38.779 advantage. And then right after that, another 00:17:38.779 --> 00:17:41.619 acquisition, Pensando for $1 .9 billion. This 00:17:41.619 --> 00:17:43.619 was for something called a DPU. A data processing 00:17:43.619 --> 00:17:45.339 unit. This is another really smart strategic 00:17:45.339 --> 00:17:47.980 move. A DPU is basically a smart network card 00:17:47.980 --> 00:17:50.400 that offloads tasks like security and storage 00:17:50.400 --> 00:17:52.359 from the main CPU. So it makes the whole server 00:17:52.359 --> 00:17:54.259 more efficient. Exactly. It shows that they're 00:17:54.259 --> 00:17:56.039 not just thinking about selling a CPU anymore. 00:17:56.200 --> 00:17:58.299 They're thinking about selling an entire optimized 00:17:58.299 --> 00:18:00.440 data center platform. And to support all this 00:18:00.440 --> 00:18:02.140 new hardware, they went on a software buying 00:18:02.140 --> 00:18:04.980 spree. This is maybe the most critical part of 00:18:04.980 --> 00:18:08.109 their plan to take on NVIDIA. NVIDIA's dominance 00:18:08.109 --> 00:18:10.930 isn't just its hardware. It's the software ecosystem, 00:18:11.410 --> 00:18:15.250 CUDA, that locks developers in. So AMD's strategy 00:18:15.250 --> 00:18:17.470 is to build a powerful open -source alternative. 00:18:17.990 --> 00:18:23.029 Yes. They bought Nod .ai in 2023, then Silo .ai 00:18:23.029 --> 00:18:26.829 for over $600 million in 2024, then a whole string 00:18:26.829 --> 00:18:30.549 of smaller AI software companies in 2025. They 00:18:30.549 --> 00:18:32.710 are buying the talent and the technology needed 00:18:32.710 --> 00:18:35.150 to make their hardware easy to program for AI. 00:18:35.390 --> 00:18:37.250 They even bought a hardware company, ZT Systems. 00:18:37.589 --> 00:18:40.569 For $4 .9 billion. This shows you the full vertical 00:18:40.569 --> 00:18:42.250 integration play. They're moving beyond just 00:18:42.250 --> 00:18:45.069 selling chips to selling entire fully customized 00:18:45.069 --> 00:18:47.329 server racks and data center hardware solutions. 00:18:47.509 --> 00:18:49.809 A one -stop shop. And all of this leads up to 00:18:49.809 --> 00:18:52.910 that bombshell announcement in late 2025, the 00:18:52.910 --> 00:18:55.430 OpenAI deal. This was the ultimate validation 00:18:55.430 --> 00:18:58.099 of the entire strategy. OpenAI, arguably the 00:18:58.099 --> 00:19:00.220 most important AI company in the world, committing 00:19:00.220 --> 00:19:03.359 to buy six gigawatts of AMD's next -gen AI processors 00:19:03.359 --> 00:19:05.640 over five years. A deal that will be worth tens 00:19:05.640 --> 00:19:07.940 of billions of dollars. It instantly makes AMD 00:19:07.940 --> 00:19:10.700 a true co -equal competitor to NVIDIA in the 00:19:10.700 --> 00:19:14.019 AI space. And the structure of the deal is just 00:19:14.019 --> 00:19:16.660 fascinating. It includes warrants for OpenAI 00:19:16.660 --> 00:19:20.509 to buy up to a 10 % stake in AMD. But only if 00:19:20.509 --> 00:19:22.970 AMD's stock price hits an incredibly ambitious 00:19:22.970 --> 00:19:26.549 target of $600 a share. It's an incredible show 00:19:26.549 --> 00:19:29.450 of confidence. It directly ties the success of 00:19:29.450 --> 00:19:31.930 AMD's hardware to the company's future market 00:19:31.930 --> 00:19:34.390 value in a way we've never really seen before. 00:19:34.710 --> 00:19:37.190 But amid all this AI excitement, what about the 00:19:37.190 --> 00:19:40.170 original fight, the server CPU market against 00:19:40.170 --> 00:19:42.329 Intel? They haven't taken their eye off that 00:19:42.329 --> 00:19:44.569 ball. Thanks to the success of Epic, they have 00:19:44.569 --> 00:19:47.750 been clawing back market share steadily. As of 00:19:47.750 --> 00:19:51.269 mid -2025, they hold about 36 .5 % of the server 00:19:51.269 --> 00:19:54.089 CPU market. Which is a huge number. Intel still 00:19:54.089 --> 00:19:57.630 has the lead, but 36 .5 % is a serious, serious 00:19:57.630 --> 00:20:00.990 position to be in. The comeback is real. So to 00:20:00.990 --> 00:20:02.450 really get the full picture of where they are 00:20:02.450 --> 00:20:04.329 today, you have to look at the whole ecosystem 00:20:04.329 --> 00:20:06.190 they've built. Right. It's not just about one 00:20:06.190 --> 00:20:08.150 chip anymore. It's the hardware, the software, 00:20:08.289 --> 00:20:10.089 everything working together. Let's start with 00:20:10.089 --> 00:20:12.430 the graphics side, the Radeon Technologies Group. 00:20:12.650 --> 00:20:15.430 They spun that out as its own unit in 2015 to 00:20:15.430 --> 00:20:18.369 give it more focus. They did. And its architecture 00:20:18.369 --> 00:20:20.849 has evolved dramatically. They moved from their 00:20:20.849 --> 00:20:23.250 older designs to what they called Graphics Core 00:20:23.250 --> 00:20:26.470 Next, or GCN. And GCN was a big philosophical 00:20:26.470 --> 00:20:29.450 shift, wasn't it? It was. It was all about using 00:20:29.450 --> 00:20:32.130 the GPU for more than just graphics, for general 00:20:32.130 --> 00:20:35.069 -purpose computing, GPGPU. Which positioned them 00:20:35.069 --> 00:20:36.990 perfectly for the machine learning room that 00:20:36.990 --> 00:20:39.230 was just around the corner. Exactly. And their 00:20:39.230 --> 00:20:43.089 newest architecture is RDNA. The RDNA2 version... 00:20:43.309 --> 00:20:45.650 which came out in 2020, was their first real 00:20:45.650 --> 00:20:49.309 high -end competitor to NVIDIA's RTX cards with 00:20:49.309 --> 00:20:51.450 full support for hardware -based ray tracing. 00:20:51.630 --> 00:20:53.869 It shows they're still serious about the consumer 00:20:53.869 --> 00:20:56.269 gaming market, not just the data center. Absolutely. 00:20:56.450 --> 00:20:58.609 And all of this ties back to that core concept 00:20:58.609 --> 00:21:00.869 we talked about earlier, the APU. The Accelerated 00:21:00.869 --> 00:21:03.609 Processing Unit. The CPU and GPU on one piece 00:21:03.609 --> 00:21:05.829 of silicon. It's the foundation of so much of 00:21:05.829 --> 00:21:08.509 their success. It simplifies design, it saves 00:21:08.509 --> 00:21:11.589 power, and it boosts performance. Because the 00:21:11.589 --> 00:21:14.130 CPU and GPU can communicate directly through 00:21:14.130 --> 00:21:15.730 that high -speed Infinity Fabric we mentioned. 00:21:15.990 --> 00:21:18.410 No more slow trips across the motherboard. Right. 00:21:18.490 --> 00:21:20.369 And beyond the core chips, they've innovated 00:21:20.369 --> 00:21:23.609 in other areas too, like FreeSync. Which synchronizes 00:21:23.609 --> 00:21:25.650 your monitor's refresh rate to your graphics 00:21:25.650 --> 00:21:28.079 card, getting rid of screen tearing. And they 00:21:28.079 --> 00:21:31.460 made it an open standard VI adaptive sync so 00:21:31.460 --> 00:21:33.759 anyone could use it without paying for expensive 00:21:33.759 --> 00:21:36.099 proprietary hardware. Which brings us back to 00:21:36.099 --> 00:21:38.640 their software strategy. It's all about open 00:21:38.640 --> 00:21:42.160 source. It's a huge deliberate choice. A direct 00:21:42.160 --> 00:21:46.180 contrast to NVIDIA's closed off CUDA ecosystem. 00:21:46.599 --> 00:21:49.119 They have GPU open for their graphics tools, 00:21:49.299 --> 00:21:52.480 things like FidelityFX, Super Resolution or FSR. 00:21:52.859 --> 00:21:55.400 FSR is that upscaling technology that gives you 00:21:55.400 --> 00:21:58.359 higher frame rates with almost no loss in visual 00:21:58.359 --> 00:22:01.500 quality. It is. And for the AI and supercomputing 00:22:01.500 --> 00:22:04.720 side, they have ROCM, the Radeon Open Compute 00:22:04.720 --> 00:22:06.859 Platform. So the idea is to attract developers 00:22:06.859 --> 00:22:09.299 who don't want to be locked into one company's 00:22:09.299 --> 00:22:12.019 proprietary system. That's the gamble. Build 00:22:12.019 --> 00:22:14.420 an open, powerful platform, and the developers 00:22:14.420 --> 00:22:16.359 will come. And if the developers come, they'll 00:22:16.359 --> 00:22:18.920 drive demand for AMD's hardware. It's a long 00:22:18.920 --> 00:22:20.940 -term play. They're even making their own open 00:22:20.940 --> 00:22:23.750 -source AI models now. That's right. In early 00:22:23.750 --> 00:22:27.450 2025, they announced Instella, a 3 billion parameter 00:22:27.450 --> 00:22:30.349 language model. It's another signal that they 00:22:30.349 --> 00:22:33.329 are all in on open standards for AI. We can't 00:22:33.329 --> 00:22:34.849 finish talking about their engineering culture 00:22:34.849 --> 00:22:37.369 without mentioning one fun fact. Ah, the Guinness 00:22:37.369 --> 00:22:39.970 World Records. Pure bragging rights. they hold 00:22:39.970 --> 00:22:42.230 the record for the highest frequency ever achieved 00:22:42.230 --> 00:22:45.809 by a computer processor they do they hit 8 .794 00:22:45.809 --> 00:22:49.789 gigahertz with their old fx 8350 processor using 00:22:49.789 --> 00:22:52.789 liquid helium for cooling of course it's just 00:22:52.789 --> 00:22:54.930 a little symbol that even when the company was 00:22:54.930 --> 00:22:57.910 struggling for its life that core drive for raw 00:22:57.910 --> 00:23:00.319 performance never went away So when you look 00:23:00.319 --> 00:23:03.440 back at it all, from what, 1969 with a few frustrated 00:23:03.440 --> 00:23:06.779 engineers. To 2025, signing these multi -billion 00:23:06.779 --> 00:23:09.059 dollar deals to power the future of AI. It's 00:23:09.059 --> 00:23:10.940 an incredible story. And the biggest takeaway 00:23:10.940 --> 00:23:13.740 for you, listening, has to be that none of this 00:23:13.740 --> 00:23:15.779 was guaranteed. It was earned through a series 00:23:15.779 --> 00:23:18.539 of huge calculated risks. Every step of the way, 00:23:18.619 --> 00:23:20.619 the legal fight with Intel forced them to become 00:23:20.619 --> 00:23:23.660 innovators. The massive debt -fueled ATI acquisition 00:23:23.660 --> 00:23:26.359 gave them the IP they needed to survive. The 00:23:26.359 --> 00:23:28.559 shift to a fabulous model freed up the cash they 00:23:28.559 --> 00:23:32.000 needed for Zen. And that Zen comeback under Lisa 00:23:32.000 --> 00:23:34.299 Su's leadership, it just changed everything. 00:23:34.640 --> 00:23:38.039 It set the stage for this massive pivot into 00:23:38.039 --> 00:23:41.559 AI with acquisitions like Zilinx. Their survival 00:23:41.559 --> 00:23:44.500 has always depended on two things. Making big 00:23:44.500 --> 00:23:47.420 architectural leaps, like with AMD64 back in 00:23:47.420 --> 00:23:50.220 the K8 days, and making smart business pivots, 00:23:50.220 --> 00:23:52.440 like focusing on the console market when the 00:23:52.440 --> 00:23:55.099 PC market was weak. It's a masterclass in adaptation. 00:23:55.720 --> 00:23:58.500 Which brings us to our final thought. AMD's massive 00:23:58.500 --> 00:24:01.099 deal with OpenAI is tied to its stock hitting 00:24:01.099 --> 00:24:04.480 an ambitious $600 a share. To get there, they 00:24:04.480 --> 00:24:07.339 have to become an undeniable leader in AI. And 00:24:07.339 --> 00:24:09.980 they've spent $50 billion on Xilinx to get access 00:24:09.980 --> 00:24:12.539 to those flexible, reconfigurable FPGA chips. 00:24:12.819 --> 00:24:14.920 Their entire software strategy is built around 00:24:14.920 --> 00:24:17.240 being open and adaptable. So here's the question 00:24:17.240 --> 00:24:19.319 for you to think about. For the last 30 years, 00:24:19.420 --> 00:24:20.880 high -performance computing has been defined 00:24:20.880 --> 00:24:23.920 by the battle between standardized CPUs and GPUs. 00:24:23.940 --> 00:24:26.299 But AMD's massive investment in customizable 00:24:26.299 --> 00:24:29.880 FPGAs suggests a different future. Is the future 00:24:29.880 --> 00:24:31.660 of computing going to be another round of the 00:24:31.660 --> 00:24:34.980 same... CPU versus GPU fight, or is the market 00:24:34.980 --> 00:24:36.920 shifting to a place where victory belongs to 00:24:36.920 --> 00:24:39.099 whoever can build the most specialized, rapidly 00:24:39.099 --> 00:24:42.359 adaptable, custom silicon for every single new 00:24:42.359 --> 00:24:45.000 AI workload? If that's the future, AMD might 00:24:45.000 --> 00:24:46.579 be betting that the era of the general purpose 00:24:46.579 --> 00:24:47.960 processor is coming to an end.