WEBVTT 00:00:00.000 --> 00:00:04.980 Intro music. On March 2nd, 2026, over $7 billion 00:00:04.980 --> 00:00:09.039 quietly vanished into one narrow, highly specific 00:00:09.039 --> 00:00:12.500 corner of the tech supply chain. No flashy AI 00:00:12.500 --> 00:00:17.320 startup. No new LLM. Just glass, light, and, 00:00:17.519 --> 00:00:21.510 uh... Welcome to your custom deep dive. Today, 00:00:21.670 --> 00:00:23.550 we are exploring a fundamental infrastructure 00:00:23.550 --> 00:00:25.910 shift happening right now beneath the surface 00:00:25.910 --> 00:00:28.089 of the AI boom. We're looking at a definitive 00:00:28.089 --> 00:00:31.600 March 2026 guide. detailing the massive pivot 00:00:31.600 --> 00:00:34.259 away from traditional copper wiring. Right. First, 00:00:34.380 --> 00:00:36.159 we'll unpack the physical limits of what engineers 00:00:36.159 --> 00:00:38.359 are calling the copper wall. And next, we'll 00:00:38.359 --> 00:00:40.939 explore the mind bending, elegant solution of 00:00:40.939 --> 00:00:43.259 silicon photonics. Finally, we'll break down 00:00:43.259 --> 00:00:45.020 the three layer framework of companies actually 00:00:45.020 --> 00:00:47.740 building this new light speed ecosystem. So let's 00:00:47.740 --> 00:00:49.920 dig in. The best place to start is tracking that 00:00:49.920 --> 00:00:52.399 massive capital flight because, you know, money 00:00:52.399 --> 00:00:54.060 of that magnitude does not move by accident. 00:00:54.079 --> 00:00:56.979 Absolutely not. In early March, NVIDIA quietly 00:00:56.979 --> 00:01:00.219 took $4 billion and split it evenly between two 00:01:00.219 --> 00:01:02.979 relatively unknown hardware companies, Coherent 00:01:02.979 --> 00:01:06.420 and Lumentum. $2 billion each. Wow. And this 00:01:06.420 --> 00:01:09.000 wasn't speculative venture capital. These were 00:01:09.000 --> 00:01:12.359 backed by massive multi -year purchase commitments. 00:01:12.680 --> 00:01:15.500 Yeah, and Marvell Technology made similar aggressive 00:01:15.500 --> 00:01:18.959 moves at the exact same time. They spent nearly 00:01:18.959 --> 00:01:24.040 $3 .8 billion acquiring optical companies. I 00:01:24.040 --> 00:01:27.340 mean, they closed a $3 .25 billion deal just 00:01:27.340 --> 00:01:30.640 for a single startup called Celestial AI. And 00:01:30.640 --> 00:01:33.120 then threw in another $540 million deal for a 00:01:33.120 --> 00:01:35.689 networking company. Right. When you see the most 00:01:35.689 --> 00:01:38.189 dominant players in artificial intelligence aggressively 00:01:38.189 --> 00:01:40.750 locking up the supply chain for a specific type 00:01:40.750 --> 00:01:43.750 of hardware, it points to a much deeper shift. 00:01:44.030 --> 00:01:47.090 It really does. This capital flight marks the 00:01:47.090 --> 00:01:49.609 definitive end of the copper era in high performance 00:01:49.609 --> 00:01:52.040 computing. The industry hasn't hit a software 00:01:52.040 --> 00:01:54.420 bottleneck. It has hit a hard physical limit 00:01:54.420 --> 00:01:57.659 of nature. The copper wall. Yeah. Let's unpack 00:01:57.659 --> 00:01:59.379 the mechanics of this wall because this is where 00:01:59.379 --> 00:02:01.439 the physics get incredibly fascinating. Let's 00:02:01.439 --> 00:02:04.239 do it. The bottleneck holding back the next generation 00:02:04.239 --> 00:02:07.219 of AI is not actually computing power anymore. 00:02:07.379 --> 00:02:09.620 We know how to make faster chips. The problem 00:02:09.620 --> 00:02:12.020 is how fast those chips can talk to each other. 00:02:12.719 --> 00:02:16.000 Every modern AI data center is essentially a 00:02:16.000 --> 00:02:19.389 vast, sprawling city of silicon. You have tens 00:02:19.389 --> 00:02:22.789 of thousands of GTUs constantly exchanging data 00:02:22.789 --> 00:02:25.889 every single second to train these massive models. 00:02:26.150 --> 00:02:28.550 And right now, the speed of that data transfer 00:02:28.550 --> 00:02:32.050 is pushing toward a critical threshold, 1 .6 00:02:32.050 --> 00:02:35.509 terabits per second, or TBPS. Okay. At that specific 00:02:35.509 --> 00:02:38.469 speed, traditional copper cables literally begin 00:02:38.469 --> 00:02:40.550 to break down. And they don't fail gracefully, 00:02:40.650 --> 00:02:43.110 right? Or in just one way? No, not at all. They 00:02:43.110 --> 00:02:46.030 fail in three distinct ways all at once. Engineers 00:02:46.030 --> 00:02:48.830 call it the three -threat failure. heat, power, 00:02:49.050 --> 00:02:51.110 and signal quality. I was thinking about this 00:02:51.110 --> 00:02:53.189 earlier. It's like trying to force a high -pressure 00:02:53.189 --> 00:02:56.009 fire hose of data through a garden hose made 00:02:56.009 --> 00:02:58.629 of copper. Exactly. But worse, because of electrical 00:02:58.629 --> 00:03:00.750 resistance, the water in this hose isn't just 00:03:00.750 --> 00:03:02.889 backing up. It is boiling. Yeah, the friction 00:03:02.889 --> 00:03:05.610 of the electrons moving at 1 .6 terabits physically 00:03:05.610 --> 00:03:07.509 threatens to melt the hardware. And the power 00:03:07.509 --> 00:03:10.270 demands explode non -linearly. The signal just 00:03:10.270 --> 00:03:14.139 degrades into pure static noise. This is not 00:03:14.139 --> 00:03:16.259 a software bug we can patch with a clever algorithm. 00:03:16.879 --> 00:03:20.240 It is a hard limit of physics. It all comes down 00:03:20.240 --> 00:03:22.479 to how electrons travel through a conductor. 00:03:22.800 --> 00:03:26.039 At lower speeds, copper is fantastic. It's cheap. 00:03:26.099 --> 00:03:28.960 It's malleable. It's reliable. Sure. But as you 00:03:28.960 --> 00:03:32.120 push data at high frequencies, you encounter 00:03:32.120 --> 00:03:35.080 phenomena like... The skin effect. The skin effect. 00:03:35.360 --> 00:03:37.560 Yeah. The electrons actually get pushed to the 00:03:37.560 --> 00:03:40.379 outer edges of the copper wire, which minimizes 00:03:40.379 --> 00:03:42.960 the usable area for the current. Oh, I see. That 00:03:42.960 --> 00:03:45.139 drastically jacks up the electrical resistance. 00:03:45.340 --> 00:03:48.319 And that resistance generates massive, unmanageable 00:03:48.319 --> 00:03:51.280 heat. Which brings up an obvious question. Data 00:03:51.280 --> 00:03:53.900 centers already consume gargantuan amounts of 00:03:53.900 --> 00:03:55.960 electricity. Oh, massive amounts. Some of them 00:03:55.960 --> 00:03:58.539 draw over 100 megawatts, which is enough to power 00:03:58.539 --> 00:04:02.280 a small city. So could this heat issue simply 00:04:02.280 --> 00:04:04.460 be solved by... building radically better liquid 00:04:04.460 --> 00:04:07.479 cooling systems instead of completely abandoning 00:04:07.479 --> 00:04:09.960 copper. I mean, it sounds like a logical brute 00:04:09.960 --> 00:04:13.300 force fix. Yeah. But the power usage at 1 .6 00:04:13.300 --> 00:04:16.120 TBP is spikes exponentially, not linearly. Wait, 00:04:16.199 --> 00:04:18.579 exponentially? Yeah. If you double the speed, 00:04:18.699 --> 00:04:20.839 you don't just double the heat, it multiplies. 00:04:21.860 --> 00:04:24.060 Pumping more extreme cooling into the system 00:04:24.060 --> 00:04:26.639 requires even more power to run the chillers 00:04:26.639 --> 00:04:29.089 and pumps. Right. Eventually, the power required 00:04:29.089 --> 00:04:32.110 just to cool the copper exceeds the power used 00:04:32.110 --> 00:04:35.209 to run the AI calculations. It becomes physically 00:04:35.209 --> 00:04:38.189 and economically unviable. So we literally cannot 00:04:38.189 --> 00:04:41.470 out -cool the fundamental laws of thermodynamics. 00:04:41.709 --> 00:04:44.129 We really can't. And even if you could somehow 00:04:44.129 --> 00:04:46.889 keep it cold, you still face the third threat, 00:04:47.069 --> 00:04:51.329 signal quality. Right. At 1 .6 TBps, electrical 00:04:51.329 --> 00:04:53.850 signals degrade over incredibly short distances. 00:04:54.569 --> 00:04:56.529 A copper connection that is just a few inches 00:04:56.529 --> 00:04:59.250 long starts losing clarity. The ones and zeros 00:04:59.250 --> 00:05:02.129 blur together into noise. Exactly. To fix that, 00:05:02.189 --> 00:05:04.230 the network has to constantly run error correction 00:05:04.230 --> 00:05:06.769 algorithms and resend dropped packets of data. 00:05:06.930 --> 00:05:08.990 Which creates latency. Which entirely defeats 00:05:08.990 --> 00:05:10.670 the purpose of buying faster, more expensive 00:05:10.670 --> 00:05:13.129 hardware. Right. So since physics is breaking 00:05:13.129 --> 00:05:16.269 copper, we clearly need a new medium. We have 00:05:16.269 --> 00:05:19.170 to bypass electrical resistance altogether. And 00:05:19.170 --> 00:05:21.310 the elegant, almost science fiction solution 00:05:21.310 --> 00:05:24.209 to this is light. We are taking the electrical 00:05:24.209 --> 00:05:26.410 signals and replacing them with beams of light. 00:05:26.550 --> 00:05:29.689 We're swapping out heavy friction bound electrons 00:05:29.689 --> 00:05:33.600 for weightless. frictionless photons. Yeah, that's 00:05:33.600 --> 00:05:36.319 the core concept behind silicon photonics. I 00:05:36.319 --> 00:05:38.399 have to admit, I still struggle to wrap my head 00:05:38.399 --> 00:05:40.860 around light replacing electricity at a microscopic 00:05:40.860 --> 00:05:43.319 packaging level. It's wild, isn't it? It really 00:05:43.319 --> 00:05:45.839 is. I understand fiber optics running under the 00:05:45.839 --> 00:05:48.579 ocean to connect continents, but shrinking that 00:05:48.579 --> 00:05:51.660 concept down to a microscopic scale inside a 00:05:51.660 --> 00:05:55.339 single computer chip feels wild. It is a profound 00:05:55.339 --> 00:05:58.420 feat of engineering. Instead of using tiny copper 00:05:58.420 --> 00:06:01.139 wires to connect the processors, the History 00:06:01.139 --> 00:06:03.779 is transitioning to wave guides. Let's define 00:06:03.779 --> 00:06:06.079 that quickly for the listener. Wave guides, tiny 00:06:06.079 --> 00:06:08.660 glass pathways etched into silicon chips that 00:06:08.660 --> 00:06:11.110 steer light beams. Yeah. That is exactly it. 00:06:11.170 --> 00:06:13.589 They are microscopic glass channels. Yeah. And 00:06:13.589 --> 00:06:15.209 because light behaves completely differently 00:06:15.209 --> 00:06:17.410 than electricity, it doesn't suffer from that 00:06:17.410 --> 00:06:19.350 electrical resistance we just talked about. Which 00:06:19.350 --> 00:06:21.829 completely eliminates the massive heat generation 00:06:21.829 --> 00:06:24.769 problem. Right. Photons don't cause friction 00:06:24.769 --> 00:06:27.410 the way electrons do. It also requires significantly 00:06:27.410 --> 00:06:29.569 less energy to travel over the same physical 00:06:29.569 --> 00:06:32.029 distance. How much less? By moving to silicon 00:06:32.029 --> 00:06:35.709 photonics, you get roughly 3 to 3 .5 times better 00:06:35.709 --> 00:06:38.449 power efficiency per individual interconnect 00:06:38.449 --> 00:06:42.220 link. Just to visualize the scale of that, a 00:06:42.220 --> 00:06:45.579 standard modern AI data center has somewhere 00:06:45.579 --> 00:06:48.680 between 40 ,000 and 50 ,000 interconnect links. 00:06:49.019 --> 00:06:52.480 Whoa! Imagine 50 ,000 interconnects in a single 00:06:52.480 --> 00:06:54.680 data center suddenly running three times more 00:06:54.680 --> 00:06:56.899 efficiently. It completely reshapes the grid. 00:06:57.060 --> 00:06:58.740 You're talking about a structural transformation 00:06:58.740 --> 00:07:00.959 of how computing centers are built from the ground 00:07:00.959 --> 00:07:03.149 up. Absolutely. The market projections are already 00:07:03.149 --> 00:07:05.230 pricing in this massive structural shift. Yeah. 00:07:05.350 --> 00:07:07.750 In 2025, the silicon photonics market was sitting 00:07:07.750 --> 00:07:11.829 at roughly $2 .65 billion. But because of this 00:07:11.829 --> 00:07:14.709 copper wall, it's projected to reach $9 .65 billion 00:07:14.709 --> 00:07:17.889 by 2030. Growing at nearly 30 % annually. And 00:07:17.889 --> 00:07:19.810 some of the more aggressive forecasts even push 00:07:19.810 --> 00:07:22.750 it toward $28 billion by 2034. It is an absolute 00:07:22.750 --> 00:07:25.209 tidal wave of capital. But replacing every copper 00:07:25.209 --> 00:07:27.889 wire with a glass waveguide is a colossal undertaking. 00:07:28.459 --> 00:07:31.139 Which raises a massive logistical question. If 00:07:31.139 --> 00:07:34.019 we are completely abandoning copper for light, 00:07:34.240 --> 00:07:37.480 does using light mean we have to invent entirely 00:07:37.480 --> 00:07:40.300 new manufacturing processes and build new factories 00:07:40.300 --> 00:07:43.100 from scratch? And that is the true underlying 00:07:43.100 --> 00:07:47.199 genius of this technological shift. We don't 00:07:47.199 --> 00:07:50.060 need to reinvent the wheel. We don't. No. The 00:07:50.060 --> 00:07:53.319 genius is using the exact same silicon chip manufacturing 00:07:53.319 --> 00:07:56.819 processes that the semiconductor industry has 00:07:56.819 --> 00:07:58.980 ruthlessly perfected over the last 40 years. 00:07:59.319 --> 00:08:01.959 Oh, wow. Yeah, we're just repurposing those billion 00:08:01.959 --> 00:08:04.420 dollar machines. We're hacking silicon factories 00:08:04.420 --> 00:08:06.939 to print glass light paths instead of copper 00:08:06.939 --> 00:08:10.339 wires. Sponsor. That is the magic of it. It allows 00:08:10.339 --> 00:08:12.980 this incredibly advanced transition to scale 00:08:12.980 --> 00:08:15.579 up rapidly without needing to spend trillions 00:08:15.579 --> 00:08:17.459 of dollars building a whole new type of industrial 00:08:17.459 --> 00:08:19.850 supply chain. OK, now that we understand the 00:08:19.850 --> 00:08:22.110 core technology and why the physics of copper 00:08:22.110 --> 00:08:24.670 are forcing this shift, we need a map. Right. 00:08:24.750 --> 00:08:27.110 We have to navigate this emerging market intelligently 00:08:27.110 --> 00:08:28.949 rather than just throwing darts at tech stocks. 00:08:29.209 --> 00:08:32.149 Exactly. The sources outline a fantastic mental 00:08:32.149 --> 00:08:34.009 model for this called the three -layer framework. 00:08:34.409 --> 00:08:37.190 This framework helps us categorize the ecosystem. 00:08:37.429 --> 00:08:39.330 We start at the foundation, which is the base 00:08:39.330 --> 00:08:42.190 layer. These are the mega caps. The trillion 00:08:42.190 --> 00:08:44.610 dollar giants dictating the architecture of the 00:08:44.610 --> 00:08:46.850 shift. For them, photonics is just one necessary 00:08:46.850 --> 00:08:49.429 component of a much larger supercomputing business. 00:08:49.690 --> 00:08:52.639 First up in that base layer is NVIDIA. They aren't 00:08:52.639 --> 00:08:54.860 just a chip designer anymore, right? No, they're 00:08:54.860 --> 00:08:57.059 the primary architect of this entire transition. 00:08:57.779 --> 00:09:01.200 NVIDIA sells entire rack -scale supercomputing 00:09:01.200 --> 00:09:04.019 systems now. It's not just a GPU you plug into 00:09:04.019 --> 00:09:07.019 a motherboard. Right. Their new Vera Rubin platform, 00:09:07.259 --> 00:09:09.600 which starts shipping in the second half of 2026, 00:09:09.860 --> 00:09:14.240 is a perfect example. It packs 72 massive GPUs 00:09:14.240 --> 00:09:17.779 into a single server rack. 72? That's insane. 00:09:18.019 --> 00:09:21.259 It holds 1 .3 million individual components per 00:09:21.259 --> 00:09:23.679 system. If you tried to wire that entirely with 00:09:23.679 --> 00:09:26.259 copper, it would incinerate itself. So Vera Rubin 00:09:26.259 --> 00:09:29.200 integrates silicon photonics natively. Exactly. 00:09:29.259 --> 00:09:31.740 To support that, their new Spectrum 6 switches 00:09:31.740 --> 00:09:34.580 are delivering five times better power efficiency 00:09:34.580 --> 00:09:36.679 than traditional networking hardware. And their 00:09:36.679 --> 00:09:39.840 SuperNex are running at exactly 1 .6 TBPS. That 00:09:39.840 --> 00:09:42.620 speed is not a coincidence. Nope. It is right 00:09:42.620 --> 00:09:44.740 at the precise threshold where copper networking 00:09:44.740 --> 00:09:47.379 breaks down, forcing their customers to adopt 00:09:47.379 --> 00:09:49.820 the optical standard. Next in the base layer 00:09:49.820 --> 00:09:52.870 is Broadcom. And they're taking a slightly different 00:09:52.870 --> 00:09:55.330 approach. They're the pioneer of what is known 00:09:55.330 --> 00:09:57.909 as co -packaging. Co -packaging. So Broadcom 00:09:57.909 --> 00:10:00.629 embeds the optics directly onto the networking 00:10:00.629 --> 00:10:03.590 chip itself. Right. Just to clarify, usually 00:10:03.590 --> 00:10:06.090 the optical laser and the processor are separate 00:10:06.090 --> 00:10:08.830 components, right? Right. And data has to travel 00:10:08.830 --> 00:10:10.870 across the board to get from one to the other. 00:10:11.049 --> 00:10:14.789 Exactly. But by co -packaging them, Broadcom 00:10:14.789 --> 00:10:17.169 drastically shortens the physical data paths. 00:10:17.740 --> 00:10:20.659 It cuts power consumption and dramatically improves 00:10:20.659 --> 00:10:23.039 signal clarity. Because the data barely has to 00:10:23.039 --> 00:10:25.759 travel before it turns into light. Yeah. Then 00:10:25.759 --> 00:10:28.539 we have Cisco. They are making the enterprise 00:10:28.539 --> 00:10:31.360 backbone play. People often think of Cisco as 00:10:31.360 --> 00:10:33.960 old -school networking. Maybe a bit boring. Right. 00:10:34.059 --> 00:10:36.019 But that actually puts them in a dominant position. 00:10:36.480 --> 00:10:39.639 Every enterprise data center, banks, hospitals, 00:10:39.899 --> 00:10:42.139 logistics hubs will eventually have to replace 00:10:42.139 --> 00:10:44.940 their aging copper infrastructure with optical 00:10:44.940 --> 00:10:48.399 interconnects. And Cisco already owns those relationships. 00:10:48.759 --> 00:10:51.799 They acquired an optics pioneer called Acacia 00:10:51.799 --> 00:10:55.139 a few years ago. And now they're manufacturing 00:10:55.139 --> 00:10:58.919 800 GBP silicon photonic transceivers to feed 00:10:58.919 --> 00:11:01.740 that enterprise demand. Finally, we have the 00:11:01.740 --> 00:11:04.320 controversial wild card of the base layer, Intel. 00:11:04.860 --> 00:11:06.840 Intel is fascinating here. They actually have 00:11:06.840 --> 00:11:09.740 25 years of quiet research and development in 00:11:09.740 --> 00:11:12.539 photonics. 25 years. They've already shipped 00:11:12.539 --> 00:11:16.840 over 8 million photonic chips to date. Wow. Their 00:11:16.840 --> 00:11:19.559 new optical compute interconnect or OCI chiplet 00:11:19.559 --> 00:11:23.559 is highly advanced. It runs at four TBPs bidirectionally. 00:11:23.600 --> 00:11:25.899 And here's the insane metric. It consumes just 00:11:25.899 --> 00:11:28.700 five picojoules per bit of data transferred. 00:11:28.879 --> 00:11:30.779 Let's contextualize that because picojoules are 00:11:30.779 --> 00:11:34.279 hard to grasp. Right. A picojoule is one trillionth 00:11:34.279 --> 00:11:37.299 of a joule. Traditional pluggable optical modules 00:11:37.299 --> 00:11:40.500 consume about 15 picajoules per bit. So Intel's 00:11:40.500 --> 00:11:42.779 chiplet represents a massive 3x power advantage 00:11:42.779 --> 00:11:45.299 in an industry where power is the ultimate bottleneck. 00:11:45.320 --> 00:11:48.120 But Intel has obviously faced real public, corporate, 00:11:48.179 --> 00:11:50.100 and execution struggles recently. Yeah, they 00:11:50.100 --> 00:11:52.159 have. Does their current corporate turmoil make 00:11:52.159 --> 00:11:54.700 their 25 -year head start in photonics irrelevant? 00:11:55.179 --> 00:11:58.360 Well, the execution risk for Intel is admittedly 00:11:58.360 --> 00:12:00.940 quite high right now. The market is highly skeptical. 00:12:01.159 --> 00:12:03.220 Understandably. But if you look strictly at their 00:12:03.220 --> 00:12:06.340 assets, Intel is uniquely positioned globally. 00:12:06.580 --> 00:12:09.039 They are the only major foundry on the planet 00:12:09.039 --> 00:12:11.659 right now, offering optics -based manufacturing 00:12:11.659 --> 00:12:14.879 to external customers at that specific scale. 00:12:15.039 --> 00:12:17.820 If Intel executes their manufacturing advantage, 00:12:18.220 --> 00:12:21.720 their upside is much bigger than expected. Precisely. 00:12:21.720 --> 00:12:24.860 And talking about manufacturing perfectly transitions 00:12:24.860 --> 00:12:27.500 us up. a level to the middle layer these are 00:12:27.500 --> 00:12:29.440 the foundries and the manufacturers right because 00:12:29.440 --> 00:12:32.240 the mega caps like nvidia and broadcom design 00:12:32.240 --> 00:12:34.840 the grand architecture but they don't actually 00:12:34.840 --> 00:12:37.039 own the machines that build the chips someone 00:12:37.039 --> 00:12:39.299 has to physically etch the glass into the silicon 00:12:39.840 --> 00:12:42.200 This layer is crucial because it reveals the 00:12:42.200 --> 00:12:44.399 actual physical constraints of the market. And 00:12:44.399 --> 00:12:47.159 we got a huge signal about this on March 18th, 00:12:47.159 --> 00:12:50.519 2026. TSMC, Taiwan's semiconductor manufacturing 00:12:50.519 --> 00:12:53.120 company, made a major announcement. Yeah, they 00:12:53.120 --> 00:12:55.379 confirmed their dedicated COUPN manufacturing 00:12:55.379 --> 00:12:58.440 line is at absolute full capacity. This line 00:12:58.440 --> 00:13:01.360 is specifically building 1 .6T optical modules 00:13:01.360 --> 00:13:04.279 for companies like NVIDIA and Broadcom. TSMC 00:13:04.279 --> 00:13:06.700 is arguably the most important foundry in the 00:13:06.700 --> 00:13:08.649 world. Without a doubt. When they dedicate a 00:13:08.649 --> 00:13:11.129 massive production line to a new technology and 00:13:11.129 --> 00:13:14.049 immediately sell out of capacity, it proves that 00:13:14.049 --> 00:13:16.389 photonics is officially an infrastructure standard. 00:13:16.590 --> 00:13:19.909 It is no longer an R &D experiment. Exactly. 00:13:20.029 --> 00:13:22.870 We also have tower semiconductor in this middle 00:13:22.870 --> 00:13:25.190 layer. They are the number one specialist foundry 00:13:25.190 --> 00:13:27.350 in the world for this tech. They focus purely 00:13:27.350 --> 00:13:30.509 on fabricating silicon photonics on silicon wafers. 00:13:30.570 --> 00:13:33.059 Yeah. They have a direct partnership with NVIDIA, 00:13:33.059 --> 00:13:35.879 and the demand is so high that 70 % of their 00:13:35.879 --> 00:13:38.679 new manufacturing capacity is already reserved 00:13:38.679 --> 00:13:41.899 through 2028. Customers are making massive non 00:13:41.899 --> 00:13:44.720 -refundable prepayments just to secure a physical 00:13:44.720 --> 00:13:47.500 space on the assembly line. Then there is GlobalFoundries. 00:13:47.580 --> 00:13:49.929 They are playing the scale game. Global Foundries 00:13:49.929 --> 00:13:52.330 uniquely integrates the optical chip and the 00:13:52.330 --> 00:13:54.690 full module together. Right. They do the entire 00:13:54.690 --> 00:13:57.370 fabrication in one single continuous process. 00:13:57.590 --> 00:14:00.190 That end -to -end capability is incredibly hard 00:14:00.190 --> 00:14:02.570 to replicate at scale, and it lowers the cost 00:14:02.570 --> 00:14:04.850 per unit. They're aiming to turn this into a 00:14:04.850 --> 00:14:09.850 $1 billion annual run rate by 2028. And we absolutely 00:14:09.850 --> 00:14:13.860 cannot forget about Fabernae. This might be the 00:14:13.860 --> 00:14:15.940 most interesting company in the entire document. 00:14:16.179 --> 00:14:18.419 They are a hidden beneficiary based in Thailand. 00:14:18.639 --> 00:14:21.720 Fabernae focuses purely on high -precision manufacturing. 00:14:22.840 --> 00:14:25.940 Specifically, they specialize in submicron alignment. 00:14:26.399 --> 00:14:28.000 Which sounds technical, but if you think about 00:14:28.000 --> 00:14:30.039 it, it's wild. You're trying to perfectly align 00:14:30.039 --> 00:14:33.600 a microscopic laser beam so it fires directly 00:14:33.600 --> 00:14:36.840 into a microscopic glass pathway. It's like threading 00:14:36.840 --> 00:14:39.240 a microscopic needle during an earthquake. Exactly. 00:14:39.340 --> 00:14:41.940 If the alignment is off by even a fraction of 00:14:41.940 --> 00:14:44.470 a hair, the light hits the silicon instead of 00:14:44.470 --> 00:14:48.350 the glass, the data drops, and a $100 ,000 server 00:14:48.350 --> 00:14:50.870 rack crashes. It is incredibly difficult, and 00:14:50.870 --> 00:14:53.009 very few companies in the world can do it reliably 00:14:53.009 --> 00:14:55.889 at scale. Right. Fabernae runs over 2 million 00:14:55.889 --> 00:14:58.830 square feet of factory space dedicated to this. 00:14:58.990 --> 00:15:02.409 And here is the kicker. They manufacture these 00:15:02.409 --> 00:15:04.669 highly sensitive components for both coherent 00:15:04.669 --> 00:15:06.909 and momentum. Those are the two competitors that 00:15:06.909 --> 00:15:10.940 NVIDIA just gave $2 billion each. Wait, if Coherent 00:15:10.940 --> 00:15:13.080 and Lumentum are technically competing for market 00:15:13.080 --> 00:15:16.179 share, but Fabrinet is manufacturing the components 00:15:16.179 --> 00:15:19.539 for both of them, why should an investor or an 00:15:19.539 --> 00:15:22.059 observer even care about picking the flashy chip 00:15:22.059 --> 00:15:24.559 designers? Because Fabrinet captures the spending 00:15:24.559 --> 00:15:27.539 across the entire ecosystem without taking on 00:15:27.539 --> 00:15:29.720 the design risk. Right. They don't need to pick 00:15:29.720 --> 00:15:32.059 which specific architectural design wins the 00:15:32.059 --> 00:15:34.379 war. The capital flows straight through their 00:15:34.379 --> 00:15:36.980 clean rooms regardless. They sell the picks and 00:15:36.980 --> 00:15:39.539 shovels. They win regardless of whose design 00:15:39.539 --> 00:15:41.679 dominates. That is the perfect way to look at 00:15:41.679 --> 00:15:43.919 it. And that finally brings us to the top layer. 00:15:44.399 --> 00:15:46.460 This is the roof of the house, the pure play 00:15:46.460 --> 00:15:49.279 photonics companies. These are the high risk, 00:15:49.320 --> 00:15:52.279 high reward specialists that are actually designing 00:15:52.279 --> 00:15:54.600 and building the specific hardware components 00:15:54.600 --> 00:15:58.039 that go into the mega cap architectures. Coherent 00:15:58.039 --> 00:16:00.580 is the vertically integrated giant in this layer. 00:16:00.759 --> 00:16:03.200 They build absolutely everything from the raw 00:16:03.200 --> 00:16:05.840 laser chip up to the final transceiver module. 00:16:06.139 --> 00:16:09.220 They keep the entire process in house, which 00:16:09.220 --> 00:16:11.799 gives them immense control over quality and margins. 00:16:12.379 --> 00:16:15.360 They are the ones who secured that massive $2 00:16:15.360 --> 00:16:18.320 billion multi -year deal from Nvidia we mentioned 00:16:18.320 --> 00:16:20.480 at the start. It was the largest single deal 00:16:20.480 --> 00:16:23.200 in company history. And yet, even with that growth, 00:16:23.480 --> 00:16:27.039 their peg ratio sits around 1 .1 right now. For 00:16:27.039 --> 00:16:29.779 you listening, a peg ratio compares a company's 00:16:29.779 --> 00:16:32.039 stock price to its expected earnings growth. 00:16:32.379 --> 00:16:35.120 A peg around 1 essentially means the stock is 00:16:35.120 --> 00:16:38.240 fairly valued, maybe slightly undervalued, compared 00:16:38.240 --> 00:16:40.960 to the hyperinflated AI software stocks. Right. 00:16:41.389 --> 00:16:44.269 Then we have lumentum. They are the laser supplier 00:16:44.269 --> 00:16:47.029 to everyone. This is a vital physical constraint 00:16:47.029 --> 00:16:49.649 we haven't touched on yet. Silicon, the material 00:16:49.649 --> 00:16:51.669 the chips are made of, is what physicists call 00:16:51.669 --> 00:16:54.649 an indirect bandgap material. In plain English, 00:16:54.870 --> 00:16:57.210 silicon cannot generate light on its own. Right. 00:16:57.629 --> 00:16:59.909 Every single silicon photonics chip, no matter 00:16:59.909 --> 00:17:02.509 how advanced, needs an external laser source 00:17:02.509 --> 00:17:05.009 to provide the actual light. So even Lumentum's 00:17:05.009 --> 00:17:07.490 direct competitors often have to buy Lumentum's 00:17:07.490 --> 00:17:09.730 lasers to power their own modules. No matter 00:17:09.730 --> 00:17:12.029 whose broader design wins out, Lumentum still 00:17:12.029 --> 00:17:14.369 gets paid for the light source. Exactly. They 00:17:14.369 --> 00:17:16.490 just lock in a massive NVIDIA supply commitment 00:17:16.490 --> 00:17:18.769 as well. And they are currently trading at a 00:17:18.769 --> 00:17:22.490 very attractive peg ratio of 0 .61. Meaning the 00:17:22.490 --> 00:17:24.589 market is severely underpricing their future 00:17:24.589 --> 00:17:28.250 growth. Yeah. Finally. We arrive at Marvell technology. 00:17:28.750 --> 00:17:32.170 They are the data interpreter and lately an incredibly 00:17:32.170 --> 00:17:35.289 aggressive serial acquirer. Marvell holds an 00:17:35.289 --> 00:17:37.890 estimated 50 % market share in digital signal 00:17:37.890 --> 00:17:41.950 processing or DSP. DSP chips are the unsung heroes 00:17:41.950 --> 00:17:44.769 of this entire revolution. Right, because when 00:17:44.769 --> 00:17:47.210 a beam of light speeds through the glass waveguide 00:17:47.210 --> 00:17:49.730 and exits the photonic chip, it is still just 00:17:49.730 --> 00:17:52.490 an analog light wave. It must be rapidly translated 00:17:52.490 --> 00:17:55.180 back into digital data. The ones and zeros so 00:17:55.180 --> 00:18:03.799 the actual GPU can process it. But interpreting 00:18:03.799 --> 00:18:06.119 data isn't enough for them anymore. They just 00:18:06.119 --> 00:18:09.900 spent $3 .25 billion to acquire a company called 00:18:09.900 --> 00:18:17.880 Celestial AI. You have to look at Celestial's 00:18:17.880 --> 00:18:21.900 specific performance specs. Celestial is pioneering 00:18:21.900 --> 00:18:24.420 something called a photonic fabric. A photonic 00:18:24.420 --> 00:18:27.180 fabric. Yeah. They use light not just to connect 00:18:27.180 --> 00:18:29.599 servers across a room, but to connect computing 00:18:29.599 --> 00:18:32.740 chips directly to memory banks at a microscopic 00:18:32.740 --> 00:18:35.400 level. Wow. Because they bypass electrical traces 00:18:35.400 --> 00:18:39.500 entirely. They boast 10x lower latency. And they 00:18:39.500 --> 00:18:42.579 claim an astonishing 25x greater bandwidth compared 00:18:42.579 --> 00:18:44.859 to current industry approaches. If those specs 00:18:44.859 --> 00:18:47.660 hold, it fundamentally changes how data centers 00:18:47.660 --> 00:18:50.240 are built. It changes the entire physical architecture 00:18:50.240 --> 00:18:53.640 of modern computing. We are moving from a wind 00:18:53.640 --> 00:18:56.589 constrained by the friction of copper. to a world 00:18:56.589 --> 00:18:59.049 operating at the speed of light. Let's step back 00:18:59.049 --> 00:19:00.789 and look at the whole board here. The copper 00:19:00.789 --> 00:19:03.670 wall is not just an academic theory or a distant 00:19:03.670 --> 00:19:06.529 warning. No, it's not. It is a physical, hard 00:19:06.529 --> 00:19:08.910 engineering constraint that slammed into the 00:19:08.910 --> 00:19:11.549 AI industry the moment we tried to push 1 .6 00:19:11.549 --> 00:19:13.549 terabits per second. Yeah, and the market reacted 00:19:13.549 --> 00:19:16.529 with violent efficiency. Over $7 billion moved 00:19:16.529 --> 00:19:18.930 in a matter of weeks. The biggest players in 00:19:18.930 --> 00:19:22.529 the world, Nvidia, Marvell, Broadcom, are frantically 00:19:22.529 --> 00:19:25.690 locking in their optical supply chains. TSMC 00:19:25.690 --> 00:19:27.490 selling out their dedicated production lines 00:19:27.490 --> 00:19:30.650 proves that this isn't a fad. Silicon photonics 00:19:30.650 --> 00:19:33.470 is officially the new inevitable foundation of 00:19:33.470 --> 00:19:36.650 all future AI infrastructure. Absolutely. For 00:19:36.650 --> 00:19:39.150 you listening, try applying this three -layer 00:19:39.150 --> 00:19:42.470 framework base, middle and top, to the next major 00:19:42.470 --> 00:19:44.950 technological shift you encounter, whether it's 00:19:44.950 --> 00:19:46.769 quantum computing or new battery chemistries. 00:19:46.849 --> 00:19:49.630 It is a fantastic mental model for calibrating 00:19:49.630 --> 00:19:52.049 risk. It cuts through the hype and truly helps 00:19:52.049 --> 00:19:54.289 you understand who actually builds the complex 00:19:54.289 --> 00:19:56.829 physical supply chains beneath the software. 00:19:57.210 --> 00:20:00.769 And here is a final thought to chew on. The physical 00:20:00.769 --> 00:20:03.490 limits of copper forced us to entirely rebuild 00:20:03.490 --> 00:20:06.089 our infrastructure from the ground up. We literally 00:20:06.089 --> 00:20:08.769 had to pivot to the physics of light just to 00:20:08.769 --> 00:20:11.349 keep the AI boom alive. If you are wondering 00:20:11.349 --> 00:20:14.130 why chat GPT or Claude might suddenly get 10 00:20:14.130 --> 00:20:16.730 times faster and cheaper in a few years, it won't 00:20:16.730 --> 00:20:18.849 be just a software update. No, it will be because 00:20:18.849 --> 00:20:22.549 of microscopic glass. So what other invisible 00:20:22.549 --> 00:20:25.950 physical limits of our current hardware are secretly 00:20:25.950 --> 00:20:28.730 holding back the next massive leap in artificial 00:20:28.730 --> 00:20:32.109 intelligence? Two sec silence. That's a great 00:20:32.109 --> 00:20:34.250 question. Something to think about. Thanks for 00:20:34.250 --> 00:20:36.190 taking this deep dive with us today. Keep learning, 00:20:36.289 --> 00:20:38.109 keep questioning, and we'll catch you next time. 00:20:38.210 --> 00:20:39.069 Out to your own music.