Semiconductors Deep Dive (Part 2): Scale and Timeline of 5G Infrastructure
Light travels at 670,616,629 miles per hour. Were it able to travel in an arc it would circumnavigate the globe in about 130 milliseconds (1/1,000th). In the time it takes you to blink, light would travel around Earth 3 times. In a single second, it would make more than 7.5 round trips. Light speed is the ultimate factor limiting how fast data can travel, but the practical limits are known as bandwidth and latency. Bandwidth describes a network’s capacity – how much data it can transfer over a period of time. Latency describes how fast the data gets there. I like the analogy of the fire hose.
Bandwidth is the amount of water coming out of the nozzle – it describes how long it would take to fill up an Olympic swimming pool. Latency is the speed at which the water is traveling through the hose – it would tell you how long it takes for a single drop of water to travel through thirty feet of hose.
5G is going to do to network connectivity what the atomic bomb did to weapons. This transition will feel different. It will make the previous advances between generations seem as insignificant as the transition between bows and arrows and muskets. 5G will wreak havoc on entire industries because mastering it will require new core competencies. As one example, I suspect 5G just might be the final nail in the coffin of legacy automakers (topic for another post).
Consider for a moment Apple and Google’s app stores. Their dominant positions and the powerful network effects from their respective ecosystems are made possible by 3G+ networks. Without 3G+ networks, there would be no apps. 3G+ didn’t directly kill off their competitors (e.g. Motorola), but it changed the game so drastically that the core competencies of legacy phone makers became comparatively irrelevant.
However, it is also going to take longer to fully transition to 5G than most people think. There simply aren’t any technologies that are game-changing enough to act as a forcing function to attract the necessary investment - yet.
After providing a brief overview of historical mobile generations we’ll dive into what changes need to be made to infrastructure in order to achieve truly ubiquitous, fully functioning 5G.
1st generation (1G) networks enabled primitive mobile phones. Voice quality was terrible and bandwidth was too poor for encryption – so anyone could theoretically listen in on a conversation. Bandwidth was on the order of 2.4 kilobits per second (kbps). For reference, the average file size of an image on a modern smartphone is 6 megabytes. Each megabyte is 8000 kilobits, so each picture is 48,000 kilobits. This means to send a modern image on a 1G network would have taken 20,000 seconds (333.3 minutes; 5.56 hours).
2nd generation technology made it possible to send texts and (low quality) pictures. Voice quality became nearly as good as what we experience today. 3G technology brought us Skype and video streaming. The iPhone launched at the tail end of 3G. 4G is where we are today (what carriers like Verizon and T-Mobile are currently marketing as 5G is not even remotely comparable to what’s coming).
New Infrastructure > Trillions of Dollars > But How Many Years?
Estimates vary widely, but the figures commonly thrown out are that 5G will improve latency by as much as 50X, speed by up to 100X, and capacity by up to 1,000X. However, hitting these figures will take time because 5G requires far higher levels of infrastructure density than 4G. In order for 4G to function optimally there needs to be a cell tower every ten miles. In order for 5G to function optimally (meaning hit the performance figures above), you don’t only need 5G towers, you need 5G nodes. This is because full speed 5G operates on much shorter wavelengths - millimeter wavelengths (mmWave). Wavelength is directly correlated to the distance a signal can travel. Shorter wavelength = more powerful signal (in terms of both bandwidth and latency) but shorter range. A 5G node has a range of about 1,500 feet (less if its line of sight is obstructed).
A Verizon 5G mmWave node. Image from Circa.com
When discussing how many semiconductors are present in a car or other device/system (e.g. 5G mmWave node), the term content is sometimes used. Semiconductors are the content. An electric vehicle with far more semiconductors than an internal combustion engine (ICE) vehicle could be said to have higher content density.
5G is going to create an explosion of content in myriad ways, but for now I just want to call attention to content density with respect to 5G nodes vs. 4G towers. The below exercise is highly simplified but provides a reasonable framework, and I believe it is on the correct order of magnitude.
A 4G tower has a ten-mile radius.
A 5G node has a 1,500 foot radius if it’s unobstructed.
How many 5G nodes do we need to cover the same ground as a 4G tower?
A circle with a radius of 10 miles is a circle with a radius of 52,800 feet. 1,078 circles with radius 1,500 feet would fit inside it, and they would still only occupy 87% of the space – so the number would actually be higher than that due to necessary overlap.
The below image is from Statista, and shows the number of mobile wireless cell sites in the US.
Let’s assume that any tower built after 2016 was for 5G (most were not, but it makes our inference conservative). In 2016 there were 308,334 cell sites. In order to blanket the US with equivalent coverage we would therefore need 1,078 X 308,334 = 332,384,052 5G nodes.
Each 4G cell tower requires a variety of semiconductors, which are located in the base transceiver station (BTS pictured below). From Wikipedia:
1. Analog transceivers: Provides transmission and reception of signals
2. Digital transceivers: Same as above but digitally
3. Power amplifiers: Amplifies the signal from the transceiver for transmission through antenna
4. Multiplexers: Separates sending and receiving signals t/from antennas
5. Alarm extension systems: collects working status alarms of various units
6. Control function: Controls and manages the various units of BTS, including any software. On-the-spot configurations, status changes, software upgrades, etc. are done through the control function
7. Baseband receiver unit: Frequency hopping, signal DSP
8. Memory chips
9. Logic chips
You don’t need to know what each of these components mean. The takeaway is that each of them represent at least 1 semiconductor (though some functions can be combined into a multi-function chip).
5G nodes will need to have all of the above chips plus more. Nvidia’s new 5G chips are just one example. From Nvidia’s website:
Nvidia GPUs make applications like cloud-based virtual reality (VR), smart cities, cloud gaming, 360-degree immersive video, connected drones, and autonomous vehicles possible. As demands on the telco industry grow, these applications will be able to take advantage of GPU computing power…
Verizon is embedding Nvidia GPUs throughout its network – at its mega data centers, at the hundreds of smaller ones that those feed, and at the thousands of smaller cell sites supported by those – to deliver the best capabilities in high performance computing (HPC)...
Some (maybe most) of you are probably wondering if building out 332,384,052 5G nodes passes the sniff test (especially if you work for a telco). To that I say…it depends on the timeline. According to the Washington Post, as of 2020 there were approximately 150 million utility poles in service in the US. Other estimates I found put the number around 180 million. Also, consider that for decades cables have been laid underground whenever possible. That’s why you see utility polls in old neighborhoods but not in new ones. So, it is not unprecedented for a type of infrastructure to be built in the hundreds of millions over a long enough time horizon.
How fast might we get there?
That depends entirely on how fast new technologies come to market that are capable of acting as forcing functions to get the requisite investment from governments, telcos, utilities, etc. One thing is certain, telcos are not going to build out ubiquitous 5G infrastructure until doing so is required in order to remain competitive. Telcos already feel like they built out 4G too fast. The below excerpt comes from an article by Beersheba Research and was posted on Seeking Alpha
I take the controlled capex spend as an indication that telco management is intent on avoiding the same “build it and they will come” mistake made during the 4G roll-out that resulted in subpar returns.
Indeed, Verizon stated in its 2021 10-K filing: we have significant discretion over amount and timing of capex…and we are not subject to any agreement that would require significant capex on a designated schedule or occurrence of designated events.
Similarly, AT&T wrote in its 2020 10-K filing: The amount of capital expenditures is influenced by demand for services and products, capacity needs and network requirements.
Consumer demand for faster app downloads and faster streaming will be enough to get telcos to build out basic 5G in dense populations, but not the level/type of 5G that results in the performance improvements quoted above. And, it certainly won’t be sufficient to get telcos to build anywhere outside of densely populated cities and suburbs. So, what technologies might be significant enough to act as our forcing functions?
My best guess is that driverless cars will be the ultimate catalyst to build out true, fully performing 5G. Their arrival will bring demand for a variety of use cases, like [vehicle 2 vehicle] and [vehicle 2 infrastructure] communication.
Once they arrive, how fast could 332,384,052 million nodes get built? Let’s go back to our utility pole analogy. Utility poles are serviced every 3-10 years. That means each year there are between 15 and 50 (150/10 and 150/3) million utility poles being serviced. Installing a 5G node will probably be as simple as screwing a pre-manufactured box onto a utility pole or rooftop. It probably won’t take much longer to install a 5G node than it does to service a utility pole. It’s entirely possible that the country could install 332,384,052 nodes in 6 years - maybe even faster if companies like Tesla start getting involved to faster accelerate driverless car adoption. The limiting factor may end up being semiconductor capacity…
I’ll conclude with one last bit of napkin math. If we assume that each 5G node requires one $250 GPU (they’re more expensive currently, but at mass scale the price might come down and I want to be conservative), what is the total addressable market just in the US for GPUs in 5G nodes?
332,384,052 X $250 = $83,096,013,000. $83 billion. If the cost doesn’t come down because supply is still limited, the figure could be double or higher. Just for a single use case of a single type of semiconductor.
All roads lead to semiconductors.