Duct Tape and Data Centers

We had an interesting discussion with a friend over the weekend. He suggested that Huawei might be able to cobble together a semiconductor supply chain by taping together a string of domestic semis fab capacity. China has become famous for improvised solutions, this is the country where a farmer built a submarine out of scrap metal after all. And while Huawei is the diametric opposite of 山寨 improvisation, maybe there is some path here.

Some background for context. Huawei is essentially cut off from building its own semiconductors. As we outlined in a post a couple weeks back, they can still source catalog parts for their telecom equipment business, but without advanced semi fabs they cannot produce their own chips which are a competitive necessity for handsets and probably their networking business as well. They need domestic sources of semiconductor manufacturing but are cut off from anything using “American technology”.

Part of the survival strategy they are pointing to for their future rests in software and the Cloud. But to be competitive in these markets they are going to need their own data centers, and their major competitors here are also building their own chips.

A lot of this strategy was revealed in a Financial Times article ($$$) two weeks ago. According to this, Huawei is working with a Shanghai foundry using second-hand semicap equipment which is theoretically unencumbered by US restrictions. The problem is that this company is struggling to reach 22nm by the end of 2022.

China does have several companies attempting to build the ‘semicap’ machines needed to build advanced fabs. And so every once in a while we see a headline about how one of them has built 5nm equipment. Dig a little in all these stories and they are too good to be true. Typically, the breakthrough is only for one part of the manufacturing line and even then it is often at greatly reduced throughput. Most people we speak with do not believe China will be able to build their own advanced-node machinery any time this decade.

On the other hand, China does have a number of small scale fabs operating at advanced nodes, usually affiliated with academic institutions. While these are probably subject to the US restrictions, from a practical point of view, they are likely untouchable by US sanctions. Nonetheless, they operate at very small volumes.

So let’s do some back of the envelope math.

If Huawei wants to build enough data centers to compete with Amazon how many chips would they need? Amazon has about 60 data centers in the US. Each of these has about 60,000 servers, give or take. To be build something equivalent in China, let’s say Huawei needs 20 data centers, each with 50,000 servers. That’s 1,000,000 CPUs, by comparison AWS probably has five or six million servers globally, and buys around 1 million a year for expansion and replacement.

The limited academic fabs in China could probably only produce a few thousand of those in a year. So what about building on trailing edge processes like 28 nm?

There are a lot of problems with this. Advanced CPUs are optimized to produce the maximum amount of compute for the least amount of power. Moving backwards in the process technology increases the number of chips needed and the amount of electrical power required. With a reversal of Moore’s law you would need a lot more chips – something like 30 times as many CPUs.

  • Instead of 20 data centers you would need 600, more than AWS, Azure and Google Cloud combined have today.
  • They would also need some high capacity foundries to get 30 million CPUs or around 200,000 wafers a year. TSMC’s global capacity is something like 13 million wafers a year, but across multiple facilities. Their proposed Arizona plant will have a capacity of 20,000 wafers a month when it starts running in a few years.
  • The average cloudscale data center uses around 500 megawatts of electricity a year. So 600 data centers requires 300 gigawatts of electricity. For comparison, China adds about this much capacity every year. So China would have to double that just to serve these data centers. (And these would likely be coal powered, so another problem there.)

All of the above are theoretically possible. China can build buildings faster than anyone. And even though a new foundry running a new process will have a lot of production issues, they could theoretically have sufficient capacity. And China can probably build sufficient electrical capacity. But there are an awful lot of assumptions (and tens of billions of dollars) built into this calculation.

To be clear, our math here is very rough. We made the assumption that you could just port a CPU to 28nm with a simple multiple, but the reality is that the design of these chips will introduce a lot of friction to the process and the actual number would likely be much higher than 30x. There is a long list of practical reasons that makes this scenario highly improbable.

Finally, it is worth pointing out that the market for this service is not wide open. Alibaba, for one, has made incredible strides building out its cloud service capacity. They have been building data centers at roughly twice the pace of AWS over the past five years, already a herculean effort, and are doing so with full access to US technology, meaning they would be much more competitive.

Our point is simply that this sort of effort, while possible, is highly impractical.

Leave a Reply