Good morning,

While the U.S. elections dominate the news — and a truly spectacular explosion happened in crypto — I wanted to check in on probably the biggest tech news story of the fall.

On to the update:

China Chip Ban Clarifications

From Bloomberg:

Washington’s restrictions on US citizens assisting China’s chip industry will be more narrowly enforced than feared, suggesting a smaller-than-expected impact on semiconductor companies doing business in the world’s second largest economy. The rules seem to be mostly targeting US persons working in certain functions for semiconductor manufacturing equipment firms, based on a document the Bureau of Industry and Security published last Friday to clarify export control measures announced on Oct. 7. Those sweeping sanctions were introduced to keep cutting-edge chip technologies out of reach for China’s military but have amplified uncertainty around its tech industry and wrought havoc on global chip stocks. A license is required for US persons — anyone with an American passport, green card or residency — conducting or authorizing the delivery of items used to develop or produce advanced chips at a plant in China, but not those who perform related clerical or administrative duties. The same controls apply to US persons who maintain, repair and refurbish those items, according to the document.

This is the exact sort of clarification that I predicted would be coming:

I expect the administration to issue some number of clarifications that a lot of chip-related activity in China is in fact permitted. I don’t think, for example, that last week’s orders from nearly every major semiconductor company that all of their U.S. employees have zero contact or interaction with any aspect of their China business will hold. It seems clear that the administration’s goal is to continue to allow trailing node production in China, including the use of U.S. equipment and technology and service by U.S. companies; trailing nodes constitute the vast majority of current semiconductor manufacturing in China, which is to say business will mostly continue as it was before the ban (which is what I was trying to express with the “loophole” line — like I said, that was a poor choice of words). The fact that that is only as far as the U.S. intends to let it go is the big shift.

Indeed, to my mind that shift was accentuated by the clarification that Americans cannot service advanced equipment — and this equipment needs a lot of servicing! At the same time, another clarification in the Bureau of Industry and Security’s FAQ was that fabrication facilities would be defined by discrete buildings, not necessarily addresses; this means that U.S. companies can sell to and service equipment at a Chinese foundry as long as the equipment they are delivering and servicing is in a different building than the advanced nodes. This is another signal that the administration is trying to do everything it can to ensure that U.S. companies continue to make money from Chinese trailing node manufacturing, which is increasingly important broadly.

Nvidia’s A800

From Reuters:

U.S. chip maker Nvidia Corp is offering a new advanced chip in China that meets recent export control rules aimed at keeping cutting-edge technology out of China’s hands, the company confirmed on Monday. Nvidia responded to Reuters’ reporting that Chinese computer sellers are advertising products with the new chip. The chip, called the A800, represents the first reported effort by a U.S. semiconductor company to create advanced processors for China that follow new U.S. trade rules. Nvidia has said the export limitations could cost it hundreds of millions of dollars in revenue.

U.S. regulations set in early October effectively banned export of advanced microchips and equipment to produce advanced chips by Chinese chipmakers, part of an effort to hobble China’s semiconductor industry and in turn the military. In late August, Nvidia and Advanced Micro Devices Inc both said that their advanced chips, including Nvidia’s data center chip A100, were added to the export control list by the U.S. Commerce Department. The Nvidia A800 can be used in place of the A100 and both are GPUs, or graphics processing units. Such advanced chips can cost thousands of dollars each.

“The Nvidia A800 GPU, which went into production in Q3, is another alternative product to the Nvidia A100 GPU for customers in China. The A800 meets the U.S. Government’s clear test for reduced export control and cannot be programmed to exceed it,” a Nvidia spokesperson said in a statement to Reuters.

There are a few interesting points about this announcement. First, notice that Nvidia wasn’t exactly broadcasting a press release about this new chip; Reuters found advertisements in China and reached out to Nvidia for comment. That seems like an implicit admission on Nvidia’s part that this announcement would not be particularly popular in the U.S.

Secondly, in case it wasn’t clear how the various parameters of the chip ban were decided upon, the new rules specify:

Integrated circuits that have or are programmable to have an aggregate bidirectional transfer rate over all inputs and outputs of 600 Gbyte/s or more to or from integrated circuits other than volatile memories, and any of the following:

• One or more digital processor units executing machine instructions having a bit length per operation multiplied by processing performance measured in TOPS, aggregated over all processor units, of 4800 or more;
• One or more digital ‘primitive computational units,’ excluding those units contributing to the execution of machine instructions relevant to the calculation of TOPS for 3A090.a.1, having a bit length per operation multiplied by processing performance measured in TOPS, aggregated over all computational units, of 4800 or more;
• One or more analog, multi-value, or multi-level ‘primitive computational units’ having a processing performance measured in TOPS multiplied by 8, aggregated over all computational units, of 4800 or more; or
• Any combination of digital processor units and ‘primitive computational units’ whose calculations according to 3A090.a.1, 3A090.a.2, and 3A090.a.3 sum to 4800 or more.

I don’t think it is a coincidence that the rules specify 600 Gb/s given the fact that Nvidia’s A100 chip has 600 Gb/s interconnects. I explained why this matters in a Daily Update:

Nvidia’s data center chips, like the current top-of-the-line A100 and upcoming H100, have the same architecture as its gaming chips (Ampere in the case of the A100 and 3000 line of gaming chips; Ada Lovelace in the case of the H100 and 4000 line of gaming chips); the biggest difference (beyond clock speed and the allocation of cores amongst shaders, tensors, etc.) is memory: data center chips not only have more of it, they also have much higher speed memory buses, both in terms of on-board memory and also interconnects for working with other chips. This, by extension, gets at the difference between training and inference: while inference (i.e. the application of machine learning models) can run on single GPUs, training (i.e. the creation of those models) is done on fleets of them; the training of AI models specifically appears to have been the primary target of the Biden administration.

As I further explained in that Update, maximizing compute is increasingly a matter of system design, not just chips; this is particularly important in terms of AI training, which is an “embarrassingly parallel” task: the speed of processing increases basically linearly with the amount of processors available, but the trick is keeping all of those processors busy, which is where the memory bandwidth comes in. In other words, the A800 chip, which is limited to 400 Gb/s interconnects, will operate as fast as the A100, but it will more likely be sitting idle more often because the interconnect is saturated moving data around.

Third, I did find the phrasing from the Nvidia spokesperson interesting: “The A800 meets the U.S. Government’s clear test for reduced export control and cannot be programmed to exceed it.” I can’t decide if this sounds like a software limit or a hardware one; the former would be faster to implement — and Nvidia sure started producing these cards quickly — but cannot be programmed to exceed it sounds like there is no software solution, which, if the limitation were software, would by definition exist.

Regardless, the broader point is that this is the other shoe to the “loophole” discussion I talked about when the chip ban first dropped: the rules would be interpreted less draconianly than first reported, and companies would move quickly to push up to the limits without exceeding them. That is exactly what Nvidia has done.

China, Intel, and Tower

Doug O’Laughlin beat me to a thought that has been bouncing around my head for a few weeks when I realized that China was the hold-up in Intel completing its Tower Semiconductor acquisition; I wrote at the time it was announced:

This is one of those acquisitions that definitely came as a bit of a surprise, but actually makes a fair bit of sense the more you think about it. First off, it is not only the case that Intel primarily makes its own designs, it also makes only digital chips (i.e. everything is a 1 or a 0). However, there is also an entire universe of analog chips, which can process gradations; this is essential for processing data from the physical world like sound, power, light, etc. Tower specializes in a whole host of specialized chips in the analog space; adding Tower’s capabilities to Intel Foundry Services (IFS) will make the latter much more of a true one stop shop for chip fabrication, matching the capabilities of TSMC or GlobalFoundries.

As I noted in that Update, the real value wasn’t just analog capabilities: Tower also brought experience in being a foundry, that is to say, a customer service organization, a capability that Intel not only lacks but which its internal culture is diametrically opposed to. Intel CEO Pat Gelsinger said in a Stratechery Interview:

Is Tower being merged into Intel Foundry or is Intel Foundry being merged into Tower? The answer is more the latter than the former. I’ve said I fully want to merge these businesses together going forward, and that doesn’t mean this is a little satellite under this thing that we’re getting started in IFS. We’re going to bring those together, and I fully expect that the outcome is a fully integrated business unit that heavily leverages the five thousand people in Tower. Its thirty years old, Israeli-centered, and we know a lot about having good Israeli discipline as part of our team. The other thing I’ve said is that, “Hey, I’d like to do a Mobileye-like spin on our foundry business at some point as well.” I’m going to keep the structure, as opposed to integrating as much, I’m going to keep it more separate to enable that, which means I’m going to leverage a lot more of Tower and the expertise that it builds over time as part of it.

In short, Tower is a really important piece to realizing The Intel Split which is an essential solution to Intel’s Problems. As O’Laughlin notes at Fabricated Knowledge, though, China is holding up the deal:

The question everyone has been grappling with [with regards to the chip ban] is how does China respond? I don’t know the broad mechanism they will choose, but I finally found a simple one that impacts public markets. SAMR deal approval. SAMR stands for State Administration for Market Regulation and is the Chinese anti-trust body that approves mergers similar to the US Federal Trade Commission. But this is China, and every regulatory agency is just an extension of the party’s will, so I think the clear way to hinder the United States and its companies are to block every deal in the approval process…

Tower is the largest independent fab that Intel could likely swallow (sorry about missing GlobalFoundries Intel). That would kickstart a foundry culture and team with industry-standard workflows and customer-facing. This is the Petri dish to grow IFS from, and I think Tower is key for long-term foundry success. China wouldn’t want that. Meanwhile, China just got kneecapped by the Bureau of Industry and Security (BIS). Almost all leading-edge efforts by China have been thwarted. The United States is directly trying to stop China’s semiconductor independence and pulled out all the stops with its recent export controls. Meanwhile, Intel is trying to bolster domestic production and reduce the United State’s reliance on Taiwan. Why would China let Intel, and by extension, the United States government, do this? They will almost certainly block the deal…

Why does an Israeli company without Fabs in China need deal approval? Well, that’s because any company that hits a revenue threshold of ~55 million dollars in China is subject to SAMR review. It’s possible to merge without Chinese approval, but then China could restrict Intel’s right to sell products in China. That’s 30% of sales! Maybe that would be impossible, given China’s reliance on global semiconductors. Still, they could be unilaterally punitive to Intel if they merged without SAMR approval to “restore competition to the market.” So I think the more likely option is a deal delay and eventual break. That would not be good for Tower…Meanwhile, I am saddened by Intel. I understand the competitive dynamics at AMD and the layoffs, but not getting Tower Semi to kickstart IFS feels like a national security travesty…Tower Semi is a quick way to make the IFS dream a reality and has to be at the top of Intel’s strategic priorities. But this is how China can strike back.

I share O’Laughlin’s concerns and analysis: first, this is a very logical place for China to retaliate, and second, this really would be devastating for Intel, which is the only option for a U.S. alternative to TSMC at the leading edge. Moreover, while it is fair to be skeptical of Intel’s ability to catch-up, that task will be far more difficult without the sort of transformation in culture around foundry services that Tower was acquired to provide.

I’m not clear what the best response here is. My initial thought is that Intel should just acquire Tower anyways, but that makes it likely that China simply bans Intel and monopolizes AMD CPUs; then again, given that AMD has indicated it won’t chase Intel in a price war for PC share, maybe it wouldn’t be the worst thing for Intel to have a huge number of AMD CPUs taken off the market in the rest of the world. Moreover, it’s not as if Intel hasn’t already leaned into its importance to U.S. national security; sacrificing the China market would at least be in line with Gelsinger’s rhetoric on the matter. Whatever Intel does, though, one thing is clear: trade wars are hard, and inefficient for everyone involved.

---

This Update will be available as a podcast later today. To receive it in your podcast player, visit Stratechery.

The Stratechery Update is intended for a single recipient, but occasional forwarding is totally fine! If you would like to order multiple subscriptions for your team with a group discount (minimum 5), please contact me directly.

Thanks for being a subscriber, and have a great day!