Daily: DeepSeek and Chips; CXMT and YMTC update
10 min read.
Housekeeping: There’s so much happening right now with DeepSeek that I thought I’d briefly return from the Lunar New Year break. I’ll be off again tomorrow and be back Friday.
This is a semiconductor-related newsletter, so I try to stay in my lane and not bleed into AI too much. But of course, there’s a lot of overlap. Every publication is pumping out DeepSeek articles, so I compiled a few representative pieces. Dig around if you’d like more. And disclaimer, this newsletter is not investment advice.
Highlights
Summary: A Chinese startup, DeepSeek, released an AI model on par with OpenAI’s ChatGPT. The rub is that it took 2 months, US$6m, and lower-quality H800 GPUs to train, and it’s more energy efficient than Western models. That makes it at least 10x cheaper than American models. There’s nuance to those numbers, but still, pretty crazy.
On the US chip ban: There’s always been debate on whether American semiconductor export controls worked in limiting China’s technological progress. Now, some commentary points to DeepSeek as evidence that the export controls failed. In general, the main criticisms seem to be 1) the controls weren’t tight enough, 2) the constraints simply accelerated China’s innovation, forcing it to become more creative and self-sufficient (this is China’s narrative), and/or 3) US is wasting time, resources, political capital on playing defence (restricting China) instead of offence (investing in domestic innovation).
On 1), the last 2 months of the Biden administration really tightened the restrictions. Besides, it seems that DeepSeek built their model with all legal chips. They were just clever about using what they had available.
On 2), this seems to be the common take with DeepSeek. I don’t really buy it. Wouldn’t the counterfactual just be that China built their AI models earlier and better? I suppose they probably wouldn’t have innovated to DeepSeek-level efficient models, simply because they didn’t need to. But still, maybe they would’ve gotten there anyway or maybe the competition wouldn’t have been about efficiency but instead about amassing GPUs or getting to AGI first. The counterfactual is the US competing with a China with no handicaps. Am I missing something?
The chip bans must be doing something, because DeepSeek is desperate for more compute. From the WSJ: “DeepSeek’s [CEO] Liang told Chinese Premier Li Qiang that while Chinese companies were working to catch up, American restrictions on the export of advanced chips to China were still a bottleneck.”
As an aside, I think a Trump administration would have done directionally the same things as the Biden administration. Both Democrats and Republicans are super hawkish on China. I doubt the new Trump admin is going to lift the export controls.
On 3), it’s not clear to me that a good defence and a good offence are mutually exclusive. You can do the CHIPS Act and chip controls. It seems to me that Washington is very aware of the importance of semiconductors, AI, and China, and so, everyone’s trying to do everything they can to tackle these issues.
So what? My take is that it is hard to blame the chip ban because we don’t know what the counterfactual would have looked like. In any case, DeepSeek made their model with legal chips. Instead, the US should double down on what they’re good at: compete, race, innovate. This probably looks like both more chip controls and more domestic investment.
I’m thinking about the Manhattan Project during WW2 and the subsequent nuclear arms race in the Cold War. The US-China AI competition is a race.
On chip stocks: Nvidia down 17%, AMD down 6%, Micron down 12%, TSMC down 13%, Broadcom down 17%, ARM down 10%, Cadence down 10%, ASML down 11%. Asian stocks down too, though China, Taiwan, Korea markets are on hold for Lunar New Year.
I guess we’ll wait and see how it all shakes out. Maybe this is a temporary blip and a new competitor will just drive more innovation and more need for more compute. So, maybe the stocks are at a discount now.
Or maybe, DeepSeek will be the first of many more competing AI models that become increasingly efficient. Satya Nadella, CEO of Microsoft, tweeted, “Jevons paradox strikes again!” Jevons Paradox is when a resource becomes so efficient that it becomes ubiquitous, thus raising overall consumption.
E.g., technological advancement meant gasoline became a cheap, efficient source of energy —> fuelling the adoption of cars —> overall consumption for oil rose.
Maybe commoditisation is a great thing for these companies, in the way gasoline was great for Standard Oil.
And/or maybe this means that there will be a paradigm shift in how we value these companies. Instead of the current Big Tech multiples, which are much higher than other companies and sectors, competition will wear down margins and multiples to be closer to S&P500 averages. Like a utility company, or something. I don’t know, this is pretty long-term speculation.
What’s next for the chip war? DeepSeek built their model off Nvidia GPUs. China’s big thing is self-sufficiency. Their Made in China 2025 project to indigenise technologies has mostly succeeded. There is already a huge push in China to have domestic AI and semiconductor innovation (see China’s new AI fund from last week or their “Big Fund” for semiconductors). I imagine, in the future, they’ll try to use domestic design firms (like Huawei’s HiSilicon, Biren Technology, Enflame etc.) and domestic manufacturers (SMIC, Hua Hong).
In other chip news, China’s CXMT DRAMs are allegedly both cheaper, and (this is new:) improving quality.
Sounds a lot like when Japanese DRAMs beat American memory in the 1970s-80s. Fun fact: Intel used to be a memory company before ceding basically all their market share to Japanese companies.
China’s YMTC has a NAND design breakthrough.
1. Policy and Geopolitics
1.1
Bloomberg (01/28): DeepSeek’s AI Model Tests Limits of US Restrictions on Nvidia Chips
Powerful artificial intelligence software from Chinese startup DeepSeek indicates that its engineers built a competitive model despite US attempts to curtail China’s tech development, raising questions about the effectiveness of Washington’s trade curbs.
DeepSeek says it used less-advanced Nvidia H800 chips, which the US government allowed to be shipped to China until October 2023, to build a model that appears on par with the best offerings from OpenAI. Nvidia on Monday called the work an “excellent AI advance” that illustrates how companies can leverage “widely-available models and compute that is fully export control compliant.”
“Export laws limited the available resources, so Chinese engineers needed to get creative — and they did,” said Pat Gelsinger, Intel Corp.’s former CEO. “Engineering is about constraints.”
1.2
The Verge (01/28): Trump says he’ll put tariffs on imported chips ‘in the near future’
Donald Trump said tariffs on foreign computer chips, semiconductors, and pharmaceuticals are coming “in the near future.”
Bloomberg reports that later, in comments to reports, Trump said he wanted a tariff rate “much bigger” than 2.5 percent.
1.3
Nikkei (01/26): India seeks 'complementary' partnership in chip industry with Japan
India aims to promote investments from Japanese companies and build a solid manufacturing supply chain in the semiconductor industry by making a "complementary" partnership between the two countries.
2. Economy, Finance, and Business
2.1
NYT (01/27): Nvidia Reels After China’s A.I. Breakthrough
On Monday, shares of Nvidia plunged 17 percent after DeepSeek showed that it could train a cutting-edge A.I. system with a fraction of the Nvidia chips that had been used in the past by OpenAI, the maker of ChatGPT. The company lost roughly $600 billion in market value, on what was its worst trading day since the pandemic sell-off in March 2020.
But DeepSeek’s apparent breakthrough has shown that the appetite for Nvidia’s chips may not be as limitless as some had imagined just a week ago. While Nvidia is still in an enviable position — there is little competition for its A.I. chips — the companies that have been buying its technology could slow down their spending.
The DeepSeek release also dragged down shares of other semiconductor companies, including Broadcom, Micron Technology and TSMC.
2.2
WSJ (01/28): Asia Tech Stocks Slump After DeepSeek’s Rise Triggers U.S. Chip Selloff
The slump in technology stocks continued into a second day in Asia after the emergence of Chinese artificial-intelligence startup DeepSeek triggered a selloff in AI-related stocks overnight, raising concerns about the potential demand for the most advanced chips and data centers.
Regional tech stocks tumbled as SoftBank Group 5.8% in Japan, and chip-testing-equipment maker Advantest plunged 12%. Australian industrial-property owner Goodman Group, which has heavily invested in data centers globally, fell 8.2% on Tuesday.
Morningstar equity analyst Phelix Lee said that DeepSeek is likely to have minimal impact on the profitability of Taiwan Semiconductor Manufacturing Co. TSMC’s existing supply remains well short of demand, its factories are expected to run at full capacity even if companies pare their AI investments, he said.
2.3
FT (01/27): Here’s what the sellside is saying about DeepSeek
DeepSeek said it took 2 months and less than $6m to develop the model - building on already existing technology and leveraging existing models. In comparison, Open AI is spending more than $5 billion a year. Apparently DeepSeek bought 10,000 NVIDIA chips whereas Hyperscalers have bought many multiples of this figure. It fundamentally breaks the AI Capex narrative if true.
Firstly, it was trained in under 3 million GPU hours, which equates to just over $5m training cost. For context, analysts estimate Meta’s last major AI model cost $60-70m to train.
Secondly, we have seen people running the full DeepSeek model on commodity Mac hardware in a usable manner, confirming its inferencing efficiency (using as opposed to training). We believe it will not be long before we see Raspberry Pi units running cutdown versions of DeepSeek. This efficiency translates into hosted versions of this model costing just 5% of the equivalent OpenAI price.
Lastly, it is being released under the MIT License, a permissive software license that allows near-unlimited freedoms, including modifying it for proprietary commercial use.
2.4
TrendForce (01/27): Decoding CXMT: DDR5 Breakthroughs and What’s Next
Starting from late 2024, South Korea’s memory chipmakers have reportedly initiated production cuts on legacy DRAMs in response to declining prices driven by increased output from China. However, it now appears that CXMT is not only pressuring with lower-priced products but also with improved quality.
According to MyDrivers, CXMT has made notable strides in DDR5 production, as the yield rate has reportedly reached 80%.
The Chinese memory giant may be aiming for something bigger, as it is reportedly considering an initial public offering (IPO) in Hong Kong as early as 2025. According to Commercial Times, citing IFR, the company is currently in discussions with potential financial advisors and plans to raise approximately $1 billion.
2.5
Korea Times (01/28): SK hynix's surpasses Samsung Electronics for 1st time in quarterly profit
Chipmaking giant SK hynix has surpassed its biggest rival, Samsung Electronics, for the first time ever in terms of quarterly profit, data showed Tuesday, as robust sales of artificial intelligence (AI) memory chips pushed up its fourth-quarter figures to record highs.
According to its quarterly earnings report, SK hynix posted a record 8.08 trillion won ($5.62 billion) in operating profit for the October-December period, up 15 percent from a year earlier.
In comparison, Samsung Electronics estimated its fourth-quarter operating profit at 6.5 trillion won, including some 3 trillion won from the semiconductor division. The tech giant has yet to disclose its final earnings results.
2.6
WSJ (01/24): Meta Spending to Soar on AI, Massive Data Center
Mark Zuckerberg announced a huge leap in Meta Platforms’s capital spending this year to between $60 billion to $65 billion, an increase driven by artificial intelligence and a massive new data center.
2.7
ArsTechnica (01/28): A long, costly road ahead for customers abandoning Broadcom’s VMware
Broadcom's ownership of VMware has discouraged many of its customers, as companies are displeased with how the trillion-dollar firm has run the virtualization business since buying it in November 2023. Many have discussed reducing or eliminating ties with the company.
3. Technology
3.1
NYT (01/27): What to Know About DeepSeek and How It Is Upending A.I.
DeepSeek is a start-up founded and owned by the Chinese stock trading firm High-Flyer. Its goal is to build A.I. technologies along the lines of OpenAI’s ChatGPT chatbot or Google’s Gemini. By 2021, DeepSeek had acquired thousands of computer chips from the U.S. chipmaker Nvidia, which are a fundamental part of any effort to create powerful A.I. systems.
When DeepSeek introduced its DeepSeek-V3 model the day after Christmas, it matched the abilities of the best chatbots from U.S. companies like OpenAI and Google.
The world’s top companies typically train their chatbots with supercomputers that use as many as 16,000 chips or more. DeepSeek’s engineers said they needed only about 2,000 Nvidia chips.
The startup’s engineers demonstrated a more efficient way of analyzing data using the chips. Leading A.I. systems learn their skills by pinpointing patterns in huge amounts of data, including text, images and sounds. DeepSeek described a way of spreading this data analysis across several specialized A.I. models — what researchers call a “mixture of experts” method — while minimizing the time lost by moving data from place to place.
3.2
SCMP (01/27): Top Chinese memory chip maker YMTC makes another design breakthrough, defying US sanctions
Yangtze Memory Technologies Corporation (YMTC), China’s leading flash memory chip manufacturer, has achieved a significant technological breakthrough despite US sanctions, amid Beijing’s push for technological self-sufficiency.
YMTC has implemented its new Xtacking4.0 memory chip design in its highest-density 3D NAND chip, which was discovered in the commercial ZhiTai TiPro9000 solid-state storage device.
The chip features a dual-deck structure – a lower deck with 150 gates and an upper deck with 144 gates – totalling 294 gates. It uses what is known as a hybrid-bonding technique to join two wafers together.
3.3
TrendForce (01/28): Rapidus Reportedly to Install 10 EUV Machines, Aiming for Early Mass Production
While foundry giant TSMC plans to raise its 2025 capex to USD 38-42 billion, marking an up to 40% year-on-year increase, Japan’s Rapidus is also making strides in expansion. With government support backing it up, the company reportedly aims to introduce a total of 10 extreme ultraviolet (EUV) lithography machines to its cutting-edge semiconductor plants.