Image source@Visual China
文 | 甲子光年，作者｜赵健
Singapore, a small country, is becoming an AI powerhouse.
On November 21, 2023, Nvidia's filings with the U.S. Securities and Exchange Commission (SEC) showed that about 15% of Nvidia's revenue (about $2.7 billion) in the third quarter of fiscal 2024 (corresponding to August ~ October 2023) came from Singapore, ranking behind the United States, Taiwan, and Chinese mainland (including Hong Kong), ranking among Nvidia's fourth largest markets in the world.
It is worth mentioning that this is the first time that Nvidia has disclosed data on Singapore as a separate consumer market in its earnings report. Prior to this, Nvidia only disclosed revenue data in the United States, Chinese mainland and Taiwan, which usually accounted for 70%~85% of the market share, and other markets, including Singapore, were classified as "other".
Recently, Nvidia CEO Jensen Huang came to Singapore on his first stop in Southeast Asia after Japan, which also confirmed Singapore's strategic position in the Southeast Asian market. Huang announced a partnership with Singapore in the field of homegrown AI models, as well as future infrastructure investments. This is Huang's second visit to Singapore in 25 years.
Singapore is the financial, economic and shipping center of Southeast Asia, but its land area is only 733.1 square kilometers (2021), which is about 1/50 of Taiwan Province. Such a small country has a small territory, but it is a big chip consumer. Last quarter, Singaporeans spent $600 on Nvidia chips per capita, compared to $60 in the United States and about $3 for Chinese.
Why is Singapore the world's fourth-largest market for NVIDIA?
GPU revenue increased by 4 times in a single quarter
In the third quarter of fiscal 2024, NVIDIA's total revenue was $18.12 billion, and the four major global markets were the United States, Taiwan, Chinese mainland, and Singapore, contributing revenue of $6.302 billion, $4.333 billion, $4.030 billion, and $2.702 billion, respectively.
Singapore's demand for NVIDIA's computing power suddenly "soared" almost in the last quarter. In the third quarter of the last fiscal year, Singapore contributed only $536 million to Nvidia's revenue, accounting for 9.04%, while in the third quarter of this year, Singapore contributed 14.91% of Nvidia's revenue.
During the same period, Nvidia's overall revenue grew by only 200%, while Chinese mainland contributed 250% and Singapore increased by 400%. Singapore has also gone from being a "national region" to becoming Nvidia's fourth-largest market.
Who is buying Nvidia's chips?
Nvidia does not directly disclose customer names. In the earnings report, Nvidia said that end customers usually do not purchase directly from Nvidia, but from Nvidia's customers, including OEMs, system builders, system integrators, retailers/distributors, automakers and Tier 1 automotive suppliers. In the third quarter of this year, Customer A's sales accounted for 12% of total revenue, and Customer B's sales accounted for 11% of total revenue in the first nine months, both of which were attributed to the Computing & Networking segment.
The Computing & Networking Group is NVIDIA's No. 1 division, which includes the most popular data center accelerated computing platforms. In the third quarter of this year, NVIDIA's data center revenue reached $14.514 billion, up 279% year-on-year.
H100 is currently the strongest chip among Nvidia's products on sale (the stronger H200 is expected to be launched in the second quarter of 2024), and it is an irreplaceable "AI fuel" for data center construction. According to Omdia Research, Meta and Microsoft bought the most H100 in the third quarter, with 150,000 copies each, followed by Google, Amazon, Oracle, Tencent, CoreWeave, Baidu, Alibaba, Lambda Labs, ByteDance and Tesla.
英伟达H100出货量，图片来自Omdia Research
Although it is not possible to directly link the list of these tech giants to the Singapore market, almost all of the cloud vendors on this list have built and operated data centers in Singapore.
Will data centers be the main demand for computing power in Singapore?
A small country with a small territory, a large data center country
"What does a small country do with these chips, of course, to build data centers," Sang Shin, a former Temasek and GIC executive, said on LinkedIn. He previously served as Director of Digital Innovation at Temasek and Head of Digital Strategy and Architecture at GIC, and is now the founder of AI startup FoFty.
Despite its small size, Singapore is a big data center country.
According to a report by Cushman and Wakefield, Singapore ranks third in the world and first in the Asia-Pacific region in terms of data center market rankings. Tied for first place are Northern Virginia and Portland in the United States, and Hong Kong in fourth place.
Due to its robust infrastructure, excellent international and regional connectivity, low risk of natural disasters, and widespread use of digital technologies, Singapore has emerged as a data center hub in the Asia-Pacific region, bringing together 60% of the region's data centers. Many large technology companies in China and the United States have chosen to build data centers in Singapore, such as Google, Amazon, Microsoft, and IBM in the United States, and Alibaba, Tencent, and ByteDance in China.
According to a report by the Singapore Economic Development Board (EDB), one-third of the world's top 500 companies have regional headquarters in Singapore.
However, data centers are a "big power guzzler". Singapore currently has more than 70 operating data centres with a total power of over 378 MW, accounting for more than 7% of Singapore's total electricity.
In order to meet the environmental obligations stipulated in the 2015 Paris Agreement and meet the challenge of reducing carbon emissions, the Singapore authorities issued a "ban" on data centers in 2019, suspended the construction of new data centers, and began to vigorously develop green energy such as solar energy and hydrogen energy, and tried to apply green energy, liquid cooling and other technologies to sustainable data centers.
After three years of "cooling", Singapore lifted the ban in January 2022. In July 2022, the Singapore Economic Development Board (EDB) and the Infocomm Media Development Authority (IMDA) announced a pilot scheme to allow the construction of new data centre projects that meet the requirements.
The advent of ChatGPT at the end of 2022 added fuel to the fire in Singapore's data center market, which is back to growth. GPU-powered accelerated computing data centers have become an indispensable engine for the development of the generative AI industry. The biggest beneficiary is, of course, Nvidia, whose CEO Jensen Huang vowed this year: "We are in the first year of a decade-long data center intelligence." ”
In July 2023, Singapore authorities allocated around 80MW of new capacity to four data centre operators, the first capacity allocation since the data centre ban was eased. The four operators are a consortium of U.S. data center operators Equinix and Microsoft, Chinese data center operators GDS, and Australian data center operators AirTrunk and ByteDance.
Singapore hopes to continue its position as a data center leader in the Asia-Pacific and global markets in this wave of generative AI. Singapore's Lianhe Zaobao said in the article "Singapore Opportunities for Global Computing Power Center" that after AI leads a new round of technological explosion, computing power will become an "upstream resource" in the global economic cycle. If Singapore can seize the opportunity to become a regional and global computing power center, it will occupy an important position in the regional and global high-tech industry chain.
In addition to the construction of data centers, Singapore is also increasing its investment in artificial intelligence. A practitioner who graduated from Nanyang Technological University in Singapore told Jiazi Lightyear that artificial intelligence has always been a direction supported by the Singapore government.
The Singapore government released an updated National Artificial Intelligence Strategy 2.0 (NAIS 2.0) on December 4, 2023, which aims to more than triple the number of local AI practitioners to 15,000 over the next three to five years and establish an "iconic" website to nurture the country's AI community.
Singapore's Deputy Prime Minister Lawrence Wong speaks at the NAIS 2.0 launch ceremony on December 4, 2023, photo: CNA/Davina Tham
Recently, Huang visited Japan, Singapore, Malaysia and Vietnam to discuss cooperation and investment with local governments and large corporations. Huang said Nvidia will deepen its cooperation with Singapore in the field of artificial intelligence, and is currently working with the Communications Media Development Authority to create a large language model in 11 languages to develop local AI for use by researchers, startups and the AI industry in Singapore.
Nvidia already has a supercomputer in Singapore, but wants to do more in Singapore, including building a bigger supercomputer and "possibly investing in a major AI iconic website".
Huang also spoke in Singapore about the need for AI chips and systems, saying that the world's largest internet and cloud service providers launched the first wave, and Nvidia sees the next wave rapidly emerging, with the core demand being the state — countries building sovereign AI infrastructure, followed by industries, companies and sectors building AI factories.
One clear trend is that the two computing power needs of data centers and artificial intelligence are driving a large number of GPU deployments in Singapore.
Is that the whole reason, though?
GPU brokers
Will Singapore's local data center construction and artificial intelligence digest all NVIDIA chips? The answer is no.
In fact, the construction of data centers in Singapore has strict carbon emission targets.
Singapore authorities stipulate that data center construction proposals must invest in hydrogen or solar panels, have a power usage effectiveness (PUE) of 1.3 and below, have a platinum certification under Singapore's BCA-IMDA Green Label, and use "state-of-the-art technology and best practices in sustainability". In addition, the proposal should strengthen Singapore's international connectivity and status as a regional hub, and make a significant contribution to Singapore's broader economic objectives.
So, that's why the first batch of data center construction capacity was allocated to a consortium of four powerful companies: Equinix, Microsoft, GDS, AirTrunk and ByteDance.
The head of the Singapore office of a Chinese company told Jiazi Lightyear: "Due to the limitation of carbon emission indicators, the demand for data center construction in Singapore is actually not that high. The chips purchased in Singapore are likely to be distributed to other countries in Southeast Asia to build data centers. ”
In a LinkedIn post by Sang Shin, a former Temasek and GIC executive on Singapore's chip market, Henry Goh, COO of MACROKIOSK, a Malaysian enterprise marketing company, made a similar point: "These chips are not necessarily consumed or used in Singapore, but are resold from entities in Singapore and redistributed throughout Asia (except China). ”
Henry Goh said that Nvidia has only one ASEAN office in Singapore, which means that all ASEAN countries will buy chips through this office. THAT'S HOW MACROKIOSK WORKS.
At the same time, he also said that if this is indeed the case, as for why Nvidia reflected Singapore instead of Southeast Asia in its financial report, only Nvidia officials may be able to answer this question.
Southeast Asia is currently a gold rush, and the demand for data center construction is also rising. Cushman & Wakefield said in its Asia-Pacific data center report for the first half of 2023: "The suspension of IT capacity in Singapore has led to unmet market demand, which has spread to nearshore markets such as Johor in Malaysia, where a large number of data center projects are being developed and land banks are being committed. Similarly, the mega data center in Jakarta, the capital of Indonesia, is driven by its central location in Southeast Asia, and the country's huge population growth maintains its attractiveness to major investors and operators. ”
It is worth mentioning that due to the US ban, the Chinese market was naturally excluded from Singapore's reselling system.
On October 17 this year, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) issued new regulations on the export ban on chips, further escalating sanctions on Chinese semiconductors. Affected by this, Chinese customers will not be able to buy Nvidia's A100, A800, H100, H800, L40, L40s and other high-performance chips.
However, the overseas business of Chinese companies has not been affected for the time being. The person in charge of the above-mentioned Singapore office told Jiazi Lightyear that their office in Singapore can purchase chips from Nvidia through normal channels. Of course, compared to the huge market demand in Chinese mainland, the demand for these overseas businesses may be insignificant.
Nvidia does not want to give up on the Chinese market, not only because China is Nvidia's third-largest market, which contributed 22% of revenue in the last quarter, but also because the ban could force Chinese chip companies to accelerate their growth and become competitors to Nvidia.
After being stuck by Nvidia's GPU, China's AI companies are not too pessimistic, but gradually realize domestic substitution.
At HUAWEI CONNECT in September this year, Huawei unveiled the Ascend AI computing cluster with a new architecture, claiming that it can support large-scale model training with more than one trillion parameters, providing customers around the world with a second computing power option. iFLYTEK is an important large-scale customer of Huawei, and its chairman Liu Qingfeng has publicly stood up for Huawei, saying that Huawei's Ascend 910B can already benchmark against NVIDIA's A100.
Jiazi Lightyear has learned that Huawei's Ascend AI computing center has cooperated with many places to deploy the Ascend 910A chip, and is expected to expand the capacity of the new Ascend 910B chip by the end of this year. A number of industry sources revealed to Jiazi Lightyear that a full set of domestic intelligent computing clusters based on the Ascend 910B chip are being deployed to relevant intelligent computing centers, "In some large model tests, both 910A and 910B can be benchmarked against NVIDIA A100, and 910B will have better performance." The relevant person in charge of a local intelligent computing center revealed.
In an interview in Singapore, Huang also publicly said that Huawei, Intel and growing semiconductor startups pose a serious challenge to Nvidia's dominance in the AI accelerator market.
In addition to Huawei, there are also a number of chip startups in China that are gradually maturing and successively bringing their AI chip products to the market and being tested by customers. At the AICC Artificial Intelligence Computing Conference held by Inspur Information in November, there was a chip wall dedicated to displaying the computing cards, acceleration cards, training cards and inference cards of various chip companies.
Domestic chip products, "Jiazi Lightyear" shooting
There is still an obvious gap between Chinese chips and NVIDIA, not only in the product itself, but also in the ecology. At the 2023 Jiazi Gravity Year-end Ceremony held by "Jiazi Light Year" recently, Yu Yi, COO of Shansi Enlightenment, said that most domestic manufacturers, if they develop their own ecology or software, may not know how to use it.
The technical person in charge of a domestic large-scale model company told "Jiazi Lightyear" that the biggest problem in adapting to domestic computing power lies in algorithm adaptation, which requires 2~3 months of engineering and a huge amount of work. For large models with extremely fast algorithm iteration, this is a time cost that AI companies cannot afford.
However, if NVIDIA chips cannot be obtained, the algorithm company will be forced to cooperate with domestic computing power companies to find problems and iterate products in practice, which will accelerate the benign development of the ecology in the long run.
If the large model is the next industrial revolution, this is not only a dangerous challenge for domestic computing power and models, but also a huge opportunity.