The bottlenecks in AI computing mainly come from the five aspects of computing power, ecology, power consumption, memory, and latency, and the current bottleneck has moved from computing power to connectivity.
The Zhitong Finance App learned that Huatai Securities released a research report saying that on March 20, Astera Labs (ALAB.US), known as “Little Nvidia,” was listed on the US NASDAQ. The company received this title due to the importance of its products to AI development and its ability to resolve AI bottlenecks. The bottlenecks in AI computing mainly come from the five aspects of computing power, ecology, power consumption, memory, and latency, and the current bottleneck has moved from computing power to connectivity. Nvidia (NVDA.US) has two major moats of high computing power and software ecosystem. It has addressed the first two major bottlenecks, while ALAB's high-speed data transmission technology is mainly aimed at solving the latter two major problems. The company's technology rival Mellanox, which was acquired by Nvidia in 2019, was also heavily “bet” on Amazon (AMZN.US).
The main views of Huatai Securities are as follows:
To address the two major AI computing bottlenecks of memory and latency, the time is right for Astera Labs to IPO
On March 20, 2024, Astera Labs (ALAB.US), known as “Little Nvidia,” was officially listed on the US NASDAQ. The company's acquisition of this title is mainly reflected in the importance of its products to AI development and its ability to resolve AI bottlenecks.
Huatai Securities believes that the bottlenecks in AI computing mainly come from the five aspects of computing power, ecology, power consumption, memory, and latency, and the current bottleneck has moved from computing power to connectivity. Nvidia has two major moats of high computing power and software ecosystem. It already addresses the first two major bottlenecks in AI computing, while ALAB's high-speed data transmission technology mainly targets the latter two major problems, which is the key to training AI models. Since Nvidia already has its own unique transmission solutions, but major tech giants are also developing their own AI chips, ALAB's solution may support its development.
Backed by a former Texas Instruments executive, the performance was remarkable after the IPO. The stock price rose by more than 70% on the first day
Astera Labs was founded in 2018 by three former Texas Instruments executives including Jitendra Mohan. ALAB issued a total of 19.8 million shares, with a listing price of 36 US dollars (above the original range of 32 to 34 US dollars). After closing on the first day, it reached 62.03 US dollars, an increase of more than 70%. The market value was close to 13 billion US dollars on March 25. The last round of financing before the company went public was in 2022, when the market capitalization was $3.15 billion. The company's investors include well-known Silicon Valley capital such as Fidelity and Intel.
Connected solutions target AI computing bottlenecks, and product pipeline improvement aims to divide equally with InfiniBand
Astera Labs mainly provides transmission connection chips for cloud and AI infrastructure to increase the data transmission speed of AI computing in data centers. The company currently has three product lines: 1) Aries smart DSP regenerators and Retimer chips to enhance connectivity between processors; 2) Taurus smart cable modules expand transmission between servers and switches via Ethernet; and 3) Leo data and memory connection semiconductors can expand the memory potential of cloud and AI infrastructure.
Against Mellanox, which was acquired by Nvidia, Broadcom also joined the team. Connectivity technology became the key to breaking the game
In 2019, Nvidia acquired Mellanox for $6.9 billion. The company's revenue for that year was US$1.33 billion, PS was 5.2x, profit was US$393 million, and PE was 17.6x. ALAB's transmission technology is comparable to that. Mellanox's Ethernet and InfiniBand technology can increase transmission speed to enable scale-out architectures (that is, increase processing power by adding more server nodes). This acquisition establishes Nvidia's leading position in data centers. Nvidia provides exclusive NVLink and NVSwitch interconnect technology to enhance transmission speed between chips and servers to address latency and memory sharing in data center networks.
Currently, Compute Express Link (CXL)'s new open interconnect technology has the potential to target this technology, and ALAB's core product lines are widely using CXL. Notably, Broadcom (AVGO.US) also launched a CXL-based PCIe Retimer chip in March of this year to target Arab's Aries in response to competition from this startup.
The product customer base is clear, and major manufacturers such as Amazon place heavy “bets”
ALAB's products are mainly aimed at three types of customers: hyperscale Internet companies, AI accelerator vendors, and system OEMs. One of the most high-profile events since the company was founded was its share warrant agreement with Amazon in 2022. Amazon can buy up to 1.5 million ALAB shares for $20.34 per share. The agreement will be amended next year, and Amazon will be able to purchase an additional 830,000 shares. Additionally, Amazon must purchase ALAB products worth 650 million US dollars within 7 years from 2022. ALAB's revenue last year was $116 million. Currently, in addition to Amazon, ALAB's disclosed customers include Intel, AMD (AMD.US), Supermicro, etc.
Risk warning: AI implementation falls short of expectations, industry competition is fierce, and competition between China and the US is intensifying.