AI Training
AI Training
A robot gets trained with focused commands with constant refinement. I used to work in business development at a beauty robotics company. We took photos of nail beds every day with as many people as possible to train the robot to paint manicures perfectly. We needed nails of different sizes, shapes, discolorations, textures, etc. The training algorithms never stopped running. We needed the perfect manicure. Our need for digital storage and computing power skyrocketed, and our energy bills were through the roof. And this was all for just the nails on your hands.
The requirement the world has for data centers, primarily driven by AI is huge. In the world of AI and all the noise currently in the market - you’d be surprised to learn this value investor has invested in a relevant AI player. As AI grows, the need for data centers will increase with it. We can’t fathom the possibilities the future of AI holds, but we do know we’ll need many data centers. The leading hyperscalers (i.e. Microsoft and friends) have already announced significant spending on these data centers which all plan to spend more in the coming years.
Currently, data center bandwidth is growing exponentially. As AI comes online, the increase in computing power grows exponentially as well. The training workloads for these AIs are computationally intensive. For larger machines and algorithms, you would need to reduce latency, which requires even more sophisticated networking requirements (like fiber-optic cables). The capital expenditure requirement for AI is astronomical.
Marvell produces semiconductor chips for various parts of the data center. They create custom computing chips for large tech companies and chips that facilitate the movement of data throughout the data center. They are the number one player for chips in the data interconnects market with a 70% market share of digital signal processors (DSP).
These high-end DSPs are critical to the transmission of data throughout the data center. The fastest way to transmit data is through an optic fiber (essentially light), but servers and switches throughout the data center communicate in electrical current rather than light. This means that the data must be converted from light to electrical and vice versa. The process of converting data from light to electrical is completed through the digital signal processors (DSP).
Meanwhile, every large tech company is currently developing its own custom chips for AI and non-AI as a way to hedge against Nvidia and create more task-specific chips. Designing advanced custom chips is complex and companies are partnering with Broadcom or Marvell. This is essentially a duopoly similar to Boeing and Airbus. Marvell has Amazon and Microsoft’s contracts as well as key chips from Google. With Broadcom price gouging their clients, some contracts end in blowups. There will likely be many more hyperscaler contracts on the way. Marvell currently has 1/10th of the business of Broadcom. However, Marvell’s custom chip team is the preferred partner relative to Broadcom. Marvell works hand in hand with their client’s R&D team while Broadcom prefers minimal interaction between their design team and their clients. Marvell is the cheaper competitor in the duopoly and still maintains a 50% margin. That is a good business.
"The development of AI is as fundamental as the creation of the microprocessor, the personal computer, the Internet, and the mobile phone. It will change the way people work, learn, travel, get health care, and communicate with each other. Entire industries will reorient around it. Businesses will distinguish themselves by how well they use it."
- Bill Gates in The Age of AI has Begun (March 2023)