The 20 companies addressing the biggest challenges in AI

Booming AI is constantly facing new challenges that reveal new major players in the industry. So today's challenges could well help us find tomorrow's leaders.

As with any technological development, advances in artificial intelligence come with a host of challenges. Today, AI is faced with ever-increasing volumes of data and ever-higher execution speed requirements. Three issues are therefore becoming central: heat, data flow management, and energy efficiency. Here are four technologies that provide concrete answers to these challenges, and the companies we believe are best positioned in these segments.

Silicon photonics

Silicon photonics (or SiPh) refers to the application of photonics to silicon semiconductors. The principle is to replace technically limited copper cables with optical links capable of transmitting data via infrared wavelengths. This is where co-packaged optics, or CPO, comes in. Much like the transition from ADSL to fiber for the internet, this technology would allow more data to be transmitted while limiting heat and energy consumption. It's a promise that is already attracting the industry's giants.

• Intel has unveiled the first fully integrated bidirectional optical chiplet, co-packaged with an Intel processor, capable of transmitting 4 Tb/s over fiber optics up to 100 meters. The technology is designed to meet the growing needs of AI infrastructures in terms of both bandwidth and energy efficiency.
• Cisco is adopting a dual strategy: on the one hand, high-speed pluggable optical modules; on the other, active work on the direct integration of optical modules into switch ASICs, in other words, the famous CPO technology.
• Broadcom, a leader in communications chips, is a key player in silicon photonics. It has already integrated this technology into its Tomahawk chips, with Bailly Ethernet switches featuring eight optical engines for maximum density and bandwidth efficiency.
• Marvell, which is lagging behind in other segments, is investing heavily in integrated photonic engines capable of replacing traditional electrical interconnects with more efficient optical solutions.
• Nvidia took advantage of GTC 2025 to unveil two ranges of photonic network switches (Spectrum-X Photonics and Quantum-X Photonics), integrating CPO technology and merging electronic and optical components into a single package.
• Lumentum Holdings, a pure player in the optical sector, derives 80% of its revenue from the cloud and networks. Its catalog includes transceivers and wavelength managers.

The fourth generation of HBM

HBM (High Bandwidth Memory) chips have become essential in the development of AI. Integrated into GPUs to form modern graphics cards, they meet the growing need for massive data processing. Three major players are vying for Nvidia's first validation of a fourth-generation HBM chip.

• SK Hynix, which designed the first HBM chip in 2013, is now the industry leader. It is enjoying impressive momentum, with growth of 70% over the last three months.
• Micron, long considered an outsider, now seems to be neck and neck with the Korean company. Its share price, up 75% over the same period, reflects its strengthened position in the race.
• Samsung, for its part, has a solid memory division but is falling behind. Its momentum remains sluggish but could shift if a strategic shift is made.

Advanced packaging techniques

Buying an Nvidia H100 chip is not just about acquiring a GPU, but a complete card combining a GPU, HBM memory, PCI Express connectivity and other components. This assembly requires very high-level technical packaging, as it directly influences heat dissipation, bandwidth and power consumption. Several companies specialize in this crucial aspect of production.

• TSMC has developed CoWoS technology, which stacks several memory chips close to the processor. The goal is to reduce latency, improve heat dissipation, and boost bandwidth. It's a particularly well-designed structure.
• Intel is banking on its EMIB and Foveros technologies, which allow chiplets to be assembled like a 3D puzzle. With traditional scaling coming to an end, this modular packaging is one of its most serious assets against TSMC.
• Amkor offers a wide range of advanced technologies. System-in-Package (SiP) integrates processors, memory, and antennas into a single module. Flip Chip ensures better heat dissipation, and Wafer-Level Packaging (WLP) enables advanced miniaturization. Finally, 2.5D/3D stacking boosts performance in AI and autonomous vehicle applications.
• ASE Technology, a Taiwanese giant in chip packaging and testing, uses several advanced techniques: Fan-Out, 2.5D/3D IC, and SiP. These approaches have become essential in AI, chiplets, and smartphones.
• JCET, one of China's leaders in the sector, is developing XDFOI, its own version of high-density fan-out, ideal for AI and automotive applications. The company is also focusing on SiP and 2.5D/3D packaging, while producing RF modules with integrated antennas suitable for 5G and IoT.

 

Cooling solutions

With GPUs becoming increasingly powerful and demanding, cooling is becoming a critical issue. It comes into play at various levels: at the very heart of the chip, in the racks that house them, and in the overall data center environment. Here are a few names to watch closely.

• Asetek, a Danish company specializing in liquid cooling, designs systems that replace fans with water for better heat dissipation. Its flagship technology, direct-to-chip cooling (D2C), is aimed in particular at AI servers.
• Schneider Electric, a French flagship company founded in 1836, offers comprehensive solutions, from power supply to liquid cooling, through its Motivair brand. It also offers a software suite (EcoStruxure) to monitor and optimize the energy efficiency of infrastructures in real time.
• Vertiv, an American company founded in 2016, supplies direct and immersion cooling equipment (CoolCenter Immersion, CoolChip CDU) as well as hybrid solutions. It supports its products with software monitoring tools and installation and maintenance services.
• Carrier, another US player, specializes in industrial cooling. It offers cooling units, heat recovery solutions, and a software platform, QuantumLeap, dedicated to real-time thermal optimization.
• Finally,Dell is not content to stop at PCs. Its PowerEdge servers and Integrated Rack Scalable Systems are designed for high-density AI. Air cooling (Smart Flow), direct-to-chip liquid cooling (DLC), and complete thermal capture with the PowerCool eRDHx door: the company is pushing energy optimization to the limit, with up to 192 Blackwell Ultra GPUs per rack.