The future is fast, and Broadcom is betting big on it. But is their strategy enough to stay ahead of the curve? At SC25, Broadcom quietly unveiled a powerful piece of the puzzle for the next generation of server technology: the PEX90144, a massive 144-lane PCIe Gen6 switch. This isn’t just a chip; it’s a glimpse into the future of high-performance computing, specifically designed for the demanding workloads of 2026-era servers.
Spotted almost hidden within the bustling SC25 expo, the PEX90144 was integrated into a Dell PowerEdge XE7745-style PCIe board. The red board color suggests it’s a pre-release version, giving us an exclusive look at what’s coming. PCIe Gen6 is crucial for handling the immense data throughput expected from next-generation NVIDIA Blackwell Ultra GPUs and ConnectX-8 NICs (Network Interface Cards). Think of it as building wider highways to prevent data traffic jams in the most powerful servers.
Let’s dive deeper. The demonstration showcased the PCIe switch board within the server chassis. The PEX90144 boasts 144 lanes, which translates to the equivalent of nine PCIe x16 slots. Considering that the Dell PowerEdge XE7745 chassis housed eight accelerators, you quickly realize the necessity of such a high-lane count. But here’s where it gets controversial…
It’s not just about the accelerators. These powerful processors need to communicate efficiently. Each accelerator typically requires one or two NICs for networking, creating even greater I/O demands. And don’t forget the essential x16 link connecting the switches to the CPU. Suddenly, those 144 lanes start to feel less abundant. This highlights the crucial balance between processing power and efficient data transfer in modern AI and high-performance computing systems. The rear NIC I/O slots further demonstrate the sheer amount of I/O required in even standard CEM (Card Electromechanical) slot AI systems. You can see how even with so many lanes, the design is still very dense.
Broadcom’s history in PCIe switches, through Avago and PLX acquisitions, has positioned them as a dominant player. Their roadmap extends beyond Gen6, with the PLX91144 (144-lane PCIe Gen7) slated for 2027 and the PLX 92144 (144-lane PCIe Gen8) targeted for 2029. This demonstrates a clear commitment to staying at the forefront of PCIe technology. And this is the part most people miss… Broadcom isn’t just reacting; they’re proactively shaping the future of data transfer.
Broadcom’s aggressive roadmap is a direct response to increasing competition from companies like Astera Labs and XConn. Microchip’s presence in the server switch market has also seemingly diminished. This competition is ultimately beneficial for the industry, driving innovation and delivering better solutions for AI, storage, and CXL (Compute Express Link) applications. However, Broadcom faces a significant challenge from NVIDIA. NVIDIA is increasingly pushing Broadcom out of key opportunities in high-speed NICs and PCIe switches within their own B300 servers, potentially impacting Broadcom’s market share. It’s a high-stakes game, and execution is paramount.
According to Patrick Kennedy’s analysis, this shift could represent a loss of over $4000 in Bill of Materials (BOM) opportunity for Broadcom per NVIDIA B300 server. This is a bold claim and could spark differing opinions. You can read more about this analysis on his Substack. Does this mean Broadcom is losing its grip on the AI server market, or can they adapt and maintain their dominance through innovation and strategic partnerships? What are your thoughts? Is NVIDIA’s integration a smart move or a potential bottleneck for future server development? Share your opinions and let’s discuss the future of PCIe switches in the comments!