Broadcom agrees to expanded chip deals with Google, Anthropic

The market for semiconductors designed for artificial intelligence is ceasing to be the domain of just one player. While investor attention is focused on GPUs from Nvidia, a quiet revolution of ASIC (Application-Specific Integrated Circuit) chips is taking place in the shadow of giant data centers. Broadcom, a key architect of network and silicon infrastructure, has just confirmed the strengthening of its cooperation with two powerhouses of the AI world: Google and Anthropic. This is a strategic move that redefines how large language models will be trained and deployed in the coming years.

Foundations for next-generation Google TPU infrastructure

Broadcom's collaboration with Google is not new, but its expansion to future generations of AI chips sends a clear signal to the industry. For years, Google has been developing its own TPU (Tensor Processing Units), striving to become independent of external providers of computing power. Broadcom plays a key role in this process as a design and manufacturing partner, providing essential intellectual property in high-speed interfaces and chip packaging technologies. The new agreement involves the production of subsequent versions of processors intended to power the Google Gemini ecosystem and Google Cloud services.

For Google, having its own silicon is not just a matter of prestige, but primarily of cost and performance optimization. ASIC circuits designed jointly with Broadcom are strictly tailored to the specific machine learning algorithms used by the Mountain View giant. This allows Google to achieve a better performance-to-energy ratio than when using universal GPU accelerators. Broadcom, acting as an "intermediary" between the design and the fab (most often TSMC), effectively becomes the guarantor of technological stability for the world's most ambitious AI projects.

The relationship between Broadcom and Google is currently one of the most important alliances in the technology industry. It allows Google to maintain the pace of innovation at the hardware layer, while Broadcom cements its position as a leader in the custom silicon segment, which is projected to grow much faster than the standard processor market.

Anthropic bets on dedicated silicon

A real surprise and confirmation of the market trend is the announcement of an expanded agreement with Anthropic. The creators of the Claude family of models, previously associated mainly with partnerships with Amazon and Google regarding the cloud, are opting for deeper hardware integration. The collaboration with Broadcom aims to develop solutions that will allow Anthropic to scale its systems more effectively. In a world where access to computing power is a scarce commodity, having dedicated silicon solutions is becoming a survival requirement for AI labs.

• Optimization for Claude: Dedicated chips will allow for faster inference of Anthropic models with lower latency.
• Network scalability: Broadcom provides XFI/SerDes technologies, which are crucial for communication between thousands of chips in computing clusters.
• TCO reduction: In-house hardware solutions drastically lower the total cost of ownership of infrastructure compared to renting power in an on-demand model.

Anthropic's decision to tighten ties with Broadcom suggests that the company intends to follow a path similar to Google or Amazon – from a software startup to a player controlling the full technological stack, from the algorithm to the transistor. For Broadcom, this is another opportunity to prove that their Jericho design platform and Tomahawk solutions are essential for building modern AI supercomputers, regardless of who provides the computing processors themselves.

Broadcom as the quiet winner of the AI arms race

Analyzing these reports, it is hard not to get the impression that Broadcom is emerging as the second most important force in the AI hardware ecosystem, alongside Nvidia. While Nvidia dominates the general-purpose segment, Broadcom has dominated the custom AI chips market. The company's business model, based on close cooperation with the largest "hyperscalers," allows them to generate billions in revenue with relatively low risk compared to building their own consumer brand.

The key to Broadcom's success is its dominance in the area of connectivity. No AI processor, regardless of its power, is useful without the ability to exchange data instantly with the rest of the cluster. PCIe Gen 5/6, CXL technologies, and Broadcom's advanced network switches form the connective tissue for Google and Anthropic systems. Expanding chip production agreements means that Broadcom is not just providing "cables and switches," but is becoming the design brain for the most demanding clients on the globe.

The current market situation shows that the era of relying solely on ready-made "off-the-shelf" solutions is coming to an end for the biggest players. Companies like Google and Anthropic understand that to gain a competitive edge, they must optimize hardware for specific transformer architectures. Broadcom, with its massive portfolio of patents and experience in 5nm and 3nm lithographic processes, is a natural and almost unrivaled partner in this process. This partnership effectively solidifies a duopoly in the high-performance computing market, where standard GPUs and highly specialized ASIC circuits compete for supremacy.