Broadcom Builds the Proprietary Side of AI — Part 1

# Nvidia Under Pressure: Clients Are Building Their Own Infrastructure

The most significant competitive pressure on Nvidia stems from the decision by its largest customers to take direct control of an increasingly substantial share of the technology stack.

This article opens a two-part deep dive dedicated to Broadcom's role in artificial intelligence infrastructure. The first part traces the evolution toward hybrid systems, the value of custom silicon, networking, and partnerships with the leading hyperscalers. A second part will follow, focused on AI-related revenues, equity valuation, and the key risks to the investment thesis.

The potential erosion of Nvidia's dominance may come from its own largest customers rather than from any single competitor. The collaboration between OpenAI and Broadcom envisages the deployment of OpenAI-designed accelerators at up to 10 gigawatts of capacity, with rollouts commencing in the second half of 2026 and full completion targeted by end-2029. A portion of workloads favours the optimal balance between performance, power consumption and total infrastructure cost over maximum operational flexibility.

This is where the investment thesis on Broadcom takes shape.

Google, Meta, Amazon, Microsoft and the leading artificial intelligence model developers continue to purchase GPUs and pour vast sums into compute capacity. What is changing is the way in which that capacity is being built. Alongside general-purpose GPUs, purpose-built accelerators are gaining ground, accompanied by proprietary networking, custom rack designs and systems in which hardware, memory, interconnects and software are co-optimised as a single, unified architecture.

Broadcom sits at the intersection of these two transformations. On one hand, it helps customers translate their requirements into custom silicon. On the other, it provides a growing share of the networking fabric that enables thousands of accelerators to work in concert.

This is an industrial story far more interesting than a simple comparison between two stocks. But it is also a story the market already knows. That is why one must distinguish between the quality of Broadcom's competitive positioning and the price investors are willing to pay for that quality.

The Era of "GPUs Only" Is Coming to an End

Nvidia GPUs remain the benchmark for artificial intelligence by virtue of their computational power, development velocity, and — above all — the software ecosystem built around CUDA. For those who must train frontier models, iterate rapidly, or manage continuously evolving workloads, the flexibility of a general-purpose platform retains enormous value.

The inflection point emerges when a hyperscaler runs similar operations billions of times and knows with precision how its models behave. At that point, flexibility objectives are joined by cost per token, consumption per rack, effective capacity utilisation, memory availability, and the ability to schedule investments through a more diversified supplier base.

The most rational strategy consists in building a portfolio of accelerators and assigning to each the workload for which it is most efficient.

Meta states it openly: the company adopts a "portfolio approach", pairing the right accelerator to each individual workload in order to achieve the optimal balance between performance and total cost of ownership. GPUs remain an essential component of the infrastructure, while MTIA chips are being developed for inference, recommendation, ranking and, progressively, generative workloads.⁵⁶

Nvidia's response also confirms that future infrastructure will be hybrid. With NVLink Fusion, the company allows hyperscalers to integrate custom CPUs and XPUs within its own rack-scale architecture. Nvidia acknowledges the growth of custom chips and is positioning them to coexist with its own GPUs, leveraging its interconnect, its racks, and its ecosystem.⁹

Big Tech continues to purchase from Nvidia while simultaneously allocating a growing share of future capacity to more specialised systems. The industrial thesis hinges on the proportion of growth absorbed by these systems and on vendors' ability to monetise the progressive fragmentation of the stack.

The new competition concerns both who sells the most powerful processor and who controls the architecture through which different processors work together.

In the infographic below, we examine how different workloads can be distributed across various platforms, while Broadcom intervenes at several levels of the technology stack.

Back to index ↑

Why Big Tech wants purpose-built chips

GPU vs ASIC: flexibility versus specialisation

A GPU is built to be extremely versatile. It must support different models, frameworks and workloads, while maintaining an ecosystem usable by thousands of developers. An ASIC, or application-specific integrated circuit, sacrifices a degree of that flexibility in order to be optimised around operations defined in advance.

When a workload is stable, repetitive and sufficiently large, specialisation can improve performance per watt, reduce certain operating costs, and make more efficient use of memory and interconnects. In inference, this potential advantage is particularly significant: on top of model training comes the need to serve an enormous number of queries with low latency and at sustainable cost.

The economics depend on the quality of execution. Designing a chip requires capital, expertise, access to advanced packaging, HBM memory, manufacturing capacity, and a software stack capable of exploiting it. A project that is delayed, runs at insufficient volumes, or is built around a workload that is likely to change rapidly may prove more costly than a solution already available on the market.

This is why proprietary chips are a strategy primarily for those who possess three elements: massive scale, direct knowledge of their own workloads, and a roadmap long enough to justify multiple generations of silicon.

The value derives above all from the ability to embed in the chip what the company has learned about its own models, memory patterns, query management, and data centre organisation.

OpenAI described precisely this rationale when presenting Jalapeño, its first 'Intelligence Processor' designed for inference. The architecture was defined around the company's models, kernels, memory, networking, and serving systems. Initial statements indicate an improvement in energy efficiency relative to reference platforms; full technical data will be released with the report announced by the companies.⁴

Broadcom, the partner behind custom silicon

Broadcom operates as a technology partner capable of guiding the customer from logical design through to the realisation of an industrialised system.

The customer knows its own workload and defines the architectural objectives. Broadcom provides intellectual property, ASIC expertise, high-speed SerDes, interfaces, advanced packaging, memory integration, networking, and the competencies developed through building products destined for hyperscale volumes.

The correct designation is fabless: Broadcom designs, develops, and delivers solutions, while the physical production of silicon remains entrusted to foundries and supply chain partners.

The difference relative to a standard chip is also an economic one. A custom accelerator is developed in conjunction with a handful of hyperscalers, on multi-year roadmaps and at volumes that can become enormous. This creates a very deep relationship with the customer, but also increases concentration and the risk that a change in the roadmap produces material variations in revenue.

Broadcom's competitive advantage derives from its ability to combine different technology building blocks. The design of the compute engine represents the starting point: it must be connected to memory, data ingress and egress must be managed, packaging must be integrated, the system must be ensured to be manufacturable at scale, and it must be inserted into a rack that communicates with the rest of the cluster.

The company is also investing in 3.5D and face-to-face packaging technologies to increase XPU density and integrate chiplets, HBM memory, and high-bandwidth interconnects. These are elements that are largely invisible to retail investors, but are decisive in determining performance, power consumption, and the economic yield of the system.¹⁰

The chip is only half the story

In an AI cluster, the processor operates within a network of thousands of accelerators that exchange data, synchronise operations, and access memory and storage with sufficiently low latency. A network slower than the compute leaves a portion of processing capacity idle.

This is why networking has become a central component of the economics of artificial intelligence.

Broadcom covers Ethernet switches, network interface cards, retimers, optical DSPs, SerDes, routing, and co-packaged optics. The Tomahawk family is designed to connect AI clusters both in the scale-up phase, within tightly coupled accelerator domains, and in the scale-out phase, as the system expands across different racks and data centres.

Tomahawk 6 delivers 102.4 terabits per second of switching capacity on a single chip. According to Broadcom, it can connect 512 XPUs in a scale-up configuration and support scale-out networks of more than 100,000 accelerators with two tiers of switching; the architecture has been designed for clusters that, over time, may scale to as many as one million XPUs.¹⁰

These figures are technical claims made by the manufacturer and must be accompanied by their source. The economic significance, however, is straightforward: the greater the number of accelerators, the more critical the network connecting them becomes.

This is where Broadcom can benefit on two fronts. It can participate in the design of custom compute and, at the same time, sell the infrastructure required to connect XPUs, GPUs, and heterogeneous systems. Networking can also be deployed in clusters that include accelerators from other manufacturers, provided the architecture adopts Ethernet and compatible components.

Broadcom's coverage encompasses a portion of the stack. Nvidia owns NVLink, InfiniBand, Ethernet, SuperNIC, DPU, complete rack solutions, and management software. Competition is shifting towards a contest between platforms, standards, and architectures, beyond the individual processor.

In the following image, we observe the role of the Ethernet network as a common layer across accelerators from different manufacturers and customers.

OFFICIAL TECHNICAL FIGURE

Ethernet Network for AI

Back to index ↑

Three partnerships that give substance to the thesis

OpenAI: from model to silicon

In October 2025, OpenAI and Broadcom announced a collaboration to deploy up to 10 gigawatts of accelerators designed by OpenAI. The systems include custom processors and Broadcom Ethernet solutions for scale-up and scale-out. Deployment is expected to begin in the second half of 2026, with completion indicated by end-2029.³

In June 2026, the two companies unveiled Jalapeño. OpenAI defined the architecture around its own inference workloads, while Broadcom contributed to silicon implementation, networking and system industrialisation. The project is described as the first step of a multi-generation roadmap.⁴

The value of the project lies in the ability to transfer directly into silicon what a company has learned about the behaviour of its own systems. If the advantage translates into lower costs and additional capacity, Broadcom will have helped transform a GPU customer into a proprietary infrastructure designer, while purchases from Nvidia or AMD are maintained.

Meta: the accelerator portfolio

Meta has extended its collaboration with Broadcom across multiple generations of MTIA. The agreement covers chip design, advanced packaging and networking. The first phase exceeds one gigawatt and is presented as the beginning of a multi-year, multi-gigawatt deployment.⁵

The framing chosen by Meta is particularly useful for interpreting the broader industry: the company matches the processor to the workload. MTIAs are optimised for recommendation, ranking and inference, while other workloads continue to run on GPUs and different platforms.⁵⁶

This is arguably the most accurate description of the industry's direction: a more selective allocation of capacity, distinct from wholesale substitution.

Google: a strategic, multi-vendor relationship

According to Reuters, Broadcom has reached a long-term agreement with Google to develop and supply future generations of custom AI chips and components for next-generation racks through to 2031.⁷

The relationship with Google is significant because TPUs represent one of the most mature cases of a proprietary accelerator developed across multiple generations. Google uses these platforms internally and also offers them through Google Cloud, building an industrial alternative already integrated into its own infrastructure.⁸

The relationship retains an open, multi-vendor character. Reuters subsequently reported discussions between Google and Marvell regarding the development of other AI chips. Supplier diversification, which underpins the broader thesis on custom silicon, simultaneously represents a specific risk for Broadcom.¹³

In the three cards below, we examine the scale, time horizon and scope of the partnerships that give concrete form to the industrial thesis.

OpenAI

• 10 GW of custom accelerators
• Systems with Broadcom Ethernet
• Rollout expected in the second half of 2026
• Completion indicated by end-2029

Official OpenAI source ↗

Meta

• Multiple generations of MTIA
• Design, packaging and networking
• First phase exceeding 1 GW
• Multi-gigawatt roadmap

Official Meta source ↗

Google

• Long-term partnership according to Reuters
• Future generations of custom AI chips
• Components for next-generation racks
• Reported horizon through to 2031

Google image and context ↗

Back to index ↑

The second part completes the analysis with AI-related revenues, the valuation framework, expectations priced into the stock and the key risk factors. Go to Part 2 →