Broadcom CEO on the Biggest AI Chip Bets

Key Concepts

• Custom-Owned Tooling (COT): The strategy where large tech companies (like Google) design their own proprietary AI chips rather than relying solely on off-the-shelf solutions.
• AI Accelerators: Specialized hardware (like TPUs) designed to perform the massive parallel processing required for machine learning workloads.
• Generative AI Compute: The infrastructure (chips, networking, and software) required to train and run Large Language Models (LLMs).
• Token Maxing: A colloquial term for the unrestricted usage of AI model tokens, often associated with high operational costs.
• Return on Investment (ROI) in AI: The metric used to justify the cost of AI compute, focusing on productivity gains (e.g., one engineer using AI to do the work of ten).
• Photonics/Optics: High-speed data transmission technologies increasingly critical for connecting AI clusters.

---

1. AI Market Dynamics and Investor Expectations

The speaker addresses the "surreal" environment of the current AI hype cycle, where companies can report massive revenue growth and reaffirm strong forecasts, yet still face investor disappointment. The core philosophy presented is to ignore stock price volatility and focus strictly on fundamentals and value creation. The speaker argues that the industry is still in the "early innings" of generative AI, and the focus should remain on the long-term utility of these tools rather than short-term market sentiment.

2. Strategic Partnerships and Custom Silicon

Broadcom’s business model involves deep, collaborative partnerships with major tech players to develop custom AI-accelerated silicon.

• Google Partnership: Broadcom co-designs Tensor Processing Units (TPUs) with Google. While Google explores "customer-owned tooling" (designing chips in-house), Broadcom views this as a natural evolution in the semiconductor industry. Broadcom’s goal is to "out-engineer" the competition by providing superior, differentiated technology.
• OpenAI Collaboration: Broadcom is currently engaged with three major partners (including OpenAI) to develop custom AI accelerators. The speaker confirmed that these projects are on track, with production expected to begin late this year. He explicitly denied rumors of "snags" or requirements for Microsoft to guarantee purchases.
• Competitive Landscape: The speaker dismisses smaller competitors as "ankle biters" and views the primary benchmark for performance as Nvidia’s GPU roadmap. Broadcom’s strategy is to provide the necessary "picks and shovels" for the AI gold rush.

3. The Anthropic Case Study

Broadcom’s partnership with Anthropic is framed as a "leap of faith" in the potential of generative AI to transform enterprise productivity.

• Risk Management: Rather than fearing "being left holding the bag," the speaker views the investment as a bet on the viability of the generative AI business model.
• Productivity Evidence: Broadcom uses these tools internally for engineering, design, and code assistance. The speaker notes that AI tools are already producing designs that rival or exceed the output of their own highly skilled engineers.

4. Operational Strategy: Productivity vs. Throttling

Regarding the "token maxing" debate, the speaker argues against imposing strict limits on AI usage.

• ROI Framework: The decision to limit usage is purely an ROI calculation. If an engineer can use AI to complete a one-week task that previously required ten engineers working for three months, the cost of the tokens is negligible compared to the massive productivity gain.
• Learning Curve: The speaker emphasizes that AI tools require a "learning process." Productivity increases as engineers become more proficient at prompting and integrating these tools into their workflows.

5. M&A Strategy and Market Positioning

• Post-M&A Phase: Broadcom is currently prioritizing organic growth over mergers and acquisitions. With revenue doubling over a two-year period (reaching $50 billion annualized), the speaker views M&A as a "distraction" that diverts focus from the "insatiable" demand for AI compute.
• Networking Dominance: Broadcom maintains a portfolio of 17 semiconductor divisions, with networking being a primary growth engine. The speaker remains unconcerned about competition from companies like Cisco, noting that Broadcom prioritizes "strategic customers" who are committed to long-term, multi-generational technology roadmaps.

Notable Quotes

• "Stop thinking about your stock price. Trouble is very hard to do that, but we try to do it."
• "We just compete against my own customer... the whole idea is to be able to create differentiated product and technology that beats what they have."
• "You can get one great very senior engineer to produce an application design in one week. What you would otherwise take 10 engineers... and take them three months to produce the same thing."

Synthesis

Broadcom’s strategy is defined by a disciplined focus on "picks and shovels"—providing the essential networking and custom silicon infrastructure for the AI revolution. By prioritizing deep, long-term partnerships with a select few "strategic" customers and focusing on the tangible ROI of AI-driven productivity, the company aims to bypass the volatility of the AI hype cycle. The speaker concludes that the industry is still in the early stages of discovery, and the primary objective remains the continuous improvement of hardware and software to meet the massive, growing demand for compute capacity.