A CIO’s Guide to Navigating Frontier AI
I’ve witnessed countless technology cycles during my career, but nothing compares to the disruptive force of frontier AI models. Today’s large language models are the product of staggering investments of trillions of dollars in hardware, data science, and human expertise. They’ve trained on vast troves of text, images and voice data, and their capabilities are nothing short of extraordinary. From generating code to automating digital tasks, these models deliver strategic advantages and tactical benefits at a speed we’ve never seen before. In this blog series, I’ll be taking a closer look at their advantages and risks, as well as the guardrails that need to be in place to ensure CIOs can help their organizations maximize these investments.
Power, Probability and the Illusion of Intelligence
Let’s be clear: There’s nothing intelligent about artificial intelligence. At its core, AI is just matrix multiplication with weights, making choices based on probability. That probability is determined by the weights, but it’s also determined by the data you’re telling it. If you tell the model that the sky is green, it will give you the answer that the sky is green. These models are powerful tools, not sentient beings. Their outputs are only as good as the data and objectives you provide.
At Broadcom, we approached frontier AI with clear, deliberate business objectives rather than chasing the technology for its own sake. Instead, we started with a simple question: what problems can we solve more effectively with these tools? Our five core objectives included:
- Protect intellectual property: Approximately 74% of our workforce is dedicated to R&D and operations, that means engineers are inventing hardware and software at scale, so safeguarding IP is our number one priority. Broadcom is vigilant about ensuring its models do not train on proprietary data in ways that could benefit competitors.
- Accelerate engineering productivity and time to market: On the hardware side, AI is being used to accelerate chip development, compressing time to market as a competitive advance. On the software side, the focus is on delivering richer, more secure and more functional releases rather than simply shipping more versions.
- Enable smart operations and self-healing IT: Broadcom is investing in AI-driven operational capabilities, including self-healing environments and reducing mean time to recovery, keeping infrastructure up and running more reliably.
- Deliver data transparency through natural language: Rather than building more dashboards, Broadcom is enabling business users to ask questions in natural language and receive direct answers, making data far more actionable.
- Improve employee productivity: From summarizing meetings to improving presentations, AI is making day-to-day work more efficient across the organization.
Choosing the Right Model Providers
Choosing the right model providers is critical. We rely on Google for natural language processing and understanding unstructured data, and Anthropic for coding and engineering tasks. Each has unique strengths and we practice adversarial review, routinely feeding outputs from one model into another to get different perspectives and patch errors, recognizing that different models carry different strengths and subtle differences in approach. I hate putting human attributes onto digital assets, but they do have personalities for want of a better description. For that reason, it’s good to have at least two, maybe three, to do the adversarial review. At Broadcom, we are still evolving. We may add a third model provider as we go through the rest of this year and into 2027.
The Security Imperative: Speed Has Changed Everything
Our journey with frontier AI began just over a year ago. No one has a long history with these models – anyone who claims otherwise isn’t being honest. We’re still learning, evolving as technology matures. The first step was understanding where our data was going, how it was being processed, what risks we were exposing, and how it was interacting with other elements in that environment. Security reviews became comprehensive and intense, covering not just our IP but any data that could be leveraged by competitors or bad actors.
I’ve been doing this for decades. I’m not young in this field, but it’s the first time that I’ve really seen tools that give us the same capability as the bad actors. Perhaps the most urgent lesson is the speed at which the threat landscape has changed. Attackers can now go from identifying a vulnerability to building a working exploit in hours. Frontier AI levels the playing field, giving enterprise defenders that same speed and capability. But it also means that staying current is no longer optional; outdated software is a liability and real-time operational response is essential.
What’s Next
Whether you take the risk of being at the bleeding edge of the latest and greatest or cautiously following, actively participating in the AI era is mandatory. The only way forward is through engagement and continuous adaptation. Remember, there is no safe zone. The safe zone is gone. You have no choice: you need to play the game. You cannot sit this one out.
This blog is the first in a series exploring frontier AI. Based on our learnings working with frontier AI, my future posts will dive into the importance of version currency, how AI agents are transforming security operations, and why speed is the new security challenge, including practical guidance for CIOs navigating the compressed attack chain. See you on the frontlines.
