When your AI agent books a meeting, drafts an email, or digs through a spreadsheet for you, do you have any idea what physical object made that happen? Most people don’t. And that gap between what we use and what we understand is exactly where a lot of investment money is quietly moving right now.
Bernstein analysts are still strongly bullish on Nvidia and Broadcom, expecting both to hold market leadership through 2026 on the back of AI and semiconductor progress. That’s the headline. But I want to talk about why that matters for anyone who’s just trying to figure out how AI agents work.
Agents Are Software Sitting On Very Specific Hardware
An AI agent feels like magic because it feels like nothing. You type, it responds, it takes action. No moving parts, no noise, no sense of effort. That’s the illusion the whole industry has worked hard to build.
Underneath, every one of those interactions runs on chips. Specialized ones. When an agent reads your request, decides what to do, calls a tool, checks the result, and decides again, each of those steps is math executed on silicon in a data center somewhere. Agents are especially demanding because they don’t just answer once. They loop. They think, act, observe, and think again. Every loop costs compute.
So when analysts get excited about chip companies, they’re partly making a bet on agents getting more common. More agents means more loops. More loops means more chips.
The Moat Argument, Translated
Bernstein’s Stacy Rasgon covers this space, and the firm’s broader thesis leans on companies with what one analyst described as “moats that are quantified by their margin structure.” That’s finance-speak worth unpacking.
A moat is a reason competitors can’t easily take your business. The line about margin structure means something specific: if a company charges high prices and still keeps customers, it probably has something customers can’t get elsewhere. Healthy margins are the receipt, not the cause.
For Nvidia, the moat argument usually comes down to software as much as hardware. One market commentator put it bluntly, calling Nvidia’s position close to a monopoly “because they own the software ecosystem,” on top of hardware lock-in. This is the part non-technical readers tend to miss. Nvidia doesn’t just sell chips. It sells the programming tools that developers have spent years learning. Switching hardware means rewriting code and retraining teams. That friction is worth more than any single product spec.
Broadcom’s story is different. Its strength sits in networking and custom silicon. Networking is the plumbing that lets thousands of chips in a data center talk to each other fast enough to act like one machine. Custom silicon means designing chips to spec for a specific customer, usually a large cloud provider that wants something tuned to its own workloads instead of buying off the shelf.
Those two roles are complementary rather than competing. One sells the general-purpose engine most of the industry already builds on. The other builds the connective tissue and the bespoke alternatives.
It’s Not A Two-Company Story
The same Bernstein analysis named four other large-cap semiconductor companies alongside Nvidia and Broadcom as part of an expected trillion-dollar chip surge in 2026. Separate analysis has pointed to AMD’s progress in GPU and CPU architecture as a source of real market share gains. And when Morningstar ranked AI stocks it considered undervalued as of July 24, 2026, the list included Microsoft, Taiwan Semiconductor Manufacturing, Broadcom, Tencent Holdings, and Alibaba Group.
Notice how spread out that list is. Chip designers, a manufacturer, cloud platforms, companies on two continents. The infrastructure behind AI agents isn’t one company’s product. It’s a supply chain with a lot of links, and different analysts disagree about which link captures the most value.
What This Means If You’re Not Buying Stocks
I’m not here to tell you what to invest in. But there’s a practical takeaway for people learning about agents.
- Compute cost shapes agent behavior. When a product limits how many steps an agent can take or how much it can read at once, that’s usually a hardware economics decision, not a design preference.
- Vendor lock-in is real at every layer. The same switching friction that keeps developers on Nvidia’s tools exists in the agent products you use. Worth knowing before you build a workflow around one.
- Follow the infrastructure to spot hype. Claims about agents doing dramatically more get more believable when the compute to support them actually exists.
Analyst optimism about 2026 is a forecast, not a fact, and forecasts miss. What’s more durable is the structural point underneath it: agents are physical. They consume electricity and silicon and space in buildings. Understanding that makes you a sharper reader of every AI announcement that follows.
đź•’ Published: