Why Nvidia And Broadcom Still Look Cheap In AI

Every few months the same argument comes back: maybe the AI buildout is about to cool, maybe the most expensive chips have already had their moment, maybe investors should wait for a cleaner entry. I keep hearing that line in conversations with people who otherwise love technology stocks. Then I look at the tape, look at how cheap some of these names have become versus their own history, and I am not sure the pause story is as tidy as it sounds.

The Case That Nvidia And Broadcom Still Have Room

A well-followed semiconductor research team recently told clients they would still own both Nvidia and Broadcom. That is not a timid stance. The group put outperform-style ratings on both names and sketched price targets that implied substantial upside from recent closes. Broadcom was framed as having room toward the mid-$500s. Nvidia was framed as having room toward $400. Those numbers can move around, of course. Targets always do. The more interesting part is the logic underneath them.

The analysts argued that both companies still have line of sight for strong growth even if the environment stays constrained. They also called the stocks extremely cheap. That phrase almost feels sloppy until you sit with the multiples. Nvidia has been trading around the high teens on the next twelve months of estimated earnings. Its five-year average sat closer to the mid-30s. Broadcom has been nearer 20 times forward earnings versus a mid-20s five-year average. The broader market multiple has been in a similar neighborhood. For companies still talking about very large growth rates, that compression is unusual.

Companies with this kind of growth used to almost never trade at a plain market multiple for long.

I have found that cheapness by itself is not a strategy. Cheap can stay cheap. Still, a lower multiple can act like a cushion if sentiment wobbles. That is the quieter part of the bull case. The louder part is demand that does not look finished.

Why The Valuation Reset Matters More Than The Headlines

Not long ago it felt strange to see these growth engines priced like ordinary large-cap industrials. Investors paid up because the story was scarce. Now the scarcity is less about the ticker and more about the physical world behind the ticker: power, packaging, memory, networking, foundry capacity, and time. When those bottlenecks stay tight, revenue visibility can stretch further than a single product cycle.

Nvidia has pointed to growth of 70 percent or more off an already huge base into 2027. Broadcom has talked about the potential to double revenues in 2027 and again in 2028, or even more, if custom AI work keeps landing. Those are management-shaped comments, not promises from the sky. They still change the debate. If even a trimmed version of that path arrives, a high-teens or low-20s earnings multiple starts to look less like a warning sign and more like a discount to the growth rate.

Perhaps the most interesting aspect is how quickly the market can forget its own history with these names. A year of choppy headlines is enough for people to treat a former premium as a permanent haircut. Sometimes that haircut is earned. Sometimes it is just fatigue.

Two Different Machines Inside The Same Boom

It helps to stop treating every AI chip company as a clone. Nvidia designs across a wide stack now, including central processing units, but its graphics processors remain the reference standard for training large models and running heavy inference. Broadcom is a different animal. It builds custom silicon for specific workloads and counts major platform companies among its customers. Both sit in the accelerator bucket. They do not sit in the same business model.

That distinction matters when investors rotate. A week of excitement around agents or a new consumer-facing tool can send money into CPU-linked names. A week of training-cluster news can send it back toward GPUs. The underlying buildout is larger than any one rotation. Still, the rotations create noise, and noise creates opportunities for people who can keep the two models separate in their heads.

• Nvidia sells a broad, software-wrapped accelerator platform that many labs already know how to use.
• Broadcom sells tailored chips that can be extremely efficient for a defined job at a defined customer.
• Both depend on a multiyear construction cycle rather than a single product launch.
• Both can look expensive on trailing numbers and cheaper on forward numbers if growth holds.

In my experience, investors get into trouble when they flatten those differences into a single “AI basket” trade. Baskets are convenient. They are also lazy. A custom-chip winner can keep growing even if the merchant GPU narrative cools for a quarter. A platform GPU winner can keep winning even if one hyperscaler leans harder into internally designed parts. The mix will shift. That does not automatically end the cycle.

Agentic Systems And The Narrowing GPU-To-CPU Story

The ratio between graphics processors and central processors has been changing as systems get better at handling tasks with less constant human steering. People now talk about agentic setups: software that can plan, call tools, check its own work, and keep going. Those systems still need accelerators. They also need a lot of ordinary compute sitting close by. For some slices of the job, a CPU is simply more efficient.

That is why a burst of product news around agents can light a fire under processor stocks. It happened earlier in the year. It happened again when a large social platform leaned into agent-building tools. The trading pattern is familiar. Money chases the newly obvious bottleneck. Then it overshoots. Then it looks for the next bottleneck.

Does that make Nvidia less relevant? I do not think so. Coordination between CPUs and GPUs is the point, not a funeral for one side. Training still loves dense accelerator clusters. Many inference jobs still love them too. What changes is the bill of materials around the cluster: more host processors, more memory, more networking, more power gear, more packaging. The pie gets wider. Some slices grow faster than others.

Agentic computing does not retire the accelerator. It changes the mix of parts that have to show up on time.

What Tight Supply Still Tells You

The same research team pointed to multiyear visibility and tight supply across almost every component needed for the buildout. That sentence is easy to skim. It should not be. If capacity is tight in wafers, advanced packaging, high-bandwidth memory, and networking silicon at the same time, the constraint is systemic. Systemic constraints tend to keep pricing firmer than a software-only boom would allow.

I keep coming back to a simple test. If customers were truly ready to slow down, you would expect more public talk about delayed racks, delayed power hookups being used as an excuse to cancel, and delayed custom programs. You still hear friction. Power is hard. Permitting is hard. Talent is hard. Cancellation talk at the top of the demand stack has been less convincing than the commentary from people who actually sell the parts.

Last week’s round of meetings with chip suppliers, including the two names at the center of this piece and a large processor manufacturer, produced a fairly consistent message. The data-center opportunity at Nvidia still looks enormous. Broadcom’s AI path looks set to accelerate over the next two years. The processor maker is getting help from server demand, and the narrative around that recovery can feed on itself for a while. None of that reads like a market that has already peaked.

Safety Debates, Slowdown Talk, And Why Demand Has Not Bent

There has been a louder argument that frontier labs should collectively pace model releases and that governments should force coordination. Safety is not a cartoon issue. Anyone who has watched these systems fail in small ways knows that. The investment question is narrower. Does a call for pacing change purchase orders for accelerators and custom chips in the next two years?

Executives at both companies have been pretty direct. They do not see demand rolling over because of that debate. One said forecasts were not being rewritten. Another said labs can take the time they need to ship safer products without a coordinated freeze or a heavy-handed intervention. That is self-interested speech. It is also consistent with order patterns that have not suddenly gone quiet.

I’ve found that markets love a moral plot twist. If the story becomes “the machines must slow down,” some investors will sell first and think later. That can create a better entry in names that still have physical backlog. It can also trap people who confuse an essay with a cancellation notice. Those are not the same document.

How Portfolio Managers Are Splitting The Trade

Not every bull lives in the same house. Some investors still rate Nvidia as a core buy while keeping a more cautious stance on Broadcom after trimming exposure. Others want more processor and memory torque because those parts of the complex were left behind for longer. A few want all three: accelerators, custom silicon, and the duller components that make racks actually run.

One high-profile portfolio has kept Nvidia as a top-rated holding with a target well above the recent price, while treating Broadcom as more of a hold with a target that still implied upside. That same group cut a Broadcom position earlier to reduce theme concentration after political pressure around the data-center build increased. Booking profit and lowering single-theme risk is not the same as declaring the cycle dead. It is risk management with a point of view.

There is room in those targets for more upside if the growth path holds. There is also room for disappointment if one hyperscaler slips a deployment year. Anyone who tells you the range is tiny is selling comfort, not analysis.

Intel, Memory, And The Rest Of The Bill Of Materials

The processor recovery story has become fashionable again. Server strength is helping a company that spent years as a punchline crawl back toward relevance. Manufacturing and packaging work for other chip designers sweetens that pitch. A previously skeptical analyst turning even a little more constructive is notable precisely because the skepticism had been durable.

Memory sits in the same broader argument. High-bandwidth stacks are not a side quest. They are one of the binding constraints. When memory is tight, accelerator shipments cannot magically float above the shortage. That is why some investors prefer a barbell: own the obvious accelerator leaders and own a piece of the scarce supporting cast. It is less elegant than a single-stock crusade. It is often more honest.

Does that mean Nvidia should be sold so the proceeds can live in processors and memory? Only if you think the platform is done. I do not. I do think position sizing matters more than slogan loyalty. A 1-rated view on Nvidia can coexist with a smaller Broadcom line and a separate sleeve for the rest of the rack. Markets reward that kind of unglamorous mix more often than they reward all-or-nothing speeches.

The Political Overlay Investors Keep Underestimating

Data centers are no longer just an earnings story. They are a local politics story, a power-grid story, and at times a national-strategy story. That overlay can knock a stock around even when the purchase orders look fine. Community pushback, energy pricing, export rules, and industrial policy can all show up in the multiple before they show up in the income statement.

That is one reason a partial trim in a custom-chip name can be rational even if the fundamental path still looks strong. Reducing theme exposure is not a confession that the thesis broke. It is an admission that one theme can dominate a book until a headline does the dominating for you. I would rather choose the size than let the news cycle choose it.

Still, politics can cut both ways. Support for domestic capacity, faster grid connections, or friendlier permitting can extend the cycle. Investors who only model chips and ignore substations are modeling half the machine.

How To Think About “Cheap” Without Getting Cute

A 17 times forward multiple on a company guiding to very large growth is not automatically a gift. Estimates can be wrong. Mix can shift toward lower-margin parts. Customers can design more of the stack themselves. Competition can get good enough to matter. All of that is real.

What cheap does give you is time. If the growth lands closer to the bull case, the multiple can expand and earnings can compound at the same time. If the growth lands closer to a muddle, you are not paying 35 times for the privilege of being early and wrong. That is the cushion. It is not armor.

1. Start with the physical constraint, not the latest product demo.
2. Separate merchant accelerators from custom programs before you compare multiples.
3. Ask whether the multiple already assumes a slowdown that suppliers are not seeing.
4. Size the position for political and power-grid noise, not just earnings season.
5. Leave room for the supporting cast: processors, memory, networking, packaging.

That list is not clever. It is usable. Clever lists tend to age badly. Usable lists survive a messy quarter.

What Would Actually Break The Bull Case

I try to write down the kill shots before I get attached. For Nvidia, a hard share loss in both training and high-end inference plus a stall in software attach would be serious. For Broadcom, a slip in a major custom program or a long freeze in networking upgrades would hurt more than a week of agent headlines. For the complex as a whole, a true collapse in capital spending by the largest buyers would swamp valuation debates.

A softer version of those risks is already in the price. That is why the multiples compressed. The market is not naive. It is tired, a little suspicious, and still unwilling to abandon the idea that computing demand can stay elevated for years. Tired and suspicious can be a better setup than euphoric and certain.

Would I bet the house? No. Houses are for living in. These are stocks. They move. They overshoot. They punish people who need to be right this week. The better question is whether the next two to three years of construction still need these parts. The supplier conversations keep answering yes.

A Practical Way To Hold The Idea

If you already own Nvidia, the valuation reset is an argument for patience more than for a victory lap. If you do not own it and you believe the data-center runway is intact, waiting for a perfect dip can become a habit that never gets funded. Broadcom is trickier because custom programs arrive in lumps. Lumps make beautiful charts after the fact and stressful charts in real time.

Some readers will want a clean winner. Markets rarely offer one. A barbell between the platform accelerator and a smaller custom-silicon line, with a satellite in processors or memory, matches the way the rack is actually being built. It also keeps you from turning every product launch into a referendum on your entire book.

Working split I keep coming back to:
  Core accelerator platform
  Custom silicon and networking
  Scarce supporting components

You can argue about the percentages. You should argue about the percentages. What you should not do is pretend the only intellectually pure stance is 100 percent in one ticker because a target price looks exciting.

The Uncomfortable Middle Ground

So where does that leave a reader who is neither a true believer nor a professional skeptic? In the middle, which is usually the honest place. Demand commentary from suppliers remains firm. Multiples are no longer heroic. Agentic systems are changing the mix, not deleting the need for accelerators. Political noise can slap the stocks around. Custom programs can bunch. Estimates can slip. All of those statements can be true at once.

I keep a simple bias. If the physical buildout is still capacity constrained, I would rather own the companies that sit at the scarce points than wait for a narrative that feels emotionally complete. Complete narratives show up late. Constrained supply shows up in lead times, packaging queues, and cautious comments about doubling revenue again.

That is not a dare to mortgage anything. It is a reminder that the market has already done a lot of the worrying for you. When a growth compounder trades near a market multiple, the next debate should be about duration, not about whether the last twelve months felt noisy. They did. They usually do.

If the next leg is slower than the loudest forecasts, the lower multiple gives you room to stay in the seat. If the next leg is closer to what suppliers keep describing, the seat was the whole point. Either way, the useful work is the same: watch the constraints, separate the business models, and refuse to let a single week of agent headlines write the entire cycle for you.