Chip stocks used to be the gritty part of the tech complex. In trading patterns and profit margins, they had more in common with cyclical commodities than software. But as with so many things, artificial intelligence changed everything. Almost overnight, chips became the accelerant for the technology—and the market.
Supply constraints compounded the excitement. During one stretch this spring, the SOX, or PHLX Semiconductor Sector Index, rose for 18 straight days, for a gain of 47%. The index is up 80% since March 30, leading to worries about a new dot-com-style bubble with chips looking like the 2026 version of fiberoptic stocks.
Indeed, the gains have been indiscriminate. While Nvidia is up 30%, low-margin chip makers like On Semiconductor and STMicroelectronics have each risen 122%.
These stocks trade at huge multiples of earnings, way above historical trends, while Nvidia looks cheap by any historical measure.
“The multiples cannot all be accurate,” Gavin Baker, chief investment officer of Atreides Management, says of the wide range of price/earnings ratios across the AI landscape. “You have memory makers at low- to mid-single-digit P/Es, you have Nvidia at a low P/E, you have other accelerator companies at reasonable multiples. And then most everything else—power, cooling, optical, and semi-cap equipment—are at dramatically higher multiples.”
The disconnect sets up an opportunity for investors. As the market corrects its math, the quality names should outperform. Investors should focus on Advanced Micro Devices, Broadcom, Taiwan Semiconductor Manufacturing, and, yes, $5 trillion Nvidia.
In the world of semiconductors, quality means having technological advantages that are durable and enable better products and high gross margins. Quality also means having nimble executives who can identify and react to changes in a fast-moving environment. Think Nvidia’s Jensen Huang and Broadcom’s Hock Tan. The quality focus becomes more important as subsidized Chinese manufacturers increase the supply of chips made using older technologies.
All indications point to accelerating demand for AI computing. This year, the five so-called hyperscalers— Microsoft, Amazon.com, Google parent Alphabet, Meta Platforms, and Oracle—could combine to spend over three-quarters of a trillion dollars on AI data centers.
Even after its massive expenditures, Microsoft recently said its cloud-computing capacity was so constrained that it had to forsake external cloud sales to run its own operations.
But the demand trend is shifting, and investors need to pay attention to the nuances.
Early in the AI boom, growth was tied to training new models, a slow, resource-intensive process. Now workloads are moving toward running those models, a process known as inference.
The inference trend is being supercharged by the rise of AI agents, software that can use AI models to complete a complex series of tasks from a simple conversational prompt. Agents chew through computing at a rate no person could match. If predictions are correct, it won’t be too long before they outnumber humans on enterprise networks.
Nvidia already won the battle for training, but inference opens the door to new competition. Moreover, agents are software and run on traditional server central processing units, or CPUs, which should also see increasing demand in the coming years.
Even AI can’t change the fundamentally cyclical nature of semiconductors, but it can—and will—lengthen the cycle. Right now, anyone close to the AI supply chain will tell you that the industry is nowhere close to satisfying demand. That’s why Micron Technology has seen its forward P/E multiple expand from an industrial-like single-digit figure. It’s still undervalued. So are the shares of Nvidia, AMD, Broadcom, and Taiwan Semi.
Relative to expected earnings growth over the next two years, all five stocks trade at a PEG, or price-to-earnings growth ratio, of less than 0.6 times. By comparison, the S&P 500 index fetches a two-year PEG of 1.
All five companies are pillars of the new economy—ones that have lasting value and staying power, even as the momentum inevitably fades from the broader chip trade.
The Leader
Nvidia has been developing hardware and software tools for AI computing for nearly two decades, giving it the pole position when generative AI caught fire. Adjusted earnings per share have grown from 33 cents in fiscal 2023 to $4.77 in fiscal 2026, which ended in January. Over the next two years—a critical period in the AI transition—Wall Street analysts expect Nvidia’s EPS to hit $12.37, giving the stock a P/E of 17 times.
Nvidia’s graphics processing unit (GPU) chips are the workhorses of AI computing, and they’re the company’s main source of sales. Just as important, though, is the company’s two-decade focus on AI bottlenecks.
Nvidia built a unique AI “stack” that layers in software and other types of chips. Its Vera Rubin AI servers, due to ship this year, have a total of five chips, each with a specific purpose.
Nvidia remains undervalued by investors, perhaps because they misunderstand the company’s true data-center dominance. The chip maker is also the largest maker of networking chips for the data center. And it’s now making CPUs too.
Nvidia recently said it would sell $20 billion in stand-alone data-center CPU chips this year, roughly the size of Intel’s own data-center business.
Nvidia glues everything together with software called CUDA, short for “compute unified device architecture.” The company provides its customers with hundreds of free software libraries and AI models that receive regular updates.
To be sure, Nvidia’s top customers would like to diversify their supply chains, and the shift to inference offers an opening. Inference is less demanding than model training, and non-Nvidia chips that thrive at the intense math work are finding a place in the data center. Amazon, Microsoft, Google, and Meta each make their own custom inference chip.
Despite its early lead, Nvidia isn’t sitting still and continues to out-innovate everyone else. It has adopted a ticktock release schedule, with major architecture advances coming every two years and upgrades in the intervening years.
Nvidia has also fought back by using its newly loaded balance-sheet muscle. It has entered into a large licensing deal with an inference-chip start-up called Groq. In March, Huang laid out a vision of data centers where Nvidia’s GPU servers are deployed side by side with new Groq servers, working together on what each does best. Nvidia’s Groq servers are set to ship later this year, after Vera Rubin.
It’s also taking stakes in a broad swath of AI start-ups. At the end of the first quarter, its private-equity investments were valued at $43 billion, up from $3.4 billion at the beginning of 2025. The company also owns 4.5% of Intel, 9% of CoreWeave, and 2.5% of Synopsys.
And there’s something for shareholders too. Nvidia has bought back $85 billion in shares in the past two years, and this past month it increased its quarterly dividend from one penny to an almost respectable 25 cents.
Heaping praise on Nvidia doesn’t mean there isn’t room for other chip makers to be part of the AI investment boom. BofA Securities analyst Vivek Arya estimates that Nvidia has a durable 70% share of the AI chip market.
In a massive market, that leaves big opportunities for other firms, particularly AMD and Broadcom, which are smaller companies with more room to grow. They are being lifted by the same tailwinds as Nvidia, and they are a hedge against Nvidia market share loss.
The Comeback Queen
A perpetual No. 2, AMD has changed the narrative under CEO Lisa Su. In the data center, AMD has caught up to Intel in CPU chips, and it’s making a play to challenge Nvidia in GPUs.
While AMD shares Nvidia’s roots in GPUs for gaming, AMD never made a meaningful effort to compete in the data center until recently. When the AI boom began, it was far behind.
Now AMD has turned its data-center GPUs into a multibillion-dollar business, including recent deals with Meta and OpenAI, which are set to drive future sales.
AMD is even more competitive when it comes to CPUs for servers. As Intel’s technical woes grew over the past decade, AMD leaned into architectural breakthroughs and an asset-light manufacturing strategy. It now has a 40% share in the market, the company says.
AMD recently doubled its estimate for the data-center CPU market to $120 billion by 2030.
“These results mark a clear inflection in our growth trajectory and a structural shift in our business,” Su said during the company’s recent earnings call. “Data center is now the primary driver of our revenue and earnings growth.”
Wall Street analysts expect AMD’s adjusted earnings per share to jump from $4.17 last year to $13.10 in 2027, a 77% annualized increase.
Investors have priced in much of the growth, with shares up 164% since March 30. At 40 times 2027 earnings, the P/E is richer than the other quality chip makers, but it’s by no means pricey given the growth.
The Custom Job
Like Nvidia and AMD, Broadcom is in position to benefit from multiple high-end AI data-center chips. But it has a different path to success, partnering with some of tech’s largest companies.
A decade ago, as Google’s AI ambitions grew, the company ran into a now familiar problem: Nvidia GPUs were expensive and hard to get. Google developed an in-house alternative, the Tensor Processing Unit, or TPU, which could quickly accomplish the complex math of AI.
Broadcom has been Google’s chip-design partner from the start, and the TPU is on its eighth generation. Google uses the chips for its own AI needs and rents them out in the cloud. Start-up Anthropic is a major customer. Google recently hooked up with Blackstone to create a joint venture that will specialize in renting TPUs out in the cloud.
Today, Broadcom is the go-to partner for similar custom chips. It has publicly talked about deals with Meta and OpenAI in addition to Google. In total, Broadcom counts six customers for custom AI chips.
Broadcom has long been strong in networking chips, another big part of its AI opportunity.
Total AI chip sales doubled in the latest quarter. The company is targeting $100 billion in 2027. The board is betting a lot on Tan. The CEO stands to receive 1.8 million shares, worth about $750 million today, if Broadcom’s AI revenue hits $120 billion by the end of fiscal 2030.
Wall Street sees Tan easily hitting that target, with semiconductor revenue projected to be $132 billion in fiscal 2027. Earnings are set to grow an annualized 63% over the next two years, to $18.17 by 2027. The stock fetches a 2027 P/E of 23.
The Chip Factory
All of the chips mentioned above, with one exception, have something in common: They are made by Taiwan Semiconductor Manufacturing.
Taiwan Semi has been the leader in chip manufacturing for years, spurred by its work for Apple, which began using TSMC to make its in-house chips starting in 2014. The so-called fabless chip model is now the industry standard. Chip makers can focus on designing the best chips, while avoiding the costly and messy business of actually producing them.
Taiwan Semi, meanwhile, can specialize in building the best factories with the most advanced manufacturing. The company is wisely conservative with its plans, refusing to overbuild.
The combination means high margins for Taiwan Semi and its customers—everyone wins.
The strategy has paid off in the form of consistent growth, even in a highly cyclical business. Taiwan Semi’s earnings have slipped just twice in the last 14 years. EPS has doubled over the past two years, and analysts expect it to grow another 90% by 2027. The stock trades at 21 times 2027 projected earnings.
Geopolitics is a known risk for Taiwan Semi, given the possibility. The company has added factories in other countries like the U.S. and Japan, but 80% of its long-term assets are still in Taiwan.
The Memory Maker
Memory chips have long been a commoditized part of the semi industry, directly exposed to consumer demand, which leads to sharp fluctuations in inventories and prices. The down-cycle of 2023 was particularly brutal, leaving Micron with a negative gross margin for four consecutive quarters.
But the latest memory cycle is being driven by an investment boom, not consumer demand. Because of the beating they took in 2023, Micron and rivals—Korea-based SK Hynix and Samsung Electronics—held back on committing capital to another expansion. Now, supply is the tightest it’s ever been and memory prices have skyrocketed. New memory factories aren’t slated to come online until the middle of next year, and the logjam may not clear until 2028.
Meanwhile, Micron and its peers are in uncharted territory. The company saw 196% year-over-year sales growth with a 74% gross margin in the most recent quarter. It guided to even better performance in the current quarter. Analysts think that adjusted EPS will grow from $8.29 to over $100 in fiscal 2027, which ends in August. Because of its highly cyclical earnings, Micron has always had a low forward P/E. At a recent $924, the stock is trading at just nine times 2027 EPS.
However long this cycle lasts for Micron, management is being opportunistic and using the cash flow to clean up the company’s balance sheet. Over the last three quarters, debt has been reduced by 33%. The company is also buying back shares for the first time since 2022.
This Too Will End
Nothing lasts forever and that’s especially true in the chip business. Data centers are still being held back by energy, land, and building constraints. And opposition to massive complexes is growing, part of an overall backlash against AI in the U.S. and elsewhere. In a March Gallup poll, 71% of U.S. respondents opposed the local construction of data centers.
War in the Middle East is crimping the supply of energy and much-needed helium. The conflict is also driving up long-term interest rates, increasing the cost of capital for data-center investments.
These headwinds could cut the chip rally short—making it all the more important for investors to be discriminating in their stock-picking. Quality always shines. Years into the AI rally, it’s more important than ever.
Comments