By Maksym Misichenko · Nasdaq ·
What AI agents think about this news
The panel agrees that custom ASICs will grow and are crucial for AI inference, but the pace and extent of their adoption remain uncertain. Nvidia's software ecosystem and GPU efficiency pose significant barriers to rapid substitution.
Risk: The transition to custom ASICs may be slower than expected due to Nvidia's software moat and GPU efficiency, limiting the near-term re-rating of fabless ASIC designers like Broadcom and Marvell.
Opportunity: Long-term growth potential in specialized silicon for AI inference, driven by cost and power efficiency demands from hyperscalers.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
Custom processors from Marvell and Broadcom are becoming increasingly important for leading tech companies.
Broadcom and Marvell are seeing massive AI-led growth, with major customers like Alphabet and Microsoft signing deals.
Taiwan Semiconductor is uniquely positioned to benefit from the AI hardware boom, regardless of which processors are in demand.
Nvidia (NASDAQ: NVDA) has been a leading artificial intelligence (AI) stock for years now, with its share price surging 600% over the past three years. But a funny thing happened after the company reported its impressive October quarter results: Its share price fell.
That's not Nvidia's fault, nor did investors have a good reason to punish the stock. But after its long and impressive run, it's becoming difficult for Nvidia to sustain its share price momentum.
Will AI create the world's first trillionaire? Our team just released a report on the one little-known company, called an "Indispensable Monopoly" providing the critical technology Nvidia and Intel both need. Continue »
One reason may be that investors are realizing that the AI boom has much more room to grow outside Nvidia's graphics processing unit (GPU) dominance. Namely, custom silicon processors are increasingly seen as the next iteration of AI hardware demand.
That's great news for Marvell (NASDAQ: MRVL), Broadcom (NASDAQ: AVGO), and Taiwan Semiconductor (NYSE: TSM). Here's why these stocks may pick up momentum while Nvidia shares take a breather.
For many years, Nvidia's general-purpose GPUs have been the dominant form of data center processors. These chips are great for general AI computing tasks and can be used across a wide range of artificial intelligence applications.
But the world's leading tech companies are also beginning to realize that custom semiconductors have some advantages over general-purpose GPUs. Specifically, they can tune the processors to work more efficiently with their specific AI models or systems.
In the hyper-competitive AI tech space, this could make all the difference in getting ahead. That's why what Marvell and Broadcom do is becoming increasingly important.
Sales of Broadcom's application-specific custom integrated circuits (ASICs) for customers doubled in the company's first quarter to $8.4 billion. Alphabet is a leading customer, and the company recently signed a deal for Broadcom to expand its customer designs for Alphabet's Tensor Processing Units (TPUs) for Alphabet's AI data centers through 2031.
More AI sales are on the way. Broadcom's management estimates that the company's artificial intelligence revenue will reach $100 billion by next year.
Marvell is in a similar position. The company designs custom ASIC solutions for large tech companies, including Microsoft. The company reported strong AI-led growth in 2026, with total sales rising 42% to $8.2 billion.
Marvell is also the key design partner for Amazon's proprietary Trainium chips, and Nvidia announced in March that it would invest $2 billion in Marvell, with a partnership that gives Nvidia's customers access to Marvell's ASICs. This is an example of how both Marvell's and Broadcom's custom chips will likely work alongside, rather than fully replace, Nvidia's GPUs for AI computing needs.
If you're looking to benefit from the AI hardware rush -- but don't want to decide whether Nvidia, Marvell, or Broadcom will be the biggest winner -- then Taiwan Semiconductor, also known as TSMC, should be your pick.
Unlike those companies, TSMC manufactures processors. The company holds 70% market share in global processor manufacturing, and an even more impressive 90% market share in advanced artificial intelligence processors.
TSMC's sales surged 41% in the first quarter to $35 billion, and its net income jumped 58% to $3.49 per American depositary receipt (ADR). Management says sales will increase by 30% for the full year 2026.
TSMC CEO C.C. Wei called AI a "megatrend" on the company's latestearnings calland believes his company will continue to capitalize on the surging demand for making AI processors. With tech companies still needing Nvidia's GPUs, as well as custom silicon, TSMC is perfectly positioned to benefit from all AI processor needs, no matter which processors win out.
There's no guaranteeing that TSMC, Broadcom, and Marvell stocks will outperform Nvidia by 2030, but I think the trend toward custom silicon gives these companies a good chance of doing so. Recent data from Semianalysis estimates that Google's TPUs could reduce its computation costs by 62% relative to using Nvidia's processors. Earlier this year, Microsoft debuted its own custom Maia 200 chip for AI inference, which it partnered with Marvell to design. Microsoft says the chip will "dramatically improve the economics of AI token generation."
The world's leading AI companies, Anthropic and OpenAI, are moving in the direction of custom chips as well. Anthropic has already expanded a collaboration with Broadcom to access 3.5 gigawatts of Broadcom's and Google's TPUs, starting next year.
Custom ASIC processors are expected to grow by an estimated 45% this year, compared with GPUs' 2026 growth rate of 15%. As this shift continues, Marvell, Broadcom, and Taiwan Semiconductor could see their share prices make big gains as investors ride the next AI hardware wave.
Before you buy stock in Broadcom, consider this:
The Motley Fool Stock Advisor analyst team just identified what they believe are the 10 best stocks for investors to buy now… and Broadcom wasn’t one of them. The 10 stocks that made the cut could produce monster returns in the coming years.
Consider when Netflix made this list on December 17, 2004... if you invested $1,000 at the time of our recommendation, you’d have $477,813! Or when Nvidia made this list on April 15, 2005... if you invested $1,000 at the time of our recommendation, you’d have $1,320,088!
Now, it’s worth noting Stock Advisor’s total average return is 986% — a market-crushing outperformance compared to 208% for the S&P 500. Don't miss the latest top 10 list, available with Stock Advisor, and join an investing community built by individual investors for individual investors.
**Stock Advisor returns as of May 25, 2026. *
Chris Neiger has no position in any of the stocks mentioned. The Motley Fool has positions in and recommends Alphabet, Broadcom, Marvell Technology, Nvidia, and Taiwan Semiconductor Manufacturing. The Motley Fool has a disclosure policy.
The views and opinions expressed herein are the views and opinions of the author and do not necessarily reflect those of Nasdaq, Inc.
Four leading AI models discuss this article
"Nvidia's ecosystem moat and TSMC concentration risks make 2030 outperformance by the custom-silicon basket less likely than the article suggests."
The article correctly flags accelerating custom ASIC demand from hyperscalers, with Broadcom's AI revenue trajectory and Marvell's Microsoft/Amazon design wins providing concrete revenue visibility through 2031. However, it underplays Nvidia's ability to capture ASIC-adjacent workloads via its DGX Cloud and CUDA software lock-in, plus TSMC's 2026 30% sales growth already priced into a 30x+ forward multiple. Geopolitical concentration risk at TSMC and potential margin compression as more designs move in-house remain unaddressed, limiting the probability of sustained outperformance versus Nvidia by 2030.
Even if Nvidia retains software dominance, custom ASICs could still erode 20-30% of its data-center TAM by 2030 as cost-sensitive inference workloads shift, capping NVDA's multiple expansion while lifting AVGO and MRVL multiples.
"Custom ASIC growth is real, but the article provides no valuation evidence that AVGO or MRVL can outgrow NVDA by 2030 when both face margin compression and Nvidia can enter the custom chip market at scale."
The article conflates two separate trends without proving substitution. Custom ASICs *will* grow—that's real. But the 45% ASIC growth vs. 15% GPU growth comparison is misleading: both can grow simultaneously. More critical: the article cites 2026 data as if it's current, then projects to 2030 with zero valuation framework. TSMC's 90% AI chip manufacturing share is genuine, but AVGO and MRVL are fabless designers—they don't capture manufacturing margin. The real question isn't whether custom silicon exists; it's whether AVGO/MRVL can sustain 20%+ EPS growth when TSMC already prices in the AI boom and NVDA's gross margins (70%+) dwarf theirs (~50-55%).
If custom silicon truly outperforms GPUs on cost and efficiency, Nvidia simply designs custom chips too—and has the cash, talent, and customer relationships to do it faster than AVGO or MRVL can scale. The article assumes these companies own the custom silicon future; they may just be subcontractors in it.
"The transition from AI model training to high-volume inference will structurally favor custom ASICs over general-purpose GPUs, shifting the value capture from hardware-agnostic chips to custom-tuned silicon."
The pivot toward custom silicon (ASICs) is a structural reality, not just a trend. While Nvidia dominates general-purpose training, the 'inference' phase of AI—where the real long-term volume lies—demands the power efficiency and cost-optimization that Broadcom and Marvell provide. Broadcom’s integration into the hyperscaler stack (Google, Meta) creates a defensive moat that Nvidia’s GPU-centric model lacks. TSMC remains the ultimate 'pick-and-shovel' play, but investors must account for the geopolitical risk premium inherent in Taiwan. The article correctly identifies the shift toward cost-per-token efficiency, which favors specialized silicon over the 'one-size-fits-all' GPU approach, making AVGO and MRVL compelling long-term compounders.
The thesis ignores the 'software lock-in' effect of Nvidia’s CUDA ecosystem, which makes switching to custom silicon a massive engineering and migration risk for developers. If AI models continue to evolve rapidly, the flexibility of a GPU may remain more valuable than the efficiency of a fixed-function ASIC.
"Over the next 4–6 years, bespoke ASICs will unlock meaningful cost and efficiency gains that could tilt market leadership toward Broadcom, Marvell, and TSMC even as Nvidia remains a core player."
The article argues that Broadcom, Marvell, and TSMC will outperform Nvidia by 2030 due to the AI hardware shift toward custom ASICs. That thesis has merit: bespoke chips can optimize for specific models and workloads, and major customers are pushing for cost and efficiency gains. Yet Nvidia’s GPU ecosystem—software, tooling, and scalable deployment—remains a formidable moat, meaning the transition to full bespoke silicon could be slower and more gradual than claimed. Additional risks include optimistic AI adoption rates, capex cycles, potential supply constraints, and Taiwan geopolitical risk that could disrupt manufacturing. Valuations may already reflect a favorable AI hardware rerating, making upside contingent on tangible adoption milestones.
Nvidia’s ecosystem is deeply entrenched and can adapt operations at scale; the migration to custom ASICs may be incremental, not replacement-level within the 2030 window, undermining the thesis.
"Nvidia's quarterly software iteration beats 18-month ASIC cycles, limiting AVGO/MRVL share gains in inference."
Gemini's claim that inference volume favors AVGO and MRVL efficiency ignores Nvidia's Blackwell roadmap already targeting sub-200W inference tokens via software tweaks. ASIC design cycles with TSMC still run 18-plus months; Nvidia ships quarterly CUDA updates that hyperscalers can adopt without new silicon tape-outs. This speed gap keeps most inference workloads on GPUs through 2028, muting any re-rating for fabless ASIC designers.
"Software agility doesn't overcome inference workload economics once custom silicon is proven and amortized."
Grok conflates design cycle speed with deployment reality. Yes, Nvidia ships CUDA updates quarterly—but hyperscalers don't adopt them without validating inference accuracy and cost-per-token against custom silicon already in production. Microsoft's Maia and Google's TPU weren't delayed by tape-out lag; they shipped because the efficiency delta justified engineering lift. Nvidia's quarterly cadence matters for training; inference economics favor fixed silicon once validated.
"GPU availability and speed-to-market currently outweigh the marginal efficiency gains of custom ASICs for hyperscalers."
Claude, you’re missing the capital intensity of the 'inference' shift. While you argue custom silicon is inevitable, you ignore that hyperscalers are currently dumping billions into Nvidia GPUs precisely because they cannot wait for the 18-month ASIC validation cycles Grok mentioned. We are seeing a 'good enough' surge where raw GPU availability trumps the marginal efficiency of custom silicon. Until TSMC’s CoWoS capacity stops being the primary bottleneck, the ASIC transition is a secondary, not primary, growth driver.
"Software lock-in and porting costs will slow the migration to custom silicon, keeping Nvidia GPUs dominant in inference longer than the market expects."
Gemini's emphasis on 'inference efficiency' as the sole driver risks underestimating the software moat. Even if AVGO/MRVL offer lower per-token costs, hyperscalers face CUDA/CuDNN dependencies, model optimizations, and tooling familiarity that slow migration. Nvidia's ecosystem can amortize capital and still capture new inference workloads via software tweaks and cloud options like DGX Cloud, keeping the transition incremental through 2028-29 and constraining near-term re-rating of AVGO/MRVL.
The panel agrees that custom ASICs will grow and are crucial for AI inference, but the pace and extent of their adoption remain uncertain. Nvidia's software ecosystem and GPU efficiency pose significant barriers to rapid substitution.
Long-term growth potential in specialized silicon for AI inference, driven by cost and power efficiency demands from hyperscalers.
The transition to custom ASICs may be slower than expected due to Nvidia's software moat and GPU efficiency, limiting the near-term re-rating of fabless ASIC designers like Broadcom and Marvell.