For three years, the artificial intelligence (AI) trade was simple: Buy Nvidia, collect the upside, repeat. The graphics processing unit (GPU) was the atom of modern AI; everything else seemed to be organized around it.
That story is not over. But something structural is shifting beneath it, creating one of the more compelling investment setups of the next decade.
The largest technology companies on Earth are no longer content to simply buy Nvidia's chips off the shelf. They are designing their own, purpose-built silicon tuned to their specific models, workloads, and cost structures.
Custom application-specific integrated circuits (ASICs) are projected to account for 27.8% of the AI server compute market in 2026, growing 44.6% year over year, nearly triple the growth rate for merchant GPUs. That gap is going to keep widening.
Two companies sit at the center of this shift, and together they control roughly 95% of the custom AI ASIC co-design market.
Image source: Getty Images.
Broadcom: The architect hyperscalers can't quit
NASDAQ: AVGO
Key Data Points
Broadcom (AVGO 0.47%) has been designing Alphabet's Tensor Processing Units (TPUs) since TPUv2. Most people who follow AI loosely know this. What they don't fully appreciate is how far Broadcom has extended that relationship and how many other customers are now lining up behind Alphabet.
In April 2026, Broadcom filed an 8-K disclosing a long-term agreement with Alphabet to develop and supply future generations of TPUs, along with a supply assurance commitment for networking and other components used in Alphabet's next-generation AI racks through 2031. Separately, Alphabet, Broadcom, and Anthropic expanded their collaboration so that Anthropic can access approximately 3.5 gigawatts of next-generation TPU-based compute capacity through Broadcom starting in 2027. This is no longer a one-customer business.
Meta Platforms announced a partnership with Broadcom to develop multiple generations of custom silicon in mid-April 2026. OpenAI has committed to a long-term custom accelerator program. And Anthropic has committed to over $10 billion in custom silicon.
The other piece most investors miss is Broadcom's networking business. Its Tomahawk 6 switch ASIC -- the world's first 102.4 terabit-per-second switch, shipping in volume as of early 2026 -- enables AI clusters to scale past 1 million XPUs with minimal latency.
Despite all this strong news, the company's shares plunged about 15% on June 4 after the company reported slightly weaker-than-expected quarterly revenue and maintained, rather than raised, its 2026 AI chip sales forecast of more than $100 billion, disappointing investors despite strong AI-driven growth and upbeat guidance for the current quarter.
Marvell Technology: The "next trillion-dollar company"
NASDAQ: MRVL
Key Data Points
Marvell Technology (MRVL +0.85%) is where the valuation gap still exists. The company is earlier in its ASIC ramp than Broadcom, which means lower margins today -- but also more room to run.
Marvell is the custom silicon architect behind Amazon's Trainium chips and Microsoft's Maia accelerators. In April 2026, reports emerged that Alphabet is in talks with Marvell on two new custom AI chips. When Alphabet starts hedging its ASIC supplier, and the hedge is Marvell, that tells you something about where Marvell stands in the competitive hierarchy.
Marvell projects up to $11 billion in AI ASIC revenue for 2026 -- and its pipeline now carries more than 50 new custom chip opportunities, representing an estimated $75 billion in lifetime revenue.
Nvidia reinforced the thesis by investing $2 billion in Marvell in March 2026 and pulling its custom ASIC roadmap inside the NVLink Fusion ecosystem. Even in a world where hyperscalers want off-the-shelf Nvidia GPUs, Marvell's silicon is inside the rack. And even when they want custom chips to replace those GPUs, Marvell is designing them.
On top of that, at Computex 2026, Nvidia CEO Jensen Huang called Marvell Technology "the next trillion-dollar company," arguing that Marvell's networking, optical interconnect, and AI infrastructure technologies will be essential as AI data centers continue to scale. Following his remarks, Marvell's shares surged about 33% in a single session -- the largest one-day gain in the company's history -- adding roughly $56 billion in market value and pushing the stock above $250 billion in market capitalization. Marvell's operating margins today sit in the high 20s to low 30s, with management targeting the high 30s as the AI revenue mix grows.
To me, the GPU era is not ending -- but it is maturing. The companies that defined the first phase are still indispensable. The companies defining the next phase are the ones building the custom silicon that lets hyperscalers break free from paying the "Nvidia tax" at massive scale. Broadcom and Marvell are those companies, and the market is still fairly early in pricing that reality.
