Over the past two years, the main storyline of US AI stocks has revolved almost entirely around NVIDIA—GPUs were hard to come by, and earnings reports repeatedly fired up the market, lifting expectations across the entire compute supply chain. But while investors focus on GPUs, another quietly growing front is reshaping the industry: cloud giants, unwilling to leave their computing lifeline entirely in NVIDIA's hands, are partnering with chip design service providers to develop custom ASIC chips. This second battleground is becoming a new gold rush within the US AI chip sector.
1. What Is a Custom ASIC? A Plain-Language Explanation
Let's explain in plain terms. A GPU is like an all-purpose computer that can handle any AI task, but its versatility comes with lots of redundant design, making power consumption and cost relatively high. An ASIC is a specialized tool—tailor-made for a specific type of AI task, stripped of everything unnecessary, delivering notably lower energy consumption and cost per unit of computing power.
For cloud giants, the significance of in-house ASICs goes far beyond saving on power bills. First, computing power is their core means of production, and they are never comfortable with that lifeline in someone else's hands. Second, training and inference workloads have distinct characteristics, and no one knows their own business better than they do—custom chips can be optimized on demand. Third, having an alternative gives them more bargaining power when buying GPUs from NVIDIA.
2. The Players at the Table: Four Cloud Giants, Each with Their Own Game Plan
Google was the first mover; its in-house TPU series has iterated through multiple generations, powering search, ads, and Gemini model training and inference—the most thoroughly validated in-house AI chip system to date. Amazon is advancing along two tracks, Trainium and Inferentia, focused on driving down the cost of cloud AI services. Microsoft's Maia and Meta's MTIA started later but are catching up fast; Meta in particular is migrating heavy inference workloads to in-house chips, making its direction very clear.
Notably, the cloud giants' strategy is not to replace NVIDIA but to build a hybrid computing mix of GPUs plus in-house ASICs: cutting-edge model training stays on GPUs, while large-scale, relatively standardized inference tasks gradually shift to custom chips. This preserves technical flexibility while diluting overall computing costs.
3. The Real Picks-and-Shovels Players: Broadcom and Marvell
In-house chip development sounds like the cloud giants' own business, but designing an advanced AI chip is extremely difficult, and most cloud providers lack complete high-end chip design teams. This has spawned a business of building chips for others, with Broadcom and Marvell as its biggest beneficiaries.
- Broadcom: With years of accumulated IP libraries, high-speed SerDes interface technology, and advanced packaging co-design capabilities, Broadcom has become the core partner for custom AI chips at multiple cloud giants and is widely seen as the top player in the custom ASIC race. Add its positions in switch chips, optical interconnects, and other data center networking segments, and Broadcom is harvesting multiple dividends from the spread of AI computing power.
- Marvell: Following a similar path, deeply tied to major clients like Amazon, with positions ranging from custom compute chips to data center interconnect optical communications—seen as the second seed in this race.
Moreover, the volume ramp of custom ASICs also benefits TSMC's advanced process and advanced packaging capacity, as well as HBM memory makers—whether GPU or ASIC, demand for advanced processes and high-speed memory is the same.
4. Will NVIDIA Be Hit?
This is the question investors care about most. Over the short to medium term, the answer may be less bleak than imagined. First, the CUDA ecosystem's moat remains deep—the vast majority of developers' workflows are built around NVIDIA. Second, frontier large model training still chases the strongest single-card performance, a piece of the pie NVIDIA can still hold. Third, total compute demand is growing so fast that even if cloud giants shift some incremental workloads to in-house chips, it will be hard to shake the GPU market's overall expansion in the near term.
But over a longer horizon, the structural impact cannot be ignored. Inference demand is rising rapidly, and inference is precisely where ASICs excel. If inference becomes the dominant share of compute consumption, custom chips will steadily gain share in the overall computing landscape, and the GPU's pricing power and growth could be gradually diluted. For NVIDIA, the real competitive pressure will come more from slowing share growth than from a sudden loss of share.
5. How Can Everyday Investors Participate?
For US stock investors, there are at least three paths to play this theme:
- Bet directly on the picks-and-shovels players: Broadcom and Marvell are the two most direct plays on custom ASICs, but that also means withstanding higher valuation volatility.
- Watch the supporting supply chain: TSMC's advanced packaging, HBM from major memory makers, and data center networking and optical module vendors will all benefit indirectly as ASIC volumes ramp.
- Diversify with ETFs: Investors less comfortable picking stocks can indirectly hold the entire chain through semiconductor ETFs like SMH and SOXX—slightly less upside, but far less hassle.
Two Key Metrics to Track
One is marginal changes in cloud providers' capex guidance; the other is the ramp pace of custom chip projects. These two clues reflect the true quality of this race better than any short-term price swings.
6. Risk Warning
It is worth noting that this race is not a guaranteed win. First, cloud giants' capex is cyclical—if AI application commercialization falls short of expectations, custom chip orders may be the first to be cut. Second, custom ASICs depend heavily on a few large clients, making order concentration risk prominent, and rumors of clients building their own design teams remain a sword of Damocles overhead. Third, advanced process and advanced packaging capacity remains tight, creating uncertainty in delivery timelines.
Overall, competition in AI computing is shifting from single-chip performance to system-level computing costs, and the co-opetition between custom ASICs and GPUs will persist for the long term. For investors, rather than agonizing over who replaces whom, watch the marginal changes in cloud providers' capex and the ramp pace of custom chips—these two yardsticks are the key to judging the value of this second battleground.
Disclaimer: This article is for industry analysis and information sharing only and does not constitute any investment advice. US stocks involve risk; invest cautiously and make independent decisions based on your own risk tolerance.
