Over the past two years, whenever AI compute came up, investors' first reaction was almost always NVIDIA and its GPUs. However, as we enter 2026, a new theme is quietly gaining strength: in-house ASIC chips—represented by Google's TPU, Amazon's Trainium, and Microsoft's Maia—are stepping out of the shadows and into the spotlight. The biggest beneficiaries behind this shift are chip design service providers like Broadcom and Marvell, who stand behind the cloud giants. Signs are emerging that the compute landscape, long dominated by a single GPU player, is starting to loosen.
What Is an ASIC? Why Are Cloud Giants Building Their Own Chips
ASIC (Application-Specific Integrated Circuit), as the name suggests, is a chip tailor-made for a specific task. Unlike general-purpose chips such as GPUs that "can compute anything," an ASIC focuses on just one thing—for example, running large-scale inference for Transformer-based large models. Precisely because of this singular focus, it can achieve higher energy efficiency and lower per-unit compute cost under specific workloads.
For cloud giants like Google, Amazon, Microsoft, and Meta—whose annual capital expenditures run into the hundreds of billions of dollars—the motivations for building their own ASICs are very practical:
- Cost reduction: At large deployment scale, the inference cost of custom chips is expected to be significantly lower than that of purchased GPUs;
- Energy efficiency: Data center power is increasingly strained, and every extra bit of computing per watt translates into real money;
- Supply chain independence: Reduce reliance on a single supplier and gain control over pricing power and supply schedules;
- Hardware-software synergy: In-house chips can be deeply integrated with their own software stacks, unlocking overall system efficiency.
Broadcom and Marvell: The Biggest Winners in the ASIC Wave
The cloud giants have the vision, but designing an advanced chip from scratch is no easy feat—which has given rise to a group of "behind-the-scenes strategists." Broadcom is the most prominent example: it is the core design partner behind Google's TPU, and it also collaborates on custom chips with several major clients including Meta, while the market continues to closely follow the progress of its partnerships with star companies such as OpenAI. Over recent fiscal years, Broadcom's revenue from AI custom chips has grown rapidly, becoming the company's most important growth engine.
Marvell, meanwhile, is deeply tied to projects such as Amazon's Trainium training chips and Microsoft's Maia, with its AI business share rising rapidly. Beyond these, Asian design service companies such as Alchip and GUC are also taking on a growing number of small and mid-sized custom chip orders, and the entire design services supply chain is seeing growth bloom across multiple fronts.
Notably, the business model for ASIC orders differs from selling standard chips: design service fees are collected upfront, and revenue continues to be recognized as chips ship after mass production begins—offering higher visibility and greater certainty. Once a client's project enters the mass production cycle, it often means sustained volume growth for years to come. This is the core logic behind why capital markets are willing to assign higher valuations to the companies involved.
ASIC vs. GPU: A Three-Front Battle over Cost, Efficiency, and Ecosystem
The Cost Battle
On a per-chip basis, custom ASICs aim to be cheaper than same-class GPUs while omitting unnecessary general-purpose features. However, ASIC development demands massive R&D investment, and taping out on advanced process nodes is extremely costly—only customers deploying at sufficiently large scale can amortize the upfront spending. That is why the in-house chip camp today consists almost entirely of hyperscale cloud providers: they have the volume to make the numbers work, while small and mid-sized players find it hard to follow suit.
The Energy-Efficiency Contest
In scenarios with fixed workloads such as inference, ASICs offer a clear energy-efficiency advantage; but in training, where algorithms iterate rapidly and model architectures change constantly, GPU flexibility remains hard to replace. A basic industry consensus holds that training will continue to be dominated by GPUs while inference gradually migrates to ASICs—the two routes may coexist in parallel for a considerable time rather than fight to the death.
The Ecosystem Contest
NVIDIA's true moat is not the hardware itself but the software barrier built up by the CUDA ecosystem over more than a decade. Each ASIC vendor, meanwhile, builds its own software stack, which raises the development barrier. That said, as mainstream deep learning frameworks steadily improve their support for hardware like TPUs, and given that inference workloads depend far less on ecosystems, this wall is slowly being chiseled away. For NVIDIA, the real test may not be a single-point breakthrough from any one rival, but the long-term trend of the inference market as a whole moving away from GPUs.
Where FPGA Fits In: The Flexible "Vanguard"
Within the programmable logic camp alongside ASICs and GPUs, the FPGA plays a different role. Xilinx, now under AMD, and Altera, spun off from Intel's structure, are the main players in this field. FPGAs can be programmed and their logic rewritten repeatedly, making them well suited to rapid algorithm validation, communication baseband, edge inference, and small-batch customization. The FPGA is more like a "testing ground" before ASIC mass production and a "standing army" for specific niche markets—complementary to ASICs rather than a replacement. Using FPGAs as a bridge while a customer's algorithms remain unsettled, then switching to custom ASICs once volume ramps up, has become a common industry path.
The Investment Perspective: Opportunities and Risks of a New Theme
For investors in US equities, the beneficiary chain of the ASIC wave is quite clear:
- Leading design services players: Broadcom and Marvell benefit directly as custom chip volumes ramp up, making them the purest plays on this theme;
- Foundry and advanced packaging: ASICs likewise rely on TSMC's advanced process nodes and CoWoS packaging capacity, so segments with tight capacity keep benefiting;
- HBM memory: Custom AI chips also need to carry high-bandwidth memory, and HBM demand keeps rising, sustaining the boom for memory makers;
- Networking and interconnect: Large-scale ASIC cluster deployments simultaneously drive demand for switch chips, optical modules, and networking equipment.
But the risks are equally hard to ignore. First, the customer base is highly concentrated—losing one or two large orders could deal a heavy blow to earnings expectations. Second, custom chips take a long time to go from project launch to mass production, with uncertainties along the way such as failed tape-outs or projects changing course. Third, if NVIDIA proactively cuts prices in the inference market and accelerates its new-product cadence, the price-performance window for ASICs will be squeezed. Fourth, valuations of related stocks are no longer cheap; if the growth of cloud providers' capital expenditure slows, there is still considerable room for a pullback. Keeping a close eye on the capex guidance in tech giants' earnings reports and the order guidance from design service providers is the key anchor for judging the momentum of this main investment theme.
Conclusion: Not Replacement, but Restructuring
Back to the original question: Will ASICs end GPU dominance? The more likely answer is "not a replacement, but a restructuring." The future AI computing landscape will most likely be a hybrid one—GPUs continuing to dominate training scenarios while ASICs accelerate their penetration in large-scale inference—with GPUs and custom chips each playing their own role at different stages. For investors, rather than agonizing over "who will kill whom," it is better to focus on the certainty that the overall computing power pie keeps expanding: whichever side wins, GPU or ASIC, the "picks and shovels" segments—design services, advanced packaging, HBM memory, and networking interconnects—are all core beneficiaries along this new main theme. Only by tracking fundamental data while chasing hot trends can investors hold their positions through volatility and go the distance.
