With NVIDIA's Rubin AI GPUs official and growing reports of custom AI chips generating interest, we decided to interview Rahul Sen Sharma, President and Co-CEO of Indxx, a global index provider, to understand the market dynamics surrounding AI computing and application specific integrated circuits (ASICs). ASICs are commonly referred to as custom-AI chips, and with the media full of reports disagreeing about the extent of their use, we asked Rahul the current state of the AI chip ecosystem.
Though a written Q&A last month, our conversation started by discussing NVIDIA's role in the current AI chip ecosystem particularly through its Blackwell GPUs. NVIDIA's Blackwell chips are the latest AI chips currently shipping, and Rahul pointed out that the CUDA software ecosystem plays a key role in NVIDIA's success. CUDA's strength in the AI ecosystem is no secret, as it has allowed NVIDIA to continue to play a key role in the Chinese market despite headwinds.
NVIDIA's role in the current AI ecosystem, especially in the context of ASICs starting to grow and Broadcom Inc emerging as a viable competitor.
NVIDIA continues to play a central and dominant role in the AI ecosystem, holding 86% of the AI GPU market in 2025. Its GPUs, particularly the Blackwell series, are widely used for training and inference of AI models, while the CUDA software ecosystem has become the de facto standard for developers building AI applications. This combination of high-performance hardware and a mature software ecosystem have made NVIDIA the backbone of AI infrastructure worldwide, giving it a significant competitive advantage and a near-monopoly over general-purpose AI computation. (1)
However, the AI hardware landscape is gradually diversifying, driven by the growing adoption of application-specific integrated circuits (ASICs). Companies like OpenAI, in collaboration with Broadcom Inc, are now developing custom ASICs specifically designed to optimize inference workloads. These chips, fabricated using TSMC’s cutting- edge 3-nanometer process, will initially be used in OpenAI’s own data centers rather than sold commercially. This strategy mirrors similar initiatives from other AI leaders—Google’s TPUs, Amazon’s Trainium and Inferentia chips, and Meta’s in-house silicon projects—all aimed at reducing reliance on NVIDIA’s GPUs, which are expensive and sometimes supply-constrained. (2)
The Broadcom-OpenAI partnership positions Broadcom as a credible emerging competitor in the AI hardware market. Broadcom’s AI revenue reached $4.4 billion in Q2 2025, a 46% year-over-year increase, fueled by its ASICs and hyperscaler-focused Ethernet switches. Strategic moves, such as the VMware acquisition and partnerships with cloud hyperscalers, further strengthen Broadcom’s position. While NVIDIA’s dominance is unlikely to be overtaken in the near term due to its mature ecosystem and broad platform adoption, the growth of custom ASICs signals that NVIDIA may no longer have exclusive control over AI infrastructure, opening the door for competitors like Broadcom to gain relevance in the inference-focused segment of the market. (3)
After the initial response from Rahul, we followed up with a series of questions to discuss the cost benefits offered by custom AI chips, what improvements ASICs could see in the future, the computer hardware sectors that might benefit from ASIC demand growth and the changes in the cooling industry resulting from NVIDIA's latest Rubin AI chips.
What is the current sentiment in hyperscalers for relying on cost-performance for AI as opposed to performance alone as shown by Amazon relying more on in-house Trainium?
Rising energy intensity, water consumption, GPU scarcity, and Return on Investment (ROI) pressures make scaling purely on performance unsustainable. So, Hyperscalers like AWS, Google, Microsoft and more are moving away from chasing peak raw performance alone. Their new focus is cost-performance optimization — achieving the best performance per dollar spent.
Amazon: By relying more on in-house Trainium chips, AWS is prioritizing control and efficiency. While Trainium may not match Nvidia GPUs on peak specs, it offers lower TCO (Total Cost of Ownership) and efficiency gains.
Similar approach is followed by Google (TPUs), Microsoft (Athena), and Meta (MTIA) by investing in their in-house silicon to reduce dependence on Nvidia. Overall, Hyperscalers are adopting a hybrid setup — still relying on Nvidia GPUs for peak performance while gradually integrating in-house chips to improve cost performance, efficiency and control.
What improvements might the ASICs see in the future that could increase their adoption?
Future improvements to ASICs like greater energy efficiency, improved security features, enhanced AI integration and more will drive their increased adoption:
Energy Efficiency: Since they’re purpose-built for specific functions, they avoid the unnecessary overhead associated with general-purpose processors. In hyperscale data centers, where energy is a significant operational cost, ASICs can reduce energy usage by up to 30% compared to CPUs. (4)
Security: Next-gen ASICs will add tamper resistance and encrypted protocols to combat rising cyber threats.
Workload specialization: Future ASICs will be optimized for specific AI tasks, making them far more efficient than general-purpose GPUs.
Artificial Intelligence (AI) Integration: ASICs will see tighter integration with AI hardware, such as TPUs, enabling sophisticated real-time content analysis and AI-powered video processing in broadcast and other fields.
Memory Disaggregation (via CXL): Compute Express Link (CXL) separates memory from processors, creating shared memory pools. ASICs will benefit from this, as it reduces stranded memory and enables AI models to run more efficiently
ASICs are evolving to become more energy-efficient, secure, specialized, and cost-effective, making them a critical driver of the next phase of AI and cloud infrastructure adoption.
Which firms can benefit from the growth in ASIC demand?
The rise in ASIC adoption will benefit a wide range of players across hyperscalers, semiconductor firms, and supporting ecosystem providers:
- Hyperscalers (In-house ASICs):
Amazon (AWS – Trainium): Reduces Nvidia dependence, improves efficiency for AI workloads.
Google (TPUs): Integrated into Google’s AI stack, powering Gemini models.
Microsoft (Athena): Custom silicon aligned with OpenAI and Azure workloads.
Meta (MTIA): Inference-focused ASICs for scaling AI-driven services.
- Semiconductor & Foundry Players:
TSMC & Samsung Foundry: Leading manufacturers of advanced ASIC designs for hyperscalers.
Broadcom & Marvell: Supply networking/AI ASICs, benefiting from growing data center demand.
AMD & Intel: Expanding into semi-custom ASICs and heterogeneous integration solutions.
- Ecosystem & Infrastructure Enablers:
EDA Companies (Synopsys, Cadence, Siemens EDA): Provide essential ASIC design and verification tools.
Memory & Interconnect Vendors (Micron, SK Hynix, Rambus): Enable CXL-based disaggregation to support AI workloads.
Cooling & Power Firms (Vertiv, Schneider Electric): Benefit from efficiency driven ASIC deployments in dense data centers.
Connectivity Providers: Firms delivering high-speed interconnects and networking solutions also gain from ASIC-driven ecosystem growth.
Both hyperscalers (with in-house designs) and semiconductor/foundry firms (manufacturing and IP) will capture major value from the ASIC growth cycle, while supporting players in EDA, memory, and cooling infrastructure will also see strong tailwinds.
In the context of the latest Rubin GPUs, any data center cooling firms that could benefit?
At GPU Technology Conference (GTC 2025), Nvidia announced its data center roadmap for 2026–2027, introducing Rubin and Rubin Ultra GPUs. 2 With per-GPU and per-rack power density continuing to rise, Nvidia’s roadmap is expected to accelerate adoption of high-performance liquid cooling and immersion cooling. This creates strong tailwinds for cooling solution providers.
Firms to watch:
- Market leader in liquid cooling for computing.
- Showcased AI-ready coldplate systems at Nvidia GTC 2025, including OMNI coldplates with Split-Flow technology, delivering up to 30% higher performance vs. standard coldplates.
- Listed on Nvidia’s Recommended Vendor List (RVL) for server manufacturers.
Asetek — long-time player in liquid cooling and featured in market reports as a major data-center cooling vendor.
Data Center Cooling Companies (Vertiv and Schneider Electric): Large-scale providers of HVAC, PDUs, and data-center infrastructure; benefit indirectly as integrators for facility upgrades and suppliers of hybrid liquid/air cooling solutions.
References:
(1): https://www.ainvest.com/news/broadcom-emerging-ai-dominance-overtake-nvidia-2509/
(2): https://techstrong.ai/aiops/openai-and-broadcom-plan-end-run-around-nvidia/
(3): https://www.ainvest.com/news/broadcom-emerging-ai-dominance-overtake-nvidia-2509/
(4): https://www.glomore.co.in/10-ways-asics-are-shaping-the-future-of-semiconductors/#:~:text=A%202025%20study%20by%20MarketsandMarkets,net%2Dzero%20goals%20by%202030/
About the author: Ramish is a seasoned technology writer and editor with more than a decade of experience. He specializes in semiconductor fabrication and market analysis. With a background in finance and supply chain management - via his bachelors in Finance and a micromasters in supply chain management from MIT - Ramish combines financial rigor with deep industry insight to deliver accurate and authoritative coverage.
Follow Wccftech on Google to get more of our news coverage in your feeds.
