It’s not often that NVIDIA CEO Jensen Huang loses his cool. But that’s exactly what happened during his recent appearance on Dwarkesh Patel’s podcast. When the usually composed, leather jacket-clad Huang was asked whether Google TPUs could threaten NVIDIA’s dominance, he pushed back hard.
The real competition, he argued, isn’t about chips. It’s about envisioning what kind of future holds for computing.
For Huang, focusing solely on AI misses the bigger picture. “Accelerated computing is much more diverse,” he said on the podcast, adding that computing extends well beyond AI. NVIDIA’s transition from general-purpose to accelerated systems, he argued, extends its reach far beyond TPUs or ASICs (application-specific integrated circuits).
That dominance reflects in the numbers. In fiscal year 2026, the company reported total revenue of $215.9 billion, with its data centre segment, driven overwhelmingly by AI accelerators, contributing a record $193.7 billion for the full year and $62.3 billion in the fourth quarter alone.
Notably, Google, Amazon, Meta, and Microsoft are set to pour $595–625 billion into capital expenditures in 2026, largely driven by the buildout of AI data centres.
Gaurav Gupta, VP Analyst at Gartner, tells AIM that the GPU total addressable market (TAM) will continue to exceed AI ASICs over the next four to five years, although AI ASICs will grow at a faster rate.
“A big part of the reason is shift in AI semiconductor spending will be higher for inference over training. NVIDIA will continue to be the most dominant player, but hyperscalers will continue to rely more on in-house chips and invest in those capabilities,” he adds.
Challenges Galore
NVIDIA’s biggest competition comes from the same companies it has been powering. AWS, Microsoft Azure, and Google Cloud have the resources to build their own chips and are increasingly doing so.
The AI ecosystem is now witnessing a wave of NVIDIA alternatives, from Google’s TPUs to AWS’s Trainium accelerators and Huawei’s Ascend chips, as cloud providers and governments push to control costs, secure supply, and reduce reliance on the dominant supplier of AI chips.
Custom ASICs offer superior price-performance for specific workloads, particularly high-volume inference, where NVIDIA’s general-purpose GPUs carry premium pricing and higher power demands.
David Wu, Research Associate at Counterpoint Research, describes custom chips as “primarily a total cost of ownership optimisation tool for hyperscalers to handle their massive, predictable internal workloads,” adding that while this shift is structural, it remains confined to a few dominant players.
He outlines a tiered landscape of the custom AI ASIC market, spanning tier 0 (Google), tier-1 (AWS), tier-2 (Meta and Microsoft), and tier-3 (others). “We expect Google alone to account for roughly half of total AI ASIC wafer demand by 2028,” he shares with AIM.
However, he mentions that neocloud providers such as CoreWeave offering high-performance GPU-as-a-service still rely on NVIDIA GPUs as the only practical option.
At the same time, hyperscalers are accelerating investments in proprietary silicon.
At Google Cloud Next 2026, the search giant unveiled its eighth-generation lineup, including the TPU 8t, optimised for training massive models, and the TPU 8i, for inference and agentic AI workloads that require rapid reasoning and multi-step execution.
And Google is already lining up who will consume these chips at scale. It has reportedly committed up to $40 billion to Anthropic, including $10 billion upfront and another $30 billion tied to performance milestones. Alongside this, it has secured roughly 5 gigawatts of TPU compute capacity for Anthropic.
At the same time, Google is pushing deeper into the semiconductor stack. It recently partnered with Cadence Design Systems to integrate its Gemini models into AI-driven chip design tools, helping automate large parts of the design and verification process.
Yet, Huang argued that this momentum may be more concentrated than it appears. Besides Google and Anthropic, he sees little evidence of broad TPU adoption. “Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic,” he said on the podcast, further adding that there isn’t an abundance of ASIC opportunities.
Interestingly, Anthropic appears to be actively diversifying its compute resources. It has signed a new agreement with Amazon to expand its partnership, securing up to 5 GW of compute capacity for training and deploying its Claude models. The deal includes new Trainium2 capacity coming online in the first half of 2026, with nearly 1 GW of combined Trainium2 and Trainium3 infrastructure expected to be operational by the end of 2026.
Huang dismissed the idea that custom chips alone can challenge NVIDIA. He argued that the cost advantage of custom silicon is often overstated, noting that the margin gap between NVIDIA GPUs and ASICs is relatively small, which makes the economic savings far less compelling than they initially seem.
However, Gupta says that as use cases evolve and the industry moves toward agentic AI, heterogeneous compute will become a core requirement. “We will see hyperscalers combining solutions with continued reliance on NVIDIA, their in-house chips, and leveraging specific chips from various AI startups.”
Huang said the US needs to stay clearly ahead in AI. He pointed to upcoming systems like Vera Rubin and Blackwell as key to maintaining that lead, and stressed that the country must have a large and ready supply of computing power. NVIDIA has started shipping early Vera Rubin systems to select customers, with full production expected in the second half of 2026 or early 2027. Microsoft Azure is among the first to deploy these systems (NVL72) for validation and early use.
Like Google, Meta is also actively developing custom chips. Meta recently unveiled its in-house AI chips—MTIA 300, MTIA 400, MTIA 450, and MTIA 500—which are part of its Meta Training and Inference Accelerator (MTIA) family and will be deployed across its data centres between 2026 and 2027.
The chips are built to handle a wide range of AI workloads, including recommendation systems, generative AI training, and large-scale inference.
The social media giant also partnered with AWS to bring tens of millions of Graviton CPU cores into its infrastructure, making it one of the largest adopters of AWS’s custom silicon. Meta framed the decision as part of a diversification strategy rather than a replacement of existing infrastructure.
DeepSeek and China Trouble
China has long been a key market for NVIDIA, accounting for around 13% of its revenue in fiscal year 2025, or roughly $25–30 billion annually. However, Chinese AI companies are increasingly turning to domestic alternatives such as Huawei’s GPUs as US export restrictions limit access to NVIDIA hardware.
DeepSeek recently introduced a preview of its V4 model—its most advanced system so far—featuring 1.6 trillion parameters and a context window of up to 1 million tokens. The release marks one of the first major frontier models built to run on Huawei’s Ascend AI processors instead of NVIDIA hardware.
Huang pushed back strongly against the idea that restricting chip exports will automatically preserve US dominance. If anything, he warned, it risks accelerating the rise of parallel ecosystems.
According to a report by Bernstein Research, NVIDIA’s share of China’s AI chip market is projected to fall to 8% by 2026, down from 66% in 2024 and 54% in 2025, while Huawei is expected to capture around 50% of the market.
Can Cuda Save NVIDIA?
If hardware is the battlefield, software is NVIDIA’s fortress. Huang repeatedly returns to one advantage, CUDA (Compute Unified Device Architecture), a parallel computing platform and programming model that hosts an entire ecosystem of libraries, frameworks, and tools. From PyTorch to Triton, much of modern AI development is deeply intertwined with NVIDIA’s stack.
“CUDA’s moat is real, but it is primarily a moat against the broader market, with strong entrenchment among developers and enterprises,” Wu says. He explains that for most companies building AI models on NVIDIA hardware, switching costs remain significant and are unlikely to ease anytime soon.
However, he adds that the top hyperscalers operate their own proprietary software stacks and can optimise for their own hardware. Wu points to Anthropic as a case in point, adding, “Many of its core engineers have deep familiarity with Google’s TPU and AWS’s Trainium ecosystem, and with both Google and AWS as major investors, the transition to non-NVIDIA hardware becomes more natural.”
The AI chip market is growing fast, but NVIDIA’s dominance is now under fierce pressure.
Your reaction