NVIDIA remains the default choice for AI researchers publishing their work, commanding 91% of tracked hardware citations in 2026. However, new data from the State of AI Compute Index reveals that competitors are finally making measurable progress, even if they haven't yet dented the company's overwhelming market position.

The latest index, released in early July 2026, tracks how often different AI chips appear in published research papers. It's a useful window into which hardware researchers actually use, even though it captures only a fraction of what happens inside private labs and frontier AI companies. The findings paint a picture of a market in motion, but one where NVIDIA's dominance remains largely unchallenged.

Which AI Chips Are Gaining on NVIDIA?

Across all tracked accelerator categories, the 2026 projection reaches 49,339 chip-citation counts, up 10.7% year-over-year. NVIDIA's chips appear in 44,715 of those counts, up 10.9% year-over-year. The company's dominance is so complete that when competitors do gain traction, it's often in narrow, specialized niches rather than as general-purpose alternatives.

The most striking gains come from unexpected places. AMD citations nearly doubled, from 251 in 2025 to 472 in 2026. Huawei's Ascend 910 processor rose 56%, from 137 to 213 citations. But perhaps most surprisingly, Apple moved from 741 to 998 citations, overtaking AMD to become the most-cited non-NVIDIA, non-Google accelerator in the open literature. This likely reflects the spread of local inference and developer workflows on consumer Macs rather than frontier training runs.

• AMD: Nearly doubled citations from 251 to 472, showing growing adoption in research but still far behind NVIDIA's scale.
• Apple: Jumped to 998 citations, becoming the leading non-NVIDIA, non-Google chip cited in published papers, driven by local inference on consumer devices.
• Huawei Ascend 910: Rose 56% to 213 citations, reflecting geopolitical diversification efforts in AI research.
• Startup chips: Groq emerged as the most-cited startup chip at 264 citations, up 49%, though NVIDIA acquired it in December 2025.

The startup chip category reveals an important pattern. Groq, Cerebras, SambaNova, and others are specializing into specific jobs rather than trying to replace NVIDIA across the board. Groq shows up around low-latency inference tasks, Cerebras around large-scale wafer-scale systems. The older acquired or de-emphasized platforms are fading from view. As the index notes, the right conclusion is not that startup chips are breaking CUDA, NVIDIA's dominant software framework. They are not. Instead, the category is specializing into niches while NVIDIA keeps the general case.

Why Is NVIDIA's Product Cycle Now Visible in the Research Data?

One of the clearest signals in the index is how NVIDIA's own product transitions are now visible in the research literature. The A100, NVIDIA's workhorse GPU from 2020, shows citations basically flat at 15,327, up just 0.4% year-over-year. It's no longer where the growth is. Instead, the H100 and H200 Hopper-generation chips have more than doubled to 9,823 citations, up 111% year-over-year, as the 2024 and 2025 build-out finally works its way into published papers.

The newer Blackwell-family chips, tracked as B100, B200, and B300, are still small at 902 citations but are up 4.5 times year-over-year, signaling the beginning of the next wave. Meanwhile, older hardware is draining out. V100 citations are down 34%, RTX 3090 down 23%, P100 down 36%, and the K80 is barely visible anymore. The RTX 4090 consumer GPU remains useful in the academic long tail, up 9% to 6,557 citations.

How to Understand the Difference Between Research Citations and Real-World Deployment

It's important to recognize that paper citations tell only part of the story. The largest model developers still publish less of their best work, and many papers built on managed APIs or shared cloud services don't specify the underlying silicon at all. But the open literature does reflect hardware diffusion patterns, and the 2025 dip in citations looks more like publication-cycle timing and API abstraction than a collapse in compute usage.

The second half of the index shifts focus from research papers to actual deployed infrastructure, and the picture becomes even more NVIDIA-dominated. Here's what the deployment data reveals:

• Hopper deployment scale: Across tracked systems, there are 460,904 deployed H100, H200, and GH200 GPUs, plus 1,328 installing. That's more than 11 times the tracked A100 total of 41,208, showing the A100 now reads like a legacy fleet chart.
• Largest private clusters: xAI Colossus 1 leads at 200,000 GPUs, almost five times the entire tracked A100 set. Tesla Cortex follows at roughly 66,000 H100-equivalent GPUs, then Meta's GenAI clusters at 49,152.
• Grace-Blackwell pipeline: The index tracks 100,128 deployed GB200/GB300 GPUs, 308,640 installing, and 1.66 million announced. Just under 5% of the tracked Grace-Blackwell pipeline is deployed, with roughly 80% still announced.

The geopolitical dimension is worth noting. National HPC systems are exact and visible, making them easy to count. Private fleets are estimated, harder to observe, and much larger. Europe now has serious machines in JUPITER, Alps, Isambard-AI, Leonardo, MareNostrum 5, and Jean Zay. But the largest private AI clusters are operating at a scale that national supercomputing programs mostly do not match.

The constraint has moved outward from GPU count to power, land, cooling, interconnect, permitting, debt, and offtake agreements. A GPU order is not a cluster. A cluster is not always available capacity. And contracted capacity is not the same as a model training run. This shift explains why the conversation around AI infrastructure is increasingly about megawatt capacity and sovereign industrial projects rather than simple chip counts.

What Does This Mean for the AI Industry Going Forward?

The 2026 data shows that NVIDIA's dominance is real but not absolute. The company remains the default choice for researchers and builders, but the kingdom is changing shape. AMD, Apple, Huawei, and startup chips are finding their footing in specific use cases and geographies. The open-weight frontier is also catching up, with models like GLM-5.2 now matching or beating closed flagships on coding benchmarks.

For businesses and researchers, the practical implication is clear. The question is no longer simply "which chip is best?" It's "which combination of chips, models, and governance lets us deploy safely, cheaply, and at scale?" NVIDIA's dominance in research citations and deployed infrastructure is likely to persist, but the company's share of the total AI compute market is no longer growing at the rate it was in 2024 and 2025. Competitors are specializing, governments are entering the model layer directly, and the open-weight frontier is closing the capability gap.