Estimates Show Nvidia's Inference Chip Share Rises to 74%; Custom Chips Have Yet to Erode Its Dominance
N.R. Finch
The Information estimates Nvidia's AI inference chip market share rose from 66% to 74% over the past year, with quarterly sales of roughly $41 billion. This means → even as Big Tech pours money into custom silicon, Nvidia is selling more, not less, in the part of AI closest to revenue.
Why has inference become the main AI chip battleground?
Inference — the computation that lets a trained AI model actually answer questions and generate content — now accounts for roughly 60% of AI workloads. Deloitte expects that to reach two-thirds by year-end.
Two years ago the figure was one-third; in 2025 it hit half. In plain terms = companies now spend more on "running the model" than on "teaching the model," and the gap keeps widening.
Nvidia CEO Jensen Huang put it in one line at Computex in Taiwan: "Inference is where the money is."
How was Nvidia's share calculated?
The Information's scope: global inference chip quarterly revenue of roughly $56 billion, covering AMD's and Marvell's data-center revenue, Cerebras hardware revenue, and 80% of Broadcom's custom-chip business.
The key assumption: 55% of Nvidia's April-quarter revenue came from inference server chips — roughly $41 billion. The basis: Nvidia disclosed that 40% of its 2024 data-center equipment revenue was inference-related, and executives said the share had risen meaningfully since.
Wedbush's Matt Bryson and Morningstar's Brian Colello independently estimate that over half of Nvidia's data-center revenue now comes from inference. This means → multiple firms converge on the same direction — the 55% assumption is not aggressive.
Even at a more conservative 50%, Nvidia's inference share still rose from 68% to 72% year-on-year — the trend holds.
Where is the money behind inference chips coming from?
Annual run-rate revenue at 34 AI startups, led by Anthropic and OpenAI, surged from roughly $7 billion to nearly $80 billion in eighteen months.
This reflects a direct link between inference and AI service revenue — every time a user calls ChatGPT or Claude, inference chips are doing the work.
Huang attributed part of Nvidia's share gains to an expanded partnership with Anthropic and the rise of new model developers such as Reflection and Cursor. In plain terms = when new entrants buy their first batch of chips, most of them are still buying Nvidia.
With so many rivals, who can actually take share?
The inference market has far more competitors than training: OpenAI, Google, Meta, Microsoft, and Amazon are both major customers and potential rivals, while AMD, Cerebras, and other independent chip makers are also fighting for share.
Summit Insights Group analyst KinNgai Chan told Reuters the firm expects Nvidia to start losing inference market share in 2027 as volumes ramp.
Broadcom is also scaling up: this week it announced a $35 billion fund with Apollo and Blackstone to finance over 20 gigawatts of AI data-center capacity, including projects tied to Anthropic and OpenAI that rely on Broadcom-designed chips.
This means → the threat is real, but the current data shows custom silicon and third-party rivals have not yet converted into share losses — Nvidia's slice actually grew from 66% to 74%.
Content is for reference only, not financial advice.