• Market share (y-axis). A designer's portion of the AI computing power sold or shipped in a given quarter, expressed as a percentage. The percentages for all designers in a quarter add up to 100%.
• H100-equivalent, H100e. The unit that shares are built from. One H100e equals the compute capacity of one H100 chip, based on dense 8-bit operations per second.
• Time (x-axis). The quarter in which chips were sold or shipped.
• Designer (color). The company that designed the chip, for example, Nvidia or Google. This is not the company that physically fabricated it; most AI chips are manufactured by TSMC regardless of who designed them.

• Coverage starts at different times by designer (Nvidia from 2022; Google from 2023; AMD, Amazon, Huawei, and Cambricon from 2024). Earlier quarters reflect mostly Nvidia, so shares before other designers enter the data overstate Nvidia's position.
• H100-equivalents measure paper specifications, not real-world performance. Real throughput depends on software, memory, and networking, so the basis is a useful common scale, not a precise performance ranking.
• Share tracks sales and shipments, not deployment.

Epoch estimates how many chips each designer has sold or shipped from financial reports, company disclosures, analyst estimates, and media coverage, with methods that vary by the evidence available.

Each chip model's volume is converted to H100-equivalents from its peak 8-bit operations per second relative to the Nvidia H100. A designer's total is the sum across the models it designed, for example Nvidia's Hopper and Blackwell chips, and its quarterly market share is that total divided by the total across all tracked companies.

Shares are relative to the designers tracked: Microsoft Maia, Meta MTIA, Tesla Dojo, consumer GPUs, and smaller players are excluded for limited scale or public information, so each tracked designer's share is slightly overstated.

Read the full methodology