Key findings
- • High-bandwidth memory (HBM) is estimated at US$23.9 Bn in 2025, reaching US$120.9 Bn by 2035.
- • Next-generation memory grows fastest: HBM4 at 35.1% a year and CXL memory at 31.6%.
- • AI chiplets grow from US$7.9 Bn to US$88.3 Bn as accelerators are built from several dies.
- • The median AI chip market grows at 21.6% a year to 2035.
AI semiconductor markets, 2025 vs 2035
| Market | 2025 (US$) | 2035 (US$) | CAGR 2025-2035 |
|---|---|---|---|
| AI Accelerator | $55.1 Bn | $370.7 Bn | 21.0% |
| AI Inference Processor | $36.4 Bn | $192.1 Bn | 18.1% |
| AI SoC | $45.0 Bn | $235.5 Bn | 18.0% |
| High Bandwidth Memory (HBM) | $23.9 Bn | $120.9 Bn | 17.6% |
| HBM4 | $286.0 Mn | $5.8 Bn | 35.1% |
| CXL Memory | $742.0 Mn | $11.6 Bn | 31.6% |
| AI Chiplet | $7.9 Bn | $88.3 Bn | 27.3% |
| Advanced Packaging | $40.1 Bn | $90.7 Bn | 8.5% |
| Co-Packaged Optical Engines | $186.0 Mn | $3.1 Bn | 32.5% |
| Silicon Photonics Packaging | $731.0 Mn | $6.5 Bn | 24.4% |
| Photonic Computing Chips | $214.0 Mn | $2.7 Bn | 28.9% |
| Neuromorphic Chip | $214.1 Mn | $2.3 Bn | 26.8% |
Source: Market-Reports.com estimates. Base year 2025, forecast to 2035.
The memory wall
An accelerator is only as fast as the data it can reach. High-bandwidth memory stacks sit beside the processor for that reason, and each generation adds bandwidth. CXL extends the idea across servers by pooling memory that several processors can share. Both grow faster than the accelerators themselves because every new accelerator needs more of them.
Chips made of chips
Large AI processors are increasingly assembled from several smaller dies - chiplets - joined by advanced packaging. That turns packaging from a back-end step into a source of performance, and creates new markets in chiplet design, testing and interconnect.
Light between chips
Copper struggles at the distances and speeds AI clusters need. Co-packaged optics and silicon photonics move data as light right at the chip, cutting power per bit. Photonic and neuromorphic computing are smaller and earlier-stage, but they target the same problem: doing more AI work per watt.
About the data
All figures are Market-Reports.com estimates of global market revenue in US dollars, with 2025 as the base year and forecasts to 2035; growth rates are compound annual growth rates (CAGR) for 2025-2035. They come from our catalogue of more than 4,600 market reports across 19 industries, in which every market is sized on its own definition. Markets overlap - an AI server is also part of AI infrastructure - so the figures in this study must not be added together. Medians are calculated across all reports whose market name matches the theme, as published in October 2026.
Use this data
Journalists, analysts and bloggers may quote these figures and reproduce the table, with credit to Market-Reports.com and a link to this page:
Source: Market-Reports.com, "Beyond the GPU: Memory, Packaging and Photonics Markets That AI Depends On" - https://www.market-reports.com/insights/ai-chips-beyond-the-gpu-2035
Need a figure for a different market or region? Email contact@market-reports.com.
Key takeaways
- • Memory grows faster than the accelerators it feeds.
- • Packaging and chiplets have become a source of AI performance.
- • Optical links at the chip are the next interconnect shift.