ASIC
Also known as: application-specific integrated circuit, custom chip, custom silicon, accelerator ASIC
An ASIC (application-specific integrated circuit) is a chip designed to do one specific job extremely well, such as mining Bitcoin, decoding video, or running AI models.
While general-purpose processors can do almost anything, ASICs trade flexibility for speed and efficiency. Their catch is upfront cost: designing an advanced ASIC can cost tens or hundreds of millions of dollars, so it only pays off at large volume.
ASIC, explained at three levels
At home
A general processor is like a Swiss Army knife: it can do many things reasonably well. An ASIC is like a specialized kitchen tool, such as an apple corer, that does one job perfectly and quickly. Your phone has several ASIC-like parts for things like video and image processing.
At work
Companies build ASICs when they need to perform one task at huge scale, efficiently. For example, cloud providers design AI chips to run models more cheaply than buying general-purpose GPUs.
At home
Because an ASIC only has the circuits needed for its job, it wastes less power and space. That's why a Bitcoin mining ASIC can be thousands of times more efficient at mining than a laptop processor, but it can't run a web browser.
At work
ASIC development involves high non-recurring engineering costs: design, verification, masks, and testing. At advanced process nodes, masks alone can cost many millions of dollars. ASICs make sense when volumes are high and requirements are stable. Otherwise, FPGAs or GPUs offer more flexibility with less upfront risk.
At home
In AI, ASICs like Google's TPUs use architectures such as systolic arrays, which pass data through grids of simple maths units, ideal for the matrix multiplications at the heart of neural networks. In my opinion, the AI chip race is really a race to guess which operations will still matter in five years.
At work
The ASIC versus GPU decision for AI depends on workload stability, software ecosystem, and scale. GPUs benefit from mature software such as CUDA, while ASICs require custom compilers and tools. Hyperscalers can justify ASICs because they run enormous, predictable workloads. Smaller organizations usually access ASIC benefits through cloud services rather than building their own.
Real-world examples
At home
Sophie's cousin tried mining Bitcoin on a gaming PC and earned almost nothing. He learned that professional miners use ASICs built only for mining, which perform the necessary calculations far more efficiently than any general-purpose computer.
At work
Thomas is an infrastructure manager at a cloud provider evaluating AI chips. His team found that a custom ASIC offered lower cost per inference for their main model, but required rewriting software. They used the ASIC for stable workloads and kept GPUs for experimentation.
How it compares
| feature | ASIC | FPGA | GPU |
|---|---|---|---|
| Flexibility | One task | Reprogrammable | Broad parallel computing |
| Efficiency for its task | Highest | Moderate | High for parallel work |
| Upfront cost | Very high | Low to moderate | Low (buy off the shelf) |
| Time to deploy | Long (design and fabrication) | Short | Immediate |
| Best for | High-volume, stable workloads | Prototyping, changing requirements | AI training, graphics, research |
ASICs win on efficiency when the job is stable; FPGAs and GPUs win on flexibility.
For analysts
Building a custom chip is a long-term bet on a workload not changing. I'd only consider it with clear volume forecasts and a fallback plan, since AI model architectures have shifted repeatedly in just a few years.
History & origin
Custom and semi-custom chips have existed since the 1970s and 1980s, when gate arrays let manufacturers tailor chips for specific products. ASICs became common in electronics such as networking equipment and consumer devices. Bitcoin mining offered a dramatic example: after miners moved from CPUs to GPUs, the first Bitcoin ASICs arrived in 2013 and quickly made other hardware uncompetitive. Google designed its Tensor Processing Unit, an ASIC for machine learning, and revealed it in 2016 after using it internally. Large cloud companies, including Amazon, Microsoft, and Meta, later built their own AI ASICs, often with design partners like Broadcom. In my view, ASICs are a bet on stability: they win big when the task stays the same and lose when it changes.