Pricing AI for small businesses

Taiwan makes the chips the world’s artificial intelligence (AI) runs on. However, how many Taiwanese companies get to use it?

The Ministry of Digital Affairs in September set out its approach: Private operators would build, own and run new AI computing centers, and the government would ease the investment rules rather than buy the capacity itself. It aimed to reach 10,000 graphics processing units (GPUs) of new capacity within a year, Minister of Digital Affairs Lin Yi-jing (林宜敬) said. If the state bought all of it and handed it out, it would distort the market, he said.

He has a point. Canada just took the other road. Ottawa on Thursday last week vowed to help Canadian AI companies access 500 megawatts of AI compute, using government guarantees and long-term purchase agreements to draw in private money. That makes Ottawa a large buyer, and puts officials in the business of deciding who gets the capacity.

Taiwan avoids that. It runs a different risk.

A market left to price itself would usually price for its biggest customers. An operator would rather sign three large tenants for five years than serve 300 small software firms month to month, because it is less paperwork and less risk. That is sensible for the operator. For Taiwan, it could mean that new capacity fills up with buyers who already have access, while small firms that turn AI into everyday products wait.

The waiting has already started. A national research supercomputer was reportedly running at full load in September.

The people are coming, too. The government wants to train 500,000 AI professionals by 2040. Plenty of them would work at small and medium-sized companies, which make up most of Taiwan’s businesses, and they would need machines to work on.

My company, Caseway, is a 40-person AI outfit headquartered in Vancouver that sells in Taiwan. We build software that turns documents into completed forms and usable data. We do not need a training cluster. We need steady inference capacity at a price we can plan around, with data kept where our customers need it.

Taiwan does not have to subsidize anyone to fix this. It already has leverage. The ministry is easing rules for the operators building these centers, and it can ask for something back.

The simplest ask is a posted price. Each center that benefits from the new rules could publish an hourly rate for small customers, with a monthly option on part of its capacity. Then a founder in Hsinchu could compare it with what they pay a foreign cloud provider today. Operators would still set the rate, and posting a price is ordinary market behavior.

The other ask is reporting. Once a year, the ministry could publish how many companies are running live products on the new capacity, how many of those are small and how long they waited. GPU totals show that a building is full. These figures would show whether Taiwan’s software sector grew because of it.

Taiwan and Canada picked opposite tools for the same goal, and both will be judged on one test: whether a small domestic company can get a quote, sign up and run a real product within weeks. Taiwan has the chips. Its own companies should be able to rent time on them.

Al Vigier is the CEO of Caseway, an AI company that cocreates products with large enterprises and shares the revenue with them.