AI Gold Rush: No need to buy GPU, no need to train models—just by selling Token you can still earn millions a month USD
BIBIBI
AT A GLANCE
Some companies rent GPU to deploy open-source models, then charge by Token, with monthly revenue reportedly reaching 1000000 USD。
Article
AI Gold Rush: No need to buy GPU, no need to train models—just by selling Token you can still earn millions a month USD Beating AI Flash News reports that Menlo Ventures investor Deedy Das discovered that some AI startups simply rent machines, deploy open-source models such as DeepSeek and Kimi,GLM and then charge by Token, with monthly revenue quickly reaching 1000000 USD. Inference service providers of this kind have recently been emerging in large numbers. They do not need to buy GPU, nor do they need to train models themselves.
There is still a considerable price spread in this business. According to Deedy’s calculations, a single B300 even if rented at an hourly rate as high as 8 USD, can still be sold for 10～20 USD of Token per hour, provided processing speed is sufficient and there are enough customer orders.
The barriers to entry are also getting lower. Some small teams can even use AI programming tools to deploy open-source models on rented GPU. Platforms such as OpenRouter and Vercel also help them aggregate customer demand. Deedy said that dozens of similar companies enter the market almost every week. Everyone is competing to rent GPU, driving up computing prices while continually cutting prices to compete for customers.
Original link https://m.theblockbeats.info/flash/373699
Key points
01
Beating, citing Menlo Ventures investor Deedy Das, reports that some AI startups rent machines, deploy open-source models such as DeepSeek, Kimi, and GLM other open-source models, and charge by Token.
02
The original text states that some inference service providers do not need to purchase GPU, nor do they need to train models themselves.
03
Deedy calculates that a single B300 even if its hourly rental rate reaches 8 USD, may still be sold for 10～20 USD of Token。
04
The original text states that platforms such as OpenRouter and Vercel help these teams aggregate customer demand.
05
Deedy said that dozens of similar companies enter the market almost every week.
06
The original text states that competing to rent GPU may drive up computing prices, while inference service providers compete for customers by cutting prices.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
These companies rent GPU, run ready-made open-source models, and then sell customers the usage of model-processed requests by Token. They do not need to purchase GPU or train models themselves, but they need enough orders to cover rental and operating costs.
Why it matters to readers
This indicates that AI the barriers to entry for inference services may be falling, and may also increase demand for GPU rentals and competition among Token services.
You can understand Token as the unit used to charge for a model processing text; whether a company is profitable depends on whether its charges can cover GPU rental and other costs.
Risks and unknowns
Monthly revenue 1000000 USD, per hour 10～20 USD and other key figures mainly come from Deedy Das’s observations or calculations, without independent verification in the original text.
The original text does not provide the specific names of the companies involved, their customer counts,GPU utilization rates, or profit margins.
If customer demand is insufficient,Token prices decline, or GPU rental costs rise, the profit margin may narrow.
Related Developments
Loading event timeline…
Related concepts
Token
This term is not in the glossary yet. Browse related concepts in the glossary.
