The competition for access to AI servers is getting tougher3. May 2023
The competition for access to AI servers is getting tougher
San Francisco, 5/3/2023
AI companies that want to train their AI systems compete for access to AI servers. Generative AI apps like chatbots require significant investments in AI infrastructure because training a model like GPT-3 is computationally intensive and therefore very expensive.
A new report from VC firm Andreessen Horowitz breaks down the cost factors for AI companies looking to run their own models on GPUs. AI servers are based on GPUs, and according to a16z, companies can spend over 80% of their total capital on computing resources to train and run AI software.
AWS, Microsoft Azure, Google Cloud and other cloud providers all offer AI servers. While pricing depends on factors like GPU availability and negotiation, a16z found that pricing varies between vendors by a factor of nearly four times. FluidStack has the lowest hourly rates and AWS has the highest. Startups may also find more chips available from smaller cloud providers.
Currently, the success of AI companies can be largely determined by their access to AI servers with GPUs. Cloud providers large and small are experiencing exploding demand for AI servers and are being forced to limit their availability. In 2023, AI server shipments are expected to increase by 8% YoY as demand booms.