With tech companies racing to build cutting edge AI models, headlines about AI infrastructure spending are dominating the discourse around AI. At the same time, some early adopters of Gen AI business models are questioning the benefits from AI deployment. It is important to critically assess what economic value is created with all these investments and the sources of value creation in Gen AI companies. For that, we need to first understand the value chain in AI and then identify the monetization opportunities that can potentially exist. An AI value chain is the sequence of activities and layers required to transform raw computational resources and data into AI-powered products and services that deliver value to end users. It represents the end-to-end process of creating, deploying, and capturing value from artificial intelligence.
I spoke to Rohini Chakravarthy, Managing Partner at NewBuild Venture Capital, about the sources of value creation with new technologies. We discussed the idea of an AI “innovation rail” – infrastructure services (which include AI knowledge, execution capabilities, and a consumption business model) from various companies — that puts AI-first applications in motion. Applications bring in context from domains such as supply chain, finance, or marketing, and reuse the AI innovation rail to rapidly deliver value.
The AI value chain can be characterized as a multi-layered stack with distinct but interdependent components. First, there is a core infrastructure layer. At the foundation, you have compute and hardware – primarily specialized chips (GPUs, TPUs, custom AI accelerators) from companies like NVIDIA, AMD, and Google. Cloud providers (AWS, Azure, GCP) package this compute into accessible infrastructure. The next is the model layer that includes foundation – large language models, image generators, video models, etc. Companies here range from OpenAI, Anthropic, and Google to open-source efforts like Meta’s Llama.
The third layer is the platform/middleware layer wherein there is value orchestration between raw models and applications. These are services such as vector databases (Pinecone, Weaviate), model deployment platforms, prompt management tools, and API gateways. This layer makes AI models practical to use at scale. Finally, we have an application layer where end-user applications are built on top of the underlying infrastructure layers. This can range from coding assistants (GitHub Copilot) to content creation tools, customer service chatbots, and vertical-specific solutions for healthcare, legal, finance, etc.
Prior generations of technological innovations offer parallels to the industry structure of AI, and point to the sources of value capture and value creation in the AI world. In particular, the AI value chain has some notable parallels to early generation of cloud computing providers.
The AI Value Chain and Early Gen Cloud Computing Models
Pre-cloud, infrastructure vendors sold enterprise editions of open source software and application developers had to build their own infrastructure to reuse the knowledge in their contexts. The cloud enabled different types of consumption-based pricing models, i.e., “pay for outcomes” and “pay for just the infrastructure used”.
Cloud computing enabled a shift to consumption-based models where customers pay based on compute hours, storage consumed, or network bandwidth used. For example, AWS EC2 charges per second/minute of compute. With Generative AI customers pay based on tokens processed (input + output), API calls, or model inference hours. Some providers also offer “reserved capacity” (like OpenAI’s enterprise deals or Anthropic’s model access plans).
Cloud computing also saw the emergence of Platform-as-a-Service Model where the AWS, Azure, GCP monetization strategy was to offer scalable infrastructure as a service, abstracting away hardware and operations. Generative AI providers such as OpenAI, Anthropic, and Cohere provide LLM-as-a-service or model APIs, abstracting away model training and maintenance. This mirrors cloud’s “rent instead of build” philosophy.
Both paradigms offered tiered pricing and freemium models that make it economical for application developers to rapidly implement new ideas. Cloud computing models offered tiers such as free credits, developer plans, enterprise support, volume discounts. For GenAI, there can be similar tiered pricing wherein providers can offer free access with limited usage (such as ChatGPT free tier), and enterprise API discounts for high-volume customers.
While these foundational infrastructure services are easy to adopt, both value chains have ecosystem lock-ins. The switching costs with cloud providers are high due to proprietary tools, and data integration challenges. Gen AI models create lock-in due to the complexities of fine-tuning and integration into workflows. Customers often stay with a provider once they adapt prompts, pipelines, or safety guardrails.
For the infrastructure providers, both models have economies of scale and scope. With cloud computing, margins improve with scale because fixed infrastructure costs are spread across more usage. Similarly, for generative AI models, the cost of model training is extremely expensive upfront, but inference becomes cheaper as usage scales. Providers rely on high utilization of GPUs/TPUs to make economics work, just like cloud datacenter utilization. Both models seem to offer several models of delivery such as on-premise AI, behind corporate firewall, hybrid cloud-based models etc.
OpenAI seems to understand this parallel as it is moving to build its own enterprise stack. Whereas cloud vendors offered private clouds to address security concerns of enterprises, AI vendors could adopt similar approaches tailored to enterprise needs.
While most of the AI hype is centered on the providers of cutting edge AI capabilities, startups and enterprises seeking to use AI for value creation need to understand where margins exist, and where strategic leverage points lie in the AI value chain.