Six ways to manage the hidden AI costs

Organisations are increasingly turning to generative artificial intelligence (generative AI) to help turn ideas into reality. These generative AI tools rely on foundation models (FMs) and can be applied to a wide range of use cases including language, coding, genomics and much more.

Creating these models is resource-intensive work and requires specific skills. You can get around these requirements and skip directly to building and scaling generative AI applications, by using Foundation Model as a Service (FMaaS).

What is Foundation Model as a Service?

FMaaS provide API-based access to frontier or enterprise‑tuned models, together with security controls and optional fine‑tuning. This allows you to integrate generative AI capabilities into your products or service without the need to manage the underlying model infrastructure.

Think of it as the generative AI analogue of Software as a Service (SaaS).

Examples include Microsoft Azure OpenAI Service, Amazon Bedrock, Google Vertex AI and Anthropic Claude API.

In my role as Co-CEO and as a consultant, I understand the importance of using these powerful services while balancing cost optimisation, efficiency and value growth.

The six key strategies to keep FMaaS costs under control

Effective cost management becomes crucial when adopting FMaaS at scale. Runway costs can quickly add up and result on a shock bill.

Here are six key strategies to ensure you get the most value from your investment:

1. Provisioned throughput: A fixed-cost, fixed-term subscription that reserves resources and ensure specific throughput for generative AI services.
   • Cost benefit: Up to 70% discount compared to on-demand usage.
   • Best practice: Match provisioned throughput to stable workloads and monitor usage to avoid over-provisioning.
2. Batch inference: Making predictions or running inference on a large set of data points, instead of processing each data point individually. 
   • Cost benefit: Around 50% cheaper than real-time calls, ideal for large asynchronous jobs.
   • Best practice: Use batch inference for high-volume, non-real-time tasks to minimise idle time and reduce costs.
3. Token / Prompt Caching: Use previously processed prompts to reduce latency and computational costs. 
   • Cost benefit: Up to 90% discount on cached input tokens and latency improvement up to 80‑85%. In Microsoft Azure OpenAI Service, if the call runs under a Provisioned Throughput Unit (PTU), the cached input tokens can receive up to a 100 % discount.
   • Best practice: Cache repeated context (system prompts, RAG prefixes) across calls to cut both cost and response time.
4. Real-time API usage: Interact with models through an API that provides immediate responses with minimal latency.
   • Cost benefit: Pay-as-you-go model allows for flexibility.
   • Best practice: Right-size models and optimise prompts to ensure efficient usage.
5. Model selection: Select the most suitable model for a specific task, based on performance metrics, complexity and generalisation ability.
   • Cost benefit: Balances cost and performance.
   • Best practice: Choose models that are fit-for-purpose to avoid unnecessary expenses.
6. Cross-team centralisation: Integrate AI tools and data across multiple teams to streamline collaboration, improve efficiency and ensure consistent decision-making.
   • Cost benefit: Pooled savings and scale.
   • Best practice: Enable shared FMaaS platforms and use tagging to track and manage costs across teams.

FMaaS platforms are perfect candidates for FinOps

FinOps, the operational framework and practices to maximise the value of cloud and technology, can be used side-by-side with FMaaS platforms.

FMaaS platforms align well with the FinOps principles of financial accountability, cost optimisation and real-time decision making in cloud environments.

Adopting an "optimisation, efficiency, growth" mindset can help you keep the cost structure under control. In turn this can ensure that your AI initiatives are both innovative and cost-effective.