Byte the Bullet: Why Failing to Monetize GenAI is More Expensive Than You Think
February 10, 2025 BlogGenAI, Meet ROI Before we can interrogate some meatier questions around how to be deliberate in seeking ROI…
GenAI, Meet ROI
Before we can interrogate some meatier questions around how to be deliberate in seeking ROI from investments in GenAI, let’s revisit some key assumptions about GenAI and LLMs:
Generative AI is the umbrella category of technology which mimics human cognitive ability (the ability to process sensory information and acquire knowledge). Large Language Models (LLMs) is a category of tools under Generative AI that focuses on processing and generating human language.
LLMs use math to find patterns in groups of words to predict what word(s) will come next. Importantly, these math models take input iteratively and learn from their mistakes.
All this math requires a lot of computational power. Computational power is expensive and these costs are a critical dependency in scaling AI in higher education (and in general).
Recent news from China’s emerging company DeepSeek challenges AI’s reliance on computational power with an engineering breakthrough and an open-source model. While this may ease costs, concerns remain over DeepSeek’s biases (e.g., Tiananmen Square censorship) and data privacy (potential surveillance like TikTok, Italy has banned it). While a downward trend in costs appears inevitable, solution providers in the GenAi space have a series of complex trade- offs to make in bringing their offerings to market.
Why has investing in GenAI been so expensive?
Creating generative AI applications has until recent weeks been assumed to be a costly endeavor, it requires investment in two categories that are incremental to typical costs in product development. Once investments are understood, a key decision needs to be made regarding whether an organization wants to own or rent their LLM.
Model
The model (e.g., GPT, Gemini, Llama) is a LLM which has been trained on vast amounts of data – public and private. Developers (e.g., OpenAI, Google, Meta) of these models have invested in researchers and data to optimize their models.
Training LLMs like OpenAI’s GPT-4 requires running trillions of mathematical operations for weeks or months. This requires access to expensive high-performance hardware, also referred to as infrastructure.
Infrastructure
With the success of large-scale AI, there is increasing demand for state-of-the-art chips that can support a combination of processing power(logic), memory, and interconnect (what moves data between logic and memory components) needed to drive performance.
Fun fact: You don’t need to be a computer expert to appreciate the change in processing power – the numbers speak for themselves. The computer that navigated the Apollo space missions to the moon had approximately 80 bytes of memory. The H100 chip (most popular for generative AI use cases) has 80 billion bytes (80 GB).
The AI hardware supply chain is highly consolidated with a few major companies controlling key components and keeping prices high. Nvidia dominates the chip design market, with buyers like Meta and xAI purchasing over 100,000 H100 GPU chips each. Chip production is limited to a few manufacturers, such as Taiwan Semiconductor Manufacturing. Large AI models are trained on clusters of these chips, known as AI supercomputers, which are hosted in data centers. These data centers require massive amounts of energy and advanced cooling systems to operate effectively – all of this is expensive to operate.
DeepSeek lowers AI costs by using an efficient design called Multi-Head Latent Attention (MLA) to make its models run with less computational steps. This reduces the computing power needed, allowing them to run with far fewer specialized chips (not a suprise given DeepSeek is operating in China where the supply of specialized chips is far lower than in the rest of the world) – in combination, DeepSeek can achieve some degree of lower operating costs.
No matter the model (and requisite cost, privacy guardrails, and/or tolerance for embedded bias), each vendor or institution seeking to leverage a LLM needs to decide how much of the infrastructure they want to rent or own: in other words, is having a proprietary LLM a priority or can the LLM be licensed or open (rent)?
What are ways to manage costs of LLM use?
Due to the amount of processing power required to support LLMs, solution providers need to invest in the infrastructure layer and model layer of their technology stacks in order to innovate with GenAI. Many vendors struggle to see a return on these investments in a market which has evolved to expect generative AI features or modules to be available to end-users (both faculty and students). In our conversations with the edtech ecosystem, we’ve seen many of the following strategies deployed to manage the development costs of generative AI features, this is a key piece of the ROI puzzle:
Control and optimize investments
The broad goal here is to right-size and monitor usage of calls to the model. Vendors we’ve spoken to have found considerable efficiency gains by optimizing prompts and ensuring valid data features or products. The table below summarizes some strategies.
| Deployment Timeline | Cost Optimization Strategy | Quick Description |
|---|---|---|
| Before Deployment | Preprocess input data | Clean input data to avoid unnecessary compute |
| Select appropriate model size | Select the smallest LLM that meets your use case (e.g., GPT-4 vs. GPT-3.5 or SageMaker models). | |
| Evaluate pricing tiers | Compare pricing tiers (e.g., OpenAI’s GPT models or AWS’s endpoint pricing - cost of deploying machine learning models as APIs using AWS services like SageMaker or API Gateway; pricing based on compute instance, uptime, scaling and other factors). | |
| Evaluate security, compliance, and reliability needs | GDPR compliance and SLAs with regards to uptime commitments(higher availability comes at a premium) can cost more | |
| Budget Allocation and Monitoring | Set budgets and cost alerts for API usage (OpenAI) or compute resources (AWS). | |
| During Deployment | Batch Processing for Inference | Send inputs in batches instead of single requests to reduce cost per prediction. |
| Optimize token use | Shorten prompts or trim unnecessary input. | |
| Save or cache commonly requested prompt results | Minimize the need to re-ping the model. | |
| After Deployment | Monitor Usage and Costs | Use OpenAI’s usage dashboard or AWS Cost Explorer to identify inefficiencies |
| Use Tiered Pricing Discounts | Utilize OpenAI’s token-based discounts or AWS’s tiered rates. |
Pass through costs where possible
Monetize services related to implementation (e.g., seperate fees for implementation outside of license fees such as outlined by Slate and many other enterprise solutions) and maintenance. AI integration, deployment, updates, retraining, and support all require investment for solution providers and precedent exists to separately delineate these costs from recurring costs to the customer.
The key question is: how much of the cost should be passed on to the customer? Pricing strategy in edtech is hard. The decision-maker is often disconnected from the end-user and as a result, the line from price to value is not straight. In addition, in education, acquiring customers is very expensive and a rational pricing strategy may suggest offering enhanced value at no additional charge in service of improving customer retention.
What are ways to generate revenue from GenAI investments?
Selecting the right monetization models for GenAI features is crucial to ensure a return on investment (ROI). In our conversations with solution providers, we’ve seen a variety of strategies, but also significant inaction on monetization. This hesitation to capitalize on AI innovation is costly, especially given the substantial investments required to develop GenAI. Without clear monetization strategies, these capabilities can quickly become an unsustainable drain on company resources.
Clearly articulate value for users and buyers
A persistent challenge in ed tech soluton pricing is that users who recieve benefits from products are often not the buyers. Being mindful of this general dynamic is the first step in developing a revenue strategy related to AI investments as the immediate pricing implication is that solution providers need to articulae value to BOTH users and buyers. Beyond this, solution providers need to clearly articulate the alignment between user benefits (e.g., a higher degree of personalized advice to students) and institutional goals (e.g., better student outcomes) to enable users (and their champions, such as faculty) to advocate for investment in products or features powered by GenAI. A clear articulation of how value to end-users drives institutional goals enables the ability to select a licensing model that appropriately addresses ROI concerns.
Select the most appropriate licensing model
AI licensing models in higher education depend on an institution’s needs, budget, and use cases. These models prioritize affordability, scalability, and flexibility to support faculty, administrators, researchers, and students. The most common models are summarized below with their key advantages and challenges for developers:
| Licensing Model | Exemplar Use Cases | Developer Advantages | Developer Challenges |
|---|---|---|---|
| Subscription-Based license: License grants access to an entire institution, dollars based on the size of the institution or the number of users | AI-powered components of LMS or other key source system | Straightforward access management; steady, recurring revenue | Can be difficult to scale as usage grows |
| Perpetual licensing: Onetime upfront payment | STEM software (e.g. MatLab) | No recurring billing, simpler long-term financial planning | High upfront cost may deter budget- constrained institutions from participating |
| Tiered licensing: Pricing based on the number of users, features, or usage volumes | Adaptive learning systems, learning analytics, and administrative tools | Allows for differentiated pricing based on user needs and matching of resource use to customer needs | Can be complex to manage across different institutional segments |
| Research partnerships: Free or heavily subsidized for research purposes, often via grants or sponsored access | Sponsored research, innovation centers | Low-to-no-cost access and promotes research with select institutions | Limited use, competitive space to get grants |
| Freemium: Free or low costs with paid upgrades | Small projects across institution, AI education experiments | Low-to-no-cost access and builds community and user base quickly | May need to support free users with minimal resources and erode end-user experience |
Addressing market demand for premium GenAI features
Charging for premium features which are most costly to implement establishes a clear value proposition that justifies cost. Service level agreements with regards to support, accuracy, uptime, or advanced automation should carry additional costs.
Monetize training and insights as well as content to train LLMs (ethically and legally)
AI knowledge gained from developing products and training users can be shared by creating educational courses, tutorials, or certification programs and host on your own or leverage platforms like Udemy or Coursera.
Thought leadership and intellectual property can be gleaned from special datasets you collect, have analyzed, or own. Companies can use insights from the data to help others. For instance, with the right permissions, you could generate reports on market trends, learning habits, or user behaviors.
Finally, there are content owners who have signed deals with LLM providers to legally use copyrighted materials to train models or other facilitate text and data mining. Key examples include Axel Springer and the Associated Press.
What are the costs of inaction?
Charging for GenAI features is not just about covering development costs; it’s essential for the long-term success of the products and companies who seek to keep pace with innovation. Delaying or avoiding monetization slows down new feature development and creates sustainability problems in the long run.
As articulated above, for the near future, supply-side economics are a headwind to reduction in costs of infrastructure and modeling layers so solution providers cannot simply wait for costs of LLM investment to fall. While DeepSeek’s innovations may be a proof point that costs can fall, it may still take time to see real reductions in development costs for applications in US-based education. We propose that all vendors which have spent time and money on GenAI innovation without monetizing pause to consider the costs of inaction on this front.
We advise companies and institutions throughout the higher education sector on pricing strategy and see GenAI as a key area of investment for service and solution providers supporting better student outcomes and institutional efficiency. If the past years have shown us anything, it’s that research and innovation in ed tech will continue to push new limits and we look forward to hearing how these forces are impacting your work.