The economics of building an AI-powered product in 2026 look nothing like the pitch decks suggest. Founders talk about foundation models, fine-tuning pipelines, and inference at scale. Investors talk about defensible moats and platform effects. What neither side talks about publicly — but both think about constantly — is the compute bill.
For the current generation of AI startups, cloud credits have become the de facto currency of early-stage infrastructure. Accelerator programmes from Y Combinator, Microsoft for Startups, Google for Startups, and AWS Activate routinely distribute credit packages ranging from $50,000 to $350,000 across providers including Azure, Google Cloud, AWS, and increasingly the API platforms of model providers like OpenAI, Anthropic, and Cohere. These credits are meant to give startups runway to build, train, and deploy without upfront infrastructure costs. In theory, the system works. In practice, it is creating a set of problems that few people in the ecosystem are willing to discuss openly.
The credit allocation mismatch
The core issue is straightforward: the credits startups receive rarely match the infrastructure they actually need.
A startup accepted into a major accelerator programme might receive $150,000 in Google Cloud credits and $100,000 in Azure OpenAI credits as part of a standard partner package. But if the founding team has built its stack on AWS Bedrock with Anthropic’s Claude as the primary model, a significant portion of those credits sits unused. The startup cannot transfer them. It cannot combine them. In most cases, it cannot even convert Google Cloud credits into Google Cloud AI API credits for a different service tier within the same provider.
The result is that AI startups across the ecosystem are sitting on tens of thousands of dollars in credits they cannot use, while simultaneously paying full price for the compute they actually need. According to estimates from secondary market participants, more than $2 billion in AI credits expire unused every year across the major cloud providers. For startups operating on eighteen-month runways, that waste is not an abstraction — it is the difference between an additional engineer and an earlier funding round.
What the credits actually cost
The sticker price of AI compute has declined on a per-token and per-GPU-hour basis over the past two years. But the effective cost for startups has not fallen proportionally, for several reasons.
First, the models themselves have become more capable but also more expensive to run at production scale. A startup using GPT-4o or Claude Sonnet for real-time inference in a customer-facing product is spending meaningfully more per request than one that used GPT-3.5 two years ago. The quality improvement justifies the cost in most cases, but it compresses margins and accelerates credit burn.
Second, fine-tuning and evaluation pipelines consume credits at rates that are difficult to predict during the experimentation phase. A team iterating on a retrieval-augmented generation architecture might burn through $20,000 in credits over a two-week sprint without producing a production-ready system. That is an expected part of the development process, but it means that a $100,000 credit package provides less effective runway than founders anticipate when they receive it.
Third, multi-model architectures are becoming standard practice. Startups increasingly use different models for different tasks within the same product — a smaller model for classification, a larger model for generation, a specialised model for code or structured output. Each model may run on a different provider’s infrastructure, multiplying the number of credit accounts that need to be funded and managed.
The secondary market response
The mismatch between credit allocation and credit consumption has created a growing secondary market. Platforms have emerged where startups and enterprises can trade unused credits — selling what they cannot use and buying what they need at discounts that typically range from twenty to forty percent below list price.
For a startup sitting on $150,000 in Google Cloud credits it cannot use, the ability to sell google cloud credits and recover even sixty to seventy percent of the face value represents a meaningful extension of runway. Conversely, for a team that has committed to Google’s ecosystem but exhausted its initial allocation, the option to buy google cloud credits at a significant discount from the secondary market is an obvious efficiency gain.
The model is not dissimilar to what happened in other enterprise software markets as cloud adoption matured. Unused software licences, reserved instance commitments, and prepaid SaaS contracts all eventually developed secondary markets as buyers and sellers recognised the inefficiency of letting paid-for capacity go to waste.
What makes AI credits slightly different is the velocity at which they lose value. Most credit packages expire within twelve to eighteen months. Unlike a reserved EC2 instance that provides predictable compute capacity over a three-year term, an AI credit package is a depreciating asset from the moment it is issued. Every month that passes without using the credits reduces their effective value — not because the credit amount changes, but because the window for extracting value from them narrows.
Provider dynamics and the credit ecosystem
The major cloud and AI providers have complex incentives around the credit ecosystem. On one hand, credits are a customer acquisition tool. Google, Microsoft, Amazon, and the model API providers distribute credits generously because they want startups to build on their platforms, creating long-term lock-in that generates revenue well beyond the initial credit period. On the other hand, unused credits that expire represent recognised revenue without corresponding infrastructure cost — a favourable outcome from a pure financial perspective.
This creates a tension that the providers have not resolved publicly. The official terms of service for most credit programmes prohibit transfer or resale. But enforcement has been inconsistent, and the practical reality is that the secondary market exists and is growing because it solves a genuine problem for both buyers and sellers.
For streaming and media technology companies — the core audience of this publication — the dynamics are particularly relevant. Media AI workloads including automated transcription, content moderation, video understanding, and recommendation systems are among the most compute-intensive applications in production today. A mid-stage media technology startup might be spending $30,000 to $50,000 per month on inference costs alone. At those consumption rates, a twenty to thirty percent discount on credits through secondary channels is not a minor optimisation — it is a material impact on unit economics.
What this means for the market
The AI credit economy is still in its early stages. As the market matures, several things are likely to happen. Credit terms will become more standardised, making them easier to value and trade. Providers may introduce official transfer mechanisms as they recognise that rigid credit allocation discourages multi-cloud adoption without actually preventing it. And startups will become more sophisticated about credit management as a financial discipline — treating compute credits with the same rigour they apply to cash management and equity dilution.
For now, the practical advice for AI startups is simple. Audit your credit portfolio regularly. Know what you have, when it expires, and whether you are actually going to use it. If the answer is no, explore the secondary market before the expiration date turns your credits into a write-off. The compute bill is already one of the largest line items on an AI startup’s P&L. Letting paid-for credits expire unused is one of the few costs that is entirely avoidable.