A troubling pattern emerges

Amazon Web Services is confronting what internal documents describe as a fundamental shift in how artificial intelligence startups are spending their technology budgets, potentially threatening the cloud computing giant’s long-held dominance in the startup ecosystem. The documents, obtained and reviewed, reveal that AI startups are increasingly delaying their adoption of AWS and instead directing their early spending toward AI models, inference capabilities and specialized developer tools.

For years, AWS represented one of the first and often largest expenses for new tech companies. Startups embraced the affordable, scalable computing services as an attractive alternative to building and maintaining their own data centers. However, the generative AI revolution has introduced what industry insiders call a Cloud 2.0 stack, featuring specialized hardware, software and tools that are capturing startup budgets before traditional cloud services enter the picture.

The internal documents, marked as confidential and dated from March and July, were created by employees on the AWS startup business team. Among the authors was someone who works directly with startups from Y Combinator, the influential startup accelerator. AWS executives responsible for managing startup and venture capital relationships also reviewed these materials.

Founders are waiting longer

According to the documents, startup founders are explicitly telling AWS representatives that they plan to adopt the platform at a later stage in their development. This represents a dramatic departure from previous patterns where AWS services formed the foundation of a young company’s technical infrastructure from day one.

The shift appears driven by how AI startups are structuring their initial technology purchases. Many now make their first buys from AI model providers like OpenAI and Anthropic, followed by newer developer platforms such as Vercel. This means founders are postponing decisions about AWS services until they require more advanced capabilities like enhanced compliance features and security measures.

Data from Y Combinator’s 2024 cohort illustrated the trend. Among those startups, 59% reported using more than three AWS services, representing a decline of over four percentage points compared to 2022. Meanwhile, 88% were using OpenAI’s models and 72% were using Anthropic’s offerings. Only 4.3% indicated they were using Bedrock, the AWS developer tool that provides access to various AI models.

The stickiness problem

AWS employees expressed concern in the July document about new categories of AI spending that can represent the majority of a startup’s cloud consumption. Unlike traditional services, these newer offerings are far less sticky, meaning customers can rapidly shift between an expanding roster of providers without significant friction or switching costs.

The documents identified three specific AI cloud services capturing early spending from startups: GPU training and fine tuning, GPU inference, and AI-as-a-service platforms. GPUs are specialized chips that power generative AI, contrasting with traditional cloud services that run on CPUs. Training and fine tuning help build and improve AI models, while inference enables those models to run and generate responses.

Earlier this year, AWS compiled an assessment of the top 1,000 AI startups for Amazon CEO Andy Jassy, attempting to determine which ones were building primarily on the AWS platform. The exercise raised difficult questions about what it means to be fully committed to AWS in the current AI era, especially as cloud spending rapidly expands beyond traditional categories like compute, storage, databases and analytics.

A telling example

The documents cited AI coding startup Cursor as an illustration of these challenges. Despite being considered fully committed to the AWS platform, Cursor’s spending on traditional infrastructure represented less than 10% of what the company spends on newer AI categories. The majority of Cursor’s budget goes toward API calls to external AI models and neocloud providers that primarily sell access to GPU servers, including companies like CoreWeave, Crusoe, Lambda Labs and Nebius.

Market share pressures mount

The concerns outlined in these documents align with recent market performance data. In the second quarter, Google Cloud and Microsoft’s Azure each grew more than 30% year-over-year, while AWS managed 18% growth. Neocloud revenue has surged by more than 200% in the past year, though from a much smaller starting point.

Analysis from CB Insights shows AWS capturing 30% of the leading 1,100 AI startups market between January 2024 and September 2025, trailing Google Cloud’s 38% but ahead of Microsoft Azure’s 7%. This marks a decline from the prior two-year period when AWS claimed 33% of that market compared to Google Cloud’s 34% and Microsoft’s 9%. Roughly 25% of startups reported using more than two cloud providers.

Amazon maintains significant advantages in this evolving competition, including a close partnership with Anthropic backed by billions in investment. Morgan Stanley estimated in July that Amazon could generate $5.6 billion in revenue by 2027 from Anthropic’s use of AWS cloud services. However, analysts note that AWS remains a step behind Microsoft and Google in driving GPU demand and selling additional services to AI customers.

An AWS spokesperson pushed back on concerns, stating that the company remains the top choice for startups and continues seeing growth in adoption as startups seek depth and breadth of services.

Story credit: BUSINESS INSIDER