Serverless is a Firework!
July 2026
Welcome to July 2026! In the USA, we celebrated our 250th Anniversary! I know that may seem a bit quaint as we have countries and societies that are way older but even still, it’s a pretty cool fact.
I did not partake in the fireworks as my dog just had surgery and she is also terrified of fireworks so I figured I would stay home and take care of her. July was also surprisingly hot in San Francisco. I know it sound like I am complaining considering most of the country was seeing 90-100 degree (F) weather but we don’t have air conditioning in San Francisco.
It’s alright though. You didn’t come here to listen to my bad comedy about summer heat. You came here to learn about serverless.
Something I want to do a little different in this article and if it works, we’ll try doing it in the future. I am creating a section called “The Signal”. Here I will take what I consider the most interesting story, dissect it, and give you my personal interpretation.
The Signal
The story I want to feature here isn’t strictly serverless but it will greatly impact serverless compute. MCP just had a major upgrade that will impact how we use AI Agents.
Now we haven’t really talked about the MCP protocol too much in this newsletter but in short, it’s a protocol that simplifies the integration of generative AI applications. It is very similar (conceptually) to making an API call but while APIs are very generic, MCP is specific to AI applications. It is like the “USB port of AI Applications”.
Okay, so in the latest update, MCP has moved to a fully stateless architecture. This is very important as “v1” of MCP required your application to maintain a persistent session with a specific server. This requirement is not useful for serverless architecture as serverless is stateless by nature and MCP was effectively stateful.
This was also bad for cloud architecture broadly as nodes are spun up and spun down all the time. This fact was one of the key bottlenecks to people putting AI Agents into production since they couldn’t reliably guarantee performance without finding a hacky way to do so.
This has addressed the option but will also make it easier to run MCP clients and servers on serverless architectures. This is a huge step for serverless AI as it will unblock AI Engineers who are trying to find reliable platforms to deploy not just MCP clients but MCP servers for their agents to leverage.
Serverless Industry
Startups
Product Launches
Runtimes
Data
AI
MangoBoost Redefines the AI Inference Race Beyond Existing Industry Benchmarks
From dataset to deployment: A practical guide to serverless fine-tuning
Serverless AI Model Fine-Tuning with Crusoe: Revolutionizing Development
Learning
Fine-tune NVIDIA Nemotron 3 models with Amazon SageMaker AI serverless model customization
Serverless Computing Architecture: Patterns, Tradeoffs & Decision Guide
Closing Thoughts
Again, a lot of serverless AI. I really do think that AI is going to be the major tipping point for serverless adoption. I know I say that all the time but every month, I see more and more companies adopt serverless architectures specifically for generative AI use cases.
Crusoe just launched a way to fine-tune models on serverless runtimes. Databricks has launched a serverless LTAP. Fly.io is growing and raising money to improve serverless generative AI. Capital One, a major global financial firm, has adopted a serverless first approach.
I think we will see more serverless adoption and it will go way beyond functions.
—Photo courtesy Trang Nè on Pexels—
Jason “Jay” Smith (The Cloud Alchemist) writes about serverless beyond FaaS. Find me at @ https://jasonsmith.io.