Updated 21 September 2026. Hetzner raised its cloud prices on 15 June 2026, and the shared CX plans this guide first recommended are not in its current price list. Every VPS price below was re-checked on the providers' own pages on 21 September 2026. Hetzner lists prices excluding VAT, with the primary IPv4 address billed separately at €0.50. netcup lists prices including 19% German VAT, with IPv4 included.
Vercel AI SDK is great. It is also locked to a specific philosophy of how AI apps should be built, and it strongly incentivizes you to host on Vercel where you will quickly discover that real AI workloads run into Vercel's pricing in unhappy ways. A $20 Vercel bill becomes a $200 bill the first month an agent loop runs hot.
If you are looking at alternatives, you are not alone. The AI framework ecosystem in 2026 has more credible options than at any point since LangChain shipped. Some are better for agents. Some are better for type safety. Some give you the AI SDK ergonomics without the Vercel gravity well. And the hosting decision for each is different. This is a working engineer's tour of the alternatives, where they shine, and where to actually deploy them without burning money.
The good parts: streaming responses out of the box, clean React hooks (useChat, useCompletion), a unified provider abstraction so swapping OpenAI for Anthropic is a one-line change. Tool calling is well-designed. The TypeScript types are tight.
The friction: the SDK assumes Next.js as the primary host. Server actions and route handlers are the recommended patterns. The moment you want to host the backend somewhere other than Vercel, you start fighting the framework. Vercel Functions also enforce a hard maximum duration. With Fluid compute, which is on by default, Hobby functions stop at 5 minutes. Pro functions can be configured up to 800 seconds, or 30 minutes with the extended-duration beta. That is plenty for a chat reply and tight for an agent that works through a long task list. Long-running agent loops need workarounds (background jobs, queue-and-poll patterns) that erase the simplicity argument.
Pricing is the second issue. Vercel charges by function invocation, GB-hours, and bandwidth. AI streaming chews bandwidth. Agent loops chew invocations. The pricing model that works beautifully for a marketing site quickly stops working for a chatbot with a thousand daily users.
Mastra is the most direct AI SDK competitor that emerged in 2024-2025. Built by the Gatsby founders. Pure TypeScript, designed for agents from day one. Workflows are first-class. Memory and RAG are built in. Deploy targets include Cloudflare Workers, Vercel, AWS Lambda, and any Node-compatible host.
What you get:
Best deploy target for Mastra: Cloudflare Workers if your agents are short-lived, a small VPS if they run long or need persistent state. A netcup VPS 500 G12 (2 vCores, 4 GB RAM) at €5.91 per month runs production Mastra agents fine.
LangChain is the obvious one. In 2026 the conversation has shifted to LangGraph as the recommended LangChain entry point for new projects, especially for agents. LangGraph treats agent runs as state machines, which makes complex flows tractable and debuggable.
The frame here is different from the AI SDK. LangChain is a Python-first framework with a TypeScript port that is always one version behind. If your team is Python-native, LangChain or LangGraph plus LangSmith for observability is a strong stack. If your team is TypeScript-native, Mastra is usually a better fit and Vercel AI SDK is competitive too.
Deploy target: LangGraph has its own Cloud hosting product (LangGraph Cloud), or you self-host the runtime on any Python-capable VPS. For self-hosting, a 4 GB VPS handles most workloads. The netcup VPS 500 G12 (4 GB) at €5.91 or a DigitalOcean Basic droplet with 2 GB at $12 are both fine starting points.
The strongest framework if your problem is "I have a lot of documents and I need agents that answer questions over them." LlamaIndex started as a RAG-first toolkit and grew into a general agent framework while keeping that document-heavy DNA.
If you are building a knowledge worker assistant, internal-docs Q&A, customer support over a help center, LlamaIndex usually beats the alternatives because the ingestion, chunking, and retrieval primitives are more thoughtful. Deploy target: any Python host. LlamaIndex itself is light. The expensive part is the vector store, which dictates your real hosting choice. If you use Postgres pgvector, your Postgres instance is the bottleneck. If you use Pinecone or Qdrant, your bill is in the vector DB.
The newest credible entrant. Built by the Pydantic team. Heavy emphasis on type safety, model-agnostic abstraction, and structured outputs. If you already use Pydantic for validation (and most Python AI teams do), PydanticAI fits naturally.
What is notable: PydanticAI was designed after the lessons from LangChain. It is smaller, less opinionated, easier to read. The tradeoff is that it provides fewer batteries-included features. You bring your own memory, your own retrieval, your own observability. Many teams prefer this.
Deploy target: any Python host. Small footprint. A 2 GB VPS at $6/month or less is more than enough for most projects. Our VPS price comparison lists the current options.
OpenAI released their own agents framework in 2025. It is well-designed if you are exclusively on OpenAI models. It supports handoffs between agents, structured tool use, and tracing through OpenAI's observability dashboard. The tradeoff is obvious: deeper OpenAI lock-in.
If you are committed to GPT models long term, the Agents SDK is the path of least resistance. If you want flexibility to swap to Claude, Gemini, DeepSeek, or self-hosted models, look elsewhere. Worth saying out loud: a lot of teams who started with a framework end up dropping it and calling the model SDK directly. The Anthropic SDK and OpenAI SDK are both small, well-documented, and stable. If your app is a single agent loop with a handful of tools, framework overhead can hurt more than help.
The Anthropic SDK in particular handles tool use, streaming, and parallel calls with a clean API. Combined with Hono or Elysia for the HTTP layer, you can ship production AI features with fewer dependencies than the AI SDK pulls in.
Deploy target: anywhere. Hono runs on Bun, Node, Cloudflare Workers, Deno, AWS Lambda. The same code deploys to all of them. We use this exact pattern at HostingSift.
| Framework | Language | Best for | Hosting fit | Lock-in |
|---|---|---|---|---|
| Vercel AI SDK | TypeScript | Next.js chat UIs | Vercel (others possible but painful) | High to Vercel patterns |
| Mastra | TypeScript | Agents with workflows | Anywhere Node runs | Low |
| LangGraph | Python (TS port) | Complex agent graphs | LangGraph Cloud or self-host | Medium |
| LlamaIndex | Python | RAG-heavy applications | Any Python host | Low |
| PydanticAI | Python | Type-safe minimalist agents | Any Python host | Low |
| OpenAI Agents SDK | Python and TS | OpenAI-committed teams | Anywhere | High to OpenAI |
| Direct SDK + Hono | TypeScript | Custom architecture | Anywhere | Minimal |
The VPS picks in this section changed after Hetzner's June 2026 repricing. Here is what the plans mentioned below cost on 21 September 2026, as each provider lists them.
| Plan | CPU | RAM | Price per month | Billing |
|---|---|---|---|---|
| [netcup](https://hostingsift.com/hosting/netcup) VPS 500 G12 | 2 vCores | 4 GB | €5.91 incl. VAT | 12-month term; €6.81 without a term |
| netcup RS 1000 G12 | 4 dedicated cores | 8 GB | €12.79 incl. VAT | 12-month minimum | | netcup VPS 2000 G12 | 8 vCores | 16 GB | €19.25 incl. VAT | 12-month term | | netcup RS 2000 G12 | 8 dedicated cores | 16 GB | €21.43 incl. VAT | 12-month minimum | | Hetzner CPX22 | 2 vCPUs | 4 GB | €19.49 excl. VAT | Hourly, no minimum | | Hetzner CCX13 | 2 dedicated vCPUs | 8 GB | €42.99 excl. VAT | Hourly, no minimum | | Hetzner CPX42 | 8 vCPUs | 16 GB | €69.49 excl. VAT | Hourly, no minimum | | Hetzner CCX23 | 4 dedicated vCPUs | 16 GB | €85.99 excl. VAT | Hourly, no minimum | | Vultr Cloud Compute | 1 vCPU | 1 GB | $6 | Hourly, no minimum | | DigitalOcean Basic | 1 vCPU | 1 GB | $6 | Hourly, no minimum |
The trade-off is commitment. netcup is far cheaper per core and per GB, but its regular prices assume a 12-month contract. Hetzner, Vultr and DigitalOcean bill by the hour, which matters when you are still experimenting and may tear the server down next week.
Cloudflare Workers is the right answer for most teams. Long-running streams work. Global edge means low latency. The free tier covers a surprising amount of traffic, and the $5 paid tier covers an unsurprising amount. Mastra, PydanticAI (via Pyodide-Workers patterns), and Hono with the Anthropic SDK all deploy here cleanly.
If you need persistent server-side state across requests, drop down to a VPS. netcup VPS 500 G12 at €5.91, Vultr Cloud Compute at $6, or DigitalOcean Basic at $6 all work. Pair with Caddy for TLS and you have a streaming-capable backend for about 6 EUR or 6 USD per month. Stay away from anything with a hard duration cap you cannot raise. Vercel Hobby stops at 5 minutes. Vercel Pro reaches 800 seconds, or 30 minutes in beta, which covers many agent runs but not open-ended ones. Most Lambda configurations are out unless you split into step functions.
What works: PaaS platforms that run long-lived processes ([Railway](https://hostingsift.com/hosting/railway), Render), or a self-hosted VPS. Railway lets you run a long-lived process for $5-10/month. For dedicated cores, the netcup RS 1000 G12 gives you 4 dedicated AMD EPYC 9645 cores and 8 GB of DDR5 for €12.79 per month, on a 12-month minimum term. If you need hourly billing instead, Hetzner's CCX13 with 2 dedicated vCPUs and 8 GB costs €42.99 per month.
For background agent jobs that take minutes or hours, queue them. BullMQ on Redis, a SQS-equivalent, or a Postgres-backed job runner like Inngest. Trigger from your API, process on a worker process. The worker can live on the same VPS.
Your hosting bill is dominated by the vector DB, not the application server. Three reasonable approaches:
Most apps with under 5 million chunks should self-host pgvector and move on. Above that, managed services start paying for themselves.
If you are doing tens of thousands of AI calls per day for paying customers, neither Vercel nor Cloudflare Workers makes sense long-term. The pricing models are designed for variable traffic. Predictable high traffic is cheaper on dedicated resources. A netcup RS 2000 G12 (8 dedicated AMD EPYC cores, 16 GB RAM) at €21.43 per month runs a production AI backend with significant headroom. The hourly-billed alternative is Hetzner's CCX23 with 4 dedicated vCPUs and 16 GB at €85.99. Add Cloudflare in front for caching and rate limiting (free tier). On the netcup box you spend under 30 EUR per month for what would be $300+ on serverless.
The catch is ops. You patch the VPS. You configure Caddy. You set up Postgres backups. Two days of upfront work, then it runs.
The 2025-2026 development worth mentioning. Cloudflare Workers AI, Vercel's AI Gateway, and Bun's local model have made it realistic to run small models (Llama 3.2 3B, Phi-4, Qwen 2.5 7B) at the edge or directly on a Bun runtime. Latency drops dramatically. Privacy improves.
For genuinely small workloads, this can collapse your inference bill to near zero. Worth experimenting with if your use case is forgiving (autocomplete, classification, simple summarization). For deeper exploration, see our self-hosted LLM guide. Should you actually leave Vercel?
If you ship a Next.js app with light AI features (a chatbot, occasional summarization, some embedding for search), Vercel is fine. Use the AI SDK. Pay the bill. Move on with your life.
If your AI use is intense, persistent, or pricing-sensitive, plan the migration now. The good news is that the AI SDK is portable. You can run it on Cloudflare Workers, on Hono on a VPS, on AWS Lambda. The Next.js part is the harder migration, not the AI SDK itself. Tools like Coolify or Dokploy can deploy Next.js on a netcup or Hetzner VPS with Vercel-like ergonomics. We wrote about [self-hosted Vercel alternatives](https://hostingsift.com/blog/self-hosted-vercel-alternatives-nextjs-2026) in detail.
The pattern we have seen most often in 2026: keep the marketing site and lightweight pages on Vercel. Move the AI backend and heavy API routes to a Hetzner or Railway-hosted service. Your bandwidth-heavy AI traffic stops counting against Vercel. Your bill drops 70-90%. The split-stack adds some complexity but pays for itself in month one.
If you are a TypeScript team building an agent product: Mastra on Cloudflare Workers or a small European VPS from netcup or Hetzner. Modern, flexible, no lock-in.
If you are a Python team building agents: LangGraph if you need complex flows and tooling, PydanticAI if you want minimal and type-safe. Either runs on any cheap VPS.
If you are RAG-heavy: LlamaIndex on Python with pgvector on the same VPS. Often the entire stack runs on one VPS under 10 EUR, such as a netcup VPS 500 G12 at €5.91.
If you are OpenAI-committed and shipping fast: OpenAI Agents SDK, host anywhere.
If your needs are simple and you hate framework churn: Anthropic or OpenAI SDK directly, Hono for the HTTP layer, deploy to whatever you already pay for.
If you are starting today, on Next.js, with light AI needs: Vercel AI SDK is still the path of least resistance. Just keep an eye on the bill.
The choice of framework matters less than the discipline to actually measure what your AI features cost per active user. That measurement is what will tell you when to migrate, where to migrate to, and whether the migration is worth your team's time. Build the dashboard before you need it.
Originally published at hostingsift.com