AI Stack Library · Start here
Rent or own your AI stack: every layer, with a verdict
To decide whether to rent or own your AI stack, own the parts that get more useful every week and have to move with you when you change providers: a dedicated always-on machine, your memory, rules and skills, and the pipelines you build. Rent the parts that are someone else's full-time job: frontier models for hard reasoning, email delivery, website hosting, payments, password management and offsite backup.
| If you are | Start with |
|---|---|
| A business starting out with AI agents | A dedicated always-on machine you own, then memory and rules as files on it; rent the models |
| A creator who makes video every week | A dedicated always-on machine you own, then your own video pipeline that rents a model per clip |
| A service business with a CRM and a website | A dedicated always-on machine you own; rent the CRM, email delivery and hosting, and own the scripts and approval steps around them |
| A practice with private client data | A dedicated always-on machine you own, with local models for the private jobs; rent frontier models only for work with no private data |
| A team whose automations run all day | A dedicated always-on machine you own, then move the busiest flows off per-task billing and onto it |
The test: does it grow, and does it have to travel?
Two questions sort most of a stack. Does this part get more useful every week you use it? And does it have to come with you when you change model providers? Yes to both, and it is worth owning. If keeping it working is somebody's whole job, it is worth renting. A third question, about volume, settles the rest.
Memory is the clearest yes. In week one the memory folder holds a few notes. By month six it holds your prices, how each client likes to be spoken to, every correction you have made and the reason for it. A new model cannot hand you that. It is the part of the stack you built.
Three questions asked in order. If a part grows more useful every week and must move with you, own it as files you hold. If keeping it working is someone's full-time job, rent it. If it runs many times a day, own it once the volume is steady. Otherwise rent it for now and check again later.
- Does it get more useful every week, and must it move with you when you change providers?Yes Own it, as files you hold
- Is keeping it working someone's full-time job, like inbox delivery, payments or uptime?Yes Rent it
- Does it run many times a day?Yes Own it once the volume is steady; rent it while it is small
- None of the aboveRent it for now, and ask again when your volume changes
Every layer, with a verdict
Here is the whole stack on one table. The sections below give the reasons, and what would change each answer.
A twelve-row table, one row per layer of an AI stack, listing the rented option, the owned option and the verdict. The always-on machine, memory, rules and skills are marked own. Email delivery, CRM, hosting, social scheduling, password manager, offsite backup and payments are marked rent.
| Layer | Rented | Owned | Verdict |
|---|---|---|---|
| Always-on machine | A cloud server or a cloud Mac | A dedicated machine the business owns | Own |
| Memory and context | Memory and projects inside a chat app | Plain files, a notes vault, your own vector store | Own |
| Rules | A vendor's settings page | A rules file every agent reads, and checks that enforce it | Own |
| Skills and workflows | Custom GPTs or a vendor's agent builder | Written procedures and scripts in your own folders | Own |
| Video generation | A monthly plan on a video platform | Your own pipeline that calls a model per clip | Own the pipeline, rent the model |
| Automation | Hosted workflow tools billed per task or credit | A self-hosted workflow tool or scheduled scripts | Rent while runs are few; own once they run all day |
| Frontier models | Pay per call through an API, or a chat plan | Local models | Rent for hard reasoning; own for private and bulk work |
| Vector search | A hosted vector database | Postgres with a vector extension | Own while it fits on one machine |
| Transcription and voice | Hosted speech services | Local speech-to-text and text-to-speech | Own transcription; choose voices by ear |
| Social scheduling | A hosted scheduler | A self-hosted scheduler | Rent |
| Email delivery, CRM, website hosting | Rented services | Possible to self-host | Rent |
| Password manager, offsite backup, payments | Rented services | Possible to self-host | Rent |
Verdicts for a small business. Your volume and your data can move a row; the sections below say how.
Own these three: memory, rules and skills
ChatGPT's own help page says it decides which information is relevant to a response, and that memory can change as your context changes. That is reasonable for a chat app. It also means the record of how your business works is kept on someone else's terms.
Moving that record is harder than it sounds. Anthropic's way into Claude is a prompt: ask your old assistant to list everything it remembers, then paste the answer into a box. Anthropic calls memory import experimental, and says Claude may not always take in what you paste.
Memory kept as plain files moves the other way. You copy a folder, and every agent you try can read it on day one.
Rules work the same way. AGENTS.md is an open format for guiding coding agents, so one rules file can steer several agents, and one edit changes them all. A vendor's settings page changes one product.
Skills are where renting has already cost people work. OpenAI has scheduled custom GPTs to retire on Dec 11, 2026. A GPT's instructions can move into a plugin, but OpenAI says custom actions do not transfer automatically. A procedure written as a file in your own folder has no retirement date.
Video: own the pipeline, rent the model
Video platforms sell convenience, and the convenience is real. Higgsfield puts many image and video models behind one login, sold as monthly or annual plans with a monthly allowance of generations. Runway sells plans with a monthly credit allowance, and its pricing page lists one month of rollover for unused credits on its Max plan.
There is a catch for a business that runs agents. Higgsfield's pricing page says the unlimited models and free generations on its plans work only on its own website, not through its MCP server or command line. The perks are for a person clicking in a browser, not for an agent doing the work.
Your own pipeline turns that around. It calls a video model per clip through an API and keeps everything else: the script templates, the brand rules, the checks a clip must pass and every finished file. When a better model ships, you point the pipeline at it and keep the rest.
Owning has its own costs. Somebody builds the pipeline, somebody keeps it running, and every clip still costs money when a model makes it. If you make a handful of clips a month, a platform plan is simpler. Once video is a weekly habit, the pipeline is the asset and the model is the part you rent.
Two panels. A video platform plan offers many models behind one login and a monthly allowance, with plan perks that work on its website and templates kept in its app. Your own pipeline pays per clip to whichever model wins, keeps scripts, brand rules and checks as your files, can be run end to end by agents, and has to be built and looked after.
A video platform plan
- Many models behind one login
- A monthly allowance of generations or credits
- Plan perks that may work only on its website
- Templates and settings that live in its app
Your own pipeline
- Pay per clip, to whichever model does the job best
- Scripts, brand rules and checks kept as your files
- Agents can run it from script to finished file
- You build it and keep it running
Automation: it depends on how often it runs
Hosted workflow tools charge by use. Zapier sells plans by tasks per month. Make charges credits for each action a scenario performs, says unused credits expire at the end of the term, and says scenarios stop running until more credits are added. n8n prices its hosted plans by workflow executions.
For a few automations that fire a few times a day, that is a fair deal. Once a flow runs every few minutes, the bill becomes a meter on your busiest work.
Owning means a self-hosted workflow tool, or plain scheduled scripts on your always-on machine. n8n publishes a self-hosted Community Edition on GitHub under its Sustainable Use License, which allows use for your own internal business purposes. Read the licence before you build a service for clients on it.
Move a flow in-house when its monthly bill grows faster than the time it takes to look after.
Models: rent the hardest work, own the bulk
The strongest models run in the cloud, and renting them is the right call for hard reasoning and for anything a customer will read. You pay per call. Anthropic's API, for example, prices input and output tokens separately, per million tokens, and takes 50 percent off both for batch work that can wait.
Own a local model for private data, for work that must carry on offline, and for small jobs you run thousands of times, such as tags, titles and sorting. Our guide to local and cloud models goes through it job by job.
Vector search, transcription and voice
For search over your own documents, pgvector adds vector search to Postgres, so the vectors sit next to the rest of your data on a machine you own. Pinecone runs vector search for you; its paid plans carry a monthly minimum and then charge as you go. Own it while it fits on one machine. Rent when the index outgrows the box, or when nobody on the team wants to look after a database.
Transcription is easy to own. whisper.cpp runs speech recognition on your own hardware, treats Apple silicon as a first-class target, and its maintainers show it running fully offline.
Voice is the honest exception. Read the same script through a rented voice and a local one, and keep whichever your listeners would sit through. Whichever wins, the script and your list of how to say each name stay your own files.
The always-on machine: own it
Every business that runs agents gets a dedicated always-on machine of its own. It keeps the jobs running after the laptop closes, keeps the agent's keys off the computer you read email on, and costs one purchase plus electricity instead of a bill whose renewal price someone else sets.
Renting a server is the exception: a short trial, or a burst of heavy work that would swamp the machine for a week.
Rent these: someone else does them full time
Some parts of a stack are a whole company's full-time job. Renting them buys that company's staff.
- Email delivery. Google requires everyone who sends to Gmail to authenticate with SPF or DKIM, and to keep reported spam rates below 0.3 percent. A delivery service watches that all day so you do not have to.
- Social scheduling. A hosted scheduler keeps every platform connection signed in and working as the platforms change their rules. Self-hosted, that upkeep lands on you.
- CRM. Your contacts and their history are worth owning, so choose one that exports them. The software is worth renting.
- Website hosting. Public sites and forms stay on hosting, never on your home or office network.
- Password manager. Its whole job is security. Rent one from a company that does nothing else.
- Offsite backup. The point is that it lives in another building. Rent the storage, and keep the backup files encrypted and your own.
- Payments. Taking cards comes with security rules that a payment company carries for you.
What owning costs, and what renting costs
Neither side is free. Owning front-loads the cost and hands you the upkeep. Renting starts small and keeps renewing, on terms you do not set.
Two panels. Owning costs one purchase and then electricity, your time on updates, backups you have proved by restoring, and outages you handle yourself. Renting costs a bill that renews at a price someone else sets, allowances that reset or expire, terms and features that change, per-seat pricing as the team grows, and rebuilding work when a product retires.
The cost of owning
- One purchase, then electricity
- Your time on updates and upkeep
- Backups you have restored from at least once
- Power cuts and outages are yours to handle
The cost of renting
- A bill that renews at a price someone else sets
- Allowances that reset or expire each term
- Terms and features that change under you
- Per-seat pricing as the team grows
- Work to rebuild when a product retires
An example chart, marked as made-up data. Owning costs most in year one and little after. Renting starts lower and is paid again every year.
Show the numbers as a table
| Option | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Owned | 10 | 2 | 2 |
| Rented | 5 | 5 | 5 |
Example data, made up to show the idea. Not a measurement.
Your own numbers decide it, not this chart. Compare a full year at your real volume, and read every price on the vendor's page on the day you compare.
The rent-or-own worksheet
Run one part of the stack through these five questions at a time.
Five checkboxes: does it get more useful every week, can you export it in an open format today, what does a year cost both ways at your real volume, who fixes it when it breaks at night, and what would leaving cost you.
- Does it get more useful every week you use it?
- Can you export it in an open format today, without asking anyone?
- What does a year cost both ways, at your real volume, with prices read today?
- Who fixes it when it breaks at 2 a.m.?
- What would leaving cost you: files, settings, or work you would have to rebuild?
Write the date beside every price. Prices change too often to print here, and an undated price is a guess by next quarter.
What goes wrong
- Renting what compounds: a year of memory kept inside one chat app.
- Owning what is someone else's full-time job: sending email from your own server, then landing in spam.
- Hosting a public form or website on the home or office network.
- Comparing a monthly plan with an owned machine on month one alone.
- Building on a plan perk that your agents are not allowed to use.
Questions
Should you self-host AI?
Self-host the parts that grow with use and must travel with you: the always-on machine, memory, rules, skills and your own pipelines. Keep renting frontier models and the services that are someone's full-time job, such as email delivery, payments and offsite backup.
Is owning cheaper than renting?
Not on day one. Owning costs a purchase and your time up front; renting starts small and renews. Compare a full year at your real volume, with prices read on each vendor's page that day.
Can you move ChatGPT memory to Claude?
Partly. Anthropic's import asks your old assistant to list what it remembers, and you paste the answer into Claude; Anthropic calls the import experimental. Memory kept as your own files avoids the problem, because you copy the folder.
Terms on this page
Sources
- OpenAI Help Center: Memory in ChatGPT, checked 2026-10-23
- Claude Help Center: Import and export your memory from Claude, checked 2026-10-23
- AGENTS.md, checked 2026-10-23
- OpenAI Help Center: Custom GPT retirement and migration FAQ, checked 2026-10-23
- Higgsfield: Pricing, checked 2026-10-23
- Runway: Pricing, checked 2026-10-23
- Zapier: Pricing, checked 2026-10-23
- Make: Pricing, checked 2026-10-23
- n8n: Pricing, checked 2026-10-23
- n8n: License (Sustainable Use License), checked 2026-10-23
- Claude Platform docs: Pricing, checked 2026-10-23
- pgvector, checked 2026-10-23
- Pinecone: Pricing, checked 2026-10-23
- whisper.cpp, checked 2026-10-23
- Google Workspace Admin Help: Email sender guidelines, checked 2026-10-23