The Spec Sheet You Are Comparing Is the Wrong One
When most people compare web hosts, they line up four numbers: disk space, bandwidth, an uptime percentage, and the monthly price. None of those numbers tell you whether the server can hold a vector index in memory, keep a long-lived Node process alive, or let outbound HTTPS traffic reach an external inference API. They tell you about the hosting plan you would have bought in 2014, which is the point. AI hosting is a different product category wearing the same price comparison table.
The argument I want to defend in this post is blunt. Most pages selling AI hosting today are describing a control panel chatbot and an “AI-optimized” badge on the same PHP-FPM pool the host has been reselling for years. The label has been applied so broadly that it has stopped meaning anything on its own, and the only way to recover meaning is to test the underlying capability directly. WPEngine’s piece on the agentic web makes the same point more politely: “True AI-ready hosting is defined by production-grade features, clear documentation, and built-in infrastructure, not just traditional hosting paired with third-party plugins” (WPEngine, AI Hosting: Infrastructure vs. Intelligence on the Agentic Web (opens in new tab)). The trouble is that documentation and production-grade features are exactly the things a sales page can claim without delivering.
Before the rest of the post, here is the skimmable version of the checks. Each one is unpacked in its own section below.
- Run
curl -sS -o /dev/null -w '%{http_code}\n' https://api.openai.com/v1/modelsfrom the account shell to test outbound HTTPS. - Query
SELECT * FROM pg_available_extensions WHERE name = 'vector';on Postgres, or check the MySQL version for vector column support. - Read
pm.max_childrenandrequest_terminate_timeoutin the PHP-FPM pool, then watch the FPM error log for child exhaustion. - Ask support whether long-running processes are permitted, and test
curl -A "GPTBot" -I https://yoursite.example/to see whether crawler filtering is real. grep -Eic 'gptbot|claudebot|perplexitybot|bytespider' /home/*/logs/*access_logon your own access logs before you decide anything.
If you only do one of these, do the first one. It is the one that quietly fails most often.
What Is AI Hosting, and What Is It Not?
AI hosting means one of two things, and they share almost nothing in common under the hood. The first is site-facing tooling that you build features with: managed vector database hosting, Model Context Protocol (MCP) endpoints, embedding pipelines, semantic search. The second is platform-level machine learning that runs the hosting environment itself: anomaly detection on metrics, automated visual regression, AI crawler bot filtering at the edge, smarter failover decisions. WPEngine’s framing puts it the same way, splitting site-based features like managed vector databases from platform features like edge bot filtering and automated monitoring (WPEngine (opens in new tab)). Both are legitimate. Conflating them is what makes the sales copy unreadable.
The exclusions are sharper than the inclusions. A WordPress plugin that calls an external API over shared hosting is not AI hosting. A billing assistant bolted into cPanel is not AI hosting. A renamed resource tier with an AI badge is not AI hosting. None of those change the hosting layer underneath, and none of them will survive a real load test.
The claim that “most legacy web stacks cannot handle AI workloads” deserves pushback. Postgres with pgvector and MySQL’s newer vector column type run fine on ordinary hardware for a few hundred thousand embeddings. I have done it. The actual failure mode in LAMP stack AI workloads is not the database, it is that a synchronous API call inside a PHP request holds a worker hostage for the entire length of the model’s response. You can watch this happen in production: a single slow embedding call locks an FPM child, the queue backs up, and your contact form starts returning 502s. The stack did not change. The request shape did.
For the rest of this post, I am going to use a single line to keep things honest. If the capability requires a process that outlives an HTTP request, shared hosting almost certainly does not have it. That single rule eliminates most of what gets marketed as AI hosting on a $4.99 plan.
What Hosts Are Actually Shipping in 2026
The cleanest signal I know of for whether a host actually does AI hosting is documentation quality. A provider with documented connection strings, documented index size limits, and pricing that maps to those limits has built something. A provider with a blog post announcing an AI partnership has not.
The cPanel/WHM situation specifically is worth naming. MCP server hosting requires a persistent Node or Python process with a stable public endpoint, and most shared accounts kill anything that idles or breaches a memory ceiling. Passenger-style app hosting on cPanel helps: it supervises a process, restarts it on crash, and exposes it through a subdirectory or subdomain. What it does not do is give you a guarantee about cold start latency, outbound rate limits, or what happens to your neighbors when your process decides to eat 1 GB of RAM. If you intend to run an MCP server on cPanel, expect to read more of the Passenger docs than you wanted to.
Crawler traffic is the one platform-side feature that has genuinely arrived in production. On my own boxes, I run this against the previous day’s access logs:
grep -Eic 'gptbot|claudebot|perplexitybot|bytespider' /home/*/logs/*access_log
The numbers vary wildly by site. A documentation site with public reference material will see hundreds of hits per day. A local bakery site will see two. robots.txt is a request, not enforcement, so any real filtering has to happen at the edge CDN or in the firewall, not in a WordPress plugin.
WordPress AI hosting features in practice today mean two things, and they are both narrow. The first is semantic site search backed by a host-managed index, where the embedding model and the vector store live on the host and the search runs as a managed endpoint. The second is recommendations built on embeddings rather than category tags. The honest case for both is the “sage windbreaker” versus “light green jacket” example from WPEngine’s piece (WPEngine (opens in new tab)): keyword search misses it, semantic search catches it. The honest case against is that a brochure site with forty pages gains nothing, because there is no corpus to search and no recommendation graph worth building.
Ten Minutes of Checks That Beat Any AI Hosting Sales Page
Run these in order. The first one is the one that quietly fails on shared hosting, so do it before you believe anything else the host tells you.
Outbound network test. From the account shell, run curl -sS -o /dev/null -w '%{http_code}\n' https://api.openai.com/v1/models. A 200 means your outbound 443 is unrestricted. A 301 to a captive portal, a 403 from a proxy, or a connection timeout means the host has an allowlist and your “AI plugin” will hang without a useful error. This is the single most common failure I see when people move a chatbot plugin from a VPS to shared hosting.
Database check. On Postgres, SELECT * FROM pg_available_extensions WHERE name = 'vector'; tells you whether pgvector is installed and which version. On MySQL, check the version and whether your host will let you create a vector column at all. Then ask, in writing, what an HNSW index build does to your CPU quota, because index construction on a cgroup-limited plan is not background-friendly and the host should know that.
Worker exhaustion check. Read pm.max_children and request_terminate_timeout in /etc/php-fpm.d/ or WHM’s MultiPHP Manager. Then trigger a few concurrent AI-driven requests from your own browser and tail the FPM error log. The line you are looking for is server reached pm.max_children setting (5), consider raising it. That single line is what a 502 on your contact form looks like from the inside of the host, and it is the proof that your AI feature is starving the rest of the site.
Persistent process and bot filtering checks. Ask support directly: are you permitted to run a systemd service or a supervised long-running process, and at what memory ceiling? Then test the edge: curl -A "GPTBot" -I https://yoursite.example/. A 403 from the CDN means filtering is real and running at the edge. A 200 served by the origin means the feature is a toggle in a dashboard that does not yet do anything.
Where AI Hosting Goes Next, and What I Am Unsure About
MCP server hosting is the obvious next product line, and the unsolved part is authorization. Handing an external agent scoped, revocable access to a customer’s content is an OAuth problem more than a hosting problem, and I have not seen a host document token scoping properly. My expectation is that the first hosts to get this right will be the ones who already ship managed OAuth flows for headless CMS use cases. Anyone else will end up bolting it on and calling it done.
Managed vector indexes on multi-tenant nodes are the thing I am genuinely uncertain about. I do not know how memory-resident indexes behave when a single node carries hundreds of tenants competing for the same NUMA node, and nobody publishes numbers on it. If a vendor publishes a benchmark with named workloads and named index sizes, treat it as a starting point, not an answer.
Pay-per-crawl and edge-level negotiation with AI agents is coming, and the tradeoff worth naming now is that blocking a crawler also removes you from the assistant that cites you. Sites that depend on discovery through ChatGPT or Perplexity are not going to flip that switch without thinking.
The boring prediction: platform-side features will land first, as anomaly alerts and automated screenshot diffs in the host’s own dashboard, because shipping those does not require giving up any customer-facing resource. Site-side vector tooling reaches shared plans later, if ever.
Your Next Step: One Support Ticket and One Log Grep
Paste this into a ticket to your current or prospective host. Ask four things: what is the outbound HTTPS policy, is the vector extension available and what is the index size limit, are long-running processes permitted and at what memory ceiling, and does AI crawler filtering happen at the edge or the origin. Read the reply with one rule. A specific answer with a limit attached to it is a yes. “You can install any plugin you like” is a no, and it is the answer I expect most people on shared hosting to get.
Run grep -Eic 'gptbot|claudebot|perplexitybot|bytespider' /home/*/logs/*access_log on your own access logs before you decide anything. If the number is small enough that you would not pay to filter it, then the rest of this post is somebody else’s problem and your hosting bill does not need to change. Switching hosts to get a vector database you will not query is a worse decision than staying put, and the cheapest way to find out is to ask the support ticket and read the access log you already have.