State of AI app hosting 2026

Agents now make things people want to open: dashboards, landing pages, team tools, decks, whole apps. Each one needs a place to live, a URL, and somebody deciding who can see it. That job used to belong to a developer with a cloud account. In 2026 it is increasingly done by the agent itself, and the hosts are changing to suit it. Vercel's CEO told TechCrunch in July that about half of its roughly 6 million deployments a day were triggered by coding agents. Lovable says the apps built on it get close to a billion visits a month, about ten times the traffic of Lovable itself.

The market has split four ways. Builder-hosts (Lovable, Replit, Bolt, Base44) generate the app and keep it on their own domain. Developer clouds (Vercel, Netlify, Cloudflare, with Supabase as the database underneath) have rebuilt their front doors for agents. Chat apps now host their own output: ChatGPT Sites and Claude artifacts. And a set of small link hosts (here.now, tiiny.host, ShipStatic, Hostsmith, Static.app) sell a link for whatever an agent made. This post collects their numbers as of 18 October 2026: the money, how much they host, what they offer, and what goes wrong. Vendor numbers are the vendor's own claims, and most of them are run rates disclosed around a fundraise.

The money#

Every revenue figure here is a run rate: a recent month's revenue times twelve, unaudited, usually given around a fundraise or leaked. Only Wix (for Base44) and Cloudflare report through filings and earnings calls.

The two largest run rates disclosed for hosting agent-made work are now about the same size. Lovable passed $500M in June and co-founder Fabian Hedin said on stage in Amsterdam on 24 September that it had crossed $600M. Vercel, where v0 apps and much of what coding agents deploy end up, reported a $340M run rate at the end of February 2026 and crossed $500M in early July. In October The Information reported $600M, up 148% from a year earlier, with coding agents about half of new business. That figure rests on one anonymous source; Vercel hasn't confirmed it. If it holds, Vercel grew its run rate 76% in about seven months, by our arithmetic.

The others are smaller or quiet. Base44, inside Wix, had about $150M of ARR as of May, per Wix's filing. Replit last gave a number in September 2025, $150M, and since then only a target: "on track to hit $1 billion in run-rate revenue by the end of 2026". Bolt's last figure was $40M in March 2025. Netlify hasn't published revenue since its 2021 round. Cloudflare, which runs under several of the small hosts, made $696.1M in Q2 2026, up 36%, across everything it sells. At the small end, tiiny.host says it makes about $1.5M a year; here.now, ShipStatic, Static.app and Hostsmith publish nothing.

Investors price the database under these apps almost as high as the apps themselves. Supabase went from $2B in April 2025 to $10.5B in June 2026, 5.25× in about 14 months.

The builders' revenue story is covered in more depth, with multiples and revenue per head, in AI app builders in 2026.

Agents now drive the deploys#

Vercel is the only big host that has put a number on how much of its work comes from agents, and it has given several, each measuring something slightly different. In April TechCrunch reported that 30% of the apps running on the platform came from agents. In June, launching its eve agent framework, Vercel said agents triggered around 29% of deployments. In July Guillermo Rauch told TechCrunch the platform handled about 6 million deployments a day, half of them triggered by coding agents. And in October The Information reported coding agents were about half of Vercel's new business.

Half of 6 million is about 3 million agent deployments a day, roughly 35 a second. No other host gives a comparable share. Netlify has given absolute counts instead: in early 2025 it said more than 1,000 sites a day were being created on it directly from ChatGPT, and over 10,000 a day from AI codegen platforms such as Bolt.

The developer counts tell the same story from below. Cloudflare said it added about 1 million developers in Q1 2026 and nearly 2 million in Q2, ending June with more than 7.4 million, and that Q2 alone beat the roughly 1.5 million it added in all of 2025. Supabase claimed nearly 10 million developers at its June round, up from 2 million in April 2025. Netlify crossed 10 million on Christmas Eve 2025, five million of them in a single year, and called it the start of the "agent experience" era.

The database layer is where agents dominate most clearly. When Databricks agreed to buy Neon in May 2025, Neon said more than 80% of its databases were created by AI agents. Supabase, which sits under Lovable and Bolt, said in June 2026 that more than 60% of its new databases are launched "by some sort of AI tool", with Claude Code the largest single contributor. Its CEO later said Supabase was deploying more than a million new Postgres databases a week, up from 40,000 a day in September 2025: about 3.6 times as many, by our arithmetic, if the two counts are defined the same way.

How much is actually hosted#

Lovable is the one builder that reports traffic to the apps it hosts, and the number has grown much faster than its project count. Visits went from 200 million a month in December 2025 to 720 million in June 2026, per its "build economy" report, then over 900 million in August and "close to a billion" in September. Hedin added that this is "an order of magnitude more than Lovable itself" gets.

The other counts are of things made, not things visited. They're big, and they measure different units.

What the public web sees is much smaller. Common Crawl, the open web archive, only finds pages someone has linked to, so it is a floor. In its September 2026 crawl, which we counted for our AI app builders post, there were 4,485 distinct lovable.app hosts, up from 14 in January 2025. Against 60 million projects, that's about one crawled app for every 13,000 projects. Most projects are drafts, private, never published, or never linked; apps on custom domains don't show up at all. The domains still climbed the popularity rankings: lovable.app reached #3,084 on the Tranco list of 7 October 2026, from #376,351 in January 2025.

So the honest summary is: agents make millions of things a month, a small fraction get published, and a smaller fraction get most of the traffic. No host discloses how many projects are ever shared or still visited a month later.

What each host offers#

On paper every host now publishes to a URL. The differences are in who can open it, what it costs to keep it private, and how an agent gets the work there. As of mid-October 2026:

The clearest dividing line is privacy. Most agent-made apps are for a team, a client or one person, and on most of the hosts below keeping them private is the upgrade.

The other change is how work arrives. In 2025 a host needed a dashboard and a CLI. By late 2026 it needs a way for an agent to deploy unattended: an MCP server, a skill, or an anonymous deploy a human can claim later.

Anonymous deploys started with the small hosts and reached the big clouds in 2026. The unclaimed lifetime varies 72-fold.

For the chat apps specifically, we walked through publishing in How to use ChatGPT Sites and How to use Claude artifacts. Who inside a company builds these tools, and for whom, is in Internal tools built by non-engineers.

The dark side#

A shared hosting domain with free, instant, TLS-protected subdomains is exactly what a phisher wants, and agent-made hosting made it faster to fill one with convincing pages. Proofpoint wrote in August 2025 that it had observed "tens of thousands of Lovable URLs detected as threats each month" in email since February 2025; one Tycoon phishing campaign using them hit more than 5,000 organizations. Okta caught v0 being used to clone its sign-in page in July 2025. Kaseya's INKY team blocked at least 6,283 phishing emails across more than 2,030 organizations that used vercel.app pages between December 2025 and May 2026. Netcraft said in July 2026 it built two working phishing pages with public vibe-coding tools and found identity checks "inconsistent".

The chat apps aren't exempt. In June 2026 Zscaler found paid Google ads leading to shared Claude chats named "Apple Support" that told Mac users to paste a command installing an infostealer. The sharing feature was abused rather than broken, and the pages inherited claude.ai's reputation. Exposure runs the other way too: in May 2026 Axios reported researchers finding exposed vibe-coded apps with plain web searches on builder domains. We collected the scans in Vibe-coded app security in 2026.

Platform responses are mostly unmeasured. Lovable said in 2025 that it blocks about 1,000 malicious projects a day and removed more than 300 sites in two weeks; we found no 2026 takedown statistics from any host.

One structural fix is the Public Suffix List, which browsers use to treat each subdomain as its own site: cookies don't cross between apps, and one bad app's Safe Browsing reputation doesn't sink its neighbours. It doesn't stop phishing or spam. GitHub Pages joined in 2013; the AI-era hosts joined between February 2025 and August 2026.

Listing isn't a cure. In an Ahrefs pull we ran on 10 October 2026, both here.now, which is on the list, and tiiny.site, which isn't, ranked for Indonesian gambling spam on user subdomains. Spam is about what search engines are allowed to index, not about cookies.

Reliability is the other risk. Most hosts in this post run on someone else's infrastructure, so the long outages of the last year mostly came from upstream.

The pattern is reassuring for anything already published: in the Supabase incident under Lovable and in Vercel's October build outage, the serving path kept working while editing or building stopped. The serving outages that did happen, like Cloudflare's on 18 November 2025, came from infrastructure that sits under a large share of this market.

What this means if you're choosing where your agent's work lives#

1. Decide who should see it before you publish. Defaults range from public and indexable to private to the owner. If the app holds team or customer data, check the default and what a private app costs; on several builders it's the most expensive plan.
2. Keep the deploy step separate from the builder if you can. A host that takes a deploy from any agent (an MCP server, a skill, a CLI) lets you change agents without moving the app. Builder-hosts are easiest when you stay in their editor and hardest to leave.
3. Check whether your viewers need an account. Since September 2026 a newly shared Claude artifact opens only for people signed in to Claude, and ChatGPT Sites start private. That's fine inside a company on one plan and a problem for clients and outsiders.
4. Know where the data lives. Most builders put the app's data in Supabase or their own backend. Find out how you export it and who can read it before people start typing into the app.
5. Prefer a host whose domain is on the Public Suffix List, or use your own domain. It won't stop phishing, but it keeps your app's cookies and reputation separate from strangers' apps on the same domain.
6. Treat anonymous deploys as previews. They vanish in an hour to three days unless claimed. Claim the ones you mean to keep, and don't put anything sensitive in them.
7. Read the fine print on what happens when you stop paying. Some hosts keep the app live on a free plan; some take it offline or delete it. The costs side is in What it costs to run a vibe-coded app in 2026.

host0 hosts what your agent made, from any agent: sites, apps and files, private by default, shared with people by email, with a shared records database and a free plan. It's built for the step after generation, so the tool you build with doesn't decide where the work lives or who can open it.

Methodology#

We gathered the numbers between 30 September and 18 October 2026 from company blogs, filings, earnings-call summaries, changelogs, docs and press coverage, and from our earlier research packs. Most vendor pages couldn't be opened from our research environment, so many rows come from search-engine extracts of the named pages; we spot-checked the 20-odd numbers this post leans on with independent searches and dropped or attributed what we couldn't confirm. Everything is dated on or before 18 October 2026.

Run rates are a recent month times twelve, self-reported and unaudited; Vercel's ~$600M is one outlet's report from an anonymous source. Visit, project, site and developer counts are each company's own definitions and can't be compared across companies. The Common Crawl host counts and Tranco ranks are our own pulls from 8 October 2026: Common Crawl only sees publicly linked pages, and Tranco ranks registrable domains, so lovable.app aggregates every app on it. The Public Suffix List dates come from the list's own Git history, read on 18 October 2026. Outage durations come only from the provider's post-mortem; where we only had a status-page mirror, we give the event without a duration. Prices and plan features are as of mid-October 2026 and change often.

Open questions#

1. How many agent-made projects are ever published or shared, and how many are still visited a month later. No host discloses it.
2. How much of Netlify's, Cloudflare's or Replit's traffic comes from agents. Vercel is the only big host to give a share, and its own figures measure different things.
3. When Vercel's agent share was under 3%. Vercel's June post says a year earlier; its CEO and The Information have said the start of 2026.
4. What ChatGPT Sites looks like after 5 million. A higher figure from OpenAI's September DevDay has circulated secondhand; we couldn't confirm it.
5. How many Claude artifacts are shared publicly, and any count after the half-billion of June 2025.
6. What the hosts take down. Lovable's 2025 numbers are the only abuse statistics we found; no host publishes them regularly.