The 2026 Mid-Year Repricing Ledger: Every Layer of Your Rented Stack Just Moved at Once

Seven vendors, six months, one pattern — and why the only price that didn't change was the one you set yourself.

Between April 1 and June 15, 2026, the hosting stack repriced itself almost in unison. Your PaaS flipped its seat model, your registry doubled its pull tax, and the bare metal underneath both got more expensive — twice — driven by the same DRAM shortage. If you rent any part of your stack, you absorbed at least two of these changes whether you noticed the invoice line or not.

The Ledger: Every 2026 Pricing Move in One Table​

This is the core deliverable. Before the analysis, here are the numbers.

Sources: Render new workspace plans and docs; Netlify "The end of seats" and pricing; Vercel June 2024 pricing breakdown and 2026 tracker; Fly.io egress IP billing and free-tier removal; Docker announcement and The New Stack; Hetzner via Gart Solutions, DEV, and Hetzner price adjustment page; OVHcloud via DecodeStack and CEO guidance The Register.

One table, seven rows, one conclusion before we even get to the prose: repricing risk is no longer a single-vendor concern. In the first half of 2026 it became a correlated event across PaaS, registry, and substrate simultaneously. The question is not whether your next bill moves — it is how many lines move at once.

The Same Shape, Two Vendors, Nine Days Apart​

On April 14, Netlify published "The end of seats" and flipped Pro from $20 per seat to a flat $20/mo for unlimited members — a 10-person team went from $200/mo in seat fees to $20. Nine days later, on April 23, Render did the same: Professional had been $19/member/month; the new map is Hobby $0, Pro $25/mo, Scale $499/mo, all unlimited. Bandwidth allotments of 5 GB / 25 GB / 1 TB apply, then $0.15/GB overage and $0.25/domain/mo beyond 15 domains. Existing workspaces auto-migrate August 1, 2026.

Two vendors, same shape, same month — a shared admission that seat-based pricing stopped working. It penalizes the customer you most want to retain (the team that grows) and breaks under AI-agent workflows where a bot opening 40 PRs looks like 40 seats even though it never opened a dashboard.

The win for renters is real but bounded. Both vendors replaced the seat meter with a usage meter that now does the real work. Netlify draws bandwidth, compute, and deploys against a monthly credit pool (20 credits/GB, ~$0.13/GB). Render's flat workspace fee sits atop per-service compute prorated by the second (Starter $7/mo for 512 MB, Standard $25/mo for 2 GB, up to Pro Ultra at $450/mo). The seat fee is gone; the bandwidth and compute meters remain — and those are the lines that now ratchet.

A self-hosted fleet on owned hardware has no members × $19 term in its cost function. Team size is a coordination cost, not a billing dimension. Two vendors converging on that deletion in the same month is the market confirming the dimension should not have existed.

The Ratchet: Vercel's Fourth Repricing and the Credit Wave​

Vercel has now repriced four times since June 2024. The cadence itself is the story.

• June 25, 2024: Granular incremental billing — bandwidth and function duration broken into smaller units so teams could optimize per-line.
• September 2025: Pro restructured from fixed allocations to a credit pool — $20 per seat buys $20 of usage credit, everything meters against it.
• 2025–2026: Web Analytics per-event billing cut to $0.00003/event ($3/100K, down 79% from $14/100K) while other meters held or rose — a rebalancing, not a cut.
• 2026: The pool model persists and the pattern is established: each revision narrows what is included and shifts cost onto a new or repriced metered line.

The included usage shrinks, a new line appears, or the overage rate moves. The 2026 tracker that logged 202 developer-tool pricing changes put it plainly — Vercel, Netlify, and Cursor all moved to credit-based pools in the same window. When every vendor migrates to "you buy credits, usage draws down," the page looks simpler while forecasting gets harder.

A team budgeting on a fixed allocation knows its ceiling. A team budgeting on a credit pool must model traffic and invocations against credit burn — then re-model after the next repricing. That is a tax on planning, not just dollars. A high price can be budgeted; a price that moves on someone else's schedule can only be re-budgeted after the fact.

A Hetzner CX box at €3.79/mo costs €3.79/mo regardless of how many credits Vercel says a gigabyte is worth next quarter — even after Hetzner's own hikes (below), it remains a single posted figure, not a pool that can be redefined.

Metering Creep and the Registry Tax: Fly.io and Docker Hub​

Two rows illustrate different forms of the same creep.

Fly.io killed its Hobby, Launch, and Scale tiers on October 7, 2024 for pure usage billing — pay per second. Through early 2026 the new lines arrived: app-scoped egress IPs at $3.60/mo per IP from January 1, 2026, and volume snapshots moving from free to metered. Per-app the delta is small; for a 10-service fleet with daily snapshots it is ~$50–100/mo of new fixed cost that did not exist a year ago — plus the free tier itself is gone for new orgs (replaced by a 2-hour trial or $5 credit).

Per-second billing genuinely wins for bursty workloads — an API that idles overnight costs near-zero where a fixed box sits idle. For the always-on workloads most teams run (API + worker + Postgres, 730 hrs/mo each), the comparison inverts. A shared-cpu-1x at ~$3.32/mo looks cheap next to a Hetzner CX22 until you add bandwidth, volumes, IPs, and snapshots. The headline compute line is no longer the bill; the sum of the small lines is — and vendors can add small lines unilaterally, as Fly.io did in January.

Docker Hub is the other tax — not creep but a primitive repricing. In November 2024 Docker raised Pro $5 → $9/mo (+80%) and Team $9 → $15/user/mo (+67%), Business flat. Personal stays free but capped (200 pulls/6 hrs unauthenticated). The hike bundles Build Cloud and Scout, but for a team that only needed private repos and pull headroom, the bundle is a forced upsell to keep the same primitive. When the registry reprices in Q4, your PaaS in Q1, and your substrate in Q2, each vendor's "we only changed one thing" is true alone and false in aggregate.

The Substrate Moves Too: Hetzner Twice, OVHcloud in Parallel​

The rows that surprised even self-hosters were at the bottom of the stack.

Hetzner raised prices twice in ten weeks. On April 1, a broad adjustment: CX/CAX/CPX/CCX lines up 30–37% in Germany and Finland, US regions up 38–40% (an AX42 dedicated server from €47.30 to €57.30/mo is the canonical example). On June 15, a second, steeper adjustment targeting the lines most production fleets actually run: dedicated-vCPU CCX instances up 107–204% on top of the already-higher April baseline, and shared-vCPU CPX up roughly 113–176%. The cumulative effect on a CCX23 — $39.99 → $102.99 — is a 2.5–3.1× increase versus pre-April in some regions. The renamed SKUs (CPX11 → CPX22, CX22 → CX23) make before/after comparisons easy to miss if you only track the name.

The driver is not Hetzner-specific. DRAM contract prices rose 90–95% quarter-over-quarter in Q1 2026 with another 58–63% in Q2, driven by AI server and HBM demand consuming ~40% of world DRAM output. Hetzner cited the memory and NVMe shock directly. Existing instances hold their old price until a rescale or new order — which is both a grace period and a trap: the moment you need to add capacity, you pay the new number.

OVHcloud moved the same day — April 1 — with a 2026–2028 path of 9–11% annual increases on new Public Cloud / Bare Metal / Private Cloud and 2–6% on pre-2025 fleets; the headline VPS SKU jumped 55–67%. CEO Octave Klaba had warned in November 2025 that 5–10% rises by mid-2026 were coming on the same RAM/NVMe curve. A published multi-year path is a commitment to keep moving.

Two implications for "cheap owned hardware":

First, pre-April Hetzner numbers are now fictional. Any comparison still using them is off by 30% at best and 2.5× on the CCX/CPX families most teams benchmark.

Second, the advantage survives the recompute — but the margin narrowed. Even post-hike, a flat Hetzner box with included bandwidth undercuts per-second PaaS billing for steady-state, always-on services once you sum the PaaS lines. What changed is not whether owning is cheaper for that shape, but how much cheaper and how confidently you can quote a number that holds next quarter. The honest pitch is not "we are cheap" but "we are predictably cheaper — one posted figure, not five meters that can move."

What to Do With a Ledger That Moves Every Row​

The H1 2026 ledger is not an anomaly to wait out. It is a single upstream cause — AI-driven memory demand repricing the physical substrate — propagating upward through every managed layer. The substrate vendor raises prices; the PaaS that rents that substrate either absorbs margin or passes it through. The registry on top does the same. Only timing varied, and in H1 2026 it converged.

Three responses, in order of ownership:

1. Budget for repricing, not just price. Treat each pricing page as a variable with a drift rate: seat → flat+usage was a one-time shift; credit pools ratchet every 12–18 months; substrate moves 9–37% per event, potentially twice a year while DRAM is constrained. Add a "repricing reserve" at 10–15% of managed spend — if you don't, the next announcement sizes it for you.

2. Isolate the meter the vendor controls. Name the conversion you don't control: Vercel/Netlify's credits-per-GB, Fly.io's per-primitive lines, Docker Hub's pull quota. Ask what happens if that one conversion doubles. If you can't absorb it, that dependency is an unhedged price exposure.

3. Own where predictability matters most. You don't have to self-host everything. Own the steady-state tier — API, worker, database — on a fixed-cost box where the bill is one number. Burst the bursty parts — builds, previews, per-PR environments — to a metered provider where elasticity wins. A git-push PaaS on your own hardware gives you that split without choosing between "single-box simplicity" and "multi-machine fleet when you need it." The cost curve is flat where your traffic is flat, elastic where it isn't.

There will be a second half to this ledger. Hetzner grandfathering expires on rescale, Vercel's fifth repricing is a matter of when, and Docker Hub's consumption-based pull pricing is still rolling out. The teams that fare best won't be those that picked the cheapest vendor in January — they'll be those whose cost structure has the fewest lines someone else can rewrite before December.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Render-compatible API, Cluster API fleet underneath, no per-seat meter and no credit pool. Star the repo on GitHub or deploy your first app today.