Platform
Everything that changed, and when
Release notes running from the first build in 2019 to revision 5.4. Network and hardware changes sit here too, because from your side they amount to the same thing: something behaves differently than it did last week.
01
How to read this
01
Platform
The panel, provisioning, billing and the rest of the control plane. Most entries land here.
02
Network
Sites, transit, filtering and routing. These change what your packets do, sometimes without any action from you at all.
03
Hardware
Silicon generations, drives, memory and ports. New hardware reaches new orders first and existing instances at rebuild, or on request.
04
Policy
Terms, retention, refunds and acceptable use. Anything altering what we hold or what you are permitted to run appears here, dated.
05
API
Endpoints, tokens and webhooks. Breaking changes carry 12 months of notice and are announced here before anywhere else.
Version numbers are ours and mean nothing outside this list. A major bump is a control-plane rewrite, a minor one is a feature. Nothing gets renumbered after the fact.
02
What is not here
Security patches applied on the normal schedule are not listed one by one, because a list of every kernel update is noise that buries the entries worth reading. Anything requiring a reboot turns up as a maintenance notice on the status page instead.
Incidents live on the status page with their own write-ups. What appears in this list is the fix that followed one, usually a week or two later and phrased less dramatically.
Nothing is ever deleted from this page. Features we withdrew are still listed as withdrawn, including the ones nobody here would mind forgetting, because a changelog you edit afterwards is a marketing page wearing a lab coat.
2026
042026-06-185.4Snapshot scheduling, and the API leaves betaPlatformPer-instance snapshot schedules with retention rules, alongside version one of the public API. Every endpoint the panel uses is documented, breaking changes now carry 12 months of notice, and nothing is left undocumented on purpose.
2026-04-295.3.4Scoped tokens and a sandboxAPITokens gained read-only and per-project scopes, replacing a single account-wide key that far too many people had committed to a repository. A sandbox token type returns real objects, provisions nothing and bills nothing.
2026-03-115.3.3Self-service IPv6 delegationPlatformDelegated /48s and per-address RDNS moved from a ticket into the panel. The previous flow involved an engineer and a text file, which held up for six years and was never going to manage a seventh.
2026-01-215.3.2JohannesburgNetworkJNB-01 racked and cabled, taking pre-orders while transit settles. Ryzen only at launch. It has taken longer than any site since Brazil, and the honest reason is that we would not buy a second-rate path.
2025
052025-11-255.3.1Retention table published in fullPolicyEvery field held against an account is now listed with its retention period and the reason it exists. Nothing changed about what we store. The point was that you previously had to take our word for it.
2025-09-095.3Storage nodes to 100 TBHardwareThe largest tier arrived behind a 25 Gbit/s port, because 100 TB on a 10 Gbit/s uplink is a number rather than a product. An NVMe write cache now sits in front of the bulk tier on every node.
2025-06-035.2API beta, read and writeAPICreate, resize, rebuild and destroy over REST, with signed webhooks on provisioning events. Beta meant we reserved the right to change response shapes, and we used that right twice, both times with notice.
2025-03-185.1Scrubbing to 12 Tbit/sNetworkEdge filtering at the three largest sites went from 8 to 12 Tbit/s, with the mitigation decision moved closer to the customer edge. Time to first drop now sits under 10 seconds.
2025-01-285.0Control plane rewrittenPlatformThe third rewrite, and the one that finally separated billing from provisioning. Median deploy time fell below a minute the same week. Snapshot restores went from minutes to seconds as a side effect nobody had planned for.
2024
052024-11-194.6Balance top-upsPlatformPay once, spend it down. Built for people running many small instances who were tired of watching a mempool at every renewal. Balances do not expire and remain refundable under the usual terms.
2024-09-034.5L40S, and two-card instancesHardwareCards with 48 GB passed straight through to one instance, in Frankfurt, Dallas and Singapore. The dual-card tier followed six weeks later with both cards on one NUMA node, because splitting them across sockets measured worse.
2024-06-254.4Provisioning rewrittenPlatformThe old queue worked serially per site, which was fine at four sites and ridiculous at thirty-four. Parallel provisioning per node took the ninety-fifth percentile from 11 minutes down to under four.
2024-04-164.340 Gbit/s portsNetworkAvailable on EPYC, storage and metal at every site with the upstream capacity behind it. Ports are dedicated rather than shared with the rack, which is the only version of that number worth printing.
2024-02-204.2Zen 5 arrivesHardwareRyzen 9 9950X nodes shipped in Amsterdam and Frankfurt first, then across the fleet within the quarter. Single-thread throughput rose by roughly a fifth against the previous generation, at the same price.
2023
062023-12-054.1First API, read-onlyAPIList instances, invoices and locations with a token. Write operations were left out until the control plane could survive somebody scripting against it, which turned out to be the correct order to do things in.
2023-10-104.0Off-node backupPlatformNightly, encrypted client-side, written to a different site than the instance it came from. It ran internally for a year before anybody was willing to sell it, which was roughly the right amount of caution.
2023-08-013.9Dubai, Santiago and a second Frankfurt floorNetworkDXB-01 sits usefully between Europe, Africa and South Asia. Santiago opened Ryzen-only and has stayed that way. FRA-02 came up with nothing older than the current generation racked in the building, a rule that still holds there.
2023-05-233.8GPU passthroughHardwareWhole physical cards handed to a single instance over PCIe, with no vGPU layer and no scheduler queue in front of them. RTX 4000 Ada in Frankfurt first, then Dallas and Singapore within two months.
2023-03-133.7Warrant canaryPolicyFirst signed statement published, and signed every month since. It states what we have not received. The month it stops being signed is the month it will finally have said something.
2023-02-213.6Stronger password hashing, and hardware tokens for staffPlatformPassword hashing moved to scrypt at N=2^16 with parameters we publish. Every internal account with production access now requires a hardware token, with no software fallback and no exception granted for convenience.
2022
072022-12-133.5Retention cut to seven daysPolicyPanel access logs went from 90 days to 7. Sign-up addresses stopped being written to disk at all rather than being written and later deleted, which is a meaningfully different promise to make.
2022-11-083.4Panama City, Seoul, Mumbai and SydneyNetworkFour sites on four continents inside five weeks, which was two more than the plan called for. Panama takes a very literal view of what a subpoena is, and that is precisely why customers asked for it.
2022-09-273.3Custom ISOPlatformUpload an image or hand us a URL and boot it. The first month produced installations of four operating systems nobody here had heard of, and one that turned out to have no network driver.
2022-07-193.2Reykjavík, Madrid, Prague and a second AmsterdamNetworkAMS-02 opened in a different building on a separate power feed, purely so that AMS-01 stopped being a single point of anything. Reykjavík runs on geothermal power with a shorter path to New York than most of Europe.
2022-05-103.1Routed IPv6 /64 on everythingNetworkEvery instance now ships with a routed /64 at no cost and nothing to request. Delegated /48s arrived the same week for anyone subnetting behind the instance rather than in front of it.
2022-03-143.0Zen 4 everywhereHardwareEPYC 9004 and Ryzen 7000 replaced everything older across the fleet in nine months. ECC DDR5 became standard on the server line, and no node two generations behind has been in production since.
2022-01-252.6SnapshotsPlatformPoint-in-time images on separate storage, restoring in seconds. Sold as a pack of five slots, because unlimited snapshots is a phrase that means somebody else is quietly paying for your storage.
2021
062021-12-072.5Hong Kong and TokyoNetworkAsian capacity beyond Singapore at last. Tokyo has the most stable round-trip time of anything we operate. Hong Kong sells out faster than any other site in the fleet and always has.
2021-10-122.4Dallas, Toronto and São PauloNetworkDallas for cheap central routing, Toronto for the jurisdiction, São Paulo because there is no honest way to serve Brazilian users out of Miami. GRU-01 was the hardest transit negotiation of the year.
2021-08-302.3Mirrored NVMe and a hot spareHardwareEvery node moved to mirrored enterprise drives with power-loss protection and one hot spare per chassis. Two drives failed over the following year and no customer noticed either of them.
2021-06-152.2Stockholm, Sofia, Riga and MilanNetworkFour European sites, all smaller than the original four and all profitable inside a year. Riga and Sofia exist because customers kept asking for somewhere quiet that nobody thinks to look at first.
2021-04-062.1The cheap tier is deletedPolicyThe €4 plan was withdrawn from sale and existing instances ran to the end of their terms. Revenue fell for a quarter. Abuse reports fell by roughly ninety percent and the support queue emptied out.
2021-02-162.0A control panel we actually wrotePlatformBilling, provisioning and support in one place, replacing three tools held together with cron. The account model finally required no name, no address and no telephone number to create anything.
2020
072020-12-221.8Out-of-band consolePlatformSerial and VNC over an authenticated tunnel, included on every plan. Built after a week in which four separate customers locked themselves out with a firewall rule inside 48 hours of one another.
2020-12-081.7Warsaw and ChișinăuNetworkTwo sites added late in the year. Chișinău was requested repeatedly by customers hosting content that attracts complaints, and it remains the jurisdiction they ask for more than any other.
2020-10-191.6Scrubbing, after 340 Gbit/sNetworkAn attack in October took AMS-01 down for 11 minutes. Permanent edge filtering was in place across the whole fleet within six weeks and has never been switched off since. The bill was educational.
2020-07-281.5Miami, Los Angeles and SingaporeNetworkMiami was chosen for the Latin American routes rather than for Florida, and half our Brazilian traffic still lands there. Singapore was the first site outside the Atlantic and immediately the busiest per rack.
2020-05-191.4Helsinki, Zurich and ParisNetworkHelsinki brought cold air, cheap hydro and the lowest power bill per core in the fleet. Zurich brought rack space at a price that still causes arguments. Paris brought southbound routes into Iberia.
2020-03-241.3Monero, and then everything elsePlatformThe first customer to ask for Monero got it six weeks later. Twelve more assets followed before the summer, and the Bitcoin-only decision was quietly retired as the mistake it had always been.
2020-02-111.2Automated provisioningPlatformThe human reading the email was replaced by a queue. Deploy time fell from “sometime this afternoon” to about 9 minutes, which felt like an enormous achievement and looks quaint from here.
2019
022019-12-191.1Frankfurt, London, Bucharest and New YorkNetworkFour sites in eight weeks, all built to the same specification as Amsterdam. New York was the one that turned this into a business rather than a hobby, and Bucharest is still the cheapest terabyte we sell.
2019-11-041.0First buildPlatformAMS-01 took its first order. Invoices came out of a script running on a laptop, provisioning meant a human reading an email, and Bitcoin was the only asset we had finished integrating. It worked, barely.
Ready when you are
Nothing ships without an entry
Subscribe to the status feed and platform changes arrive alongside incidents. If something behaved differently this week and you cannot find it here, that is a ticket worth opening.