Koyeb in 2026: an honest review for teams leaving AWS or Azure

Koyeb is a Paris serverless platform: give it a container image or a Git repository and it runs an autoscaling HTTPS service, with GPUs, PostgreSQL and AI agent sandboxes on the same bill. Since spring 2026 it has been a product line of Mistral AI. It replaces App Runner, Fargate or Azure Container Apps, not an AWS account. It is quick to a running service, scales to zero, bills by the second and rents an H100 for USD 2.50 an hour without a sales call. Against that: no object storage, durable block storage or queues; a data path that runs through US companies at the edge (by our DNS checks), the control plane and the database; every H100-class GPU in North America; no SLA below the USD 299 a month Scale plan; and contracts that still name a company struck off the French register in May 2026.

This article is one of a series comparing 24 European cloud providers with AWS and Azure; the overview puts them side by side. Everything below was checked on 15 and 16 October 2026 against Koyeb's documentation, legal terms, public price catalogue and status page, the French company register, DNS records and user reports, each at its source. We did not open an account or run workloads.

At a glance

What you can build

There are no virtual machines to log into and no disks of useful size, so the fair comparison is with the serverless container services of AWS and Azure.

Deep Sleep wakes a scaled-to-zero service in "1-5 seconds", Light Sleep in about 200 ms. Services have no public IPv4 address of their own to pay for.

Missing outright: object, file and durable block storage; every managed database except PostgreSQL (the docs point to Aiven, MongoDB Atlas, Upstash and others); queues, pub/sub and scheduled jobs; Kubernetes; DNS; static outbound IP addresses ("We do not currently provide a list of static outbound IP addresses"); KMS; and backup. Of workload B, four lines have no equivalent: the 300 GB of durable block storage, the 1 TB of object storage with its requests, the database standby and the Kubernetes control plane.

The tooling is maintained: a REST API with a public Swagger document, the koyeb CLI and a partner-tier Terraform provider, also on the OpenTofu registry, whose version 0.2.0 came out on 24 September 2026 after a 21-month gap. Pulumi has had no release since December 2024. The public changelog stops at 27 February 2026, after weekly or fortnightly entries through 2025.

What it costs

Koyeb bills in US dollars, VAT excluded. Prices are list prices from its public catalogue API and pricing page, checked 15 and 16 October 2026, converted at the ECB rate of 15 October 2026 (1 EUR = 1.1208 USD) at 730 hours a month. Every new organisation pays a plan fee on top of usage:

Workload A, a starter setup: one instance with 2 vCPU and 4 GB, and 1 TB of egress. The 40 GB disk and the 100 GB of object storage have no equivalent.

Eco instances get half the cores of Standard ones and are "Designed for development and non-production environments"; with a Standard large (4 vCPU, 4 GB) the total is USD 61.05 (EUR 54.47). AWS charges EUR 112.37 and Azure EUR 116.15 for workload A in Frankfurt, most of it for the egress.

Workload B, a small production system, priced as the nearest platform equivalent in Frankfurt: three app instances of about 4 vCPU and 16 GB, the managed PostgreSQL and 2 TB of egress.

Two readings sit in the total. That the credit offsets database and bandwidth charges as well as instances is our assumption: the page says "$10 included compute". And at the Pricing FAQ's USD 0.04 per GB after 100 GB, the 2 TB of egress would cost USD 76 instead of USD 20. On Eco instances, which cannot autoscale or scale to zero, the Pro total falls to USD 573.94 (EUR 512.08).

On paper the Pro total is about three quarters of AWS's EUR 985.89 and Azure's EUR 1,017.63 for workload B on burstable VMs, and Scale about 90%. It is not like for like: AWS's bill includes S3, EBS and an EKS control plane (USD 135.76 together) and a standby in another zone, which Koyeb does not sell, while Koyeb's includes the load balancer, TLS and 1 TB of egress, which AWS bills separately.

Egress. The pricing page's table includes 1 TB a month on Pro and Scale, then USD 0.02 per GB in the EU and US and USD 0.04 in Asia; the same page's FAQ says "$0.04/GB transferred"; the Pricing FAQ says "Outbound bandwidth is not charged yet. We provide 100GB of outbound bandwidth per month for free, usage after this allowance will cost $0.04 per GB." Even the higher figure is under half of AWS's USD 0.09. Outbound addresses are shared and change "without notice".

Billing. "Only pay for what you use, by the second", charged monthly, by card only. There is no hard spending cap; billing alerts are off by default. One small free web service and a free Postgres allowance remain on every plan. Reservations get "up to 50% off" through sales. Volume and snapshot prices are not published.

GPUs, per GPU-hour, billed per second:

AWS charges USD 7.36 per H100-hour in Stockholm, USD 2.33 for an L40S and USD 4.29 per A100 80 GB in Frankfurt, and every AWS GPU family starts at a quota of zero. The catch is location: the RTX 4000 SFF Ada is the only GPU Koyeb sells in Europe. "Restricted" means not orderable without Koyeb's involvement, and GPU concurrency is capped at 20 on Pro and 40 on Scale.

Price changes. GPU prices fell on 13 December 2024 and again on 13 November 2025 (H100 from USD 3.30 to 2.50, A100 from 2.00 to 1.60, L40S from 1.55 to 1.20). In February 2026 the no-fee Starter plan closed to new organisations, making Pro at USD 29 the entry price. We found no increase to instance or bandwidth prices; the MSA allows plan price changes on 30 days' notice.

Reliability

What a region is. Koyeb sells no availability zones. Its catalogue lists three data centres in Frankfurt (fra1, fra2, fsn1) and one in Paris (par2). A 2023 engineering post says a data-centre failure is handled by "rescheduling Instances in other data centers of the same region", but no document says whether Frankfurt's three are separate buildings, or whether a service's instances are spread across them. Instances also move during "automatic software upgrades which are performed at least on a weekly basis", losing their local disk. Volumes sit on one machine: "there is no redundancy here and fault-tolerance is on you". To survive a region failure you deploy in several, and the edge then routes requests to another region "If no healthy Instances are available within the region".

The database. Koyeb's docs say nothing about standby or failover; its marketing says "Fully managed, highly available". Neon's own documentation describes the design: the write-ahead log is replicated across zones, compute runs in one zone with no standby, a failed node is rescheduled in "1-2 minutes" and a zone failure recovers in "1-10 minutes". That is Neon's design, not a Koyeb commitment, and no Koyeb SLA names the database. Backups are not documented, and "We cannot assist you in recovering data from a deleted database".

The SLA, one platform-wide document dated 28 June 2024:

• It measures "the network availability of the Platform and Services" per month, not a service, region or instance.
• Excluded: scheduled maintenance announced "at least three (3) working days in advance", "emergency maintenance operations without notifying the Customer", and failures of the customer's environment "including a failure of a third-party provider".
• The credit is "a pro rata discount" of the month's fees, deducted from the next invoice, capped at the month's fees, and "the Customer's only recourse".
• You start the clock: the customer must have "notified the Service Failure and requested a Service Credit". No deadline is stated; a month paid late earns nothing.
• Liability overall is capped at what Koyeb received in the six months before the event.

AWS commits to 99.99% for EC2 across two or more zones, with claims until the end of the second billing cycle; Azure gives 99.99% across zones and 60 days to claim.

Status page and incidents. status.koyeb.com showed 90-day uptime on 16 October 2026 of 99.75% for builds and provisioning, 99.96% for the API and 99.97% for Europe. We found no postmortems; entries carry a one-line cause at most.

Problems with builds, deployments and scheduling are the most frequent kind and often span regions; once, on 8 September 2025, running services in both European and both US regions were hit together.

Support

Support comes with the plan: email and chat in business hours on Pro, a Slack cross-connect and "priority support" on Scale, and "24×7×365 premium support" on Enterprise. No plan has a published response target: the SLA refers to the pricing page, which gives none. A response "is deemed to be provided as soon as Koyeb sends an e-mail confirming that the request has been taken into account", so an acknowledgement stops the clock.

After the Mistral deal Koyeb said customers could "continue to reach us through the same support channels", naming "The community forum, as well as the in-app chat"; on 16 October 2026 the forum was behind a login. The team was 13 people at the sale, all moving to Mistral Compute; how many still staff Koyeb support is not stated. For comparison, AWS Basic cannot open technical cases, and AWS Business Support+ costs the greater of USD 29 or 9% of the bill for a one-hour target on a production outage.

Getting an account

The MSA admits only professional customers: the customer "cannot be qualified as a consumer". Since early 2026, "You'll need a valid payment method and must subscribe to one of our paid plans to get started." Koyeb places a USD 29 pre-authorisation on the card; validation "usually only takes a few minutes and but occasionally can take up to 3 business days", and Koyeb "may need to ask you for additional information". No identity document or deposit is documented.

Pro allows 10 members, 100 services, a CPU concurrency of 100 and a GPU concurrency of 20; Scale 50, 1,000, 1,000 and 40. Koyeb does not define the unit of CPU concurrency. If it is vCPU, workload B's 48 vCPU fit inside Pro on day one; on AWS the default of 5 vCPUs for standard instances is below workload B's 12.

Koyeb may suspend "ipso jure, without notice, and without liability" for a serious breach, an authority's request, late payment or "a recognized risk to the security and/or stability of the Platform". Unpaid bills follow a schedule: new resources blocked on day 7, "all of your Instances are paused" on day 14, and from day 21 resources "will be permanently deleted".

Compliance and data protection

Mistral AI's own certifications, if any, do not automatically cover Koyeb.

The Data Processing Agreement is dated 27 April 2021. It says "Customer Data will stay in the region selected by the Customer". Its sub-processors are Google Cloud Platform for the control plane in Belgium; Equinix Metal, IBM Cloud, Scaleway and Amazon Web Services for workloads; and Intercom and Slack in the USA. Customers get at least seven days to object to a change, though the MSA reserves "the right to modify at any time ... the choice of its technical suppliers". The DPA has no standard contractual clauses or breach notification deadline.

The list does not match the platform. Equinix Metal was to shut down by June 2026, according to Network World, and the list omits every company the evidence below points to: OVHcloud, Hetzner, Latitude.sh, Vultr, Cloudflare, Neon and AWS for Postgres, and Stripe. Koyeb's FAQ still sends customers to this DPA for "Which Cloud Service Providers are used to run my applications?". We found no newer list.

Who controls it

The owner. Koyeb SAS was registered in 2019 in Boulogne-Billancourt, next to Paris, by Yann Léger, Édouard Bonlieu and Bastien Chatelard. On 17 February 2026 it announced "a definitive agreement to join Mistral AI". The register shows Mistral AI becoming president of Koyeb SAS and a "transmission universelle du patrimoine" (the transfer of all its assets and liabilities to its sole shareholder), both published on 10 April 2026, and the company struck off on 19 May 2026.

Mistral AI SAS is a private Paris company whose president is Arthur Mensch. Its largest disclosed shareholder is ASML, at about 11% after the September 2025 round, according to press reports. On 8 September 2026 it raised EUR 3 billion in a round "Samsung Electronics led"; US investors such as a16z and NVIDIA hold stakes of undisclosed size. No percentages after that round are public, and we found no right of a non-European party to take control.

The contracting party. On 16 October 2026 the Master Services Agreement (28 June 2024), the SLA and the DPA still named Koyeb SAS, number 850 183 948. We found no document naming Mistral AI, and we did not check which company issues invoices today. The MSA chooses French law and the Paris courts.

Whose infrastructure. Koyeb owns no data centres that we could find; it runs its own scheduler on rented servers. It has not published which provider runs which region, except that its blog said in March 2025 that San Francisco is "backed by Vultr", and the aws-us-east-1 region is AWS by name. The rest of this table is our own inference from the networks that announce each region's addresses (DNS and RIPEstat, 16 October 2026): strong evidence, not a statement by Koyeb.

Three layers sit with US companies in every region. App URLs and custom domains resolve to Cloudflare addresses and Koyeb's docs say "TLS connections are terminated at the edge"; Koyeb never names Cloudflare, so the conclusion that TLS ends on Cloudflare's network is ours. The control plane runs on Google Cloud in Belgium, per the DPA. And a Frankfurt Postgres hostname is an alias of Neon's eu-central-1 endpoint on AWS; Databricks announced on 14 May 2025 that it would acquire Neon. So a Koyeb app in Frankfurt or Paris computes on servers rented from OVHcloud, Hetzner or Scaleway (in Frankfurt partly from Latitude.sh, by our inference), but its public traffic, control plane and database pass through US companies.

Leaving

The application is an OCI container image and runs on any container platform. Postgres is PostgreSQL, so pg_dump and logical replication are the way out. Koyeb holds little data by design, so exit egress is mostly the database, at USD 0.02 or 0.04 per GB depending on which Koyeb page you read. Service definitions, scaling rules, mesh names and secrets have to be rebuilt elsewhere, and volumes and their snapshots have no documented export.

The contract is short on time. "Upon expiration or termination of the MSA, Koyeb will retain the Content for a period of seven (7) days", during which the customer may ask for a copy "in a state-of-the-art format"; any complete export beyond that single one is charged by quotation.

The larger lock-in is the product's direction. Koyeb is to "gradually transition to become a core part of Mistral Compute", with the focus on "Inference, Sandboxes and serverless capabilities for MCP Servers"; the no-fee plan went first. A team building general web services on Koyeb is betting that this product line stays as it is.

What users report

User reports are thin. We read Hacker News since October 2023; Reddit gave nothing usable, and review sites do not count as evidence here.

The free plan is gone for new users. Users report "Koyeb doesn't seem to have a free tier" (47172646, February 2026) and that it had a hobby tier "before they joined mistral" (47176008, February 2026); one reader of the announcement wrote "Better start migrating" (47059047, February 2026). Archived copies of the pricing page show the Starter plan disappearing between 19 and 26 February 2026.

Easy to deploy. Users report it was "absolutely flawless to deploy there" (47051227, February 2026) and that it "has affordable GPU instances" (45616263, October 2025). These are short endorsements, not accounts of production use.

Seen as a European Fly.io. Users call it "an EU (France) alternative to fly.io" (43397151, March 2025) and point out it is "based in Paris" (45605174, October 2025); one user looking for European options after Airbus's sovereign cloud plan found it (46335412, December 2025). The data path above shows what sits behind that perception.

On reliability and support we found no two recent first-hand reports either way.

Who it suits, and who should look elsewhere

Koyeb suits small teams shipping stateless APIs, workers, inference endpoints, MCP servers or agent sandboxes who want Heroku-style simplicity with per-second billing and scale to zero, who keep object storage and other services elsewhere, and who are content to run H100-class inference in the US at a fraction of AWS's price.

Look elsewhere if you need the whole data path on European providers or a current sub-processor list; if your workload keeps state on disk; if your GPU inference must stay in Europe on anything larger than a 20 GB card; if you need certificates in hand; or if you need a broad AWS-style catalogue, an SLA without paying USD 299 a month, or a roadmap commitment for general web hosting.

Sources

Checked on 15 and 16 October 2026.

Pricing

• Pricing page and Pricing FAQ
• Public catalogue API: instances, regions and data centres
• "Slashing prices on A100, H100 and L40S", 13 November 2025
• "New Eco Instances", 12 December 2023
• Pricing page archived 19 February 2026 and 26 February 2026
• AWS and Azure figures: our baseline from the AWS Price List bulk API and the Azure Retail Prices API, converted at the ECB rate of 15 October 2026

Terms and SLA

• Master Services Agreement, 28 June 2024
• Service Level Agreement, 28 June 2024
• Data Processing Agreement, 27 April 2021

Documentation

• Regions, Instances, Edge Network, Domains, High Availability
• Volumes, Local SSD Storage, Databases, Scale-to-Zero, Autoscaling, Sandboxes
• Organizations, Accounts, General FAQ
• Changelog, "Building a global deployment platform is hard", 25 October 2023
• Terraform provider, OpenTofu registry, Pulumi registry, CLI
• Neon documentation, High availability

Status and incidents

• Status page and history feed
• Incidents of 8 September 2025, 25 May 2026, 27 May 2026, 16 June 2026, 22 June 2026 and 20 July 2026

Company and ownership

• "Koyeb is Joining Mistral AI", 17 February 2026
• "Partnering with Vultr", 13 March 2025
• BODACC notices B202600692855 and B202600692856 (10 April 2026) and B202600932437 (19 May 2026)
• French company register: Koyeb SAS and Mistral AI
• Mistral AI, Series D, 8 September 2026
• CNBC, ASML's stake in Mistral, 9 September 2025
• Databricks agrees to acquire Neon, 14 May 2025
• Network World, Equinix to end its bare metal service, 17 December 2024
• RIPEstat prefix lookups of Koyeb's ingress and outbound addresses, 16 October 2026

User reports

• Hacker News: 43397151 (18 March 2025), 45605174 (16 October 2025), 45616263 (17 October 2025), 46335412 (20 December 2025), 47051227 (17 February 2026), 47059047 (18 February 2026), 47172646 (26 February 2026), 47176008 (27 February 2026)