Network report 2025: 10 Gbps uplinks, two new regions, 99.98% delivered

Every January since 2021 we publish what we built, what we broke and what it cost. Here is 2025.

The numbers

Metric	2024	2025	Change
Measured fleet uptime	99.97%	99.98%	+0.01 pt
Active instances	28,400	41,800	+47%
Locations	19	20	São Paulo
Median first support reply	17 min	12 min	-29%
Average delivery time	6 min 40 s	4 min 05 s	-39%
Refund rate	1.6%	1.4%	-0.2 pt
Largest DDoS mitigated (cumulative record)	612 Gbps	1.14 Tbps	set June 2024

What we shipped

10 Gbps uplinks on Titan and Ultra

Every Titan and Ultra instance moved from 2 Gbps to a 10 Gbps port in Q2, at no extra cost to existing customers. The bottleneck had shifted to the vSwitch, so we also moved those nodes to SR-IOV passthrough, which cut p99 network latency inside the node by 38%.

São Paulo, region twenty

A region customers had asked for since 2022 and we kept postponing until we could do it properly — meaning our own IPv4 allocations and at least three transit providers, not a single-homed rack. It went live in August. Latin America had been served from Panama City since 2023 and from Dallas before that; Brazilian and Southern-Cone users now get 31 ms from Buenos Aires instead of 100+.

Provisioning rewrite

The old provisioner built machines serially. The new one parallelises image write, IP assignment and DNS registration, which is where the 39% delivery-time improvement came from. Windows now uses a pre-syspreped differencing image per version, cutting first-boot time from about 90 seconds to 22.

Two-factor everywhere

TOTP became mandatory for client-area accounts with active services in October. Adoption is 100% because we did not offer an opt-out. Account-takeover tickets went from 31 in 2024 to 2 in 2025, both of which were email compromises, not credential stuffing.

What broke

We publish incidents in full on the status page. The three that mattered:

14 March — Frankfurt, 47 minutes

An upstream router at our transit provider failed to withdraw a route during maintenance, blackholing about 30% of Frankfurt-bound traffic. Our monitoring saw it in 90 seconds; the provider took 44 minutes to roll back. We have since added a third transit provider in Frankfurt and moved BFD timers down to 300 ms so we can fail away in under a second.

2 July — Singapore, 3 h 12 m degraded

A storage controller firmware bug caused elevated write latency on four nodes. Not an outage, but customers noticed and were right to. We rolled back the firmware and now stage storage firmware for 30 days on internal nodes before any production rollout.

19 November — client area, 22 minutes

A bad database migration during a deploy locked the invoices table. Servers were entirely unaffected; nobody could pay or open a ticket. Migrations are now gated behind a lock-duration check that aborts anything estimated over two seconds.

Total customer-visible downtime across the fleet: 1 h 41 m weighted by affected instances, against an annual SLA budget of 4 h 22 m at 99.95%. Credits were applied automatically without anyone having to ask.

What we did not do

We did not launch a Kubernetes product, an object store or a managed database. Every year we consider it and every year we conclude that the reason customers stay is that a machine they bought in 2021 still works exactly as it did. Boring infrastructure is the product.

We also did not raise prices. Hardware costs fell enough to absorb higher power costs in Europe. Customers on 24-month terms are price-locked regardless.

2026 plan

• IPv6 by default — a /64 on every instance, no request needed. In progress, Europe first.
• Two more regions — Seoul and Johannesburg, subject to getting proper transit diversity.
• Self-service migration between locations with a snapshot restore, targeted for Q3.
• Public API v2 with Terraform provider. The v1 API stays supported until at least 2028.

As always: if something on this list matters more to you than the others, tell support. We weight the roadmap by what people actually ask for.