● iaas · compute & networknot OpenStack

Supercharged computeand network. No legacyfrom 2014.

(translation pending)

EPYC · Xeon Platinum · NVMe Gen5 · live migration · per-second billing
my-app · c5.large · msk-1
running
h3 vm create
$

Why another IaaS — when OpenStack already exists

OpenStack clouds are assembled from 2010-era parts. The result is a layer of duct tape: Keystone breaks, Neutron is slow, Heat doesn't work. We kept only what was good (KVM/QEMU, OVS) — and rewrote the rest.

alternative A

An OpenStack cloud

  • Open-source, in theory
  • Keystone breaks on every release
  • Neutron is a latency bottleneck
  • Heat doesn't work in real life
  • Cold-start VM: 60–120 seconds
h3llo · iaas

h3llo compute and network

  • KVM/QEMU 9 + Open vSwitch — no OpenStack wrapping
  • EVPN-VXLAN spine-leaf, ≤ 50 µs between VMs in an AZ
  • VM in 11 seconds from POST /vms to SSH
  • Live migration with no application downtime
  • NVMe Gen5 replicated, 2.5M IOPS
  • Terraform, CLI, REST API — take your pick
alternative B

VMware on-prem

  • Solid for 20 years
  • Licenses in USD, 5-year vendor lock
  • Your own hands, your own DC, your own team
  • vSphere API — a universe of its own
  • Scaling = a new hardware order
By the way, Kubernetes is a separate product. If your stack is already on K8s, start with/mk8s →
● dx

VM in 11 seconds. CLI or Terraform — your call.

Pre-warmed images, parallel bootstrap, no-op cloud-init by default. SSH is ready before Slack finishes loading.

h3 vm create · CLI11 seconds
# spin up a VM with one command
$ h3 vm create my-app \
    --flavor=c5.large \
    --image=ubuntu-22.04 \
    --region=msk-1 \
    --network=default

→ Creating VM (4 vCPU · 16 GB · 100 GB NVMe)
→ Allocating SDN port (vxlan-vni 8421)
→ Bootstrap cloud-init (no-op default)
✓ VM ready in 11 seconds

$ ssh root@95.183.21.44
Welcome to Ubuntu 22.04 (5.15.0-h3llo)
root@my-app:~#
Terraform · IaCprovider
terraform {
  required_providers {
    h3llo = { source = "h3llo/h3llo", version = "~> 0.4" }
  }
}

resource "h3llo_vm" "app" {
  name     = "my-app"
  flavor   = "c5.large"
  image    = "ubuntu-22.04"
  region   = "msk-1"

  network {
    name      = "default"
    public_ip = true
  }

  disk {
    size_gb = 100
    type    = "nvme-replicated"
  }
}
● flavorsall available

VM flavors — no hidden limits

Prices are per-second pay-as-you-go. Reserved (1 year) — −30%, dedicated host — separate. You can switch between flavors live (no downtime).

FlavorvCPURAMNetworkPrice
c5.small1 vCPU2 GB1 Gbps1.30 ₽/hr
c5.largebest · workhorse4 vCPU16 GB5 Gbps5.40 ₽/hr
c5.xlarge8 vCPU32 GB10 Gbps10.80 ₽/hr
m4.compute8 vCPU32 GB10 Gbps8.00 ₽/hr
m4.memory16 vCPU128 GB25 Gbps26.40 ₽/hr
r4.4xlarge16 vCPU256 GB25 Gbps38.00 ₽/hr
g1.gpu16 vCPU96 GB100 Gbpsfrom 250 ₽/hr (B300 MIG)
bare-metal · m664 vCPU512 GB2× 100 Gbpsfrom 28,000 ₽/month

Prices for compute. Disks, IPv4 and traffic are billed separately — full pricing in /price.

● benchmarksour numbers vs typical legacy

Concretely: what we have and what others have

Not "modern infrastructure" from marketing brochures. Real numbers on hardware, network and traffic — compared with typical clouds on Xeon gen 2/3 and 25G uplinks.

Compute · CPU + RAM

Xeon 5 + DDR5 vs Xeon 2/3 + DDR4

h3llo · Xeon Platinum 8480+ · DDR5-4800247
Other · Xeon Gold 6248 · DDR4-2933162

+52% IPC on SPECrate2017, +63% memory bandwidth. Your DB on DDR5-NVMe-NVLink — not on 7-year-old 2018 hardware.

Network · uplink per rack

400G spine vs 25G legacy

h3llo · 400G spine + 100G leaf · EVPN-VXLAN400 Gbps
Other · 25G legacy aggregation25 Gbps

16× more uplink per rack, p99 ≤ 50 µs between VMs in the same AZ. DPDK fast path, SR-IOV, RDMA — no bottlenecks in the network plane.

Egress · outbound traffic

Free egress vs paid

h3llo · up to 1 TB/month · internal traffic 0 ₽0 ₽
Other · 1 TB egress at typical pricing8,400 ₽

Up to 100,000 ₽/year in savings on a typical SaaS stack. Beyond 1 TB — 1.50 ₽/GB (6× cheaper than AWS egress).

CPU numbers — SPECrate2017 single-thread, geomean. Network — fabric uplink per rack. Egress — competitors' public pricing for outbound traffic up to 1 TB/month.

● use cases

What people typically spin up VMs for

01 / web
Web and API hosting
c5.large is the typical pick. SDN with private VPC, public IP, security groups. Auto-healing, snapshot in 4 seconds.
02 / db
Databases
r4 for in-memory + NVMe Replicated disk. 3× replication, auto-failover ≤ 60 sec, snapshot to S3.
03 / bastion
Bastion and edge
c5.small + dedicated public IP. SR-IOV port for line-rate firewall. Audit log to S3 by default.
04 / hpc
HPC and compute-heavy
Bare-metal m6 for a cluster or dedicated host. 100 Gbps RDMA network, NUMA-aware scheduler.
● securityall included

Security isn't an add-on — it's the foundation

Hypervisor-level isolation, encryption by default, 800 Gbps DDoS protection — included for every VM, no upgrades, no extra fees.

ISO

Hardware-level isolation

Every VM gets its own SR-IOV port on the NIC and NUMA-aware placement. No noisy neighbors, no shared disk queue.

KVM/QEMU 9 · SR-IOV · NUMA pinning
KEY

Encryption by default

Disks — LUKS on NVMe self-encrypting drives. Traffic inside the VPC and outbound — TLS 1.3. Keys — in our KMS (HSM-backed) or in yours.

AES-256 at-rest · TLS 1.3 · KMS / HSM
DOS

800 Gbps DDoS protection

L3/L4 protection at the network edge — on for every VM by default, no extra cost. The "you're under attack" scenario is closed by automation in tens of seconds.

800 Gbps · L3/L4 · auto-mitigation
GOV

Compliance out of the box

FZ-152, FZ-242 (data in-country), FSTEC (K2), PCI DSS 4.0, ISO 27001/27017, GOST R 57580. Ready to take part in your audits — checklists, policies, evidence base.

8 certifications · copies under NDA

Full list of certificates and attestations (8 of them) — at /infrastructure → compliance. Copies — on request under NDA.

● pricing

Pay however suits you — per-second or reserved

Pay-as-you-go for dev and variable load. Reserved for production (−30%). Dedicated host for compliance and regulated markets.

pay-as-you-go
Per-second billing (rounded to the minute), no minimums, no reservations. Ideal for prototypes and dev/staging.
from 1.20 ₽ / vCPU·hr
  • All flavors available
  • Per-second billing
  • Snapshot and backup to S3
  • Public IP + DNS
  • 8×5 support
popular
reserved
A year of reserved instances — −30% off list. Guaranteed capacity, fixed invoice. For production workloads.
from 0.84 ₽ / vCPU·hr (−30% year)
  • Guaranteed capacity
  • Fixed invoice
  • 1 / 3 years · discount up to 45%
  • 24×7 support · response ≤ 1 hr
  • Rule-based auto-scaling
dedicated host
A whole bare-metal box at your disposal. Isolated compute, for compliance and regulated markets.
from 28,000 ₽ / month · m6 host
  • A whole physical server
  • Isolated hypervisor
  • FZ-152, FSTEC, PCI DSS
  • SLA 99.99% · 10× compensation
  • Dedicated SRE
● materialsfree

Guides and case studies on IaaS

Real benchmarks, migrations, architecture deep dives. No fluff, no marketing.

All materials →
● quickstart

From install to SSH — 60 seconds

No portal, no YAML manifests, no container builds. Just handler → deploy.

Let's go →
1

Install the CLI

brew install h3llo/tap/cli · or curl get.h3llo.cloud | sh
2

Create a VM

h3 vm create my-app --flavor=c5.large · 11 seconds to SSH-ready.
3

Connect

h3 vm ssh my-app or direct SSH to the public IP. No bastion setup.
4

Scale via Terraform

A ready h3llo_vm provider. From the outside — your usual Terraform flow.
● faq

What people usually ask

Is this really not OpenStack?
Really. With OpenStack — Keystone breaks on every release, Neutron is a bottleneck, Heat doesn't work in real life. We kept only what was good (KVM/QEMU, Open vSwitch as a forwarder, the hypervisor API) and rewrote the orchestrator and control plane from scratch. No Nova/Neutron/Cinder/Heat surprises.
Which hypervisor is it?
KVM with QEMU 9.x. SR-IOV for networking, virtio-blk for disks, vhost-user for NICs. On GPU nodes — full passthrough or MIG partitions on B300.
How long does a VM take to start?
Median 11 seconds from POST /vms to SSH ready. Pre-warmed images, parallel bootstrap of network and disk, no-op cloud-init by default.
And the network — what's under the hood?
Spine-leaf on Arista with EVPN-VXLAN overlay. Each VM gets an SDN port via the DPDK fast path. Between VMs in the same AZ — bare-metal latency (≤ 50 µs p99). No Neutron bottleneck.
Does live migration work?
Yes, with no application downtime. Pre-copy + Post-copy fallback. We've migrated workloads at 4K TPS without a single dropped connection.
How much does it cost?
vCPU 1.20 ₽/hr, RAM 0.50 ₽/GB·hr, NVMe 16 ₽/GB·month, SSD 7.50 ₽/GB·month. Reserved (1 year) — −30%, dedicated host — from 28,000 ₽/month. Exact prices in /price.
● compute, rewritten after 2020

IaaS without OpenStack pain —
just VMs and a network that work

No Keystone glitches, no Neutron bottlenecks. KVM/QEMU 9, EVPN-VXLAN spine-leaf, NVMe Gen5. Per-second billing, live migration, FZ-152 compliant.

Create a VM →Documentation