“We were essentially able to reduce the cost of that cluster by about 75%. On AWS, DevZero demonstrated they could achieve significantly higher savings than we initially thought possible.”
Mihir Nair
Head of Architecture, Databahn
Kubernetes infrastructure operations
that actually scale
DevZero continuously analyzes your Kubernetes clusters and cuts infrastructure costs without touching application performance. Unlike Karpenter, it adjusts CPU and memory in place using MPA. No node replacement. No restarts.
Initial State
Applying Policy
Optimized
Before
After
c5.4xlargeOn-demand
$0.680/hr
c5.2xlargeOn-demand
$0.340/hr
c5.xlargeOn-demand
$0.170/hr
+3more nodes
6 nodes all
on-demand
on-demand
c5.4xlargeOn-demand
$0.680/hr
c5.2xlargeOn-demand
$0.340/hr
c5.xlargeOn-demand
$0.170/hr
+3more nodes
On-demand — all 6 nodes
avg utilization: 28% · 72% wasted28% utilized / 72% idle waste
All 6 nodes running on expensive on-demand instances
Companies who slashed their Kubernetes
spend using DevZero
Every resize stays inside your policy
DevZero adjusts CPU and memory based on observed workload behavior. Choose Conservative, Balanced, or Aggressive per namespace. Hard min/max boundaries define what recommendations can never cross. autoApply is off by default. Changes queue for approval until you enable it. Models are per workload, not cluster-wide averages.
Cost Overview
Current CostUsage Cost
Avg. Monthly Projection
$136.84K/ mo▼2%
Average estimate based on request patterns during the period.
Right-Sized Projection
$44.36K/ mo▼68%
Estimated monthly cost if requests were optimized to usage.
Period Cost Delta
▼2.0%
Comparison of total cost against the previous period.
$210.00$180.00$150.00$120.00$90.00$60.00
Apr 1 at 00:00Apr 1 at 04:00Apr 1 at 08:00Apr 1 at 12:00
Cost attribution
Spend mapped to the workload consuming it
cluster / month
$14k
total infra spend
saved / month
$6.3k
after rightsizing
cluster efficiency
83 /100
up from 41/100
Cost by team
ml-team↓ 31%
$5,060
data↑ 4%
$3,840
platform↓ 59%
$1,820
web↓ 64%
$0
$2,500$5,000+
$6,340 attributed savings this monthTrend data: 30-day lookback per team
Spend mapped to the workload consuming it
Cost attribution follows the same namespace filters and label selectors in your policies. The data matches exactly how your team scoped their rules. Drill from cluster to individual pod. Trend data surfaces cost drift before it becomes a problem.
One policy or fifty the fleet adapts
Target by namespace, label, kind, or name pattern. Precedence is clear when rules overlap. autoApply is set per policy; production waits for approval while dev runs automatically. Node group policies are separate. Instance families, spot vs on-demand, and consolidation aggressiveness are independent from workload rules.
The fleet adapts
Most specific policy wins · precedence is always clear
Priority1By nameapi-server
WINSPriority2By labelapi-server
2nd matchPriority3By namespaceproduction
3rd matchPriority4Defaultfallback
last resortautoApply behavior per namespace
autoApply: false
production
waits for manual approval
autoApply: truemode: aggressive
dev / staging
applies changes automatically
↻WorkloadRequests Based
CostOptimizationsHealthNamespace
CostOptimizationsHealthNamespace
ml--nar-earth-performan…
CPU: $5.3172 · Mem.: $0.7060 · GPU: $21.80
ml--nar-earth-staging-u…
CPU: $5.5440 · Mem.: $0.3409 · GPU: $15.8C
ml--nar-mars-performan…
CPU: $4.7951 · Mem.: $0.6284 · GPU: $16.32
ml--mars-staging-use1
CPU: $5.4751 · Mem.: $0.3218 · GPU: $13.65
ml--mercury-staging-use1
CPU: $8.4314 · Mem.: $0.6993 · GPU: $0.941
nuwa-embedding-servic…
CPU: $2.4192 · Mem.: $0.9325 · GPU: $0.00
ml--apac-mars-perform…
CPU: $4.0049 · Mem.: $0.5248 · GPU: $20.15
ml--earth-performance-…
CPU: $4.0644 · Mem.: $0.5397 · GPU: $20.17
Different workloads per workload, not per cluster
Write scoped policies per workload type. Tight boundaries and manual approval for production. Aggressive P90 rightsizing for dev and staging. A dedicated GPU policy targets nvidia accelerator labels directly. Node group policies handle spot preference and consolidation aggressiveness per pool, fully independent from workload rules.
Built for platform engineers
The controls your team actually needs, not a dashboard that requires a PhD to interpret.
cgroup v2 · live
CPU throttle
78%Mem PSI stall
43msOOM events
0sub-second
prometheus · 15s avg
CPU usage
38%Mem usage
51%Throttle visible
—averaged out
P99 latency spike · api-server-7f6d9
throttle_pct=0.78 · caught in 180ms window
Detected
DevZero reads CPU throttle rates, memory pressure (PSI), and OOM events directly from cgroup v2, not 15-second Prometheus averages. Sub-second anomalies that cause P99 latency spikes are visible and actionable.
EKS
841pods managed
GKE
519pods managed
AKS
302pods managed
On-prem
178pods managed
Single tenant1,840
pods · 4 clusters$41kMonthly Savings
3Cloud Providers
4Clusters
1Tenant
One DevZero tenant manages rightsizing across EKS, GKE, AKS, and on-prem clusters. Monthly savings and optimization metrics are aggregated in a single view across all cloud providers.
DEVZERO
cpu.request: 2000m → 680m · conf=0.94SLACK · #PLATFORM-RIGHTSIZING
Advisory mode · saves $2.84/hrGITHUB · HELM-VALUES PATCH
resources.requests.cpu: "680m"Advisory mode surfaces recommendations where your team already works. GitHub PR integration proposes changes before anything touches production. Slack integration posts to the on-call channel, keeping the right people informed.
What our customers say
Frequently asked questions
Stop paying for CPU and memory you're not usingStart optimizing your infrastructure today
Connect your cluster in under 30 minutes. No code changes. No pod restarts. First savings visible within 24 hours.