the optimizer · ~8%
Trajectory: declining
Active remediation. FinOps embedded in engineering. One Optimizer reduced daily cloud spend by 50% in 9 weeks — no migration, no rebuild, just hygiene.
The State of
Cloud 2026
An Emerging Markets view of global cloud. How enterprises in India, Southeast Asia, the Middle East, and Latin America actually run cloud, and what the rest of the world should learn from them. Grounded in telemetry from 25 production cloud accounts.
publishedJune 2026 · ZopDev reading time23 min read
about · why the lens matters
Most “State of Cloud” reports are written from Silicon Valley, for Silicon Valley.
They assume a USD-denominated cloud bill, US labor costs, US regulatory pressures, and US patterns of adoption. Their case studies are American. Their benchmarks are American. Their advice is implicitly addressed to an American CIO. That coverage is fine. It is also incomplete.
India is now the third-largest cloud market in the world and the fastest-growing. The Middle East is investing over $50B in sovereign infrastructure. Southeast Asia is leapfrogging traditional IT entirely. Latin America runs cloud under currency conditions that fundamentally change the economics. None of this is visible in the existing reports, not because the topics are unimportant, but because the analysts writing the reports are not close to the data.
We are. This report is grounded in primary telemetry from 25 production cloud accounts across 13 enterprises, predominantly India-headquartered and India-presence multinationals, with operations spanning Asia-Pacific, the Middle East, Europe, and the Americas.
The 16.9% waste finding holds whether you operate from Mumbai, Manchester, or Mountain View.
How that waste accumulates is what differs by geography
executive summary
Cloud has stopped being new. So why is the bill still a mystery?
After fifteen years of migration, the question is no longer whether to be on the cloud. It is why the cloud bill is what it is, and what it could look like if we treated it like a product.
the five findings that matter
What this report found, in one screen.
| # | Finding | What it means |
|---|---|---|
| 1 | Waste is hygiene, not strategy | 16.9% of every cloud dollar is recoverable waste, and 87% of those items are forgotten resources, not bad architectural decisions. |
| 2 | Detection vastly outpaces action | Cloud teams identify waste roughly 30× faster than they fix it. The bottleneck is execution, not awareness. |
| 3 | Three personas, three trajectories | Optimizers cut spend 50% in 9 weeks. Drifters grow 10% a quarter without intent. Scalers grow 4× and bake in tomorrow’s waste. Most enterprises are Drifters. |
| 4 | The Emerging Markets lens reveals different mechanics | Currency exposure, sovereign cloud pressure, and labor-cost economics shape cloud decisions in ways US-centric reports systematically miss. |
| 5 | The GPU shift is the next inflection | Today, AI/GPU spend is under 2% of enterprise cloud bills. By Q2 2027 it crosses 15%. The same anti-patterns return, with three more zeros on every waste number. |
the three industry archetypes
Every enterprise we examined fell into one of three trajectories.
the drifter · ~50%
Trajectory: flat-rising
Detection in place, action absent. Waste grows quietly, 5–10% a quarter. The recommendations exist but nobody owns them. The default state of enterprise cloud.
the scaler · ~42%
Trajectory: 2–4× growth
Growth-stage spend. New regions, workspaces, accounts every week. Each new service spins up its own NAT gateway, scratch volume, snapshot policy.
Roughly $1 in every $6 of cloud spend is recoverable through hygiene alone.
Before any architectural change, renegotiation, or repatriation
how to read this report
CTO
Ch 6 (Architectural Picture), Ch 7 (Emerging Markets View), Ch 9 (GPU Shift)
Cloud Architect
Ch 4 (Anti-Pattern Catalog), Ch 6 (Resource Sprawl), Appendix B
FinOps Lead
Ch 2 (the 16.9% number), Ch 5 (the Action Gap), Ch 3 (the three personas)
Board / Investor
Ch 1 (Hypothesis Scoreboard), Ch 7 (Emerging Markets View), Ch 8 (Predictions)
CFO
Ch 7 (currency exposure section), Ch 2 (16.9%)
methodology · at a glance
The sixty-second version. Full method in Appendix A.
- cloud accounts
- 25
- enterprises
- 13
- observation window
- 69 days
- cloud regions covered
- 32
The data pyramid
The findings rest on three independent layers of evidence, ordered from most to least reliable.
| Layer | Source | What it tells us |
|---|---|---|
| Telemetry | Live cloud account APIs across AWS, Azure, GCP | Ground-truth resource inventory, billing data, and utilization from 25 production accounts. Cannot be biased by respondent intent. |
| Recommendations | A FinOps analytics platform with 315+ detection rules | Programmatic detection of waste, compliance gaps, and architecture anti-patterns, applied uniformly across all accounts. |
| Secondary | Hyperscaler earnings, FinOps Foundation, CNCF, public 10-Ks | Industry context to validate that the cohort patterns generalize beyond the sample. |
What we measured: $420,554 of spend across the 69-day window ($2.22M annualized), 25,225 resources across 36 service types, 32 active regions, spanning FMCG, e-commerce, consumer internet, B2B logistics, sporting goods, global beverages, and creator-economy SaaS.
chapter 1 · hypothesis scoreboard
Ten hypotheses, written before analysis began.
Most industry reports describe what the data showed. We did the opposite. Before opening a single dashboard, we wrote ten falsifiable hypotheses. This is the only honest way to do research.
1 · each scored against the evidence
10 / 10Every hypothesis, written before a single dashboard was opened, was confirmed by the evidence.
- 1 Confirmed
AWS dominates enterprise spend, even where account count favors others
53% of accounts; 65.5% of spend on AWS
- 2 Confirmed
Multi-cloud is rarer than the narrative suggests
Only 1 of 13 ran production across all three hyperscalers
- 3 Confirmed
Waste concentration follows a power law
76% of savings sit in 3 of 13 environments
- 4 Confirmed
Orphan resources outnumber strategic waste
87% of all flagged items are orphans
- 5 Confirmed
Customers detect waste 10× faster than they act
Actual ratio: ~30×. More extreme than predicted
- 6 Confirmed
Dev/test runs 24/7 by default in prod accounts
67% potential savings on dev VMs through scheduling
- 7 Confirmed
Databricks creates a predictable Azure anti-pattern
Premium SSD in non-prod, confirmed across 2 enterprises
- 8 Confirmed
EBS snapshot sprawl is the universal AWS cost
3,807 snapshots in one customer; orphans in 4 of 6 AWS users
- 9 Confirmed
Most accounts have a Single-AZ production database
36+ flagged in one customer; pattern visible everywhere
- 10 Confirmed
AI/GPU spend is still a single-digit share
Under 2% of cohort spend on AI services
All ten hypotheses confirmed. The patterns we predicted are real and well-documented, and the industry’s conventional wisdom about cloud is broadly correct — what’s missing is the will to act on it.
chapter 2 · the defensible number
If you take only one statistic from this report, take this one.
2 · the defensible number
16.9¢
of every cloud dollar is recoverable waste, sitting in plain sight.
For an enterprise spending $10M annually on cloud, that is $1.69M sitting on the table — recoverable without architectural changes, vendor renegotiations, or workload repatriation. Across the global enterprise cloud market (near $680B in 2026), the implied addressable waste is north of $100B annually.
Annualized cloud spend · cohort
$2.22M
$2,222,930 total
Recoverable savings · proven
$375K
$375,217 annualized
Waste-to-spend ratio
16.9%
$375,217 ÷ $2,222,930
In Indian-rupee terms, the addressable waste at a single mid-sized enterprise can exceed ₹50 crore per year — and the currency conversion makes it sharper, because the cloud bill is USD-denominated while the revenue defending it is not.
This is the floor, not the ceiling
We only counted recommendations we could prove with high confidence. We excluded architectural rebuilds, workload repatriation, commitment renegotiations, storage-tier migrations without access evidence, and application-layer efficiency. Including these would likely double the number — the true recoverable waste is probably 30–40% of cloud spend. But we refuse to claim a number we cannot prove.
It is not “can we cut cloud spend?” It is “why haven’t we already?”
16.9% is the headroom you don't know you have
chapter 3 · the state of enterprise cloud
Three views: spend, footprint, and trajectory.
3.1 · spend vs market share
Spend by provider tells a different story than account count.
Azure has fewer accounts in our cohort but commands disproportionate enterprise spend. GCP is the long tail. AWS still takes two-thirds of every cloud dollar.
- AWSDominant in spend & accounts · 53% of accounts65.5%
- AzurePunches above its weight · 28% of accounts25.7%
- GCPConcentrated in product / SaaS · 19% of accounts8.7%
AWS still takes two-thirds of every cloud dollar. Account count and spend are not the same shape — Azure commands disproportionate enterprise spend on far fewer accounts.
3.2 · multi-cloud
A strategy more talked about than practiced.
In our cohort, exactly one enterprise in thirteen ran production workloads across all three hyperscalers. Of the rest:
- 6 of 13 are single-cloud AWS (the modal pattern)
- 4 of 13 are single-cloud Azure (concentrated in regulated industries)
- 2 of 13 are single-cloud GCP (concentrated in product/SaaS)
- 1 of 13 is genuinely multi-cloud across all three
The “multi-cloud strategy” in industry conversation is often code for multi-cloud failover plans nobody tested, not parallel production workloads. The latter remains rare.
3.3 · footprint is a choice
Most enterprises have not made it consciously.
| Footprint pattern | Example (anonymized) | Regions | $ / region / mo |
|---|---|---|---|
| Global sprawl, one tenant | Global sporting goods, Azure | 32 | ~$60 |
| Concentrated production | Global CPG, Azure | 3 | ~$5,900 |
| Single-region monolith | Consumer internet, AWS | 2 | ~$12,500 |
| Spread without intent | E-commerce, AWS | 19 | ~$560 |
Two of these patterns reflect deliberate architectural choices. Two reflect accident. Footprint is now a choice, but most enterprises have not made that choice consciously.
the most important chart in this report
Some enterprises cut spend by half in a quarter. Others quietly inflate by 4×.
the optimizer · ~8% · declining
$1,565/day → $786/day in 69 days
Active FinOps embedded in engineering. Recommendations get triaged, owners get assigned, savings get tracked. No migration. No rebuild. Pure hygiene.
the drifter · ~50% · flat-rising
$5,787/week → $6,361/week over 8 weeks
Tooling in place; nobody owns the output. Cost grows quietly, 5–10% a quarter from accumulated unfixed waste. No new product launch. Just drift.
the scaler · ~42% · 2–4× growth
$1,429/week → $5,918/week in 6 weeks
Growth-stage spending. Each new team gets its own workspace, region, NAT gateway, snapshot policy. The growth is legitimate. The waste embedded in it is not.
3.4 · the bimodal economy
The average masks the story.
Optimizers shed waste fast. Scalers accumulate it fast. Drifters stay stuck in between. Every cloud bill is on one of these three trajectories.
The Optimizer · daily spend
−50%
$1,565 → $786 · 69 days · pure hygiene
The Drifter · weekly spend
+10%
$5,787 → $6,361 · 8 weeks · no launch
The Scaler · weekly spend
4.1×
$1,429 → $5,918 · 6 weeks · embedded waste
The “average enterprise cloud bill” is a meaningless number. The story is in the spread — one customer halves spend while another quadruples it, in the same quarter.
Which trajectory are we on, and is it the one we chose?
chapters 1–3 · recap
So far
- 10/10hypotheses confirmedAll ten pre-registered hypotheses were confirmed — the conventional wisdom is right; the will to act is what’s missing.
- 16.9%recoverable waste16.9% of every cloud dollar is provably recoverable waste; the true figure is likely 30–40%.
- 65.5%share on AWSAWS takes 65.5% of spend; genuine multi-cloud is 1 in 13; footprint is mostly accidental.
- 5-10%quarterly driftThree trajectories — Optimizer, Drifter, Scaler. Most enterprises are Drifters, inflating 5–10% a quarter without intent.
chapter 4 · where the waste lives
87% of cloud waste isn’t strategic. It’s forgotten.
Vendor-led narratives focus on big strategic moves: Reserved Instances, Savings Plans, Spot. These account for only 6% of the recommendations we flagged. The real waste is mundane: things nobody remembers creating.
4 · the waste category distribution
Orphans — unattached volumes, orphan snapshots, abandoned IPs — take ~5 minutes each to fix. The biggest cloud waste problem isn’t that you bought the wrong thing. It’s that you forgot to delete the right thing.
Why orphans dominate
1
Provision generously, clean up rarely
The friction of provisioning is near zero; the friction of de-provisioning is enormous.
2
No one owns the lifecycle
Accounts are organized by team, environment, or application. Almost never by lifecycle.
3
Snapshots compound silently
Every dev creates a snapshot before risky changes. Almost no one deletes them. Snapshots can exceed live volumes by 14×.
4 · the hall of shame
Fix only these ten things, and you recover 80% of the waste.
| # | Anti-pattern | Universality | $/yr at stake |
|---|---|---|---|
| 1 | Orphan EBS snapshots — no source volume, charged forever | 100% of AWS users | $80K+ |
| 2 | Unattached EBS volumes — single volumes leak $2,500/yr | 100% of AWS users | $25K+ |
| 3 | Premium SSD in non-prod Databricks (Azure) | 100% of Databricks on Azure | $66K+ |
| 4 | EC2 not covered by Savings Plans | ~90% of prod fleets | $50K+ |
| 5 | Single-AZ production RDS — not cost, risk | Every AWS enterprise | Resilience risk |
| 6 | Dev/test workloads with no schedule | 100% of dev environments | $40K+ |
| 7 | Windows VMs not using Azure Hybrid Benefit | Common in regulated industries | 40% on license |
| 8 | On-demand for stateless / batch workloads | Common in cloud-native shops | $30K+ |
| 9 | x86 where Graviton/ARM would work | 80% of compute fleets | $25K+ |
| 10 | Stopped instances still incurring EBS charges | Found in every account | $5K+ |
None of these are strategic decisions to revisit. They are hygiene to install. If your platform team cannot recite this list from memory, you have a 16.9% problem.
chapter 5 · the detection-to-action gap
Detection has been solved. Execution has not.
Cloud teams identify waste roughly 30× faster than they act on it. The cloud-cost industry has spent a decade building detection tools. Action remains an open problem.
5 · the funnel
- waste signals detected
- 8000+
- auto-resolved by churn / turnover
- 6300+
- explicitly reviewed by a human
- ~250
- formally applied via remediation
- <24
<1%
of detected waste is formally remediated
~80%
auto-resolves through churn, not intent
30×
faster to detect waste than to act on it
Fewer than 1% of detected waste signals are formally applied through a remediation workflow. The platform stops at the dashboard. The work happens — or doesn’t — in the cloud console.
Why the gap exists — all human, none technical
Owner ambiguity
The account belongs to platform. The workload belongs to the app team. The cost belongs to finance. Nobody owns the recommendation.
Risk asymmetry
Deleting a snapshot has a non-zero chance of breaking something. Not deleting it has a 100% chance of costing money. The first risk is visible and personal; the second is diffuse.
No apply-path
Most FinOps platforms show recommendations. Almost none execute them safely.
The Optimizer in our cohort cut spend 50% in 69 days. Their secret was not better detection — they had the same tools as everyone else. Their secret was that someone owned the list, every day, and worked through it.
The market for FinOps that detects is mature. The market for FinOps that acts is wide open.
the state of cloud 2026 · chapter 5 · the detection-to-action gap
chapter 6 · the architectural picture
Resource sprawl is the new technical debt.
For CTOs and cloud architects. The visible compute layer of any production account is dwarfed by its metadata.
6.1 · the metadata iceberg
Metadata scales even when workloads don’t.
- EBS snapshots vs volumes
- 14×
- CloudWatch alarms per EC2
- 22
- IAM roles per workload
- 3
- snapshots / volume / year
- 2.3
The compute layer is the tip. The metadata is the iceberg. And the metadata, unlike the compute, has no lifecycle. The bill grows linearly with metadata, even when workload is flat.
6.2 · the fragmentation problem (azure)
Consolidators vs Fragmenters.
consolidators
3–5 large subscriptions
Cost-per-subscription is high (~$15–20K/month). Governance is tractable.
fragmenters
20–35 small subscriptions
Cost-per-subscription is low (~$500–1,500/month). Governance scales linearly with team count. Cleanup is hopeless.
One customer ran 33 distinct Azure subscriptions in a single tenant. Another ran 3 for comparable spend. The first is governance debt waiting to compound. The second is governance done right. The most visible symptom of fragmentation: Databricks workspace proliferation, each spinning up its own NAT gateway and Premium SSD scratch volumes.
6.3 · resilience theater
Every AWS enterprise had at least one Single-AZ production database.
The fix is trivial. The cost is a small uptick in storage. The benefit is the difference between a 5-minute outage and a multi-hour outage when an AZ fails. The reason is never “we made a deliberate choice.” It is always “we ran the migration script in 2021 and never went back.”
6.4 · The Observability Tax
One production AWS account ran 6,922 CloudWatch alarms — the vast majority unmonitored, with no SNS destination, no escalation, no on-call rotation listening. They were noise generators paying their own bill. Treat your monitoring stack like your production stack: lifecycle policies, deprecation windows, ownership tags.
chapter 7 · an emerging markets view
Where the global narrative breaks down.
Everything in the first six chapters applies universally. But the conditions surrounding those findings — currency, regulation, labor economics — vary enormously across geographies. And they change which interventions are realistic.
7.1 · currency exposure is a cloud risk
A 10% currency swing is a 10% cost increase.
Hyperscaler bills are denominated in US dollars. Revenue at most emerging-market enterprises is denominated in local currency. A 10% adverse currency movement — which has happened to the rupee, real, and rand multiple times in the last five years — translates directly into a 10% cloud cost increase, without a single new resource being provisioned.
Multi-year RI commitments look attractive on US dashboards because they lock in dollar pricing — and risky on Indian dashboards because they lock in dollar exposure.
Repatriation conversations begin earlier in emerging markets, because the local-currency math turns favorable sooner. And sovereign cloud offerings denominated in local currency command a premium even at higher list price, because the currency hedge is itself valuable.
7.2 · sovereign cloud is not just european
The more aggressive activity is happening outside Europe.
India has formalized data-localization for banking, payments, and increasingly healthcare. The UAE and Saudi Arabia are both investing tens of billions in sovereign infrastructure, with explicit goals of becoming regional cloud hubs. Indonesia is moving similarly for Southeast Asia.
Data-residency requirements at emerging-market enterprises are stricter and more rapidly evolving than at US enterprises. Workloads must be region-pinned. Backup and DR topologies cannot rely on cross-border replication that worked five years ago. The pace of regulatory change still exceeds the pace of region availability.
7.3 · AI economics break differently when labor is cheap
One US developer, or seven in India.
US developer
$150–300K
fully-loaded, annual
India developer
$20–40K
fully-loaded, annual
The same role. For the fully-loaded cost of one US developer, an enterprise can staff about seven in India.
In a US enterprise, almost any GPU spend is justified if it amplifies modest human output. When the same role costs a fraction, the math inverts — the break-even threshold for AI adoption is higher in absolute productivity terms, even though the technology costs the same.
The corollary: AI use cases that win in emerging markets are different. Customer-facing applications scale on user volume, where India’s 800M+ consumers create extraordinary economics. Internal-productivity AI tools must clear a higher ROI bar.
7.4 · the leapfrog pattern
Being late to a technology can be an advantage.
Most US enterprises sit on twenty years of accumulated on-premise infrastructure they have spent the last decade migrating. Most enterprises in Southeast Asia and Latin America do not — they are building cloud-native from the start, with no meaningful legacy footprint to migrate from.
The Optimizer that achieved a 50% cost reduction in 69 days has ~11,800 cloud resources, and almost none are migration artifacts. Their architecture is internally consistent because nothing was carried over from a 2010-vintage data center. Emerging-market enterprises are, on average, more architecturally coherent than their developed-market peers.
7.5 · where mainstream reports get it wrong
The consensus stops at the US border.
the consensus
the contrarian read
Multi-cloud is the dominant strategy
Only 1 of 13 runs production across all three hyperscalers
Reserved Instances are universally attractive
EM CFOs increasingly avoid 3-year commitments due to FX exposure
AI adoption is universal and accelerating
Under 2% of EM cohort cloud spend on AI services as of June 2026
Sovereign cloud is a European topic
India, UAE, Saudi Arabia, Indonesia are all driving aggressive mandates
Cloud migration is the dominant project type
In SEA and LatAm, cloud-native greenfield dominates
The mainstream cloud narrative is not wrong. It is partial. It accurately describes US-headquartered enterprises with US-centric operations. The experience of the rest of the world looks meaningfully different.
chapter 8 · predictions for 2027
Falsifiable, dated, and gradeable in next year’s report.
We will grade each of these in the 2027 edition. Confirmed, rejected, or partial — no retconning, no reframing.
8 · the seven predictions
What we expect by the 2027 edition.
| # | Prediction | Why it matters |
|---|---|---|
| 1 | GPU share crosses 15% | By Q2 2027 the average bill is more than 15% AI/GPU, up from <2%. Inference, not training, drives the shift. |
| 2 | Auto-remediation becomes RFP standard | By Q4 2027, automated remediation (not just detection) is a standard procurement requirement. |
| 3 | EM CFOs treat cloud as treasury risk | By Q4 2027, a majority of large Indian and Brazilian enterprises manage USD exposure through formal FX hedging. |
| 4 | Negative-resource pricing emerges | By 2027, two of three hyperscalers offer pricing variants that bill for managed waste reduction. |
| 5 | A $1M+ idle GPU cluster goes public | The first enterprise-scale shutdown of an idle training cluster costing $1M+/quarter is disclosed in a 2027 earnings transcript. |
| 6 | Sovereign cloud becomes board-level | By Q4 2027, sovereign AI infrastructure is a board-level question across Europe, India, the Middle East, and Brazil. |
| 7 | Multi-model routing is table stakes | By Q2 2027, multi-model routers are a standard layer in every production AI stack. |
chapter 9 · the gpu shift
The next FinOps inflection has three more zeros.
For 15 years, cloud waste was a CPU and storage story. Over the next 36 months, it becomes a GPU story.
9.1 · the order-of-magnitude shift in stakes
| Scenario | Cost of a forgotten weekend |
|---|---|
| An idle CPU box (m6i.4xlarge) | ~$0.77/hour. Forgotten for a weekend: $37 wasted. Annoying but manageable. |
| An idle GPU cluster (p5.48xlarge, 8× H100) | ~$98/hour. Forgotten for a weekend: $4,704 wasted. Same mistake. ~127× the cost. |
GPU utilization rates become the new EBITDA leading indicator. Most enterprises will discover their GPUs run at under 30% utilization.
The same way they discovered their EC2 fleets did in 2015
Three forces converging in 2026–2027
1
The Blackwell transition
H100 → B200. New silicon unlocks 2.5× perf/W. Existing 3-year H100 reservations look increasingly stranded.
2
Silicon polyculture
AMD MI300X, AWS Trainium2, Google TPU v6. The era of NVIDIA-only inference is ending — routing complexity rises.
3
Inference overtakes training
By 2027, inference is ~80% of enterprise AI compute. Training is bursty; inference is 24/7. Commitment models break.
9.2 · the 2027 AI infrastructure stack
The durable advantage has moved up the stack.
Durable advantagelayers 4–6 · where the margin lives
6
AI Governance & Spend Control
Cost-aware routers · eval gates · audit logs · FinOps for GPUs · per-tenant budgets
5
Agent & Workflow Orchestration
LangGraph · AutoGen · MCP servers · Bedrock Agents · multi-step reasoning
4
Inference Stack
vLLM · TensorRT-LLM · SGLang · managed endpoints · serverless inference
Commoditizinglayers 1–3 · capacity is abundant
3
Model Routing & Caching
Multi-model routers · prompt caching · distillation · semantic cache
2
Vector / Retrieval Layer
Turbopuffer · Vespa · pgvector · Pinecone · Qdrant
1
GPU Substrate
Hyperscaler · Neoclouds (CoreWeave, Lambda, Crusoe) · Colo · Sovereign GPU clouds
Layer 1 is becoming fungible — capacity is abundant, pricing curves are converging. The durable architectural advantage in 2027 lives in layers 4 through 6: the runtime that serves efficiently, the router that sends simple queries to small models, and the governance plane that enforces budgets.
9.3 · the CTO checklist
Seven actions before the GPU bill exceeds the EC2 bill.
You probably have 12–18 months. After that, the cost shape of your cloud bill will be unrecognizable.
- 1Audit your AI spend by service lineMost enterprises cannot — AI spend is buried inside “compute” with no per-model, per-team allocation. If you can’t break it out, that’s the first project.
- 2Establish a GPU utilization SLOAnything under 40% in production should be flagged like a P1. Most will discover their initial measurement is 15–25%. The goal is to know the number, and make it someone’s job to drive it up.
- 3Build a model router, even a simple oneA two-tier router can cut inference cost 60–80% with negligible accuracy impact. Don’t wait for the perfect solution. Ship the obvious one.
- 4Negotiate inference commitments, not just trainingMost enterprises over-buy training reservations and under-buy inference. The economics have inverted.
- 5Treat AI cost as a product KPICost-per-active-user, cost-per-completion, cost-per-thousand-tokens belong on the same dashboard as DAU and retention.
- 6Architect for portability at the inference layerLock-in is moving from the model to the inference stack. Model APIs are converging; inference frameworks differ wildly.
- 7Put a sunset policy on every AI experimentMost idle GPU clusters started as Friday-afternoon proofs of concept. Every experiment gets a max-lifespan tag at creation, with an automated reaper. Friction is the feature.
The teams that adopt all seven in 2026 will look like Optimizers in 2027. The teams that adopt none will look like Drifters and Scalers, and their GPU bills will be unrecognizable in 18 months. There is no third option.
chapters 4–9 · recap
The core of the report
- 87%forgotten resources87% of waste is forgotten resources; fixing the anti-pattern top 10 recovers ~80% of the total.
- <1%formally remediatedDetection is solved — under 1% of signals are formally remediated. The next decade of FinOps is about who acts, not who finds.
- 6,922unmonitored alarmsMetadata sprawl, subscription fragmentation, Single-AZ databases, and 6,922 unmonitored alarms are the architectural debt of cloud.
- FXtreasury riskThe Emerging Markets lens changes the math: currency is a treasury risk, sovereign cloud is global, and cheap labor raises the AI ROI bar.
- 7CTO actionsThe GPU shift carries every anti-pattern forward with three more zeros. Seven CTO actions separate the Optimizers from the rest.
closing · what to watch next
Three trends. Three counter-trends. One bet.
what to watch next
three trends
- 01GPU spend crosses 15% of enterprise cloud
- 02auto-remediation displaces detection-only tooling
- 03sovereign cloud becomes board-level globally, not just in Europe
three counter-trends
- 01Repatriation gains traction for steady-state workloads in currency-pressured economies
- 02EM enterprises increasingly question 3-year RI commitments
- 03“quiet” Drifter cost growth outpaces flashy AI bills as the dominant cause of budget overrun
By the 2027 edition, more than half of our cohort will have crossed into Optimizer territory — not because tooling improved, but because the next downturn made cloud cost a board priority.
The bet is on the cycle, not the technology
The most important thing you can do with this report is challenge it.
the state of cloud 2026 · we published our methodology so you can replicate, contest, and improve on it
appendices · the fine print
Methodology, caveats, and how to challenge this.
A · cohort & window
Telemetry from 25 cloud accounts across 13 enterprises, October 2025–June 2026. All cost and trend data was collected over a 69-day window, March 20 to May 27, 2026. Instruments: native AWS/Azure/GCP resource-discovery and cost APIs, plus a 315+ rule recommendation engine.
- FMCG / Consumer goods30%
- Consumer internet18%
- Beverages / Hospitality14%
- B2B SaaS / Logistics13%
- Sporting goods11%
- E-commerce retail8%
- Creator economy / Other6%
B · anti-pattern catalog (excerpt)
| Rule | Anti-pattern | Detection trigger |
|---|---|---|
| RC-001 | Stopped EC2 accruing EBS charges | status = stopped for ≥ 7 days AND has attached EBS volume |
| RC-002 | Unattached EBS volume | state = available AND last-attached > 14 days ago |
| RC-005 | Single-AZ production RDS | MultiAZ = false AND identifier matches production pattern |
| RC-021 | Orphan EBS snapshot | source volume deleted AND not referenced by AMI |
| RC-091 | EC2 not covered by Savings Plan | on-demand AND running ≥ 30 days AND no covering plan |
| RC-1386 | Non-prod Premium SSD disk (Azure) | SKU = Premium_LRS AND RG matches non-prod pattern |
These 15 rules account for ~78% of the dollar value of all recommendations generated across our cohort.
C & D · caveats and independence
The cohort is 13 enterprises, not 1,300 — it is deliberately deep rather than broad, and it skews toward India-headquartered and India-presence multinationals. Findings about mechanism generalize; findings about distribution should be read as indicative. No vendor paid for inclusion, no finding was reviewed by a hyperscaler before publication, and every number in this report is reproducible from the instruments named in Appendix A.
colophon
The State of Cloud 2026 — An Emerging Markets View
Set in Space Grotesk and JetBrains Mono. Industry research grounded in primary telemetry from 25 production cloud accounts across 13 enterprises — an Emerging Markets view of global enterprise cloud.
Ten chapters · Four appendices · ~40 minutes · interactive experience