Cloud Cost Ledger

• 1: Model contract
• 2: Compute: EC2, ECS, and owned servers
• 3: Block storage: EBS and ESSD
• 4: Object storage: S3, OSS, and owned topology
• 5: Egress and network paths
• 6: Managed-service markup
• 7: Cloud price history
• 8: Dataset releases

1 - Model contract

Two cost functions

The comparison period $t$, currency date, tax treatment, financing cost, and residual value must be explicit. Cash cost and economic cost are shown separately rather than blended.

Like-for-like rule

Match usable capacity, required availability, durability, backup, recovery objective, operations coverage, security controls, region, and demand shape. When parity cannot be established, publish two bounded scenarios rather than one false-precision ratio.

Required outputs

• Monthly and cumulative cost curves.
• Break-even scale and break-even time.
• Sensitivity to utilization, hardware life, labor, power, financing, and growth.
• Cash, economic, and exit-cost views.
• Applicability boundaries and omitted costs.

The claim is never “self-hosting is X times cheaper” without the workload and assumptions that make $X$ true.

2 - Compute: EC2, ECS, and owned servers

Direct answer

Inside a mainstream instance family, price is almost linear in vCPU. Buying a larger shape reduces instance count, not unit-core price. The variables that move unit cost are memory ratio, CPU architecture, generation, region, purchase term, and—most importantly—useful utilization.

For stable owned compute, the material-cost curve falls rapidly while the first server fills, jumps when active and spare servers are added, then flattens. The economic question is not “how large is the company?” but “can this workload keep the minimum reliable pool usefully full for the asset life?”

Current 2C8G reference shapes

The raw cash-price ratio is not a procurement conclusion. m8i and g9i differ in CPU, network, storage channel, regional context, tax, and business performance. Normalize with application TPS, latency, compile throughput, or another useful output before choosing.

The family skeleton

Same-vCPU prices reveal a repeatable memory-ratio structure:

Moving from large to xlarge or 2xlarge inside the same family remains nearly linear. Exceptions—burstable CPU, local disks, dedicated hosts, bare metal, GPUs, NUMA, and commercial licenses—carry separate fixed or step meters.

Architecture, generation, and region

The snapshot’s price factors are meaningful only as provider pricing intent:

• AWS m7g/m7i = 0.810; m8g/m8i = 0.848. Graviton is 15–19% cheaper in list price, but application compatibility and useful performance decide value.
• Alibaba Yitian g8y.large is ¥192/month, versus Intel g8i.large at ¥251.16; current long-term discounts can still change the total ordering.
• AWS m8i.large regional factor ranges from 1.000 in core US regions to 1.250 Singapore, 1.292 Tokyo, and 1.594 São Paulo.
• Alibaba g9i.large is at factor 1.000 in Hangzhou/Shanghai/Beijing, 0.900 Ulanqab, 1.770 Singapore, 1.798 Tokyo, and 1.929 Hong Kong.

Region is not a coupon code. Data sovereignty, latency, service coverage, disaster-recovery topology, and cross-region traffic can dominate the instance factor.

Elasticity and commitment

Alibaba g9i.large PAYG at 730 hours costs ¥362.88, or 1.521× the ¥238.60 monthly price. The time-only crossing is:

For any commitment factor $f$ and realized utilization $u$:

A five-year term and a purchased server both transfer demand risk to the customer. Compare unused commitment, generation lock, cash opportunity cost, residual value, failure replacement, and exit terms—not just the discount.

The owned-compute curve

The scenario uses a DHH-class 2U server with 192 vCPU, 384 GiB RAM, fleet-average 19.2 TB local NVMe, purchase price about ¥169,500, five-year life, and 70% target fill. A spare is added from the second active server onward. Internal labor, migration, application changes, and business risk are excluded.

The first server filling from 16 to 128 vCPU cuts unit cost eightfold. Expanding from 4,000 to 64,000 vCPU—16× more scale—cuts it only another 17%. Most scale economy arrives at the beginning; later results are driven by utilization and hardware life.

First and durable crossings

“First” means one fill interval becomes cheaper. “Durable” means later server, spare, rack, and network steps no longer push cost back above the cloud line. The AWS RI row uses an m7i discount factor as a proxy for c7a, not a quoted c7a RI offer.

Sensitivity: why the five-year comparison is fragile

Against expensive on-demand compute, reasonable scenarios cross early. Against a deep five-year term near hardware amortization, utilization and asset life can change the answer by orders of magnitude.

Decision zones

Cloud remains strongest for unknown duration, high peak/mean ratio, global small footprints, interruptible work, and rapid release. Owned compute is strongest for stable, long-lived, standardized workloads with existing procurement and operations capability.

Historical context

AWS ~2C8G advertised-capacity price fell 73.5% from 2007 to 2026, but almost all list-price decline occurred before 2017. Alibaba 2C8G monthly price fell only 21.0% from 2016 to 2026 while PAYG fell 62.1%, largely by narrowing the elasticity premium. Mature FinOps therefore means benchmarking and migrating, not waiting for automatic cuts. See Cloud price history.

Data and sources

• Current EC2 reference shapes
• Current ECS reference shapes
• Compute scale curve
• Compute breakpoints
• Compute sensitivity
• Full pricing report
• Full build-versus-buy report

3 - Block storage: EBS and ESSD

Direct answer

“Price per GB” is not a valid block-storage comparison by itself. It can rank gp3 as cheaper at baseline and then reverse the answer once performance is matched. A usable bill must price at least four things:

The final usable performance is also bounded by the instance:

Paying for IOPS that the attached instance cannot deliver creates cost without performance.

Two pricing philosophies

AWS EBS gp3: capacity and performance are explicit meters

The snapshot uses:

• Capacity: $0.08/GB-month (¥0.5425).
• Included: 3,000 IOPS and 125 MiB/s.
• Extra IOPS: $0.005/IOPS-month.
• Extra throughput: $0.04/MiB/s-month in the exact us-east-1 snapshot.

So:

All three are provisioned meters. Actual low usage does not refund unused provisioned IOPS or throughput.

Alibaba ESSD: capacity often carries a performance curve

Traditional PL0–PL3 bind capacity to a product level. AutoPL separates capacity from optional provisioned and burst performance, but capacity still includes a rising baseline:

At 200 GiB this yields 11,800 IOPS and 220 MB/s; at 2,048 GiB the baseline reaches 50,000 IOPS and 350 MB/s. Alibaba’s documentation also notes that the attached instance can impose the lower I/O ceiling.

Capacity meter alone

At this layer, gp3 is near PL0 and about 54% of PL1/AutoPL. That statement is true but incomplete.

Performance-matched scenarios

This is the core block-storage insight: the answer flips because one product charges extra performance separately while the other bundles a capacity-driven baseline. It does not prove equal latency, durability, burst behavior, or instance delivery.

The exact rows are in block_storage_scenarios.csv.

What the price table still hides

Snapshots and recovery

EBS snapshots bill changed blocks in S3-backed storage; cross-region copy, archive retrieval, fast snapshot restore, time-bounded copy, and provisioned initialization can add separate meters. ESSD snapshots, backup products, and cross-region recovery have their own lifecycle and transfer costs. A recovery budget needs restore time and restore throughput, not just snapshot GB-month.

Topology and network

A single cloud disk is an availability-zone resource. Cross-AZ database or application topologies may duplicate volumes and add network charges. Managed database storage can price capacity, IOPS, and throughput at different markup factors from raw block storage.

Provisioned versus consumed

Both capacity and performance can be paid before they are used. Track:

• provisioned vs used GiB;
• provisioned vs observed P95/P99 IOPS;
• provisioned vs observed throughput;
• instance channel utilization;
• snapshot growth and orphan retention;
• commitment coverage and unused term value.

Extreme-performance layers

At the high end, capacity is no longer the dominant meter. AWS io2 prices provisioned IOPS by volume-level tiers. Alibaba ESSD PL-X similarly separates capacity and IOPS; an official example for 2,048 GiB and two million IOPS totals ¥152,048/month, of which ¥150,000 is performance. Low-end disks monetize capacity; high-end disks monetize performance.

Historical insight

AWS gp2 to gp3 reduced capacity price only 20%, from $0.10 to $0.08/GB-month. For a 200 GiB volume, however, included baseline IOPS rose from 600 to 3,000, cutting price per baseline IOPS by 84%. Alibaba’s 2016 SSD and 2026 ESSD PL1 both anchor at ¥1/GB-month; performance, reliability, and product semantics improved while the capacity meter stayed flat.

This is why “storage prices did not fall” and “storage price-performance improved sharply” can both be true. See Cloud price history.

A minimum bill worksheet

Before comparing disks, fill one row per volume:

If any of these is missing, a capacity-price ratio is a lead, not a conclusion.

Boundaries

• GB and GiB are retained exactly as providers publish them.
• AutoPL baseline does not imply the attached instance can deliver it.
• This comparison does not equalize latency, durability, burst, or SLA.
• AWS rates are regional and tax-exclusive; Alibaba China-site rates are tax-inclusive.
• Prices may have changed after 2026-08-22; refresh before purchase.

Data and sources

• Block-storage rates
• Performance-matched scenarios
• Full pricing report
• Pricing source ledger
• Checksums
• Original Chinese essay

4 - Object storage: S3, OSS, and owned topology

Direct answer

Object-storage capacity is now cheap and surprisingly close across providers. The bill diverges elsewhere:

For self-hosting, the corresponding denominator is not raw disk TB:

That is why object-storage economics do not have one universal “cloud exit size.” The first question is the minimum reliable topology.

Current capacity rates

The rows are price-adjacent, not service-equivalent. Retrieval latency, availability, durability topology, restore workflow, and feature semantics differ.

The small-object tax

Cold tiers often bill a minimum object size:

A 1 KB object can therefore consume 128× or 64× its logical size on the bill; a 10 KB object consumes 12.8× or 6.4×. Before moving data to a colder tier, measure object-count distribution, average and percentile size, overwrite/delete rate, and retention—not just total TB.

Activity meters

Requests

Snapshot reference rates:

• S3 Standard PUT/COPY/POST/LIST: $0.005 / 1,000 requests.
• S3 Standard GET: $0.0004 / 1,000 requests.
• OSS Standard: first 5 million PUT and 20 million GET per region-month free, then ¥0.01 / 10,000; IA/Archive PUT/GET commonly ¥0.10 / 10,000.
• OSS Deep Cold PUT: ¥3.50 / 10,000.

Request policy can change quickly, and console browsing itself can issue GET or LIST requests. High object count can matter even when capacity is small.

Retrieval and early deletion

Snapshot examples:

• S3 Standard-IA retrieval: $0.01/GB; Glacier Instant: $0.03/GB.
• OSS IA: ¥0.0325/GB; Archive: ¥0.06/GB; Cold Archive standard retrieval: ¥0.06/GB; Deep Cold standard: ¥0.018/GB.

Deleting, overwriting, or transitioning before the minimum duration can charge the remaining term. An archive restore can also create a temporary hot copy that is billed while accessible.

Transfer: the exit and delivery meter

The same download has a different price from ECS, OSS busy time, OSS off-peak, or CDN. Draw the full data path before pricing. For provider exit, also check documented waiver eligibility and operational lead time; a conditional waiver is not the same as a zero-rate default.

Packages and commitment risk

Alibaba OSS Standard LRS 100 GB illustrates the package ladder:

Unused capacity expires; region, redundancy type, and meter mismatch can prevent deduction. A package lowers nominal unit price only when its scope and usage fit.

Self-hosting begins with topology

The released scenario uses 75% erasure-code efficiency, 80% maximum safe fill, 5% spare-drive inventory, 5% monthly data change, and one practical 1 Gbps port per 100 TB/month of remote replication.

None is automatically equivalent to S3 Standard’s service level. Durability, availability, API semantics, operations, audit, and recovery must be tested separately.

Unit cost by topology

At 100 TB the eight-node pool is mostly empty. Around 200 TB it approaches an OSS one-year package; at 1 PB its single-site capacity cost is about one fifth of that package. Adding sites changes the answer because hardware and replication network are duplicated.

First and durable capacity crossings

These are capacity plus replication-network material crossings. S3/OSS requests, retrieval, and small-object overhead are excluded on the cloud side; commercial software, support, load balancers, security, and internal labor are excluded on the owned side.

Sensitivity versus OSS one-year package

Single-site capacity is insensitive to write churn; multi-site storage is not. High-churn data can make network replication dominate and may belong in a database, block store, or shared filesystem instead.

Historical insight

S3 Standard fell from $0.15/GB-month in 2006 to $0.023 in 2016, then stayed flat through the snapshot. Alibaba OSS Standard directory price reached ¥0.12 in 2018; a durable public effective rate of ¥0.09 appeared in 2024. Capacity followed a hardware-like early decline, then became a platform meter. Egress and activity did not follow the same curve. See Cloud price history.

Decision zones

Cloud exit does not require rejecting CDN, edge, or managed security. A common efficient design owns stable origin capacity and continues buying statistically pooled distribution.

Boundaries

• Capacity crossings are not SLA or durability certifications.
• Four-node MinIO is retained only as a lower-bound reference.
• Commercial object software and internal operations are excluded from the baseline.
• DHH’s observed S3 contract uses billed footprint; primary logical data gives a different denominator.
• Request policy, packages, and prices can change after 2026-08-22.

Data and sources

• Current storage-class rates
• Current transfer rates
• Object scale curve
• Object breakpoints
• Object sensitivity
• Full pricing report
• Full build-versus-buy report
• Original Chinese essay

5 - Egress and network paths

Direct answer

Network is its own bill, not an attachment to compute or storage. The same byte can be free on ingress, charged twice across an availability-zone round trip, processed by NAT, and charged again on internet egress.

Draw the path before applying a rate.

Current reference paths

AWS aggregates egress tiers across many services and regions; eligibility and exceptions matter. Alibaba same-region VPC product communication is often free, but NAT, CEN, EIP, load balancers, cross-region links, and public paths introduce their own meters.

Fixed bandwidth versus per GB

The self-transit scenario uses a public 1 Gbps line at ¥12,500/month and a conservative practical load of 100 TB/month—about 30% of theoretical line rate. This is a procurement anchor, not a universal market price.

Break-even volume is:

The cloud curve starts at zero and scales with bytes. The port curve starts with a fixed step and then has near-zero marginal bytes until the next port. Bursty, uncertain, global, or DDoS-exposed traffic may justify the cloud/CDN premium; stable bulk transfer crosses early.

Alibaba fixed public bandwidth

The public Hangzhou example prices 1 Mbps fixed bandwidth around ¥23/month versus ¥0.8/GB by traffic. They cross at 23 / 0.8 = 28.75 GB/month. One Mbps can theoretically deliver about 328.5 GB/month, so fixed bandwidth becomes cheaper at roughly 8.8% utilization. Peak entitlement and service quality remain different from a by-traffic peak cap.

Architecture consequences

• Use an S3 Gateway Endpoint instead of routing S3 traffic through AWS NAT when applicable.
• Avoid accidental cross-AZ round trips for chatty application/database paths.
• Price ECS, OSS busy/off-peak, and CDN-origin delivery as separate exits.
• Keep stable origin and backup capacity on fixed or private paths; keep global edge, CDN, DDoS, and volatile traffic on statistically pooled services.
• Track IPv4, idle NAT, orphan load balancers, and cross-region replication as first-class meters.

Exit capability

Provider exit waivers can materially change C1, but only when eligibility, covered services, destination, notice, transfer window, and approval process are documented. A negotiated or conditional waiver is recorded separately from the default meter. The operational export path and time limit remain part of the test even when the rate is zero.

Historical insight

AWS marginal egress fell from $0.20/GB in 2006 to $0.09 in 2014, then remained flat through the snapshot. Alibaba ECS Hangzhou stayed at ¥0.8/GB across the report’s 2016–2026 anchors. Silicon and disk gains do not automatically flow into the network meter. See Cloud price history.

Boundaries

• Port procurement lacks the cloud platform’s global backbone, elasticity, routing, DDoS, operations, and zero-startup option.
• The 1 Gbps quote and 100 TB practical load are scenario inputs.
• Taxes, installation, private contracts, and regional fees can move crossings.
• All rates must be refreshed after the snapshot before procurement.

Data and sources

• Cloud transfer rates
• Egress scale curve
• Pricing source ledger
• Build-versus-buy source ledger

6 - Managed-service markup

Split price from value

The model first estimates equivalent resource cost, then itemizes operations the managed service supplies: provisioning, patching, backups, failover, monitoring, support, compliance, control plane, and service risk. The residual price is not automatically waste; it is the amount that must be justified by delivered value.

Questions the page must answer

• Which responsibilities actually move to the provider, and which remain shared?
• Which features are used, and what would they cost to reproduce at the required coverage?
• What restrictions, missing extensions, or control-plane dependencies reduce value?
• How do commitment, outage remedy, and exit cost change the apparent markup?

Seed essay: Is a cloud database an intelligence tax?. Current model outputs will replace its historical prices while keeping the original argument and period visible.

7 - Cloud price history

Direct answer

Cloud prices fell substantially, but they did not follow a stable “halve every two years” path. Early competition produced large step changes; mature core meters often plateaued. Recent gains are more likely to require moving to a new instance generation, architecture, storage class, or commitment model than waiting for an old SKU to become cheaper automatically.

What actually halved—and how slowly

The values come from long_run_metrics.csv. Inflation-adjusted figures are published separately; they do not correct for cross-generation performance.

Four different histories

Compute: price cuts became migration work

AWS same-advertised-capacity compute fell rapidly through 2017, then roughly plateaued: m4.large was $0.108/h, m5.large and m6i.large $0.096, m7i.large $0.1008, and m8i.large $0.10584. Price-performance improved, but realizing it required moving generation and validating the workload.

Alibaba’s monthly and PAYG curves diverged. A 2C8G monthly shape fell only 21% from 2016 to 2026, while PAYG fell 62.1%. Much of the latter was a narrowing of the elasticity premium: a full PAYG month went from 3.17× monthly to 1.52×.

Object storage: closest to hardware—then a ten-year plateau

S3 Standard’s first tier fell from $0.15 in 2006 to $0.023 in 2016, then stayed nominally flat through the snapshot. Alibaba OSS Standard moved from ¥0.30 in 2014 to a ¥0.12 directory rate; a durable public effective rate of ¥0.09 appeared in 2024. The large early fall did not remove request, retrieval, minimum-duration, or egress meters.

Block storage: capacity plateau, performance-density gain

EBS gp2 to gp3 reduced capacity price only 20%. For a 200 GiB volume, however, baseline IOPS rose from 600 to 3,000, so price per baseline IOPS fell 84%. Alibaba’s 2016 SSD and 2026 ESSD PL1 both anchor at ¥1/GB-month; product semantics and performance changed even though the capacity meter did not.

Egress: the least silicon-like meter

AWS first-paid-tier egress reached $0.09/GB in 2014 and then remained flat; the 2021 free allowance improved small-account average cost but not large-volume marginal cost. Alibaba ECS Hangzhou egress stayed at ¥0.8/GB across the 2016–2026 anchors. Path engineering—CDN, private connectivity, caching, regional placement, or self-hosted transit—matters more than waiting.

Full bills move slower than headline compute

The capacity basket is approximately 2 vCPU / 8 GiB compute + 200 GB/GiB general block storage + 1 TiB monthly internet egress, plus current AWS public IPv4. It is a capacity comparison, not a performance or SLA comparison.

In the 2026 AWS basket, egress is 46.2% of the bill—larger than compute at 42.9%. In the Alibaba monthly basket, unchanged egress is 65.1%. A workload’s bill weights decide whether a new compute generation matters.

Budgeting and FinOps implications

1. Do not budget automatic 10–20% annual list-price cuts for mature core SKUs.
2. Benchmark new generations every two to three years; migrate first, commit second.
3. Keep separate budgets for compute, storage capacity, storage performance, object activity, and network paths.
4. Preserve public price as the contract anchor, but plan using invoice-level effective rates and commitment waste.
5. Model high-egress workloads independently; their economics are not governed by the compute curve.
6. For self-hosting, hardware also arrives in steps. New server purchases capture technology gains only at refresh time, while facilities, power, and network set a floor.

Boundaries

• Public price is not enterprise net price.
• Same-advertised-capacity generations are not performance-equivalent.
• S3, OSS, EBS, ESSD, and self-hosted systems differ in durability, latency, topology, and operations.
• Alibaba’s exact long-run compute series starts in 2016 because earlier full SKU tables could not be reconstructed reliably.
• The endpoint may change after 2026-08-22; refresh before purchase.

Data and sources

• Full report
• Price anchors
• Normalized indexes
• Workload baskets
• Source ledger
• Checksums

8 - Dataset releases

Every number highlighted in the ledger points to a dated release. A release is immutable: a correction creates a new version rather than overwriting the old bytes.

“Released snapshot” means the files ship with the site together with sources, methodology, schema, and checksums. It is still not a live provider quote.

Release contents

Each version contains:

README.md             scope, dates, units, and reuse warning
release.json          machine-readable version and status
report.md             complete human-readable analysis
methodology.md        comparison and normalization rules
schema.json           table keys and unit conventions
tables/               all CSV and JSON inputs/outputs plus source ledger
charts/               reproducible report charts where available
checksums.txt          SHA-256 for every released file

Three complementary datasets

Cloud pricing · 2026-08-22-v1

Current-reference AWS and Alibaba Cloud tables for EC2/ECS, EBS/ESSD, S3/OSS, network, PostgreSQL RDS, commitments, and managed-database curves.

• Release notes
• Full report
• Source ledger
• Checksums

Build versus buy · 2026-08-24-v1

Discrete cost curves, first and durable crossings, and sensitivity tables for compute, PostgreSQL, object storage, and fixed-port egress. Baseline internal labor and migration cost are deliberately excluded.

• Release notes
• Assumptions
• Full report
• Source ledger
• Checksums

Price history · 2026-08-24-v1

Long-run public-price anchors, nominal and real indexes, workload bill baskets, physical-input context, and historical source records from 2006–2026.

• Release notes
• Full report
• Long-run metrics
• Source ledger
• Checksums

Status semantics

released-snapshot means the files ship with the site and pass checksum and headline-value verification. It does not mean the price is current today, that a provider supplied a formal quote, or that two services are fully equivalent. Use the snapshot date and applicability note whenever quoting a value.

The machine-readable release index is at /data/index.json.