Audit Cutoff: 2026-10-23 · Verified Specs & Benchmarks [S01–S55] · Fight Complexity

How 10 Years of Hardware Killed Legacy System-Design Complexity

Most system-design interviews still optimize for 2014 hardware (SATA SSDs, 16-core VMs, 10 Gbps NICs). In 2026, local PCIe Gen6 NVMe (28.4 GB/s, 6.8M IOPS, 100 µs) is 5.9x faster than a cross-AZ Redis hop (590 µs, $0.02/GB).

Explore Interactive Graphs ↓ 10 Outdated Myths Corrected

PCIe Gen6 NVMe [S03]

28.4 GB/s · 6.8M

75x 2016 SATA IOPS · 100 µs 8KiB read [S32]

Cross-AZ Cache / RPC Tax

590 µs · $97.9k

590 µs RTT [S30] + $97.9k/mo @ 2 GB/s [S34]

Same-AZ ENA Express [S30]

52.2 µs · $0/GB

11.3x faster than cross-AZ · up to 25 Gbps/flow

1 Node Compute [S10, S38]

192–256 Cores

m9g.48xlarge ($6,857/mo) · EPYC 9006 1.6 TB/s

Max Single-Node RAM

32 TiB – 48 TB

AWS U7inh [S13] · Google X5 Bare Metal [S20]

1. Interactive Hardware Multipliers, Latency Ladder & Complexity Taxes

Click or hover any bar, or switch metrics on any graph to inspect verified 2026 numbers

The 10-Year Hardware Multiplier (2016 → 2026)

Click any hardware layer to see why 2016 sharding and caching rules inverted.

[S03, S16] Storage Random Read IOPS 75x IOPS (90k IOPS (SATA III Enterprise SSD) → 6.8M IOPS (Samsung PM1763 PCIe Gen6) / 20.7M (Azure Laosv5)) Single-node Postgres on NVMe absorbs read concurrency that required 16-shard clusters in 2016.

2026 Storage & Network Latency Ladder (µs)

Local NVMe random reads (100–243 µs) are faster than a cross-AZ network RTT (590 µs).

[S32] Local NVMe 8 KiB Random Read (GCP C4) p50: 100 µs · p99: 180 µs Refreshed March 2026 local SSD 8-KiB random read reference.

2016 Frankenstack vs. 2026 Radical Simplicity

Compare monthly cloud cost, p99 latency, and number of moving parts across 4 workloads.

Comparison High-Scale API & Web Tier (100k–1M RPS, 2 kB Payload) Saves $125,500/mo (28 → 4 boxes) 1M RPS × 50 µs = 50 CPU-seconds/sec. Avoiding one unnecessary 2 GB/s cross-AZ RPC hop saves $97,901/month [S34, S38].

Where the Cloud Bill Actually Goes (Napkin Sec. 8)

Scale RPS to see how cross-AZ chatter ($0.02/GB) and egress dwarf compute.

1,000,000 RPS 1x Avoidable Cross-AZ RPC Hop (2 GB/s) Cost: $97,901/mo [S34] 14.3x the price of the 192-core m9g.48xlarge ($6,857/mo)

2. Interactive Back-of-the-Napkin Capacity & Cost Simulator

Verified 2026 AWS US East (N. Virginia) rates [S34, S38, S40] · 730h = 2,628,000s/month

Size Any Workload on 2026 Hardware

Payload Throughput

2 GB/s

16 Gbps · 5256 TB/mo (4,895,031 GiB)

Active CPU Cores

50 Cores

1x m9g.48xlarge (192 vCPU) w/ 2x headroom

Compute Bill [S38]

$6,857/mo

m9g.48xlarge @ $9.39264/hr (730h)

Cross-AZ Hop Tax [S34]

$97,901/mo

$0.02/GiB per cross-AZ hop ($0 if zonal)

Public Egress [S40]

$248,638/mo

AWS tiered schedule ($0.09 → $0.05/GiB)

2026 Architecture Recommendation 2x m9g.48xlarge (N+1 Static Stability) + Partitioned Zonal Cells + Edge CDN Shield

3. Ten Outdated System-Design Dogmas Corrected for 2026

Showing 10 of 10 corrections

5.9x Faster · Zero Cache Drift Ch. 03, 06, 16 · [S10, S30, S31, S32, S34]

1. Just Use Postgres (Skip the Default Redis Look-Aside Cache)

-4 boxes

2016 Outdated Wisdom

Put a distributed Redis/Memcached cluster in front of PostgreSQL on day one because disk reads take 10 ms.

2026 Hardware Reality

Cross-AZ Redis RTT is 590 µs (+ serde + $0.04/GB bidirectional tax), while local NVMe random reads take 100–243 µs and Postgres shared buffers hit in ~0.5 µs up to 1.5 TB RAM.

P99 Latency1.4 ms → 0.15 ms (9.3x)

Cloud Bill$4,200 → $1,143/mo

Monthly Savings$3,057/mo saved

When your working set fits in a 128 GB–1.5 TB Postgres instance or local Gen5/Gen6 NVMe (6.8M IOPS), a remote cross-AZ cache is slower than the database itself.

Evidence, Math & Exceptions (5) Caching & Queues

Cross-AZ network RTT (590.4 µs p50, 677.4 µs p99 [S30]) is 2.4x–5.9x slower than a local NVMe random read (100–243 µs [S31, S32]).
Cold-cache restarts at 95% hit ratio trigger a 20x read amplification storm that topples the backing database.
Eliminates dual-write race conditions, TTL stampedes, and $0.02/GB per-direction cross-AZ transfer fees [S34].

Still use when: Pre-computed expensive aggregations or rendered fragments that burn >20 ms of CPU per request.
Still use when: Same-AZ or in-process (local memory) hot-key absorption for extreme celebrity fan-out.

Saves $9.7k/mo · No 2PC Ch. 01, 05, 13 · [S10, S13, S20, S38, S50]

2. Scale Up Vertically Before Sharding (192 Cores, 1.5–32 TB RAM)

-14 boxes

2016 Outdated Wisdom

Horizontally shard your relational database at 500 GB or 10,000 QPS using application-level shard routers or Vitess.

2026 Hardware Reality

Single nodes now reach 192–256 cores, 768 GiB–32 TiB RAM, and 45–120 TB of local NVMe at 28.4 GB/s. One HA Postgres pair handles 80k–120k QPS without cross-shard 2PC.

P99 Latency25 ms → 0.8 ms (31.3x)

Cloud Bill$12,000 → $2,285/mo

Monthly Savings$9,715/mo saved

Premature sharding trades a $1,143/mo VM upgrade for years of broken foreign keys, scatter-gather pagination, and 2PC deadlocks.

Evidence, Math & Exceptions (4) Database & Storage

AWS m9g.8xlarge (32 vCPU, 128 GiB) is $1,142.77/mo; m9g.48xlarge (192 vCPU, 768 GiB) is $6,856.63/mo [S10, S38].
Extreme scale-up reaches 32 TiB on AWS U7inh [S13] and 48 TB on Google X5 bare metal [S20].
PostgreSQL 18 group commit and io_uring saturate modern NVMe queues without coordinator hops [S50].

Still use when: Multi-region active-active data residency mandates or sustained write concurrency exceeding a single primary's lock/WAL pipeline.

$0 Extra Infra · ACID Atomic Ch. 06, 17 · [S03, S50, S51]

3. Postgres SKIP LOCKED Instead of Kafka for <20k Jobs/sec

-6 boxes

2016 Outdated Wisdom

Deploy a multi-broker Kafka or RabbitMQ cluster + ZooKeeper for every background job queue, webhook processor, or transactional outbox.

2026 Hardware Reality

PostgreSQL FOR UPDATE SKIP LOCKED processes 10k–20k jobs/sec atomically inside your existing ACID transaction with zero dual-write drift.

P99 Latency15 ms → 1.2 ms (12.5x)

Cloud Bill$3,400 → $0/mo

Monthly Savings$3,400/mo saved

Adding a separate message broker means writing to the DB and publishing to the queue can diverge unless you build a CDC/outbox pipeline anyway.

Evidence, Math & Exceptions (4) Caching & Queues

A 10,000 TPS ledger or job stream at 1 kB/row is only 10 MB/s — 0.048% of a Gen6 NVMe drive's 21 GB/s sequential write rate [S03].
INSERT into job_queue + business state update in one ACID commit eliminates ghost events and consumer rebalance storms.
When you truly exceed 50k+ msgs/sec, 2026 S3-backed zero-disk Kafka (WarpStream/Bufstream) avoids cross-AZ broker replication taxes.

Still use when: High-throughput telemetry (>50k msgs/s), multi-day replay for dozens of independent downstream consumers, or cross-org streaming.

1 Store Instead of 5 Ch. 05, 22, 23 · [S02, S14, S31, S32]

4. Collapse the 5-Database Frankenstack into Unified Postgres

-11 boxes

2016 Outdated Wisdom

Run Postgres (OLTP) + Elasticsearch (BM25 search) + Pinecone/Milvus (vectors) + MongoDB (JSON) + Redis (cache) with CDC glue.

2026 Hardware Reality

PostgreSQL 18 + JSONB + BM25 + pgvector (DiskANN/HNSW) handles relational, document, lexical, and vector hybrid retrieval in one ACID store up to tens of millions of vectors.

P99 Latency42 ms → 4.5 ms (9.3x)

Cloud Bill$8,900 → $1,650/mo

Monthly Savings$7,250/mo saved

Every secondary database you sync via CDC adds a lagging failure domain, schema drift, and permission-filtering mismatches.

Evidence, Math & Exceptions (4) Search, Logs & AI

DiskANN/Vamana compressed PQ vectors keep index footprints in RAM while reading full vectors from 100 µs local NVMe, cutting RAM needs by 15x.
Single-query relational WHERE clauses + vector cosine distance + BM25 ranking eliminate post-filtering N+1 hydration.
1.0–1.6 TB/s socket memory bandwidth [S02] makes SIMD vector distance scans blisteringly fast on standard compute nodes.

Still use when: Billion-scale vector corpora or petabyte log search where stateless S3-backed engines (Quickwit, ClickHouse, Turbopuffer) take over.

Saves $97.9k/mo @ 1M RPS Ch. 11, 14, 15 · [S30, S34, S35, S38, S54]

5. Kill the Cross-AZ Chatty Microservice Tax ($97.9k/mo at 2 GB/s)

-25 boxes

2016 Outdated Wisdom

Decompose into 40 microservices calling each other round-robin across 3 Availability Zones over REST/JSON.

2026 Hardware Reality

AWS charges $0.01/GB at each endpoint across AZs ($0.02/GB one-way, $0.04/GB round-trip) and adds 590 µs RTT per hop. Keep hot paths modular in-process or pinned to same-AZ cells (52 µs RTT, $0 transfer).

P99 Latency140 ms → 4 ms (35.0x)

Cloud Bill$97,901 → $0/mo

Monthly Savings$97,901/mo saved

A 20-hop microservice fan-out across AZs turns a 1% single-hop tail glitch into an 18.2% request failure rate while burning six figures in AWS inter-AZ fees.

Evidence, Math & Exceptions (4) Sharding & Services

At 1M RPS × 2 kB (2 GB/s = 4.895M GiB/mo), one cross-AZ hop costs $97,901/mo — 14.3x the price of the 192-vCPU m9g.48xlarge ($6,856.63/mo) doing the work [S34, S38]!
Same-AZ ENA Express RTT is 52.2 µs ($0/GB) vs 590.4 µs ($0.02/GB) across AZs [S30, S34].
Reserve cross-AZ traffic strictly for asynchronous/quorum WAL replication and client failover.

Still use when: Independent organizational deployment boundaries — but route RPCs with strict AZ-affinity and binary Protobuf/ConnectRPC.

98% Cheaper Storage Tier Ch. 19, 20, 21 · [S32, S39, S43]

6. Object Storage + NVMe Cache Over $122k/mo SaaS Log Ingest

-18 boxes

2016 Outdated Wisdom

Ship 100% of raw un-sampled application logs to a SaaS observability vendor or a 30-node stateful Elasticsearch hot cluster.

2026 Hardware Reality

At 100k lines/s (1 kB/line = 244,752 GiB/mo), a $0.50/GiB ingestion fee costs $122,376/mo. 5:1 compressed Parquet on S3 Standard is $1,126/mo for storage + stateless NVMe query nodes.

P99 Latency800 ms → 120 ms (6.7x)

Cloud Bill$122,376 → $2,800/mo

Monthly Savings$119,576/mo saved

S3 concurrent range GETs deliver ~2 GiB/s per node [S32], turning object storage + local NVMe cache into a primary analytical & log store.

Evidence, Math & Exceptions (4) Search, Logs & AI

Worked Example E [S39]: 244,752 GiB/mo raw logs compressed 5:1 = 48,950 GiB. At S3 Standard ($0.023/GiB-mo), raw storage is $1,126/mo.
Stratified tail sampling keeps 100% of errors and high-latency traces while sampling 1–5% of 200 OK noise.
Columnar Parquet/Iceberg + ClickHouse/DuckDB scans billions of rows on local NVMe in milliseconds.

Still use when: Small engineering teams with low log volume (<500 GiB/mo) where managed SaaS convenience outweighs infrastructure setup.

Zero Silent Data Loss Ch. 05, 09, 10, 11 · [S32, S50, S51]

7. ACID SQL & Group Commit Over Eventually Consistent NoSQL

-18 boxes

2016 Outdated Wisdom

Abandon ACID transactions, joins, and foreign keys for Cassandra/MongoDB Last-Write-Wins (LWW) to achieve write scale.

2026 Hardware Reality

LWW silently drops concurrent updates under NTP clock skew. Modern PostgreSQL group commit or Raft/Paxos Distributed SQL delivers 50k–100k ACID TPS without application-side repair logic.

P99 Latency85 ms → 3.5 ms (24.3x)

Cloud Bill$8,500 → $2,400/mo

Monthly Savings$6,100/mo saved

Giving up ACID forced application developers to re-implement buggy two-phase commits, client-side joins, and anti-entropy repair jobs in user code.

Evidence, Math & Exceptions (4) Database & Storage

PostgreSQL group commit amortizes WAL fsync across concurrent transactions on 300 µs NVMe [S32, S50].
Eliminates Cassandra tombstone scanning timeouts and 50% disk compaction slack.
When multi-node write scale is genuinely required, 3-voter quorum across 3 AZs provides linearizable ACID survival of an entire AZ loss [S51].

Still use when: AP-first edge telemetry capture where temporary disconnection must accept writes locally and merge via CRDTs.

18.1x Lower EBS Bill Ch. 03, 07 · [S14, S26, S27, S31, S41, S42]

8. Right-Size Block Storage: gp3 ($542) vs io2 ($9,824) vs Local NVMe

-2 boxes

2016 Outdated Wisdom

Provision maxed-out enterprise SAN / io2 volumes for every database, or assume cloud block storage can deliver 2,000 MiB/s at 4-KiB random I/O.

2026 Hardware Reality

1 TiB gp3 at 80K IOPS / 2,000 MiB/s costs $541.92/mo vs $9,824/mo for 1 TiB io2 at 256K IOPS (18.1x total bill). And 80K × 4 KiB is physically 312.5 MiB/s — check block size math!

P99 Latency2.5 ms → 0.24 ms (10.4x)

Cloud Bill$9,824 → $542/mo

Monthly Savings$9,282/mo saved

Don't pay $9,824/mo per TiB for provisioned IOPS when replicated local NVMe (I8g: 243 µs, millions of IOPS) or gp3 ($542/mo) fits the SLO.

Evidence, Math & Exceptions (4) Database & Storage

Verified AWS billing math [S41, S42]: 1 TiB gp3 baseline is $81.92/mo; maxed at 80K IOPS & 2,000 MiB/s is $541.92/mo.
1 TiB io2 Block Express at 256K IOPS costs $9,824.00/mo ($9,696 in IOPS charges alone — 25.18x the IOPS fee).
WiscKey/BlobDB key-value separation reduces LSM write amplification from 28x to 1.3x, eliminating artificial IOPS bloat.

Still use when: Single-instance legacy commercial databases that cannot replicate across local NVMe nodes and require <500 µs shared SAN durability.

Provably Safe Under GC Pause Ch. 10, 12 · [S43, S50]

9. Monotonic Fencing Tokens in Storage Over Redis Redlock

-5 boxes

2016 Outdated Wisdom

Run a 5-node Redis Redlock cluster with wall-clock TTL leases to guarantee mutually exclusive distributed locks.

2026 Hardware Reality

A GC pause or kernel scheduling stall longer than the TTL breaks Redlock mutual exclusion. Use storage-enforced monotonic fencing tokens (UPDATE ... WHERE version = expected) directly in Postgres/S3.

P99 Latency4.2 ms → 0.6 ms (7.0x)

Cloud Bill$1,500 → $0/mo

Monthly Savings$1,500/mo saved

Locking for correctness requires the storage engine itself to reject stale writes via an incrementing epoch/version column.

Evidence, Math & Exceptions (4) Caching & Queues

Wall-clock leases assume bounded process pauses and synchronized clocks — violated in real cloud environments.
S3 conditional writes (If-None-Match / If-Match ETag) and Postgres optimistic concurrency check the token atomically at commit time.
Eliminates a 5-node external lock cluster from the critical path.

Still use when: Advisory efficiency locks (preventing duplicate expensive work) where an occasional double-execution is harmless because storage is idempotent.

12 Boxes → 2 Nodes + CDN Ch. 27, 28, 29 · [S10, S38, S40]

10. Classic Interview Archetypes Fit on 1–2 Modern Nodes + CDN

-10 boxes

2016 Outdated Wisdom

Design a URL shortener, paste bin, or dispatch service with 12 boxes: Snowflake ID fleet, Cassandra ring, Redis cluster, Kafka, and 20 stateless app nodes.

2026 Hardware Reality

100M new URLs/mo is 40 writes/s; 10B redirects/mo is 3,800 reads/s. One 32-vCPU m9g.8xlarge ($1,143/mo) + Cloudflare/CloudFront CDN handles 100x that volume on a single Postgres database.

P99 Latency35 ms → 2 ms (17.5x)

Cloud Bill$14,500 → $1,143/mo

Monthly Savings$13,357/mo saved

Interview whiteboard diagrams routinely deploy planetary-scale distributed machinery for workloads that represent 0.5% of a single 2026 server.

Evidence, Math & Exceptions (4) Sharding & Services

Worked Example A [S38]: Even 1,000,000 RPS (2.6 trillion reqs/mo!) at 50 µs CPU/req is only 50 CPU-seconds/sec.
Edge CDN caching (max-age=60s) absorbs 90%+ of read redirects before they ever hit origin egress ($0.09–$0.05/GiB [S40]).
64-bit random/sequence Base62 IDs generated in Postgres or application memory need zero ZooKeeper coordination.

Still use when: Multi-continent active-active write ingestion where sub-20ms write latency is required in Tokyo, London, and Virginia simultaneously.

4. 2026 Radical Simplicity Reference Architectures

High-Scale API & Web Tier (100k–1M RPS, 2 kB Payload)

1M RPS · 2 GB/s Payload · 50 µs CPU/req

$16,500/mo vs $142,000/mo

2016 Multi-AZ Microservice Mesh + Redis + Sharded DB (28 boxes · 68 ms p99)

40x Small x86 VMs across 3 AZs ($14k/mo)
Cross-AZ Chatty RPC Tax at 2 GB/s ($97,901/mo [S34])
12-Node ElastiCache Redis Cluster ($8.5k/mo)
16-Shard MySQL/Vitess Fleet ($21.6k/mo)

2026 Zonal-Affinity Cell + Graviton5 M9g + Direct NVMe Postgres (4 boxes · 3.2 ms p99)

2x AWS m9g.48xlarge (192 vCPU, 768 GiB RAM, 100 Gbps) ($13,713/mo [S38])
Strict Same-AZ Routing (52.2 µs RTT, $0 Cross-AZ RPC Tax [S30, S34])
HA Primary/Standby Postgres 18 on Local NVMe ($2,787/mo)

Napkin Math 1M RPS × 50 µs = 50 CPU-seconds/sec. Avoiding one unnecessary 2 GB/s cross-AZ RPC hop saves $97,901/month [S34, S38].

ACID Financial Ledger (10,000 TPS, 315 TB/yr)

10k ACID TPS · 10 MB/s · 3x Quorum

$3,850/mo vs $28,400/mo

2016 Sharded NoSQL + Kafka + Saga Compensators + io2 SAN (19 boxes · 45 ms p99)

24-Node Cassandra Ring + Kafka Saga Orchestrator ($14.2k/mo)
Provisioned io2 256K IOPS Volumes ($9,824/mo per TiB [S42])
Nightly Reconciliation Batch Cluster ($4.4k/mo)

2026 Single-Primary HA Postgres 18 (Group Commit) + S3 Cold Tier (3 boxes · 1.8 ms p99)

3-AZ Quorum Postgres 18 on Local NVMe / gp3 ($3,100/mo [S41, S51])
Group Commit on 300 µs NVMe WAL (10 MB/s = 0.048% of Gen6 SSD [S03, S50])
Immutable Historical Partitions Tiered to S3 Standard ($750/mo [S39])

Napkin Math 10,000 TPS × 1 kB = 10 MB/s logical writes (315.36 TB/yr). Group commit batches concurrent fsyncs across one 677 µs cross-AZ quorum RTT [S30, S50].

Observability & Log Pipeline (100,000 lines/s, 263 TB/mo)

100k lines/s · 1 kB/line · 30-Day Retention

$2,950/mo vs $122,376/mo

2016 Un-Sampled SaaS Log Ingest ($0.50/GiB) or 30-Node Elastic (22 boxes · 950 ms p99)

Raw SaaS Ingest: 244,752 GiB/mo × $0.50/GiB ($122,376/mo [Sec 8.E])
Dedicated Log Forwarder & Kafka Buffer Fleet

2026 Tail-Sampled OTel + S3 Standard Parquet + NVMe ClickHouse (3 boxes · 110 ms p99)

S3 Standard 5:1 Compressed Parquet: 48,950 GiB × $0.023 ($1,126/mo [S39])
Stateless ClickHouse / Quickwit Query Pair with Local NVMe Cache ($1,824/mo)
OpenTelemetry Stratified Tail Sampling (100% errors, 2% baseline)

Napkin Math S3 concurrent range GETs stream ~2 GiB/s per instance [S32], scanning 1 TiB of compressed Parquet in ~8.5 minutes when cold and sub-second when cached on NVMe.

Hybrid Lexical + Vector AI Search (20M Docs, 5k QPS)

20M Vectors · BM25 + DiskANN + RBAC

$1,850/mo vs $11,400/mo

2021 5-DB Frankenstack (Postgres + Elastic + Pinecone + Redis + Kafka) (15 boxes · 54 ms p99)

Managed RAM-Resident Vector DB Cluster ($4,800/mo)
Stateful Elasticsearch Cluster ($3,200/mo)
RDS Primary + Debezium/Kafka CDC Sync + Redis ($3,400/mo)

2026 Unified Postgres 18 (pgvector DiskANN + BM25) on Local NVMe (2 boxes · 5.5 ms p99)

HA AWS I8g / M9gd Instance Pair with Local Gen5/Gen6 NVMe ($1,850/mo [S10, S14])
Single ACID Store: Relational RBAC + BM25 + DiskANN PQ Vector Index

Napkin Math DiskANN stores compressed PQ vectors in RAM and fetches full vectors from 100–243 µs local NVMe [S31, S32], cutting memory cost by 15x with zero CDC lag.

5. Verified 2026 Hardware, Cloud Maxima, Ten Rules & Primary Sources

10 Evidence-Based Rules from the 2026 Audit

#	Rule	Engineering Implication	Refs
#1	Model latency, bandwidth, and cost separately	A 600-Gbps link still has 590 µs cross-AZ RTT and $0.02/GB one-way transfer charges ($0.04/GB bidirectional).	[S11, S30, S34]
#2	Large machines exist, but maxima are not one composite VM	A 192-core Graviton5 VM (M9g), a 32-TiB memory instance (U7inh), and a 48-TB bare-metal server (GCP X5) are distinct SKUs.	[S10, S13, S20]
#3	Distinguish component peaks from application throughput	A 28.4 GB/s PCIe Gen6 SSD does not automatically establish database TPS; WAL fsync, locks, and indexes govern real throughput.	[S03, S50]
#4	Check instance, per-flow, and destination network limits together	A 600-Gbps VM is still capped at 5 Gbps per single ordinary TCP flow outside a cluster placement group, or 25 Gbps with ENA Express.	[S23, S24, S25]
#5	Treat AZ latency as a measured distribution, not a global constant	52.2 µs same-AZ and 590.4 µs cross-AZ p50 RTT come from a specific c8gn.8xlarge ENA Express test in us-west-2.	[S30]
#6	Never turn 1 / RTT into a database-wide TPS ceiling	1 / 0.6 ms = 1,667 ops/s applies only to a single serialized client thread without batching. Concurrent clients and group commit achieve 50k–100k+ TPS.	[S50, S51]
#7	Price memory and storage with dated SKU quotes	Conventional DRAM contract prices rose 93–98% QoQ in Q1 2026, while cloud instance rates follow published provider price feeds.	[S38, S46–S49]
#8	Compare storage at the actual I/O size and queue depth	At QD1, 243 µs latency caps a single thread at ~4,115 IOPS. Reaching 6.8M IOPS requires high concurrency (QD256) and 4-KiB blocks.	[S03, S16, S31]
#9	Include transfer, requests, retention, and restore in cost models	At 1M RPS (2 kB response), public AWS egress ($248.6k/mo) is 36.3x the cost of the 192-core compute VM ($6,857/mo).	[S38, S40, S43]
#10	Design cells, quorums, and static stability explicitly	99.99% availability allows 4m 22.8s/month downtime; 99.999% allows 26.28s/month. Use 3-AZ majority quorums and zonal affinity.	[S51, S54]

Verified 2026 Silicon & Cloud Instance Maxima

Dimension	AWS Verified	Azure Verified	Google Cloud Verified
High-Throughput VM Networking [S11, S17, S23]	C8in up to 600 Gbps (up to 384 vCPUs)	Esv7 up to 400 Gbps	C4N up to 400 Gbps (95M PPS; 50 Gbps single VPC flow)
Scale-Up Memory Instances [S13, S19, S20, S21]	U7inh-32tb: 32,768 GiB, 1,920 vCPUs, 200 Gbps	Mdsv3 VHM: 30,400 GiB (marketed 32 TB), 185 Gbps	X5 (x5-2064-48t-metal): 49,152 GB (48 TB), 2,064 vCPUs, GA Sep 2026
Local NVMe Capacity & IOPS [S14, S15, S16, S18, S22]	I8g up to 45 TB ($12,026/mo 48xl); I8ge up to 120 TB	Laosv5 up to 138.24 TB & 20.7M read IOPS; Edsv7 53.8 GB/s read	Z3 bare metal up to 72 TiB local NVMe SSD
Remote Block Storage Path [S12, S17, S27, S28, S29]	R8idb: 300 Gbps & 1.44M IOPS; io2: 256K IOPS / 4 GiB/s (<500 µs)	Esv7: 800K IOPS / 20 GB/s; Ultra Disk: 400K IOPS / 10 GB/s	Hyperdisk Extreme: 350K IOPS / 5,000 MiB/s; Balanced: 160K IOPS
Single-Flow & Inter-AZ Rules [S23, S24, S25, S34, S36, S37]	5 Gbps ordinary flow (10 Gbps cluster PG); 25 Gbps ENA Express; $0.02/GB 1-way cross-AZ	Inter-AZ transfer free on documented path (peering billed separately)	3 Gbps external flow; 50 Gbps C4N VPC flow; $0.01/GiB sent cross-zone

2026 Server CPU, Gen6 NVMe & Switch Silicon

Component	Verified Specification	Audit Qualification
AMD EPYC 9006 (Venice) [S01, S02]	Up to 256 cores / 512 threads; 16-ch DDR5/MRDIMM up to 12,800 MT/s (~1.6 TB/s/socket); 128 lanes PCIe Gen6	Announced platform maxima (SP7 16-ch, SP8 8-ch). 16 × 12,800 MT/s × 8B = 1,638.4 GB/s peak interface.
AWS Graviton5 (M9g / M9gd) [S10, S38]	192 cores, DDR5-8800, PCIe Gen6; GA 10 June 2026 (m9g.48xlarge: 192 vCPU, 768 GiB, $9.39264/hr = $6,856.63/mo)	Production cloud compute in US East (N. Virginia); 100 Gbps network and 72 Gbps EBS on 48xlarge.
Samsung PM1763 (PCIe Gen6 SSD) [S03, S04]	Up to 28,400 MB/s seq read, 21,000–21,900 MB/s write, 6.8M random-read IOPS, 950K random-write IOPS	High-queue-depth catalogue maxima (not QD1 latency). 6.8M × 4 KiB = 27.85 GB/s payload.
Micron 9650 & 6600 ION SSDs [S05, S06, S07]	Micron 9650 (Gen6): 28 GB/s read, 5.5M IOPS. Micron 6600 ION (Gen5 QLC): 245.76 TB usable capacity per drive	Separates ultra-fast Gen6 TLC performance from ultra-dense 245.76 TB Gen5 QLC capacity tiers.
NVIDIA ConnectX-8 / ConnectX-9 [S08]	ConnectX-8: up to 800 Gbps (PCIe Gen6 x16). ConnectX-9 SuperNIC: up to 1.6 Tbps aggregate	Adapter/fabric capability; check per-VM and per-flow cloud entitlements separately.
Broadcom Tomahawk 6 Switch [S09]	102.4 Tbps switching capacity per chip; production volume shipment 12 March 2026	Datacenter fabric switch silicon enabling non-blocking 400G/800G east-west clusters.

Primary Vendor & Benchmark Sources (23 Oct 2026 Cutoff)

Ref	Source & Link	Verified Scope
[S01]	AMD EPYC 9006 Series Specifications	Up to 256 cores, PCIe Gen6, 16-ch SP7.
[S02]	AMD EPYC 9006 Architecture White Paper (2026)	1,638.4 GB/s peak socket bandwidth with 12,800 MT/s MRDIMMs.
[S03]	Samsung Enterprise SSD Catalogue (PM1763)	28.4 GB/s seq read, 21 GB/s write, 6.8M random read IOPS.
[S06]	Micron 6600 ION 245.76 TB SSD	245.76 TB usable QLC enterprise SSD shipping June 2026.
[S08]	NVIDIA ConnectX-8 & ConnectX-9 SuperNIC	800 Gbps PCIe Gen6 x16 and 1.6 Tbps aggregate.
[S09]	Broadcom Tomahawk 6 102.4-Tbps Switch Release	Production-volume shipment announced 12 March 2026.
[S10]	AWS EC2 M9g & M9gd (Graviton5) GA	192 cores, DDR5-8800, PCIe Gen6; GA 10 June 2026.
[S11]	AWS EC2 C8in / C8ib 600-Gbps Instances	Up to 600 Gbps network and 384 vCPUs.
[S13]	AWS U7inh-32tb (32,768 GiB Memory)	32 TiB RAM, 1,920 vCPUs, 200 Gbps network.
[S16]	Microsoft Azure Laosv5 Storage-Optimized VMs	Up to 138.24 TB local NVMe and 20.7M random-read IOPS.
[S20]	Google Cloud X5 48-TB Bare Metal GA	x5-2064-48t-metal: 49,152 GB RAM, 2,064 vCPUs (25 Sep 2026).
[S25]	AWS ENA Express (SRD) Documentation	Up to 25 Gbps single flow including eligible cross-AZ paths.
[S27]	AWS EBS Provisioned IOPS io2 Block Express	Up to 256K IOPS, 4,000 MiB/s, <500 µs average target.
[S30]	DevelopersIO ENA Express Same-AZ vs Cross-AZ RTT Benchmark	27 May 2026 c8gn.8xlarge: 52.2 µs same-AZ p50 RTT vs 590.4 µs cross-AZ p50 RTT.
[S31]	DevelopersIO AWS I8g Local NVMe fio Benchmark	243.11 µs mean completion latency for 128-KiB QD1 random reads.
[S32]	Sirupsen Napkin-Math (March 2026 Refresh)	100 µs 8-KiB NVMe read, 300 µs fsync write, 2 GiB/s concurrent S3 range GET.
[S34]	AWS EC2 On-Demand & Inter-AZ Data Transfer Pricing	$0.01/GB at each endpoint across AZs ($0.02/GB one-way, $0.04/GB bidirectional).
[S38]	AWS EC2 Linux Rate Feed (US East N. Virginia, 22 Oct 2026)	m8g.xlarge $131/mo, m9g.8xlarge $1,143/mo, m9g.48xlarge $6,857/mo, i8g.48xlarge $12,026/mo.
[S39]	AWS S3 Regional Price List (15 Oct 2026)	S3 Standard $0.023/GiB-mo first 50 TiB; S3 Express One Zone $0.11/GiB-mo.
[S41]	AWS EBS Volume Types & Reference Rates	1 TiB gp3 (80K IOPS, 2,000 MiB/s) = $541.92/mo vs 1 TiB io2 (256K IOPS) = $9,824/mo.
[S46]	TrendForce Q1–Q4 2026 Enterprise DRAM & NAND Market Reports	Q1 2026 conventional DRAM contract prices rose 93–98% QoQ.
[S50]	PostgreSQL 18 WAL Configuration, Group Commit & Replication	Group commit amortizes fsync across concurrent transactions; synchronous_commit semantics.