Failed Startup Stack Autopsy
What did dead startups spend on infra, and what should they have used instead? We analyze 34 real post-mortems — exposing how over-engineering and wrong tool choices accelerated failure.
Data from public founder retrospectives, HN posts, and blog post-mortems. Every dollar of burn documented.
NFTVault Pro
2021–2023 · $8M Series A$31.5K/mo
on $35K MRR at peak (crypto winter → $8K) revenue
Stack
mistakeAWS (multi-region EKS)
$18000/mooverkillAlchemy (Growth plan)
$5000/mooverkillMoralis
$2500/mooverkillDatadog (Enterprise)
$4200/mooverkillAuth0 (Enterprise)
$1800/moCause of death: Built for crypto bull market scale that evaporated. $31.5K/mo infra burn with 90% of capacity sitting idle after NFT market crashed. Couldn't downscale fast enough — annual contracts.
What they should have done: Web3 startups should use pay-per-use infra, not enterprise contracts. When your market can crash 80% in a month, annual Datadog and Auth0 contracts become anchors. Serverless + usage-based pricing is insurance against market volatility.
Source: Founder keynote at ETHDenver (2023)
AgentForge
2024–2025 · $3.2M seed$27.2K/mo
on $22K MRR revenue
Stack
mistakeGPT-4 Turbo (agent orchestration)
$18000/momistakeAnthropic Claude 3 Opus (code gen)
$7500/mooverkillWeaviate Cloud (vector store)
$1500/mogoodVercel (Pro)
$120/mogoodSupabase
$75/moCause of death: AI agent platform running multi-step reasoning chains with GPT-4 and Opus. Each agent run cost $0.50-2.00 in API calls. Users expected unlimited runs on $49/mo plan. 10 power users = $5K/mo in LLM costs.
What they should have done: AI agent products have non-linear cost curves. Each reasoning step multiplies cost. Use smaller models (Haiku, GPT-4o-mini) for planning steps, big models only for final output. Hard cap agent steps. Charge per-run, not per-seat.
Source: CTO post-mortem (2025)
PromptLayer
2023–2025 · $1.8M seed$19.7K/mo
on $15K MRR revenue
Stack
mistakeOpenAI GPT-4 + GPT-4 Turbo
$14000/mooverkillAnthropic Claude 3 Opus (fallback)
$4200/mooverkillPinecone (vector DB)
$1200/mogoodVercel
$200/mogoodSupabase
$75/moCause of death: LLM cost death spiral. Every new user increased API costs faster than revenue. Opus as 'fallback' cost more than GPT-4 primary. No caching, no model routing, no cost controls.
What they should have done: LLM apps need a cost architecture from day one. Semantic caching saves 40% on repeated queries. Route 80% of traffic to Haiku/GPT-4o-mini. Use Opus/GPT-4 only for complex reasoning. If your LLM bill > 50% of revenue, you don't have a business.
Source: Founder post-mortem (Substack, 2025)
TokenGate
2022–2023 · $3.8M seed$17.0K/mo
on $12K MRR at peak revenue
Stack
goodVercel
$80/momistakeAlchemy (Enterprise)
$8000/mooverkillAWS (Lambda + DynamoDB)
$3500/momistakeChainalysis (compliance)
$5000/mooverkillAuth0
$450/moCause of death: Web3 gating platform that spent $13K/mo on blockchain and compliance tools for a product with 200 active users. Chainalysis alone cost more than total revenue.
What they should have done: Compliance tools priced for exchanges (Chainalysis) aren't meant for seed-stage startups. At 200 users, manual review + basic heuristics work fine. Don't buy institutional compliance for a consumer app.
Source: Founder tweet (2023)
CrowdLinker
2020–2023 · $3.5M Series A$16.6K/mo
on $45K MRR at peak revenue
Stack
overkillAWS (full stack — Lambda, DynamoDB, S3, SQS, CloudFront)
$12000/mooverkillDatadog (APM + Logs + Infra)
$2400/mooverkillSalesforce
$1800/mogoodStripe
$450/moCause of death: Over-architected from day one. 37% of MRR went to infra. Couldn't reach profitability.
What they should have done: DynamoDB + Lambda + SQS for a CRUD app is engineering theater. A single Postgres on Railway would have handled 10x the load at 1% of the cost.
Source: CTO retrospective (Medium, 2023)
ChainVault
2022–2023 · $5M seed$15.5K/mo
on $2K MRR revenue
Stack
overkillAWS (EKS + Lambda + DynamoDB)
$8500/mooverkillAlchemy (blockchain node)
$3200/mooverkillThe Graph (indexing)
$1800/mooverkillAuth0
$600/mooverkillDatadog
$1400/moCause of death: Web3 wallet SaaS built for millions of users that never came. Alchemy costs scaled with blockchain reads, not revenue. Team of 4 burning $15K/mo on infra for 80 users.
What they should have done: Web3 infra costs are decoupled from user growth — blockchain reads cost money whether you have 10 users or 10K. If your revenue model is SaaS but your cost model is per-read, you're building a charity.
Source: YC post-mortem thread (2023)
MeetingAI Pro
2023–2025 · $2.5M seed$14.3K/mo
on $18K MRR revenue
Stack
overkillOpenAI Whisper API
$3800/momistakeGPT-4 (summarization)
$6200/mooverkillAWS (MediaConvert + S3)
$2400/mooverkillMux (recording)
$1800/mogoodVercel
$80/moCause of death: AI meeting notes tool competing with Otter.ai, Fireflies, and a dozen others. Costs scaled linearly with meeting hours but revenue was flat-rate per seat. Power users cost 10x what they paid.
What they should have done: Meeting AI is a race to the bottom. Self-host Whisper for transcription ($0.10/hr vs $0.36/hr API). Use Claude Haiku for summaries instead of GPT-4. Cap meeting hours per plan tier. Or just don't enter this market.
Source: Founder post-mortem (2025)
DocuBrain
2023–2025 · $2M seed$13.6K/mo
on $11K MRR revenue
Stack
mistakeAnthropic Claude 3 Opus
$9800/mooverkillWeaviate Cloud
$1200/mooverkillAWS Bedrock
$2400/mogoodVercel (Pro)
$120/mogoodClerk
$50/moCause of death: Document analysis AI using Claude Opus for everything — including tasks Haiku could handle. When Anthropic raised Opus pricing, monthly costs jumped 40% overnight with no way to absorb it.
What they should have done: Single-model dependency is as dangerous as single-cloud dependency. Use model routing: Haiku for classification/extraction (90% of queries), Sonnet for summarization, Opus only for complex reasoning. Diversify across providers for pricing resilience.
Source: CTO post-mortem blog (2025)
StackPilot
2023–2025 · $2M seed$12.8K/mo
on $16K MRR revenue
Stack
mistakeAWS (multi-service: ECS, RDS, ElastiCache, SQS, SNS, CloudWatch)
$8200/mooverkillTerraform Cloud (Team & Governance)
$700/mooverkillDatadog (Enterprise APM)
$3400/mooverkillLaunchDarkly
$450/mogoodClerk (Pro)
$100/moCause of death: DevTools startup that 'practiced what they preached' with a full enterprise stack for 3 engineers and 200 users. 80% of MRR eaten by infrastructure.
What they should have done: DevTools founders are the worst offenders — they over-engineer because they love infrastructure. Your 200 users don't care that you're running Terraform Cloud. Railway + simple deploys = same outcome at 5% of the cost.
Source: CTO lightning talk at DevOps Days (2025)
GridOps
2020–2022 · $4.2M Series A$11.9K/mo
on $28K MRR revenue
Stack
mistakeConfluent (Kafka managed)
$3800/mooverkillDatabricks
$4500/mooverkillAWS MSK
$1200/mogoodTerraform Cloud
$350/mooverkillDatadog
$2100/moCause of death: Big data infrastructure for a small data problem. Kafka for 50 events/second. Databricks for reports that could run in 2 seconds on Postgres.
What they should have done: Kafka is for millions of events per second. At 50 events/second use Postgres + BullMQ. Databricks is for petabyte-scale ML pipelines. At under 1M rows use DuckDB or just Postgres with a good index.
Source: CTO talk at local meetup (2022)
SynthWave
2023–2024 · $750K pre-seed$9.5K/mo
on $4K MRR revenue
Stack
mistakeOpenAI GPT-4 (music generation)
$5400/mooverkillReplicate (Stable Audio)
$2200/mooverkillAWS S3 + CloudFront
$1800/mogoodVercel
$60/moCause of death: AI music generation startup where each song cost $2-5 in API calls but subscription was $9.99/mo. Users generated 10+ songs per session. Unit economics were impossible.
What they should have done: Generative AI with per-generation costs and flat-rate subscriptions is a trap. Either charge per-generation, implement strict quotas, or use self-hosted models. $5/song on a $10/mo plan means you lose money on every active user.
Source: Founder Twitter thread (2024)
DevPilot
2023–2024 · $1.5M seed$9.4K/mo
on $12K MRR revenue
Stack
mistakeOpenAI GPT-4
$8500/mooverkillPinecone
$700/mogoodVercel
$120/mogoodSupabase
$75/mogoodClerk
$50/moCause of death: OpenAI API costs consumed 70% of revenue. Couldn't find margin even with paying users.
What they should have done: LLM wrapper businesses can't survive on GPT-4 margins. Use smaller models (Haiku, Mistral) for 80% of queries, route only complex ones to GPT-4. Or fine-tune an open model.
Source: Founder blog post (2024)
DataPulse
2021–2024 · $2.8M seed$9.2K/mo
on $20K MRR revenue
Stack
overkillSnowflake
$4200/momistakeFivetran (12 connectors)
$2400/mooverkilldbt Cloud (Team)
$600/mooverkillSigma Computing
$1200/mooverkillSegment
$800/moCause of death: Analytics startup used a full enterprise data stack to analyze their own product data. 46% of MRR went to understanding their users, not serving them.
What they should have done: The modern data stack (Fivetran + Snowflake + dbt + BI tool) costs $8K+/mo minimum. PostHog does product analytics, event tracking, and session replay for $0 at your scale. Don't build a data team before you have product-market fit.
Source: COO retrospective on LinkedIn (2024)
Novu Labs
2022–2024 · $1.8M seed$8.2K/mo
on $14K MRR revenue
Stack
overkillGCP (GKE + Cloud SQL + Cloud Run)
$5600/mooverkillElastic Cloud (logging)
$900/mooverkillSegment
$1200/mooverkillIntercom
$500/moCause of death: Hired a GCP architect at month 2. Infra became a vanity project. 58% MRR burn on tooling.
What they should have done: Segment at $14K MRR is paying $1.2K to pipe data to tools you barely use. PostHog does event tracking and product analytics for $0. Intercom at $500/mo can be replaced by Crisp ($25/mo) at early stage.
Source: HN comment by former employee (2024)
KubeCart
2021–2023 · $3.2M seed$8.2K/mo
on $11K MRR revenue
Stack
mistakeGKE (3 clusters, 3 envs)
$6800/momistakeIstio + Linkerd (both!)
$800/momistakeArgoCD + Flux (both!)
$400/mooverkillHelm + Kustomize
$0/mogoodCloudflare
$200/moCause of death: Kubernetes cosplay taken to the extreme. Running TWO service meshes and TWO GitOps tools 'to evaluate which is better.' One engineer spent 100% of their time on K8s. Zero time on product.
What they should have done: If you're running Istio AND Linkerd in production, you're not doing engineering — you're writing a blog post. Pick one tool for each job. Better yet, skip K8s entirely until you need horizontal scaling across regions.
Source: DevOps engineer post-mortem (Medium, 2023)
PredictML
2023–2024 · $1.2M seed$8.1K/mo
on $6K MRR revenue
Stack
mistakeAWS SageMaker
$5600/mogoodAWS S3 (training data)
$800/mooverkillMLflow (self-hosted on EC2)
$600/mooverkillWeights & Biases
$400/mooverkillDatadog
$700/moCause of death: ML platform startup using SageMaker for training when 90% of their models were XGBoost that could train on a $20/mo instance. Paid enterprise ML tooling for a scikit-learn problem.
What they should have done: SageMaker is for teams training large neural networks on GPU clusters. If your models train in under 10 minutes on a laptop, use a simple EC2 instance or Modal. Don't buy a Formula 1 pit crew for a bicycle.
Source: ML engineer blog (2024)
ContextAI
2023–2024 · $900K seed$7.8K/mo
on $9K MRR revenue
Stack
mistakeOpenAI GPT-4 Turbo
$6200/mooverkillWeaviate (cloud)
$800/mooverkillTemporal Cloud
$650/mogoodVercel
$80/mogoodSupabase
$50/moCause of death: LLM costs non-linear with usage. As beta users came on, OpenAI bill tripled. No cost guardrails. Invoice shock at month 3 forced cap-then-shutdown.
What they should have done: LLM APIs need hard cost guardrails from day one. Set per-user token budgets. Use caching (Redis + semantic cache). Route simple queries to GPT-3.5 or Claude Haiku. GPT-4 for everything is a cost cliff.
Source: Founders' shared post-mortem (2024)
PulseMetrics
2021–2023 · $1.2M seed$7.6K/mo
on $18K MRR revenue
Stack
overkillSnowflake
$2800/mooverkilldbt Cloud
$600/mooverkillFivetran
$1200/momistakeLooker
$3000/mogoodVercel
$40/moCause of death: Modern data stack for a 3-person team. Looker alone cost $3K/mo — more than their hosting costs combined. Data stack ate 42% of MRR.
What they should have done: Snowflake + Fivetran + dbt + Looker is a $500K/year enterprise stack. At under $50K MRR use Postgres + Metabase ($500/mo). The modern data stack is for post-Series-B.
Source: Founder LinkedIn post (2023)
LeadBlitz
2022–2024 · $1.4M seed$5.7K/mo
on $8K MRR revenue
Stack
overkillApollo.io (data enrichment)
$2400/mooverkillInstantly (cold email)
$900/mooverkillSmartlead
$600/mooverkillOpenAI (personalization)
$1800/mogoodVercel
$40/moCause of death: Sales automation tool that sold to other sales teams. Their own GTM stack cost more than their product revenue. Ironic: a sales automation startup that couldn't automate its way to profit.
What they should have done: Don't spend $5.7K/mo on outbound tools when your MRR is $8K. At early stage, manual outreach on LinkedIn + personal email gets you to PMF faster and cheaper than three-tool cold email stacks.
Source: Founder tweet (2024)
KubeFlow Analytics
2022–2024 · $800K pre-seed$5.4K/mo
on $6K MRR revenue
Stack
overkillKubernetes (self-managed EKS)
$4500/mooverkillPrometheus + Grafana (self-hosted)
$600/momistakeIstio service mesh
$300/mogoodSupabase
$25/moCause of death: Two engineers spent 60% of their time on Kubernetes maintenance. Product velocity died. Competitors shipped 3x faster.
What they should have done: Kubernetes at 100 users is not engineering — it's cosplay. Railway or Fly.io handles 100K users without a dedicated DevOps engineer. Save K8s for when you have 10 engineers and real traffic.
Source: YC Hacker News post-mortem (2024)
SyncDesk
2022–2024 · $2.4M seed$5.4K/mo
on $22K MRR revenue
Stack
mistakeSalesforce (Enterprise)
$2400/mooverkillHubSpot Marketing Hub
$800/mooverkillZendesk Suite
$600/mooverkillSlack Business+
$400/mooverkillAWS
$1200/moCause of death: GTM tooling for a B2B startup with 40 customers cost more than their hosting. Salesforce had a 6-month onboarding. Team never used 90% of the features.
What they should have done: Salesforce at 40 customers is a cargo cult move. Use Notion for CRM until $500K ARR. HubSpot free tier handles outbound for most early-stage companies. Don't buy the enterprise suite because enterprise feels legitimate.
Source: CRO retrospective post on LinkedIn (2024)
MetricHQ
2022–2024 · $1.1M seed$5.4K/mo
on $7K MRR revenue
Stack
mistakeDatadog (full stack)
$3200/momistakeNew Relic (backup monitoring)
$1400/mooverkillPagerDuty
$450/mogoodSentry
$280/mogoodRailway
$40/moCause of death: Monitoring an app with 200 users using TWO enterprise monitoring platforms. $4.6K/mo on observability for an app that Railway's built-in metrics could have covered.
What they should have done: Running Datadog AND New Relic is not 'defense in depth' — it's burning money in stereo. At $7K MRR, use BetterStack ($30/mo) + Sentry free tier. Save enterprise monitoring for when you have an on-call team.
Source: CTO blog post (2024)
Zestio
2021–2024 · $2.1M seed$4.7K/mo
on $8K MRR at peak revenue
Stack
overkillAWS (ECS + RDS + ElastiCache)
$3200/mooverkillDatadog
$800/mooverkillAuth0
$450/mogoodVercel
$40/mogoodStripe
$180/moCause of death: Ran out of runway. Infra costs consumed 58% of revenue.
What they should have done: At $8K MRR, use Railway ($50/mo) not AWS ECS ($3.2K/mo). Use PostHog (free) not Datadog ($800/mo). Use Clerk ($25/mo) not Auth0 ($450/mo).
Source: Founder post-mortem on HN (2024)
FleetOps
2022–2024 · $1.5M pre-seed$4.7K/mo
on $5K MRR revenue
Stack
overkillTerraform Cloud (Team)
$350/mooverkillHashiCorp Vault
$500/momistakeConsul
$300/momistakeNomad
$400/mooverkillDatadog
$900/mooverkillAWS (ECS)
$2200/moCause of death: Bought the entire HashiCorp suite because the CTO came from a company that used it. Team of 3 running Vault + Consul + Nomad + Terraform for a single-service app.
What they should have done: Enterprise infrastructure tools exist for enterprise problems. HashiCorp suite is for managing 100+ services across teams. For a single app: Railway (deploy) + .env files (secrets) + simple DNS (service discovery). Done.
Source: Engineering retrospective (dev.to, 2024)
RepoInsight
2023–2024 · $500K angel$3.5K/mo
on $1.8K MRR revenue
Stack
goodGitHub API (heavy usage)
$0/mooverkillVercel
$400/mogoodNeon (Postgres)
$69/mooverkillOpenAI GPT-4o
$2800/mooverkillClerk
$250/moCause of death: GitHub analytics dashboard that analyzed repos with GPT-4o. GitHub's free API rate limits hit quickly, Vercel costs scaled with cron jobs. Revenue never justified the AI analysis cost per repo.
What they should have done: Build-on-top-of-platform businesses are fragile. GitHub API rate limits constrain your scale. Pre-compute insights in batch jobs, cache aggressively, and serve static dashboards instead of real-time AI analysis.
Source: Founder indie hackers post (2024)
Tracify
2023–2024 · $300K angel$2.7K/mo
on $4K MRR revenue
Stack
overkillDatadog (full APM + logs + synthetics)
$1800/mooverkillPagerDuty
$450/mooverkillLaunchDarkly
$350/mogoodRailway
$30/mogoodSupabase
$25/moCause of death: Pre-revenue, team of 2, spending $1.8K/mo on Datadog to monitor a hobby-scale app. Zero ROI from observability tooling.
What they should have done: Datadog is for teams with on-call rotations and SLAs. At pre-revenue, Sentry free tier ($0) + Railway built-in metrics ($0) is enough. Don't buy enterprise monitoring to feel legitimate.
Source: Founder tweet thread (2024)
Loomify
2021–2022 · $600K pre-seed$2.2K/mo
on $5K MRR revenue
Stack
mistakeHeroku (multiple dynos)
$800/mooverkillCloudinary (high tier)
$650/mooverkillSendGrid (Marketing tier)
$400/mooverkillMongoDB Atlas M30
$380/moCause of death: Heroku removed free tier in 2022, forcing a migration mid-growth sprint. Three weeks lost to infra migration. Growth momentum broken.
What they should have done: Free-tier dependency is technical debt. When Heroku killed its free tier, startups on it had to drop everything to migrate. Railway, Render, or Fly.io offer better pricing without the rug-pull risk.
Source: Blog post (2022)
Meadow
2022–2024 · $500K pre-seed$2.2K/mo
on $3K MRR revenue
Stack
mistakeFirebase (Firestore + Auth + Hosting)
$1200/mooverkillAlgolia
$350/mogoodTwilio
$600/mogoodVercel
$20/moCause of death: Firebase costs exploded after launch. One bad Firestore query cost $2K in a single day. Lost 3 weeks migrating mid-crisis.
What they should have done: Firestore per-read pricing is a landmine. If your data model isn't perfect, one recursive query bankrupts you. Use Supabase (Postgres) — costs are predictable.
Source: Twitter/X thread by founder (2024)
Calmly
2022–2024 · $400K angel$1.9K/mo
on $600 MRR revenue
Stack
overkillLiveblocks (multiplayer)
$500/mooverkillMux (video)
$800/mooverkillDaily.co (video calls)
$600/mogoodSupabase
$25/mogoodVercel
$20/moCause of death: Meditation + therapy group video platform. Built impressive multiplayer + video stack, but only 30 active users. Cost-per-user was $65/mo on a $14.99/mo subscription.
What they should have done: Video and real-time collaboration infrastructure is expensive per-user at low scale. Validate with Zoom/Google Meet links first. Don't build custom video until you have 500+ paying users who specifically need it.
Source: Founder blog (2024)
CollabNote
2022–2024 · $600K angel$1.6K/mo
on $1.2K MRR revenue
Stack
overkillLiveblocks (collaboration)
$800/mooverkillPlanetScale (Scaler Pro)
$299/mooverkillVercel (Pro)
$200/mooverkillClerk (Pro)
$100/mooverkillLaunchDarkly
$250/moCause of death: Built a premium collaborative note-taking app in a market dominated by Notion, Google Docs, and free alternatives. Stack was reasonable per-tool but total exceeded revenue.
What they should have done: In commoditized markets, even moderate infra costs kill you. $1.6K/mo infra on $1.2K revenue means every user costs you money. Use free tiers aggressively until PMF is proven: Supabase free, Vercel free, Yjs (self-hosted) instead of Liveblocks.
Source: Founder reflection on Indie Hackers (2024)
ShipFast AI
2023–2024 · Bootstrapped$1.1K/mo
on $2K MRR revenue
Stack
overkillShopify
$299/momistakeShopify apps (5 apps)
$350/mooverkillKlaviyo
$250/mooverkillGorgias
$180/moCause of death: SaaS stack cost exceeded gross margin. Shopify + apps + email + support = $1K/mo on $2K revenue.
What they should have done: The Shopify app ecosystem is a hidden tax. Each $30-100/mo app adds up fast. At low revenue, use Medusa.js (free) + Resend (free) + plain email support.
Source: Indie Hackers post (2024)
FormWave
2022–2023 · $150K angel$599/mo
on $800 MRR revenue
Stack
overkillNext.js on Vercel (Pro)
$200/mooverkillPlanetScale (Scaler Pro)
$299/mooverkillClerk (Pro)
$100/mogoodResend
$0/moCause of death: Not a cost problem — pure PMF failure. But over-provisioned database for a product with 50 users.
What they should have done: PlanetScale Scaler Pro at 50 users is paying for traffic you don't have. Supabase free tier handles the first 500MB without a credit card. Match your database tier to your actual data volume.
Source: Indie Hackers forum (2023)
NoteStack
2022–2023 · $200K angel$50/mo
on $1.5K MRR revenue
Stack
goodRailway
$25/mogoodSupabase
$25/mogoodResend
$0/mogoodVercel
$0/moCause of death: Product-market fit failure. Not an infra problem — just no demand.
What they should have done: Good stack choices. $50/mo infra is exactly right for a startup exploring PMF. They failed for product reasons, not technical ones. This is how it should work.
Source: Founder AMA on Reddit (2023)
InstaAPI
2023–2024 · Bootstrapped ($40K savings)$15/mo
on $200 MRR revenue
Stack
goodRailway
$15/mogoodSupabase
$0/mogoodVercel
$0/mogoodResend
$0/moCause of death: Perfect stack choices but no market. Built an API aggregation tool nobody needed. Failed for product reasons, not technical ones.
What they should have done: You can get infra costs to near-zero with modern tools. $15/mo is achievable. If you still fail, at least you know it was the idea, not the infrastructure. This is the right way to explore PMF.
Source: Founder reflection (personal blog, 2024)