bentoml.com/llm/inference-…

GitHub Actions cache keyed on package-lock hash cuts npm install from ~3min to ~40s — 4–5× CI speedup, one config block. Most repos still re-fetch every run because hashFiles() was never wired into the cache key. Target a 70–90% hit rate. eastondev.com/blog/en/posts/…

100% JSON-schema compliance with OpenAI's Structured Outputs (gpt-4o-2024-08-06), up from ~93% schema-following without it. Most LLM apps still parse-and-retry on malformed JSON because they never enabled the feature. One API param, deterministic output. openai.com/index/introduc…

OpenAI's text-embedding-3-small: $0.02 per million tokens, 5× cheaper than ada-002 — AND scores 44% on MIRACL retrieval vs ada-002's 31.4%. Most RAG apps still ship ada-002 because the upgrade is a one-line model-name change. openai.com/index/new-embe… text-embedding-3-small: $0.02

Annual SaaS subscribers churn at ~1/3 the rate of monthly subscribers (ProfitWell, 22K-company dataset). Default new signups to annual with a 15–20% prepay discount and churn typically drops 40–60%. The cheapest retention lever in SaaS is a checkbox. baremetrics.com/blog/annual-vs…

the "studios to systems" line hits hard - that's exactly what's happening

Cloudflare R2 egress: $0/GB. AWS S3 egress: $0.09/GB. A 10 TB/mo media-heavy SaaS pays $891/mo on S3 — $15/mo on R2. Same files, 60× cheaper. AWS now even waives transfer when you migrate off S3. developers.cloudflare.com/r2/pricing/

15% vs 30% — Apple's App Store cut for iOS devs under $1M annual proceeds. On $500K/yr that's $75K back in your pocket. Many indie devs still pay 30% because they never enrolled in the Small Business Program. One form, effective next calendar year. developer.apple.com/app-store/smal…

cloudzero.com/blog/lambda-pr…

Stripe Atlas: $500 to incorporate a Delaware C-corp (vs $2K–5K via lawyer). Includes EIN, bylaws, founder stock, 83(b), registered agent. Plus $50K+ in partner credits — Mercury, AWS, Xero, $2,500 Stripe credits. The partner stack alone clears the fee. stripe.com/atlas

vLLM's PagedAttention recovers 60–80% of KV-cache memory wasted by traditional serving — up to 24× throughput vs HuggingFace Transformers on the same GPU. Self-hosted LLM stacks still on raw transformers leave the multiplier on the table. blog.vllm.ai/2023/06/20/vll…

Bun install runs ~7× faster than npm — and the cause is ~165K syscalls vs npm's ~1M for the same job. Binary manifest cache, OS-native file copies, multi-core tarball extraction. Systems engineering, not JS optimization. bun.com/blog/behind-th…

A Standard-1x Heroku dyno ($25/mo) costs ~$3.19/mo as an equivalent Fly.io VM running continuously — ~8× cost gap. Fly.io scale-to-zero on idle takes staging environments to cents/day. 40% off with a machine reservation for steady prod.

Bunny.net CDN egress (EU/NA): $0.005/GB. AWS CloudFront: $0.085/GB. 17x cheaper, same files. A 5 TB/mo media app pays $25 on Bunny vs $425 on CloudFront. Most indie SaaS still default to CloudFront on the AWS-everything reflex.

Two things about Postgres Row-Level Security in multi-tenant SaaS: 1. With a composite index on (tenant_id, ...), RLS overhead is 2-4% - basically noise. 2. Without it, RLS can be 100x slower. The lead column on the index has to be tenant_id.

Stripe Tax: 0.5% per transaction, 100+ countries, US sales tax all states. Tax Complete adds filings in 93 countries — still 0.5%. Solo SaaS at $100K ARR pays ~$500/yr vs Avalara/TaxJar's $20K+/yr base fees. stripe.com/tax/pricing

The government subsidy optimization angle is wild. In the US we just argue about whether programs should exist - Europeans are building entire businesses around maximizing them. Different world.

DeepSeek R1: $0.55/M input, $2.19/M output. OpenAI o1: $15/M input, $60/M output. Same chain-of-thought capability at 27× lower cost on both axes — and R1 leads o1 on AIME, MATH-500, SWE-bench. A $100/day o1 workload runs ~$3.60/day on R1. api-docs.deepseek.com/quick_start/pr…

Neon Postgres cold-starts in ~500ms vs Aurora Serverless v2's ~15s — 30× faster. Scale-to-zero on dev/staging cuts the bill to ~24% of always-on. Caveat: disable scale-to-zero on prod primary if queries already complete in <10ms. neon.com/docs/guides/be…

Why is Aurora failover ~3x faster than RDS Multi-AZ? Aurora replicas read from the same shared storage volume - promotion takes <30s, no catch-up. RDS Multi-AZ has a synchronous standby that isn't queryable, so failover waits 60-120s on DNS + instance promotion.