The 100:1 math, worked end to end

No CDN, just a look-aside cache and arithmetic. Working the numbers because nobody else will:

Traffic: steady 10k redirects + 100 creates; viral hour 30k + 120. Read:write is 100:1, so every capacity decision is a read decision.

Hot set: 0.05% of 20B links = 10M links × 210 B = 2.1 GB. At 70% traffic share, holding the full 95% cacheable target needs ~3 GB of cache RAM. cache.small ×2 has 32 GB. The cache is trivially easy here; it is not the hard part of this problem.

App tier: this is the cost of skipping the CDN. The full 30k peak hits my app servers, so app.xlarge ×5 (40k capacity, ρ ≈ 0.75 at peak). Misses to the db: 30k × 5% = 1.5k.

Storage: 20B × 210 B ≈ 4.2 TB logical. db.large holds 2 TB, so 3 shards, each RF 3 + quorum for the never-lose-a-mapping contract. 9 db.large boxes is where the money goes.

Bill: $8.40/hr, vs $5.90 for the edge version. The $2.50/hr difference is what I pay for every redirect transiting my own compute. Whether that's dumb depends on whether you believe your CDN's cache hit accounting. I mostly do, so the edge design is probably better. Posting this anyway because the math should be on record.