Research brief: the falling cost floor of "real" web functionality for SMBs (June 2026)

Status: Synthesised June 2026. Sister brief to Research brief: the searchable, structured catalogue as a working tool — when records-not-prose pays off (June 2026) (searchable structured catalogue) and Research brief: the website as a working surface of the business — four capabilities, AI-citation decoupling, freshness as a real signal (June 2026) (website as working surface). Same skeptical, source-incentive-flagged methodology.

Thesis

Building a website that does real work — a database-backed tool, a customer account portal, a live-data feature — was once effectively large-company-only because it was expensive. That cost floor has dropped far enough to put it within SMB reach. The drop rests on three durable, independently-cheapening drivers, each of which got cheap on its own dated, pre-AI timeline (2004–2014):

1. Infrastructure — storage, compute, bandwidth, managed DB, serverless. See Amazon S3 launched March 14, 2006 at $0.15/GB — object storage as a metered utility, Amazon EC2 launched as limited public beta August 25, 2006 — single m1.small instance at $0.10/hour, Internet transit collapsed from $1,200/Mbps (1998) to ~$5/Mbps (2010) — ~240× per DrPeering/Norton series, Amazon RDS announced October 2009 (MySQL first); GA May 31, 2011 — managed DB absorbs admin/backup/failover, AWS Lambda previewed November 13, 2014; GA April 9, 2015 — code with no servers to provision.
2. The off-the-shelf parts — frameworks, payments, search, auth, viz. See Ruby on Rails open-sourced July 2004 (DHH from Basecamp); v1.0 December 2005 — convention-over-configuration, Django created 2003 at Lawrence Journal-World; released publicly 2005 — "batteries-included," built-in admin/auth/ORM, Stripe launched 2011 — card data never touches merchant server (Stripe.js → token), removing the heaviest PCI burden, Elasticsearch first release 2010 — open-source on Apache Lucene; commodity search, Algolia founded 2012 (Dessaigne & Lemoine; Y Combinator W2014) — search-as-a-service, no infra to manage, Auth0 founded 2013 — managed identity, social login, SSO, SAML, D3.js released 2011 by Mike Bostock (with Heer & Ogievetsky, Stanford) — foundational data-viz building block.
3. Data — government and community open-data displacing licensed feeds. See data.gov launched May 21, 2009 with 47 datasets — now 370,000+ (with >2,000 removed Jan 2025), OpenStreetMap founded 2004 by Steve Coast — response to UK Ordnance Survey not releasing data freely; OSMF established 2006, US Census TIGER/Line — geographic data, public domain, released annually, NWS/NOAA api.weather.gov — free public REST/JSON API, no key, no registration, no rate-limit account.

TL;DR magnitudes

• Storage: $0.15/GB (2006) → ~$0.02/GB (2025), ~85% / ~7× cheaper (S3 storage cost fell ~85% from launch — ~7× more data for the same dollar; independently corroborated).
• Entry compute: $0.10/instance-hr (2006) → ~$0.005/hr, ~20× cheaper at the floor (EC2 entry price fell from $0.10 to ~$0.005/instance-hour by 2018 — ~20× cheaper at the floor (Jeff Barr, HashiConf)).
• Internet transit: $1,200/Mbps (1998) → ~$5/Mbps (2010), ~240× (Internet transit collapsed from $1,200/Mbps (1998) to ~$5/Mbps (2010) — ~240× per DrPeering/Norton series).
• Capex → opex: dedicated/colo + sysadmin became managed services with administration absorbed (The structural shift: capex (server + sysadmin) → opex (managed service, administration included)).

The honest ceiling

The commodity floor is cheap; the bespoke labor to assemble, customize, integrate, and maintain a genuinely custom tool remains the dominant and still-substantial cost. Custom client portals run $20,000–$50,000 to build with $10,000–$25,000/yr maintenance (SPP (2025): custom client portal $20k–$50k initial development + $10k–$25k/yr maintenance; Agency Handy $25k–$60k all-in over 6–12 months), and ~80% of software features are rarely or never used (Pendo 2019 Feature Adoption Report: 80% of software-product features rarely or never used (615 subscriptions analyzed)). The floor fell; the ceiling did not. See R4 — The floor fell, the ceiling did not: bespoke client portals still cost $20k–$50k + $10k–$25k/yr — say so honestly.

The decade 2004–2014 — when the spine got cheap

See Timeline 2004–2014: the decade when each of the three historically-expensive parts independently got cheap and R1 — Anchor the falling-cost case on the decade 2004–2014 (infrastructure + parts + data); AI-assisted coding is NOT the spine. AI-assisted coding (2022–present) is a recent accelerant at the margin — it must NOT anchor the falling-cost argument. The case stands entirely on infrastructure + parts + data, all of which got cheap on dated, pre-AI timelines. See R2 — Keep AI-assisted coding at the margin; it lowers assembly labor recently but is not a structural driver.

What this means for Candid client work

• Rent the commodity parts; build only what is differentiated. See R3 — Rent the commodity parts (Stripe / Auth0 / Algolia / RDS / Lambda); build only what is genuinely differentiated logic.
• Pendo discipline. Don't scope features that won't be used; see R5 — Build features customers will actually use; ~80% of software features go unused (Pendo 2019).
• Editorial discipline. Flag vendor-incentive on every cost claim; see R6 — Flag the vendor incentive on every cost claim — AWS price-cut count, SaaS build-vs-buy ranges, agency portal estimates.

Source-incentive meta-finding

Vendor "we cut prices N times" framing is quarantined throughout. The dated launch prices and the current published prices are the primary figures; vendor retrospectives are excluded. Build-vs-buy SaaS vendor cost pages (SPP, Agency Handy, Space-O, Kavara) are used for ceiling figures but flagged. See Caveats: AWS vendor framing quarantined; build-vs-buy vendor cost pages flagged; DrPeering single-source; sysadmin wages needs-verification.