A Berlin founder launched an AI application into production last month. Total infrastructure cost: $10.83. In Austin, someone built a newsletter platform that charges subscribers but pays exactly zero dollars for hosting—just the domain registration. Singapore, meanwhile, is home to a solo developer running a full-stack SaaS product serving thousands of users for $21 a month. Domain included.
These aren't weekend experiments. They're actual businesses generating actual revenue, built atop an economic convergence that would have seemed implausible even three years ago: AI inference costs have collapsed by roughly 600-fold between 2020 and 2026, while cloud providers wage a quiet war to give away compute, storage, and services that used to cost hundreds monthly.
The new math of starting a company has arrived. And it's strange.
When the Floor Falls Out
The data behind this shift is almost violent in its implications. An academic study published in March analyzing more than 3,555 AI models across OpenRouter and Epoch datasets documented what researchers called a "Tiered Super-Moore's Law"—a 600-fold decline in token pricing between 2020 and 2026.
That's not improvement. That's obliteration of the old cost structure.
Current pricing reflects the wreckage. OpenAI's o3-mini model, as of March 2026, costs $1.10 per million input tokens and $4.40 per million output tokens. Anthropic's Sonnet class sits at $3 input and $15 output per million tokens—the latest Sonnet 4.6 maintains those prices even as performance climbs. Google's Gemini API offers multiple tiers. xAI launched speech-to-text at $0.10 per hour for batch processing in April.
Gartner projects that by 2030, running inference on a one-trillion-parameter model will cost providers more than 90% less than it did in 2025. The mechanisms vary—better chips, algorithmic improvements, raw competition—but the trajectory holds.
There's complexity lurking beneath the headline numbers, though. A second March paper documented what it termed a "Price Reversal Phenomenon." Cheaper list prices can actually cost more in practice when you factor in variable "thinking" tokens in reasoning models. In 21.8% of pairwise model comparisons, the apparently cheaper option ended up more expensive. Some swings reached 28-fold.
The lesson? Test your actual workload. Pricing pages lie by omission.
The Great Giveaway
While AI got cheaper, infrastructure providers began racing to give away the basics. It started gradually, then all at once.
Cloudflare Workers offers 100,000 requests daily per account on its free plan—updated as of April. Cloudflare Pages maintains a soft limit of 100 projects per account with effectively unlimited static asset requests. Their D1 serverless SQL database includes a free tier, though exact limits shift as the product evolves.
Vercel's Hobby plan remains free for personal projects. Netlify shifted to a credit-based free tier in September 2025, generating some grumbling in developer communities but keeping the entry point at zero. Neon offers free serverless Postgres with typical allocations around half a gigabyte of storage and limited compute hours. Upstash provides 256 MB Redis and 500,000 commands monthly. Qdrant Cloud advertises a "free forever" 1GB single-node vector database tier.
Authentication through Clerk costs nothing up to 50,000 monthly retained users per app on the Hobby plan, though sessions cap at seven days. Brevo delivers 300 emails daily at no charge. Amazon SES runs $0.10 per 1,000 emails plus ancillary charges—new customers receive $200 in free tier credits starting July 2025, though the pricing structure has grown more Byzantine through late 2025 and into 2026.
Cloudflare's registrar sells domains at cost. No markup. A .com typically runs around $10 annually.
The economics resemble something between a land rush and a loss-leader strategy that never quite converts to paid. Perhaps it's both.
Real Projects, Real Ledgers
These aren't thought experiments. A February case study documented a full-stack AI application built entirely on Cloudflare's infrastructure—Workers, D1, Durable Objects, Queues, R2 storage. AI inference via Google's Gemini 2.5 Flash (as of February 2026). Production-ready. No external database required.
A March 25 community post detailed a newsletter platform running on Cloudflare Pages, Workers, D1, and KV. Total hosting cost: zero. Domain: ten dollars annually. Another developer in March built a complete marketing automation toolkit on Supabase's free tier.
Multiple indie developer threads from 2025 through early 2026 cite infrastructure costs between zero and fifty dollars monthly using combinations of Vercel, Supabase, Resend, and Stripe. One Rails-based SaaS reported $45 monthly. Others climb to $150 as revenue grows, but the starting point has fundamentally changed.
The pattern repeats: domain ($10/year), maybe an AI subscription if you're using a chat interface for development—Claude Pro at $20/month or equivalent—and everything else on free tiers until usage forces your hand. For many workloads, that threshold sits higher than founders expect.
Anatomy of the $21 Stack
The minimal viable production stack in April 2026 looks roughly like this:
Frontend via Cloudflare Pages or Vercel Hobby. Backend on Cloudflare Workers (up to 100,000 daily requests) or Vercel's free tier. Database through Neon's free Postgres or Cloudflare D1. Authentication via Clerk for up to 50,000 users. Email via Brevo at 300 messages daily or SES for pay-per-use at tiny volumes. Analytics through Cloudflare Web Analytics. Vector search—if your application requires RAG implementations—via Qdrant's free 1GB tier.
Domain from Cloudflare Registrar at cost. Roughly $0.80-$1.20 monthly when you amortize the annual fee. Claude Pro or ChatGPT Plus for development assistance, $20/month. API usage pay-as-you-go if the application requires programmatic AI calls—OpenAI's o3-mini at current $1.10/$4.40 per million tokens often costs negligible amounts for early-stage products.
Total: around $21 monthly, depending on domain TLD and which development tool you prefer.
The infrastructure itself can run at zero until you exceed free tier limits. Which brings us to the fine print.
Stripe or Paddle handle payments without monthly fees—2.9% plus $0.30 per transaction for Stripe's US standard pricing as of March. You pay only when customers pay you.
Where the Free Lunch Ends
Free tiers come with constraints that actually matter. Cloudflare Workers enforce daily caps—hit 100,000 requests and traffic stops until the next day. Firebase Spark has daily quotas on reads and writes. These are hard stops, not overage billing.
You design for failure modes or you wake up to angry users.
Vendor changes happen without much warning. PlanetScale eliminated its free Hobby plan in March 2024. Netlify restructured pricing in September 2025, creating community friction despite maintaining a free tier. GitHub Actions began charging $0.002 per minute for self-hosted runners on March 1, 2026, after years of free access.
Hidden costs lurk in the details—they always do. Email deliverability at scale often requires dedicated IPs or verification tools that trigger billing. Storage egress outside Cloudflare's ecosystem can cost real money. Image transformation, CDN usage, background jobs. Each has thresholds where free becomes paid, sometimes abruptly. The Anthropic API pricing page warns that some features exclude free trial usage. AWS SES grew more complex through 2025-2026 changes around verification, deliverability monitoring, and IP allocation.
Regulatory considerations are emerging, particularly in Europe. The EU AI Act begins general application August 2, 2026, with obligations for general-purpose AI systems having started August 2, 2025. Small builders using AI in high-risk categories face transparency and risk management requirements. The compliance overhead may be minimal for many use cases but deserves attention before launch, particularly if you have European users.
That March paper about variable inference costs in reasoning models? It matters. Testing with production workloads beats spreadsheet projections every time.
What Changed, What Didn't
The $21 startup isn't a hack. It's not a temporary arbitrage that will close when someone at Amazon or Google realizes they're giving away too much. This is the new arithmetic of building software. AI costs are projected to fall another 90%-plus by 2030, per Gartner. Cloud providers compete on free tiers as moats, not loss leaders. The infrastructure to build, deploy, and scale to initial revenue now costs less than two movie tickets.
This doesn't mean every startup should bootstrap on free tiers until breaking even. Some domains need managed services, compliance tools, or compute that exceeds hobby limits from day one. But the decision is now strategic rather than forced by capital constraints. That's the shift.
Sequoia noted in January that new capacity coming online in 2026 could drive further compute price declines, though timing remains uncertain. The direction is consistent even if the slope varies month to month.
The founders launching for $10.83 or $21 aren't cutting corners. They're building on infrastructure that didn't exist at these price points five years ago, using AI models that cost one six-hundredth what similar capability would have required in 2020. The barriers to starting have shifted from "can I afford to build this" to "is this worth building."
That's perhaps the harder question.
But it's the better one to face.
