There's a thesis circulating in AI-infrastructure engineering that's worth dragging out of the investor circles and into the small-business conversation. It comes from an engineer named Kusireddy who runs $100K of his own money against it. Short version:
Wall Street is pricing the model layer. The money is in everything underneath it. Power, chips, data centers, observability, cost discipline. The model itself is commoditizing every six months.
If you run a 30-person business, you probably don't care which energy ETF is going to outperform. But the same gravity that shapes the investment story shapes your AI bill. The model layer is getting cheaper. The infrastructure around it (the integrations, the safeguards, the prompt engineering, the human-review process, the cost controls) isn't.
The agencies and vendors selling you "AI implementation" are pricing themselves around which model they use. The actual value, and the actual cost, lives somewhere else.
Here's the translation.
Investor thesis: the model is the iceberg's tip
The headline AI story is OpenAI vs Anthropic vs Google. Benchmarks, enterprise deals, who's hiring whose researchers. It's a compelling narrative; it also misses where the money is actually going.
Kusireddy's argument, distilled:
- Power. Microsoft restarted Three Mile Island because a single large data center can draw as much electricity as a small city. Sam Altman invested in Oklo (a nuclear reactor startup) because he's buying a stake in every watt OpenAI ever consumes.
- Chips. Everyone knows Nvidia. Almost nobody tracks ASML, the Dutch company that builds the only machines on earth that can print the most advanced chips. No ASML, no TSMC. No TSMC, no Nvidia. No Nvidia, no ChatGPT. One company. Total chokepoint.
- Data centers. Someone has to build them. Cool them. Power them. Vertiv makes the cooling infrastructure. Eaton handles power distribution. Boring names. Unsexy businesses. Non-negotiable to every dollar spent on AI.
- Observability and cost control. The software layer that keeps an AI agent from doing $47,000 worth of nothing at 3 a.m. Early, fragmented, and the market hasn't figured out how to price it.
The argument: when models commoditize (and they are, fast), the pricing power moves down the stack. The infrastructure layer is durable. The model layer is replaceable.
You can disagree with the investment thesis. The translation to your business doesn't depend on whether you do.
Same gravity, smaller stage: where your AI bill actually goes
Your business isn't trying to win the model race. You're trying to make AI pay for itself on a specific workflow. The same shape shows up at your scale.
The model is the cheapest part of your AI spend. A Claude Pro subscription is $20/month. The actual API cost of running invoice-chase on a 200-customer list is roughly $5/month. The model layer for the typical SMB AI workload, all-in, is under $30/month.
The integration layer is where the price tag is. Connecting Claude to QuickBooks, Gmail, Stripe, and Slack so the skill can actually do its job. Building the human-approval workflow. Setting up the cost caps. Writing the lint-clean skill text. Watching the agent for the first 30 days. This is what an agency quotes you $1,500-$5,000/month for; this is also what an in-house operator can do for $20/month plus their own time.
The observability layer is what saves you when something goes wrong. Daily spending caps. Iteration limits. Alerts when usage doubles. Lint checks on every new skill. Read-only by default permissions. Most SMB AI deployments don't have any of this. They run skills wide open and hope. When it breaks, they get a $4,700 bill they didn't expect.
The investor takeaway is the same as the SMB takeaway:
Don't pay for the model. The model is cheap and getting cheaper. Pay for the layer underneath. The layer underneath is where the durable value sits.
Three concrete implications for the 25-person company
Implication 1. Stop optimizing for "which model is best"
A meaningful share of vendor pitches in 2026 hinge on "we use the best model." Sometimes it's framed as "GPT-4," sometimes as "Claude Opus," sometimes as "a proprietary fine-tune of the latest frontier model."
For 90% of SMB workflows, the model doesn't matter that much. Invoice-chase, calendar triage, simple drafting, classification, basic Q&A: all of these run fine on the cheapest reasonable model tier (Claude Haiku, GPT-4o mini, etc.). The marginal performance you get from a frontier model isn't worth 10x the cost on these workloads.
If a vendor's pitch depends on the model tier, run the math on the AI Vendor Cost Reverse-Calculator. Pick the cheapest tier that actually does the job. The savings show up in the bottom line.
Implication 2. Pay for the integration, not the AI
When you do hire help, hire it for the boring layer. The "what we'll do for you" list that should justify the price:
- Build a Salesforce-to-Claude bidirectional sync that handles 12 custom fields cleanly.
- Set up the human-approval workflow in Slack so every send-money action pauses for the right person.
- Configure the cost caps, alerts, and monitoring dashboard.
- Write the lint-clean skill files and customize them for your specific edge cases.
- Train the team on the human-review loop for the first 30 days.
Nothing on that list is about AI per se. Each item is about the durable infrastructure that makes AI actually work in your business. This is what's worth paying for. The "we use AI" part of the value proposition is the part the market is going to commoditize out from under any vendor that hangs their hat on it.
The pre-agency self-audit walks through the 11 questions to ask. Notice that exactly one of them is about the model. The other 10 are about the layer underneath.
Implication 3. Invest in your own observability layer, even if it's tiny
The investor version of this implication is "buy Vertiv and Eaton." The SMB version is much smaller and free.
What "your own observability layer" looks like for a small business:
- Daily spending cap on your AI account ($20/day for most SMBs is plenty). Set it. Never think about it again.
- Alerts at half the daily cap. Five seconds in the billing dashboard.
- Lint every new skill before turning it on. Use the Claude Skill Linter. Sixty seconds.
- Read-only by default on every connector. See the connector permission cheat sheet for the per-platform recommendations.
- A 20-minute weekly review of what your skills did. Just read the logs. You'll catch 90% of the drift before it costs you anything.
Total time investment: about an hour to set up, 20 minutes a week to maintain. Total cost: zero beyond the Claude Pro subscription.
If you skip this layer, you'll eventually get an unexpected bill or an embarrassing email or a customer reply that makes you look bad. If you build it, you'll never know about the disasters you didn't have.
Why the boring layer is where the SMB advantage is
There's a strategic implication that runs deeper than cost.
The vendors that price themselves around the model layer are fighting on commoditizing terms. The agencies that quote you $3,500/month for "we use Claude" are selling you the part of the stack that will be free in 18 months.
The SMB that builds its own boring layer (skills, connectors, observability, cost discipline) is investing in the part that compounds. Three years from now, the models will be different. The skills you've written, the connectors you've wired, the human-review processes you've trained, the cost-control dashboards you've set up: those are all still there. You can swap the underlying model without rebuilding the layer.
That's the same shape as the infrastructure investment thesis, just at your scale. The model is replaceable. The infrastructure you build around it is your moat.
The 25-person company that figures this out in 2026 has a structural cost advantage over the equivalent 25-person company that's still paying an agency for "AI implementation" in 2028. The cost advantage compounds. The agency's margin doesn't.
Where this comes from, attributed honestly
I'm not taking credit for the investor framing. Kusireddy's original argument is sharper than my translation, and worth reading in full if you're investing your own money against the AI thesis. He's also published the reverse-engineering study on 200 funded AI startups and the $47K agent-loop incident report, both of which inform the more tactical posts in this archive.
What I've added is the translation. The same gravity that's moving real institutional capital toward picks-and-shovels infrastructure plays is the same gravity that should move a small-business owner's AI budget toward the boring durable layer. Different stakes. Same physics.
The deeper version
The full argument for treating the AI stack as a capital allocation problem (not a vendor selection problem) is the spine of The $100 Network (Digital Empire series, $9.99 on Kindle). The book covers the broader pattern: that the under-$100 AI stack is now more durable than most agency offerings because it's the part that compounds while the model layer commoditizes.
Related reading
- Before you pay an agency $3,500/month for proprietary AI, the tactical version of "don't pay for the commoditizing layer."
- How a small business runs AI agents without a $47K surprise bill, the observability layer at SMB scale.
- Connector permission cheat sheet, the access-control discipline that compounds.
- Invoice-chase skill deep dive, the skill layer that survives model swaps.
- Pre-agency self-audit (Part 1) and post-agency migration (Part 2), the bridge from agency-bought to in-house built.
Fact-check notes and sources
- The Wall Street infrastructure thesis is Kusireddy's: I Build AI Infrastructure for a Living. Wall Street Is Betting on the Wrong Layer, Level Up Coding, March 2026.
- The 73% wrapper finding: I Reverse-Engineered 200 AI Startups. 73% Are Lying, Towards AI, November 2025.
- The $47K agent-loop incident: We Spent $47,000 Running AI Agents in Production, Towards AI, October 2025.
- Microsoft's Three Mile Island restart for AI data center power, September 2024.
- Sam Altman's stake in Oklo, the nuclear reactor startup, public via Oklo's SEC filings.
- Model price compression (~10x in 18 months) per public pricing-page archives at OpenAI and Anthropic.
- Claude for Small Business launch details per Inc.com's coverage of the May 13, 2026 announcement.
This post is informational, not investment advice. The capital-allocation framing applies to operational decisions about your own business AI spend, not to whether to buy any specific company's stock. Mentions of OpenAI, Anthropic, Microsoft, Nvidia, ASML, Vertiv, Eaton, Oklo, and other third-party companies are nominative fair use. No affiliation is implied.
