Four Architectures Compared
James had been staring at TutorClaw's $60/month infrastructure cost and its $15,750/month revenue. The margin was extraordinary. But something nagged him.
"We designed Architecture 4 in Chapter 58," he said. "But you mentioned there were three other ways to build this. Were they all this cheap?"
Emma shook her head. "Not even close. Let me show you the full picture."
You are doing exactly what James is doing. You have one architecture's numbers memorized. Now you need to see all four side by side to understand why Architecture 4 was the right choice.
The Comparison Table
Here are all four architectures evaluated across eight dimensions, using TutorClaw's real numbers at 16,000 learners:
| Arch 1: Custom Brain | Arch 2: NanoClaw | Arch 3: Hybrid | Arch 4: MCP-First | |
|---|---|---|---|---|
| Monthly infra | $200-300 | $575-1,600 | $200 to $1,600 | $50-70 |
| Monthly LLM | $12,000 (you) | $12,000 (you) | $12,000 (you) | $0 (learner) |
| Infra + LLM cost | ~$12,300 | ~$12,575-13,600 | ~$12,300 to $13,600 | ~$50-70 |
| Revenue | $15,750 | $15,750 | $15,750 | $15,750 |
| Infra + LLM margin | ~22% | ~14-20% | ~14-22% | ~99.5% |
| Code execution | No | Yes (container) | No (Phase 1) | Yes (MCP submit_code) |
| IP protection | Server-side | Server-side | Server-side | MCP server (strongest) |
| Time to ship | 2-3 weeks | 2-4 months | 2-3 weeks | 2-4 weeks |
Look at the margin row. Three architectures cluster between 14% and 22%. Architecture 4 sits at 99.5%. That is not an incremental improvement; it is a fundamentally different business model.
These margins compare infrastructure and LLM costs only. Every architecture also pays Stripe fees (2.9% + $0.30 per transaction), which add roughly $1,650/month at this revenue level. Including Stripe, Architecture 4's gross margin drops to ~89% (see the Python calculator in Lesson 4). The point still holds: 89% versus 22% is a structural advantage, not a marginal one.
Where the Difference Lives
The infra + LLM cost row tells the story. Architectures 1, 2, and 3 all carry ~$12,000/month in LLM costs because the operator pays for every learner's model usage. Architecture 4 shifts that cost to the learner: they choose their model in OpenClaw, they pay their own API provider, and the operator's LLM line item drops to $0.
Infrastructure costs vary across architectures. NanoClaw (Architecture 2) runs containers for each learner, which pushes infrastructure to $575-1,600/month. The Custom Brain (Architecture 1) and Hybrid (Architecture 3) range from $200 to $1,600 depending on configuration. Architecture 4's MCP server, R2 storage, and PostgreSQL database cost $50-70/month total, and that cost stays flat whether you serve 1,000 learners or 100,000.
But infrastructure differences are noise compared to the LLM cost. The gap between $12,000/month and $0/month is what creates the margin gap.
Compute the Margins Yourself
The infra + LLM margin is: (Revenue minus Infra and LLM Cost) divided by Revenue, multiplied by 100.
For each architecture, plug in the numbers:
Architecture 1: ($15,750 - $12,300) / $15,750 = $3,450 / $15,750 = 21.9%
Architecture 2 (low infra, $575): ($15,750 - $12,575) / $15,750 = $3,175 / $15,750 = 20.2%
Architecture 2 (high infra, $1,600): ($15,750 - $13,600) / $15,750 = $2,150 / $15,750 = 13.7%
Architecture 3 (low end): ($15,750 - $13,600) / $15,750 = $2,150 / $15,750 = 13.7%
Architecture 4: ($15,750 - $70) / $15,750 = $15,680 / $15,750 = 99.6%
These calculations exclude Stripe fees, which every architecture pays equally (~$1,650/month at this revenue). Including Stripe, Architecture 4's gross margin drops to ~89% (see the Python calculator in Lesson 4). The comparison still holds because Stripe fees are the same across all four architectures; the differentiator is infrastructure plus LLM cost.
The Capability Tradeoffs
Cost is not the only dimension. The table includes three capability columns:
Code execution matters for a programming tutor. Architecture 2 (NanoClaw) provides sandboxed containers where learners can run code. Architecture 4 provides code execution through the MCP server's submit_code tool, which the student built in Chapter 58. Architectures 1 and 3 (Phase 1) have no code execution at all.
IP protection describes how well your pedagogical intelligence stays private. All four architectures protect content server-side. But Architecture 4's MCP protocol adds an extra layer: the learner's OpenClaw instance can only access content through MCP tool calls, never by direct file access. The MCP server is the sole gatekeeper.
Time to ship ranges from 2 weeks (Architectures 1 and 3) to 4 months (Architecture 2, NanoClaw). Architecture 4 falls at 2-4 weeks, which is what the student experienced building TutorClaw across Chapters 57 and 58.
The Honest Limitation
Architecture 4 dominates on every dimension except one: it requires learners to have OpenClaw installed. For learners who do not have OpenClaw, a lightweight Architecture 1 instance ("TutorClaw Web") serves as the onramp. This is a real constraint, not a footnote. Your onboarding funnel must account for the subset of learners who arrive without OpenClaw, especially in the early stages of a product launch.
The book teaches OpenClaw installation in Chapter 56, so for students following the curriculum, this subset shrinks rapidly. For a public product, you would need both: Architecture 4 as the primary experience and a minimal Architecture 1 fallback for new users.
Try With AI
Exercise 1: Verify the Margins
Give your AI assistant the comparison table and ask it to compute the gross margin for each architecture.
Here is TutorClaw's architecture comparison at 16,000 learners:
Architecture 1 (Custom Brain): Total cost ~$12,300, Revenue $15,750
Architecture 2 (NanoClaw): Total cost ~$12,700, Revenue $15,750
Architecture 3 (Hybrid): Total cost ~$12,300 to $13,600, Revenue $15,750
Architecture 4 (MCP-First): Total cost ~$50-70, Revenue $15,750
Calculate gross margin for each: (Revenue - Total Cost) / Revenue * 100.
Which architecture has the highest margin, and by how much does it
beat the second-best option?
What you are learning: Gross margin calculation from revenue and cost inputs. The exercise forces you to see that the margin gap between Architecture 4 and the others is not 10% or 20%; it is roughly 77 percentage points. That size of gap means a completely different business, not just a cheaper version of the same one.
Exercise 2: Break-Even for Architecture 1
Ask your AI assistant to find where Architecture 1 stops being profitable.
Architecture 1 costs ~$12,300/month (mostly LLM). Revenue is $15,750
at 16,000 learners (75% free, 19% paid at $1.75/mo, 6% premium at
$10.50/mo). If LLM costs stay at $12,000/month but the number of
paying learners drops, at what total learner count does Architecture 1
break even (revenue equals costs)? What if LLM prices drop 50%?
What you are learning: Break-even sensitivity. Architecture 1's margin is thin enough that modest changes in conversion rate or LLM pricing can push it into loss. Architecture 4, with $50-70 in fixed costs, stays profitable even at very low learner counts. This is the practical meaning of "margin as a moat."
Exercise 3: A Fifth Architecture
Ask your AI assistant to model a variation on Architecture 4.
What if I used Architecture 4 (MCP-First) but hosted the MCP server
on a managed platform at $500/month instead of a $40-60 VPS?
Revenue stays at $15,750/month. Compute the new total cost, gross
margin, and the margin difference compared to the original $50-70
Architecture 4. Is the managed platform worth it?
What you are learning: Even with 10x the infrastructure cost, Architecture 4's margin barely moves because the dominant cost in other architectures (LLM at $12,000/month) is still $0. This shows that Architecture 4's advantage comes from the cost structure, not from penny-pinching on servers.
James leaned back and studied the table. "This is like evaluating shipping contracts," he said. "At my old warehouse, we would get quotes from four carriers. Most people compared the per-package rate. But the total landed cost, that is the real number. You have to include fuel surcharges, insurance, dimensional weight fees. One carrier looked cheap per package but had a fuel surcharge that doubled the cost at volume."
"That is exactly what this table does," Emma said. "Per-package rate is like looking at infrastructure cost alone. Total landed cost is the total cost row, including LLM."
"I would have just listed the features," Emma added after a moment. "Code execution: yes or no. IP protection: strong or weak. Your framing, evaluating it like a supplier contract with total landed cost as the headline number, that is sharper than how I would have presented it."
James grinned. "Operations teaches you one thing: the line item that looks small on the quote is the one that kills your margin at scale. Here, LLM is that line item. Twelve thousand dollars a month, hiding in Architectures 1 through 3. Architecture 4 just removes it entirely."
"And the content that makes Architecture 4 possible, R2 delivers it for free. That is what the next lesson covers."