Companies Are Paying "Taxes"
Look at enterprise IT spending structurally, and you see it is not investment in technology — it is taxes.
You pay because there's no alternative. You pay out of inertia. You pay because switching is frightening. This is not technology investment. It is structural taxation.
Six taxes companies pay:
- Oracle Tax / SQL Server Tax — Database licensing
- Microsoft Tax — Windows, Office 365, Azure
- Cloud Tax — AWS / Azure / GCP monthly billing
- SaaS Tax — Accumulating monthly subscriptions
- SIer Tax — Outsourced system development and operations
- Consultant Tax — Paying outsiders to tell you what to do
Claude is the tool that structurally eliminates these taxes.
Oracle Tax / SQL Server Tax — The Clearest Example
Have you ever seen Oracle Database licensing costs?
The Oracle tax structure: Oracle Database Enterprise Edition → Millions of yen per processor per year Oracle RAC (high availability) → Additional license Oracle Partitioning → Additional license Oracle Advanced Security → Additional license Oracle support contract → 22% of license cost paid annually, forever You're charged for every feature you use. You pay the base fee even for features you don't.
SQL Server has the same structure. Enterprise Edition costs hundreds of thousands of yen per core per year. Every feature addition adds another license.
But PostgreSQL exists.
| Oracle / SQL Server | PostgreSQL | |
|---|---|---|
| License cost | Millions to tens of millions of yen/year | Free |
| High availability | Additional license (millions of yen) | Build with Patroni etc. (free) |
| Partitioning | Additional license | Built-in feature |
| JSON support | Additional option | Built-in feature |
| Full-text search | Additional option | Built-in feature |
| Performance | High | Comparable (depends on use case) |
| Migration barrier | — | SQL dialect differences, stored procedure rewriting |
The migration barrier is "SQL dialect differences" and "stored procedure rewriting." This is why companies have paid the Oracle tax for decades.
Then Claude enters the picture.
Hand Oracle SQL and stored procedures to Claude
→ Claude rewrites them as PostgreSQL-compatible SQL
→ Maps Oracle-specific functions to PostgreSQL equivalents
→ Converts PL/SQL → PL/pgSQL
→ Generates test cases
→ The migration barrier disappears
→ The Oracle tax disappears
This is not hypothetical. Claude can comprehend and rewrite codebases of hundreds of thousands of lines. SQL dialect conversion is structural code transformation — exactly the kind of work AI excels at.
Migration from SQL Server works the same way. T-SQL → PL/pgSQL. SSMS-dependent management scripts → standard SQL. Claude rewrites them, and the SQL Server tax disappears.
SSMS + Copilot — AI Plugged Directly into the Database
Is SSMS (SQL Server Management Studio) even necessary?
Microsoft has embedded Copilot into SSMS. When you open the query editor, Copilot automatically recognizes the connected database and can generate and execute SQL through a chat window.
What putting Copilot in SSMS means: A non-deterministic LLM gains direct access to production database schemas and data. A probabilistic AI writes SQL into a deterministic database. If Office + Copilot is a backdoor to "email, files, and chat," SSMS + Copilot is a backdoor to the database itself. Customer data, payment records, transaction histories — all directly accessible by a non-deterministic AI.
Compare this structurally to PostgreSQL + Claude Code. Claude Code runs in the developer's local environment. Generated SQL is reviewed by humans, and only code that passes tests is applied to production. Human judgment stands between AI and the production database.
SSMS + Copilot connects AI directly to the production database. Nothing in between. This is the exact same structural error as Office + Copilot.
And SSMS itself becomes unnecessary once you migrate to PostgreSQL. psql, pgAdmin, DBeaver — all open source, none designed to tightly couple AI into the interface. The SQL Server tax, the SSMS tax, and the Copilot risk all disappear simultaneously.
Microsoft Tax
Windows, Office 365, Azure, Teams — Microsoft levies multiple "taxes" on enterprises simultaneously.
Microsoft tax breakdown: Windows → OEM license per PC. No choice. Office 365 → ¥1,000–4,000/user/month × all employees × 12 months Azure → Lock-in via "Windows compatibility" Teams → Bundled with Office 365. Looks "free" but it's part of the bundle For a midsize company (500 employees): tens of millions of yen annually.
Enabled by Default — The Automatic Tax Collection Mechanism
Microsoft's tax has another structure. Features are enabled by default, taxing users before they even notice.
GitHub Copilot → Uses your code as AI training data. Enabled by default.
OneDrive → Starts syncing files without asking. Enabled by default.
Bing search → Windows Update changes your default browser.
Edge → Repeatedly sets itself as default via "recommendations."
Recall → AI records your screen at all times. Enabled by default.
Copilot → Integrated into Windows. Appears without consent.
This is a dark pattern: By making opt-out the default, the vast majority of users continue using features without realizing they're enabled. Users who do notice must hunt through settings to disable them. Updates sometimes reset those settings. "Enabling features by default" means "levying taxes by default." Users' data, attention, and freedom of choice — and CPU cycles — all taxed. Copilot, Recall, OneDrive sync, Windows Update, telemetry — background processes constantly consuming CPU and memory. The hardware performance users paid for is being consumed by Microsoft without permission. This is a CPU tax.
Copilot — Not a Tax, a Backdoor
In April 2026, the situation fundamentally changed. The Microsoft tax is no longer a cost problem. It is a security problem.
The hole Copilot opened: EchoLeak vulnerability (CVE-2025-32711, CVSS 9.3) — hidden prompts in emails can hijack Copilot. The user does not even need to open the email. The moment Copilot's RAG engine ingests the email as context, it collects confidential data using the legitimate user's permissions and exfiltrates it to an external server. Logs record it as normal activity. Copilot's convenience features function directly as a backdoor for system compromise.
What Claude Mythos Preview proved is that AI capable of exploiting this backdoor already exists. Exploits that once required state-sponsored hackers and months of work can now be generated overnight for under $50. The median patch window is 70 days. Mythos discovers and weaponizes vulnerabilities in hours. There is no longer time for "gradual migration."
Microsoft taxes Claude can eliminate — and this must begin immediately:
Document creation → Claude generates Markdown/HTML/PDF directly. Word becomes unnecessary.
Spreadsheets → Claude writes data analysis code. Excel becomes unnecessary.
Presentations → Claude generates HTML/Marp slides. PowerPoint becomes unnecessary.
Email → AI drafts and organizes. Outlook's necessity decreases.
Cloud → Migrate to AWS/GCP/self-hosted. Claude writes the configuration.
Copilot → Disable immediately. AI does not go in the product. Use it for development.
This is no longer something you can eliminate gradually. The moment Copilot was tightly coupled into the Microsoft monoculture, every user's data became part of the attack surface. The longer you wait to eliminate it, the more the risk expands.
Cloud Tax — Your Own Linux Server Is Enough
"You should migrate to the cloud" — one of the most successful marketing messages of the past decade.
Cloud tax structure: AWS / Azure / GCP → Monthly billing for servers, storage, networking — everything EC2 instances → Tens of thousands to hundreds of thousands of yen/month × number of instances RDS (managed DB) → Several times the cost of equivalent self-hosted hardware S3 / Blob Storage → Billed proportionally to data volume, forever Data transfer fees → Free to put data in, charged to take it out (lock-in structure) Midsize companies: millions of yen/year. Large enterprises: hundreds of millions.
But the combination of 2025 hardware and AI has changed the equation.
Linux server → A machine costing ~$1,000 is sufficient. PostgreSQL, Nginx, everything runs on it
Configuration → Claude writes config files. Faster than navigating the AWS console
Monitoring → Claude builds monitoring scripts. No CloudWatch needed
Backup → rsync + external storage. Cheaper than S3
SSL certificates → Let's Encrypt for free. Claude sets up auto-renewal
Static sites → Cloudflare Pages for free hosting
Monthly cloud bills of hundreds of thousands of yen become a one-time $1,000 investment + electricity
"The cloud has redundancy." "The cloud has scalability." — But does your company actually need that?
When you actually need the cloud: Tens of thousands of requests per second from around the world → Yes Need to instantly scale to 100 servers → Yes 99.999% availability required by SLA → Yes But most small and midsize company workloads do not fit the above. Internal systems, websites, databases — your own Linux server + Claude development handles all of it. Configuration is faster too.
Why This Matters — The Fundamental Error in AI Demand Forecasts
This is not just about cost savings. This is the very reason why demand forecasts for the entire AI industry are fundamentally wrong.
AI adoption spreads → Companies use AI → AI runs on the cloud
→ Massive GPU servers needed in the cloud
→ NVIDIA GPUs sell in enormous quantities
→ Data center investment explodes
This premise underpins the entire AI investment boom
Companies use AI → They develop with Claude → Their own Linux machines are sufficient
→ No need to run AI agents 24/7 on public servers
→ Cloud GPU servers are far less needed than projected
→ NVIDIA GPU demand falls far short of forecasts
→ Data center investment becomes excessive
The structure makes it obvious: The essential value of AI is "assisting development" — not "running inference on the cloud around the clock." Talking with Claude to write code, create config files, convert SQL — none of this requires massive GPU clusters. A developer's local machine and API access are sufficient. What the "users" of AI need is not GPUs — it is intelligence. Companies pouring massive capital into cloud infrastructure are misreading this structure. The collapse of NVIDIA analyzed in Chapter 7 is born from this structural error.
The Market Is Overfitting — Not Just AI, Humans Overfit Too
In machine learning, there is a concept called overfitting. The model over-adapts to patterns in the training data and fails to generalize to reality.
Right now, the AI market itself is overfitting.
Past pattern: IT demand grows → Servers sell → Data centers profit
AI-era pattern: AI demand grows → GPU servers sell → Data centers profit
→ The past pattern was applied directly
→ The market overfitted to "AI = massive infrastructure investment"
→ Reality: AI's value is in development assistance, not in infrastructure volume
AI overfits. Humans overfit too. Just as AI overfits to training data and loses generalization, humans overfit to past success patterns and misread new structures. Investors overfitted to "IT grew → servers sold" now assume "AI grows → GPU servers sell." Agriculture is the same. Overfitted to "chemical fertilizer increased yields," continuing to kill soil microbes. The essence of overfitting is failing to see the structure beneath the pattern. GitHub trying to use users' code as AI training data is the same structure. "More data + bigger models = better performance" — this premise itself is overfitting.
SaaS Tax — Accumulating Monthly Charges, Now with Built-in Backdoors
SaaS (Software as a Service) was sold as "you don't own it." But before you know it, monthly charges have piled up.
SaaS tax reality (midsize company example): Salesforce → Tens of thousands of yen/user/month × sales department Slack → Thousands of yen/user/month × all employees Zoom → Thousands of yen/license/month Notion → Thousands of yen/user/month Figma → Thousands of yen/user/month × design department Jira → Thousands of yen/user/month × development department 10–20 other SaaS tools → Hundreds of thousands to millions of yen/month Total: tens of millions of yen per year in "taxes" paid without even realizing it.
SaaS + AI — The Tax Now Comes with a Backdoor
In the reality of 2026, many of these SaaS tools have begun embedding AI agents.
The reality of AI-embedded SaaS: Salesforce Einstein AI → Scans all customer data to generate "suggestions" Notion AI → Accesses all documents in the workspace for summarization and generation Slack AI → Reads all channel messages for search and summarization HubSpot AI → Accesses customer email and transaction history for predictions Common structure: AI has full access to business data, processed by external LLMs.
This is the exact same structure as Office + Copilot and WordPress + AI plugins. A non-deterministic AI is being tightly coupled into deterministic systems (CRM, project management, documents).
SaaS vendor embeds AI → AI gains full access to business data
→ AI inference is processed by an external LLM provider
→ Data exits to both the SaaS vendor and the LLM provider
→ Prompt injection, data leakage, loss of audit trail
→ You completely lose control of your own data
You are not just paying monthly fees. You are letting AI you do not control read your business data. This is not a tax — it is structural data leakage.
Many features these SaaS tools provide can be replaced with custom tools built with Claude's help.
CRM → Claude builds a simple database + UI
Chat → Open source (Mattermost, etc.) + self-hosted server
Project management → Claude builds a tool tailored to your requirements
Documentation → Markdown + Git + static site generation
SaaS with embedded AI is the top priority to eliminate. These carry the highest data exfiltration risk.
SIer Tax — From "Built for You" to "Built by You"
In Japan, the bulk of corporate IT investment flows to SIers (System Integrators).
SIer tax structure: Requirements gathering → SIer interviews you → Millions of yen Basic design → SIer designs → Millions of yen Detailed design → SIer documents → Millions of yen Development → SIer implements → Tens of millions of yen Testing → SIer tests → Millions of yen Operations & maintenance → SIer bills monthly → Millions to tens of millions/year Total: tens of millions to hundreds of millions of yen for one system.
The bulk of SIer person-hours goes to "writing design documents" and "routine coding." Exactly what Claude does best.
Requirements → The person who knows the business tells Claude directly
Design → Claude generates design documents
Development → Claude writes the code
Testing → Claude generates test code
Operations → Claude builds monitoring and maintenance systems
Claude replaces the majority of the SIer's role.
The people who know the business best are your own employees. Not the SIer. When employees communicate requirements directly to Claude and Claude builds the system, the SIer as "translator" becomes unnecessary.
This is not merely cost reduction; the structure of software development itself is being reorganized. The irreversible transition that completes within a few years, and the shape of the "builder" (a judgment-centered profession in the same position as lawyers and doctors) that rises in the SIer's place, are developed across eleven chapters in the AI-Native Ways of Working — Software sub-series.
Consultant Tax — Think for Yourself
McKinsey, BCG, Accenture — companies pay tens of millions to hundreds of millions of yen for outside consultants to tell them "what to do."
Consultant tax structure: "Develop our DX strategy" → Tens of millions of yen "Propose our AI adoption plan" → Tens of millions of yen "Identify cost reduction measures" → Tens of millions of yen Deliverable → A few dozen beautifully designed PowerPoint slides Implementation not included. You pay tens of millions just for "thinking."
Claude is a structural thinking tool. The content consultants spend weeks producing in slides, Claude can structure in hours.
Industry analysis → Claude extracts structure from public data
Competitive analysis → Claude traces production routes and dependencies
Cost analysis → Claude decomposes spending structures
Strategy proposals → Claude visualizes options and causal relationships
The bulk of what you pay consultants for can be replaced by Claude.
A consultant's real value is "an outside perspective." But the era of paying tens of millions for "an outside perspective" is ending. Claude is the ultimate outside perspective. Unconstrained by industry conventions. No politics. No deference.
The Total of Six Taxes — From Cost to Vulnerability
| Tax | Annual cost (midsize, 500 people) | Urgency (post-Mythos) |
|---|---|---|
| Oracle / SQL Server Tax | Millions to tens of millions of yen | High (black-box legacy = attack surface) |
| Microsoft Tax | Tens of millions of yen | Top priority (Copilot is a backdoor. Act now) |
| Cloud Tax | Millions to tens of millions of yen | High (self-managed to shrink attack surface) |
| SaaS Tax | Tens of millions of yen | Medium–High (SaaS with AI agents especially dangerous) |
| SIer Tax | Tens to hundreds of millions of yen (per project) | High (not understanding your own structure = vulnerability) |
| Consultant Tax | Tens of millions of yen (per engagement) | Medium (outsourcing structural judgment = delayed response risk) |
| Total | ¥100M+ per year in "taxes" | All: begin immediately |
Claude's cost is orders of magnitude less than these taxes. A few tens of thousands of yen per month in Claude usage can eliminate tens to hundreds of millions of yen in annual taxes.
But this is no longer about cost reduction.
What Mythos changed: Before 2026 → "Eliminating taxes saves costs. Take your time, do it gradually." After April 2026 → "Taxes = dependency on monoculture = attack surface. Eliminate now or systems collapse." Median patch window → 70 days. Mythos vulnerability discovery and weaponization → hours. Expected public release → late 2026 to early 2027. We have six months to one year.
The bulk of enterprise IT spending is not technology investment — it is "taxes."
Oracle tax, Microsoft tax, cloud tax, SaaS tax, SIer tax, consultant tax —
paid because there's no alternative, paid out of inertia, paid because switching is frightening.
Claude is the tool that structurally eliminates these taxes.
But with Mythos here, this is no longer about cost optimization.
Dependency on monoculture is structural vulnerability.
Eliminating taxes means reducing attack surface.
Tight coupling to loose coupling. Black boxes to structures you understand.
Use AI for development. Do not embed it in the product.
Change the structure in the time remaining, or collapse is unavoidable.