AI Is Starting to Dismantle Its Own Moats

2026-11-11·6 min read·OKCodex Editorial

AI Is Starting to Dismantle Its Own Moats

Four seemingly unrelated signals this week point in the same direction: the barriers that used to be defended by scarcity — training data, chip software stacks, verification capacity — are being weakened by AI itself.

1. Data: RL environments became a tradable commodity

According to a ChinaTalk investigation, US data brokers including Mercor, Surge AI and AfterQuery have sold roughly half a billion dollars' worth of AI training data to Chinese AI labs — including the same RL (reinforcement learning) environments previously built for OpenAI or Anthropic. At least nine firms are involved.

The amount matters less than the shift in asset form. Training-data deals used to mean question-and-answer corpora; now they mean complete environments where a model can repeatedly try, fail and receive reward signals. These environments are usually custom-built for a specific customer and embed task design and scoring logic — reused elsewhere, they amount to selling someone else's training methodology.

For Chinese teams, the takeaway: procurement channels exist, but compliance risk and vendor dependence are equally real. The question shifts from "can we buy it" to "does the scoring logic we bought still leave us any differentiation".

2. Compute: a notch carved out of the CUDA moat

Anthropic claims its in-house model Mythos ported Claude to AMD GPUs in just three days, with almost no human intervention — work that typically takes months. Anthropic says Claude now runs on four chip families, aiming to cut costs and reduce vendor dependencies.

This is a vendor's own account with no third-party reproduction, so discount it. But the direction is clear: the main cost of chip migration has long sat not in hardware but in software-stack adaptation and tuning — exactly the kind of work current models do best, with clear goals, immediate feedback and automatically verifiable results.

3. Verification: from "years" to "four weeks"

Google DeepMind's AlphaProof Code mathematically verified Signal's cryptographic core in four weeks, a process that typically takes years. Agents powered by Gemini Carbon wrote more than six million lines of proof code and flagged only about one percent for human review.

The lesson for Chinese teams may be the most direct one: formal verification used to be a quality investment only large companies could afford. Once the cost falls, security and compliance turn from pure cost items into negotiable trust assets.

4. The business side: spending still rising, but slower

The Ramp AI Index shows US corporate AI spending rose just 1.65 percent in October 2026, still a new high. The more important number is structural: one percent of companies drive about 80 percent of revenue for OpenAI and Anthropic. Customer support is emerging as a new growth driver, but it may bring lower margins than the coding business that has driven growth so far.

By contrast, TechCrunch reports OpenAI's quarterly revenue crossed $10 billion for the first time, an almost 50 percent jump from the $6.7 billion recorded in the second quarter. The leaders are still expanding fast, but "1% of customers behind 80% of revenue" shows that paying conversion among smaller customers remains unsolved.

Conclusion: three actionable steps

1. Treat RL environments as assets to inventory, not items to procure. Decide which scoring logic must be built in-house and which can be sourced externally, and audit the provenance and licensing of anything brought in.

2. Hand the dirty work of migration and verification to agents first. Chip adaptation, formal verification and dependency upgrades have clear correctness criteria — the highest return per unit of effort.

3. Don't bet your revenue structure on a handful of large accounts. When slower spending growth coincides with highly concentrated revenue, calculate your top-1% customer share before talking about expansion.

Moats haven't disappeared. They have moved from "owning resources" to "being able to keep rebuilding yourself with AI".

Sources

Share:Weibo

Subscribe to weekly briefing

Comments

Loading comments…

Leave a comment

0 / 500

Comments appear after review.

Find tools for your next task

Start with scanned documents, code review or research. Get a shortlist with matching evidence; no model is run.

Find tools for my task →
Analyze this topic with two models

DeepSeek and Qwen are preselected with the article topic. You choose when to run them.

Try it →

Keep reading

Mentioned

Agents mentioned in this article