跳到正文
DEV Community· ArkNet Digital·· 12 小時前AI 評分57

AI 產品如何在模型每年便宜 13× 時建立護城河

Your AI Strategy Is a Wrapper. Here's What a Moat Looks Like When Intelligence Gets 13 Cheaper Every Year.

AI 導讀

ArkNet Digital 主張,AI 產品的護城河應建立在隨使用累積的能力,而非可租用且快速降價的模型之上。文章將可防禦架構分為五層:專有評測、結果數據、工作流與執行權限、可信任機制,以及推理成本管理。引用 Epoch AI 的數據指出,自 2023 年起,達到固定 AI 表現的成本每季下降約 47%,相當於每年約 13×;作者亦提醒,基準測試不等同實際工作。

正文

Your AI Strategy Is a Wrapper. Here's What a Moat Looks Like When Intelligence Gets 13× Cheaper Every Year.

Most "AI strategies" I see have the same architecture: a prompt, a frontier model API, a nice UI, and a slide that says proprietary.

That isn't a strategy. It's a reseller agreement with extra steps.

This is not a knock on wrappers. Some great businesses started as one. It's a warning about which part of the stack is depreciating under your feet, and how fast.

The fastest-deflating input in economic history

Epoch AI measured what it costs to reach a fixed level of AI performance over time. Their finding: since 2023 that cost has fallen about 47% per quarter, roughly 13× per year. They call it faster than any other transformative technology they compared, including DNA sequencing, compute, lithium batteries and electricity (Epoch AI, "The plunging price of thought"). An earlier Epoch analysis found price declines ranging from 9× to 900× per year depending on the task (Epoch AI).

Epoch is careful about the caveats: benchmarks aren't real work, and real users don't switch to the cheapest model every quarter. The direction is still not in doubt.

Now run the first-principles consequence. If the core capability in your product gets an order of magnitude cheaper every year, then:

  • Your cost advantage over a competitor is gone in a quarter.
  • Your capability advantage is gone at the next model release, because they rent the same model.
  • Your "prompt engineering IP" is a few kilobytes of text that a capable engineer can approximate in an afternoon.

Here's the test I use. Call it the Depreciation Test:

If a part of your product gets better automatically when the foundation model gets better or cheaper, that part is not your moat. It's a tailwind you share with every competitor.

Run every component of your AI product through that test. What survives is your actual strategy.

Why most deployments stall, according to the data

MIT's NANDA initiative published a widely cited report in 2025, The GenAI Divide. Its headline claimed that 95% of organizations were getting zero return from generative AI (MIT NANDA, State of AI in Business 2025). The headline number has been debated, and you should read the methodology before you quote it.

The diagnosis in the report is the part worth keeping:

"The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time."

That sentence says the same thing as the Depreciation Test, from the other side. A wrapper is a system that doesn't learn. It's exactly as good on day 400 as on day 1, except that by day 400 the model underneath has become a commodity and everyone has it.

Defensibility in AI is the slope of your improvement curve, not its starting point.

The five layers that compound

A defensible AI architecture has layers that get stronger with use, not with model releases. I count five. Each is something a competitor can't rent from the same API.

Layer 1: Proprietary evals (the layer almost everyone skips)

Your evaluation suite, meaning the tasks, the edge cases, the grading rubrics and the failure taxonomy specific to your domain, is the most underrated asset in AI.

Why it compounds:

  • It lets you swap models in days instead of quarters. When a cheaper model ships, you know within hours whether it's good enough for your workload. A competitor without evals is either stuck on the expensive model or shipping regressions blind.
  • It turns the 13× per year deflation from a threat into a margin engine. You capture the price drop, and your competitors pass it on to customers because they can't tell what's safe to switch.
  • It encodes domain judgment that took years of customer contact to learn.

If you have no evals, you don't have an AI product. You have a demo with uptime.

Layer 2: Outcome data, not text data

Everyone has text. What almost nobody has is labeled outcomes: did the suggestion get accepted, did the claim get paid, did the ticket stay closed, did the code pass review, did the customer churn?

Outcome data is the only training and retrieval signal that's both proprietary and causally tied to value. Design your product so that every interaction closes a loop: the agent acts, the world responds, and you capture the response attached to the action. That's the "retain feedback" capability the MIT report found missing.

Architecturally, it means treating your event log as a first-class product surface: action, context, outcome, linked and queryable. Most teams log prompts and completions. Almost none log what happened next.

Layer 3: Workflow ownership and the right to act

Models generate. Businesses act. The durable position is owning the system of record and the permission to change it: the ERP write path, the claims adjudication step, the deployment pipeline, the procurement approval.

A model that can draft a purchase order is a commodity. A system that's integrated into approval chains, holds delegated authority and is trusted to execute is very hard to replace. Rip it out and you have to rebuild the integrations, the permissions and the organizational trust from scratch.

Distribution plus authority beats intelligence. Intelligence is the part you rent.

Layer 4: The trust surface

As agents take real actions, buyers will pay for provable control: audit trails, policy enforcement, human-in-the-loop at the right points, and evidence that holds up with regulators.

That isn't a cost center anymore. In the EU, the AI Act's transparency duties under Article 50 have applied since 2 August 2026, and the high-risk obligations now arrive on 2 December 2027 for Annex III systems and 2 August 2028 for AI embedded in regulated products, after the Digital Omnibus amendment (Regulation (EU) 2026/1744). In regulated industries, the vendor whose product already produces the evidence wins the procurement fight before it starts.

Compliance done as architecture, not as a PDF, is a moat. Few founders treat it as one, which is exactly why it works.

Layer 5: Routing economics

The final layer is your inference cost structure: model routing by difficulty, caching, distillation onto smaller models for your common cases, and batch-versus-realtime decisions per feature.

On its own this layer is copyable. Combined with Layer 1 it compounds: evals tell you which traffic a small model can handle safely, and outcome data (Layer 2) gives you what you need to distill a specialist model for your domain. A team running 80% of its traffic on a cheap specialist model with measured quality has structurally better margins than a team sending everything to a frontier API, and that gap widens every time prices fall.

The anti-patterns, stated bluntly

  • "Our prompts are our IP." Prompts leak, get reverse-engineered, and get made obsolete by the next model. They're configuration, not a moat.
  • "We fine-tuned a model." On what? If the answer is public data, a competitor can do the same next month for less money. Fine-tuning is only defensible when the data behind it is.
  • "We're model-agnostic." Good, but that's hygiene, not a strategy. Being model-agnostic without evals just means you can switch models without knowing whether you should.
  • "We'll add the feedback loop later." The loop is the product. Data you didn't capture in year one is data you'll never have.
  • "Our UX is the moat." UX is necessary and copyable. It buys you months, not years.

A one-page architecture review

If I were reviewing your AI product, these are the questions I'd ask, in this order:

  1. Depreciation Test: List the components. Which ones get better when the model improves? Those are tailwinds. What's left?
  2. Evals: How long would it take to safely move to a model that's 10× cheaper? Hours, weeks or "we don't know"?
  3. Loop closure: For each AI action, do you capture the real-world outcome and link it back to the action?
  4. Authority: Does your product execute inside a customer's system of record, or does it hand them text to paste?
  5. Trust: Could you hand a regulator or a Fortune 500 security team the evidence for why your system took a given action?
  6. Unit economics: What share of your traffic genuinely needs a frontier model?

If the answers are mostly "no" and "we don't know," that's fine for a prototype. It isn't a company yet.

The thesis

In a world where intelligence gets an order of magnitude cheaper every year, the model is the one thing you can be sure everyone will have.

So stop building on the part that's deflating. Build on the parts that appreciate with use: your evals, your outcome data, your authority to act and your provable trustworthiness. That's the difference between a wrapper and a company. A wrapper rents intelligence. A company compounds it.


Michael, ArkNet Digital: AI product architecture, agent systems and AI governance engineering. arknet.digital

If you're deciding what to build versus rent in your AI stack, or need an eval and feedback-loop plan, I'm happy to talk it through. Think the Depreciation Test is wrong? Argue with it below.

來源:DEV Community · dev.to