Opinion: In the AI gold rush, the “selling shovels” logic has already become obsolete.
TechFlow Selected TechFlow Selected
Opinion: In the AI gold rush, the “selling shovels” logic has already become obsolete.
The AI companies that will truly survive are not those selling tools, but “jewelers” who treat AI as a raw material and apply it in vertical domains.
Navigating Web3 tides with focused insightsAuthor: Ben Basche
Translation & Editing: TechFlow
TechFlow Insight: “Sell shovels during a gold rush” was once a golden rule in startup circles. But in the AI era, this logic has broken down—because the miners have opened their own hardware stores. OpenAI, Anthropic, and Google are systematically absorbing startup categories like middleware layers, coding assistants, and browser automation. Author Ben Basche argues that the AI companies truly built to last aren’t tool vendors—they’re “jewelers”: domain-specific practitioners who embed AI as raw material, dive deep into vertical industries, master local knowledge, and possess irreplicable context.
Full Text Below:
A saying became gospel among founders around the time of the first internet bubble: “Sell shovels and pickaxes during a gold rush.” The idea is that real money isn’t made by the miners—it’s made by those supplying them. Levi Strauss got rich—not the prospectors.
It’s a good framework. And for a while, it worked.
But in AI, it’s wrong. If your company is built on this logic, you need to take a hard look at what’s happened over the past twelve months.
The Lab Is the Entire Stack
Here’s what actually happened—first quietly, then all at once.
OpenAI launched Operator, a computer agent capable of browsing the web, filling out forms, and executing end-to-end tasks. Then came the Responses API and Agents SDK, giving developers native tool-calling, memory, and orchestration capabilities—no third-party frameworks required. Next up: Codex, a cloud-based programming agent that autonomously writes, tests, and iterates software—and Deep Research. Any one of these products, just two years ago, would’ve been enough to found a well-funded startup.
Anthropic released Claude Code, Computer Use, Projects (with persistent memory), and the Model Context Protocol (MCP)—almost overnight becoming the de facto standard for connecting AI with external tools and data. Then Anthropic donated MCP to the Linux Foundation, ensuring it becomes infrastructure—not a product. Following that came Claude in Excel, Claude in Chrome, and Cowork.
Google launched Gemini 2.0, natively embedding tool-calling and multimodal perception capabilities—and integrating them into Vertex AI as an enterprise-grade agent control plane, offering organization-level policies and orchestration out of the box.
Each of these moves is eating territory once claimed by startups.
The “sell shovels” logic rests on a hidden assumption: labs will stay in their lane—building foundational models and APIs, and leaving the tooling layer, orchestration layer, and application layer to the ecosystem. That assumption is dead.
The Middleware Massacre
Let’s examine what’s happening specifically in the middleware layer.
LangChain was the quintessential “sell shovels” bet of the 2023 AI boom—a framework for orchestrating LLM calls, connecting tools, and managing memory. Thousands of teams built products on it; its GitHub stars surpassed 100,000. By 2024, teams began publishing blog posts explaining why they were ripping it out of production—not because it’s bad, but because the underlying models have become smart enough to render it unnecessary. LangChain’s abstraction layer solves yesterday’s problems.
Meanwhile, OpenAI launched its own Agents SDK. Microsoft released AutoGen and Semantic Kernel. Labs—and their parent companies—didn’t acquire LangChain. They simply built LangChain’s functionality natively into their platforms.
The same script plays out across every layer: agent frameworks, prompt management tools, RAG pipelines, evaluation frameworks, observability tools. All are being absorbed into native offerings from the vendors running the underlying models.
The cruelty lies here: when OpenAI or Anthropic bake orchestration directly into their APIs, they don’t need to win on features. They only need to be “good enough”—and already there. Developers default to the path of least resistance. That clever middleware startup must not only leap far ahead—but sustain that lead as models evolve rapidly, all while competing against adversaries with infinite capital and control over the foundational infrastructure. That’s not a business—it’s a research project with a countdown timer.
The Miners Opened Their Own Hardware Store—So You Can’t Sell Shovels Anymore
The “sell shovels” analogy fails in AI due to one critical structural difference. In 1849, Levi Strauss and other hardware merchants didn’t mine for gold themselves. Miners and suppliers were separate, interest-divorced actors.
In AI, labs both mine and sell shovels—and build roads, and print maps. They have every incentive to own the entire stack: each additional layer controlled means another lock-in point, another profit expansion opportunity, another distribution moat.
Anthropic donating MCP to the Linux Foundation wasn’t charity. It was ensuring a standard they designed becomes universal infrastructure—just as Ethernet did. Standards are tech’s most powerful moats, because they’re invisible and permanent.
So if your startup’s value proposition is “We sit between developers and models, making X easier,” you face a hard truth: the entity you’re sitting between has noticed you, has the resources to replicate you—and has structural reasons to do so.
So What Does Work?
Back to the gold rush metaphor. If you can’t sell shovels anymore—what should you sell?
Sell jewelry.
Or better yet: treat gold as industrial feedstock—making things the miners themselves have no interest in building.
In the real 1849 gold rush, the businesses that survived the boom weren’t general-tool vendors. They were artisans who treated gold as raw material, applying deep expertise to craft specialized products: jewelers, dentists, later electrical engineers. Their grasp of specific use cases ran so deep that generalists simply couldn’t match it.
The AI version is building applications in vertical domains—ones requiring real-world context that labs lack and struggle to acquire.
Consider what OpenAI, Anthropic, and Google are structurally ill-equipped to do:
- They don’t deeply understand your industry’s workflows.
- They have no relationship with your customers.
- They cannot cheaply access the private data that makes models truly effective in specific scenarios.
- They’ll never dive deep into why individual artisans in South Africa invoice the way they do—or why mobile payment integration in Kenya is nontrivial—or why U.S. healthcare prior authorization is a uniquely thorny, deeply embedded operational problem.
Labs build horizontal infrastructure. Opportunity lives in verticals—domains where geographic, regulatory, cultural, and industry-specific local knowledge is essential to make anything work at all.
That’s why fintech in emerging markets, legal AI tailored to specific jurisdictions, compliance tools for regulated industries, and workflow automation for niche professional domains are far more defensible than “building a better LangChain.”
The moat isn’t in the model. The moat is in the context.
The Industrial Use of Gold
This idea has a second, equally important variant: using AI like gold in industry—not as a store of value or showpiece, but as a component embedded inside systems that generate durable economic value.
Gold’s conductivity is nearly unmatched. So it’s in every circuit board. Nobody talks about it. Nobody hypes it in that context. It works silently—as a critical input within a larger system.
The most durable AI companies being built today treat models as components—an input enabling a product that solves a real problem—not as the product itself. AI is the gold in the circuit board—not the gold in the display case.
Operationally, it looks like this: you pick a domain with real pain points, real workflow complexity, and real difficulty accessing data—and build a product where models happen to be used underneath to make it vastly better. AI is an implementation detail. The product is what replaces the painful manual process.
This is the exact opposite of “we put a wrapper on GPT-4.” A wrapper is the display case. A circuit board is invisible.
Recent Startup Categories That Got Killed
To be explicit: below are startup categories labs have been systematically absorbing since late 2024:
Agent orchestration frameworks. Now native features of OpenAI’s Agents SDK, Anthropic’s toolchain, and Google’s Vertex Agent Builder.
AI coding assistants. OpenAI’s Codex now performs full-repository autonomous coding. So does Claude Code. GitHub Copilot is Microsoft’s native solution. The standalone coding-assistant category has been sharply compressed.
Browser and computer automation. OpenAI’s Operator, Anthropic’s Computer Use, and Google’s Gemini Astra—all three frontier labs now offer products in this space. Every LLM-powered RPA startup is playing defense.
RAG pipelines and vector search tools. Largely commoditized. Most model APIs now include native retrieval capabilities. Differentiation at the framework level has vanished.
General-purpose AI assistants and productivity tools. Directly absorbed by Claude, ChatGPT, and Gemini.
Prompt management and evaluation tools. Increasingly baked into native offerings. LangSmith retains some breathing room—but it’s racing against time.
The pattern is remarkably consistent: labs spot a category gaining meaningful developer attention, judge it adjacent to their core product—and ship a version. Not necessarily better—but integrated, cheaper by default, and distributed at scale no startup can match.
What Should You Do Now?
If you’re building an AI startup today, don’t ask: “Is there demand?” Demand is everywhere. Ask instead: “Could this be killed by a single product launch from a lab with $10B+ in the bank?”
If the answer is “yes”—or even “maybe”—then it’s not a business. It’s a feature.
Durable strategies share these traits: deep vertical specificity (labs can build generic solutions—but not yours), proprietary data or relationships impossible to replicate by scraping the public web, regulatory and compliance complexity that makes “just calling the API” insufficient, and distribution channels in communities where trust and local context matter more than raw capability.
The gold rush is real. Gold is everywhere. But the miners now run stores—and they do so with infinite capital.
Sell jewelry. Treat gold as industrial feedstock. Build things the miners themselves have no interest in—because it’s too niche, too localized, too deeply embedded in domain knowledge they’ll never possess.
That’s the play I believe in.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@basche42
