Applications Can’t Compete with Models: The Winners Will Be the Shovel Sellers for Frontier Labs

In mid-May 2026, Brendan Foody, CEO of Mercor, posted a blunt prediction on X: “The next 12 months will be dramatically better for infrastructure companies upstream of Anthropic and OpenAI than for application-layer companies downstream of them.” It was a short tweet, but it crystallized a growing consensus in the AI investment community.

The frontier labs aren’t just building better models—they’re quietly eating the lunch of the very startups that once hoped to ride their coattails. And the real money, Foody argues, will flow to the companies selling the “shovels”: compute, data, and everything else the labs need to keep scaling.

Here’s why the logic is compelling—and where it might still be wrong.

---

1. Frontier Labs Are Becoming Direct Competitors to Apps

The pattern is now unmistakable. Anthropic and OpenAI are no longer content to be raw-model providers. They are shipping polished, domain-specific products that solve entire workflows natively.

• In April 2026 Anthropic launched Claude Design — a text-to-production-ready-prototype tool powered by Claude Opus 4.7. Within 24 hours Figma’s stock (already down ~50% over the prior year) dropped another 5–7%. Adobe and Wix took hits too. Designers didn’t need another plugin; they got a better designer.
• In May Anthropic released ten ready-to-run AI agents for financial services — pitchbook builders, KYC screeners, month-end closers — complete with domain knowledge, connectors, and sub-agents. Banks no longer need a third-party “AI finance wrapper”; they can spin up Anthropic’s templates in minutes.
• OpenAI, for its part, rolled out ChatGPT Health in January, a dedicated health workspace that securely ingests medical records and wellness data. It’s not a vague “health AI startup” promise—it’s the model itself becoming the product.

Every time a frontier lab ships a vertical solution, the application-layer moat shrinks. The labs already own the intelligence layer; they are simply extending it one domain at a time.

Meanwhile, on the supply side, demand for raw ingredients is exploding. Anthropic just signed a landmark deal to rent the entire output of xAI’s Colossus 1 data center — 300 MW — for $1.25 billion per month, potentially $40 billion+ through 2029.

That’s not a one-off; it’s a signal that the labs will pay almost any price for reliable, high-quality compute and expert-curated data. Infrastructure companies that can deliver either (or both) are sitting on the right side of the trade.

---

2. But Is the Moat Really That Fragile?

Not so fast. Several counter-forces could still favor well-positioned application-layer companies—especially in enterprise.

Capital vs. Context. Infrastructure moats are mostly capital-intensive (data centers) or data-intensive (expert datasets). Capital is straightforward. Data is trickier. As models approach AGI territory, the labs themselves say the next leap won’t come from bigger pre-training corpora but from rich, real-time context and workflows.

That context lives in the application layer: proprietary processes, organizational memory, compliance guardrails, and deeply embedded user habits.

Today it feels like “anyone can build context” because coding agents and vibe-based tools make wrappers trivial. But as AI starts disrupting truly complex enterprise processes — multi-month regulatory filings, cross-functional supply-chain orchestration, or highly specialized R&D pipelines — generating high-fidelity context on the fly becomes exponentially harder. The companies that already own those workflows have a structural advantage the labs can’t simply prompt their way around.

Inference Economics Are Moving in Favor of Apps. Menlo Ventures’ recent analysis showed inference costs have fallen dramatically — by orders of magnitude since GPT-4 — while enterprise usage volume has exploded.

Jevons’ paradox in action: cheaper inference doesn’t shrink margins; it expands total spend because usage goes parabolic. That dynamic actually widens the addressable market for applications rather than commoditizing them.

Local Models Change the Game. If the industry shifts toward on-device or private-cloud inference (and the trajectory of open-source and enterprise privacy demands suggests it will), context windows and workflow integration become even more valuable than training data. An application that owns the user’s day-to-day data flywheel suddenly looks like the defensible layer.

Of course, this only applies to real applications with proprietary context. If you’re building yet another thin wrapper around Claude or GPT, the labs will eat you alive—and they’re already doing it.

---

3. Enterprise vs. Consumer: Different Battlefields

The discussion above is mostly about enterprise software, where workflows are complex, data is sensitive, and switching costs are high. Consumer applications face an even starker reality. Distribution and brand still matter, but the speed at which frontier models can replicate (or surpass) consumer features is terrifying. The moat there is distribution and habit, not technology.

---

4. One Important Disclaimer

Mercor, Foody’s company, sells evaluation and benchmarking infrastructure to the very frontier labs. His view is professionally optimistic about the upstream layer. That doesn’t make the thesis wrong—it just means it’s worth pressure-testing.

Also read:

• Accenture Song Acquires Whalar in What Is Being Called the Largest Deal in Creator Economy History
• How to Go Viral on YouTube in 2026: Help an 83-Year-Old Man Sell His $200,000 LEGO Collection… and End Up as the Target of a Criminal Investigation
• Image Generator Updates 2026: No Revolution, Just a Gentle Correction for the Flagship-Chasers
• YouTube’s Total Dominance: For the First Time Ever, It Has Overtaken Netflix in Daily Viewing Time

---

The Bottom Line

Foody is almost certainly right for the next 12–18 months. The capital and data flywheels of the frontier labs are spinning faster than any application-layer startup can match. The shovel sellers—compute providers, data platforms, observability and eval companies, specialized inference hardware—will print money while many application-layer bets get repriced.

But the longer-term picture is more nuanced. The ultimate moat won’t be raw intelligence; it will be the irreplaceable combination of intelligence + proprietary context + distribution. Frontier labs can keep climbing the stack, but they can’t own every company’s internal knowledge graph or every user’s daily workflow.

The winners won’t be the pure model companies or the pure app companies. They’ll be the ones that figure out how to own the layer between the model and the real world—whether that layer is infrastructure, context engines, or something we haven’t named yet.

In the gold rush, sell shovels. But keep an eye on the miners who actually own the claim.