Built on Borrowed Power

The panic is the signal

Every hot market invents a vocabulary for its own anxiety. In AI, one of the dumbest recurring phrases is also one of the most revealing. A new feature appears, a rumor starts moving, and suddenly a product that looked unstoppable on Monday is being described as obsolete by Wednesday. The literal claim is usually nonsense. The emotional content is not.

That is what made the reaction around Lovable, n8n, and Figma worth taking seriously. Not because any of those companies were actually erased, but because the market response exposed what people now believe about power in AI.

Lovable became shorthand for a particular kind of startup vulnerability after reports suggested Anthropic was testing a Lovable-like app-building feature inside Claude. That mattered because Lovable is not some obscure experiment. It is one of the best-known names in natural-language app creation, the kind of company that helps define a category rather than merely participate in it. Once chatter began that Claude itself might absorb that kind of capability, the usual market ritual kicked in. People did not ask whether the rumored product was finished, whether the feature would match Lovable’s product depth, or whether customers would immediately switch. They jumped straight to the conclusion that the layer above the model had become unsafe.

n8n triggered a related but slightly different reaction. Anthropic’s routines push made the threat feel less like rumor and more like directional evidence. Claude was no longer just answering questions or writing code. It was being extended into scheduled execution, API-triggered automation, and GitHub-connected actions that run on Anthropic-managed infrastructure. That was enough for many people to start talking as though dedicated workflow tools had just received an expiration notice. Again, the literal claim was weak. But the fear underneath it was real. The moment a frontier lab starts moving into orchestration, a lot of the market assumes the independent workflow layer is standing on rented ground.

Then came Figma, and the tone changed.

This was no longer just startup chatter or social-media melodrama about something being “obsolete.” When Anthropic launched Claude Design, the reaction was immediate and materially different. Coverage framed the move as a challenge to established design platforms, and Figma’s stock dropped sharply on the news. That matters because the public market is not reacting to vibe or metaphor. It is reacting to a shift in expected future pressure. Investors were not saying Claude Design had already replaced Figma. They were saying something more interesting and more serious: a frontier lab had just signaled that another commercial layer of software might now be inside its field of expansion.

That reaction tells you more about the structure of the market than about any single launch.

A few years ago, success in software generally strengthened the case that you had found territory worth owning. In AI, success increasingly does something else. It proves demand for a workflow that the platform underneath you may decide to absorb.

That is the real story hiding beneath the melodrama. The market is no longer just asking whether a startup has product-market fit. It is asking whether the company has any durable right to remain in the value chain once the model provider, interface owner, and distribution layer all decide the category looks attractive.

This is why the language of obsolescence appears so quickly, even when it is plainly premature. People are not really describing product failure. They are describing platform fear.

The lab is no longer just the lab

The old mental model treated frontier AI companies as suppliers. They built the model, improved the capability, offered an API, and let the application layer convert raw intelligence into actual products.

That model is breaking down in full view.

The leading labs are not behaving like neutral providers of underlying intelligence. They are behaving like expanding platforms. They are building the model, the interface, the enterprise relationship, the billing layer, the partner ecosystem, the workflow surface, and increasingly the surrounding tools that used to belong to the application layer. Look at the sequence closely and the shape becomes hard to ignore.

Anthropic announced a major partner push backed by a large investment in the Claude Partner Network. It launched a marketplace through which enterprise customers can apply existing Anthropic commitments to partner products, including products like Lovable. It introduced routines that let Claude Code execute on schedules, API triggers, and GitHub events from Anthropic-managed cloud infrastructure. Then it launched Claude Design, extending the same environment into prototypes, presentations, one-pagers, and visual work. Around the same period, reporting suggested Anthropic was also testing a Lovable-like app-building feature inside Claude.

This is not a company staying politely in the model layer. This is a company widening its strategic perimeter.

It is also a company teaching the market a new lesson. The more the lab owns the interface, the workflow logic, the enterprise contract, and the procurement channel, the less stable the old supplier-ecosystem narrative becomes. The lab is not just the engine beneath the market anymore. It is trying to become the operating environment through which the market is accessed.

That is why Lovable, n8n, and Figma belong in the same article even though they do not belong to the same category and did not experience the same kind of reaction.

Lovable represents the startup version of the problem. A company proves that natural-language software creation is not a toy, grows explosively, and then finds itself facing the possibility that the model provider under the hood may want to own more of that layer directly.

n8n represents the workflow version. A company that built real value around orchestration, visibility, and automation suddenly has to operate in a world where the lab is moving into triggered execution and cloud-run operational flows.

Figma represents the public-market version. A category leader sees investors immediately mark down its stock because a frontier lab has entered adjacent visual territory strongly enough to make future compression look plausible.

Those are not identical stories. They are different manifestations of the same structural shift.

Marketplace today, rival tomorrow

One of the oddest features of the current AI market is that partnership and competitive exposure now arrive together.

A company can appear inside a platform’s marketplace and in the next breath become an obvious candidate for native substitution. That sounds contradictory only if you are still thinking in the older language of supplier relationships. It makes much more sense if you think in platform terms.

A marketplace is not just a sales surface. It is also an intelligence surface. It tells the platform owner which categories are attracting demand, which workflows matter, which products resonate with enterprise buyers, and where value is gathering above the model. In older technology cycles, this was already familiar. The platform let third parties explore the frontier, then decided which layers were too strategically important to leave outside the core product. AI is now reproducing that pattern under conditions of much greater dependency and much faster release cycles.

That matters because the application layer is not building on a stable substrate with slow boundaries and predictable interfaces. It is building on top of rapidly improving models whose owners are shipping new workflow abstractions, new commercial surfaces, and new native capabilities at a pace that shortens the life span of weak moats.

This is why the psychology of the market feels so unstable even when revenues look extraordinary. Product momentum no longer guarantees strategic safety. It can simply make you legible to the platform beneath you.

In a normal market, growth is validation. In this market, growth is also reconnaissance.

Why the app layer still matters

This is the point where shallow analysis usually overcorrects. Someone sees the lab moving up the stack and concludes that everything above the model is now disposable. That conclusion is just as lazy as the webinar copy.

The app layer still matters because turning raw model capability into usable, trusted, governed, and embedded workflow is real work. Product design is real work. Integration is real work. Governance is real work. Compliance is real work. Operational fit is real work.

That is why the existence of Claude Design does not make Figma irrelevant, just as rumors about a native app builder do not make Lovable irrelevant, and routines do not suddenly dissolve the value of workflow platforms like n8n. But those companies do not matter for the reasons the market liked to celebrate a year ago.

The old AI story rewarded surface magic so aggressively that many people confused delight with durability. If a product felt astonishing, spread quickly, and made something hard suddenly easy, that seemed enough. The closer the labs move to the user, the weaker that story becomes. What survives is not just the ability to generate output. What survives is the ability to own execution under real constraints.

Lovable is not just interesting because it can turn prompts into applications. It is interesting because it has built a productized environment for iterating, deploying, and governing app creation in a way that can become part of how people actually work.

n8n is not interesting because natural language can help assemble workflows. It is interesting because serious organizations increasingly care about visibility, traceability, control, infrastructure choice, and the ability to run automation inside environments they can actually govern.

Figma is not defensible because design is fashionable or because investors once loved collaborative interfaces. It is defensible to the extent that it remains deeply embedded in design systems, team workflows, approvals, handoffs, and the organizational fabric of how product work gets done.

That distinction matters. A native feature can attack convenience. It cannot automatically replace institutional embedment.

The new moat is not magic

The moat conversation in AI is changing because the stack is changing.

A polished prompt layer is not enough. A clever wrapper is not enough. A viral workflow is not enough. Even fast growth is not enough if that growth mainly proves demand for a category the platform can absorb.

The stronger defenses are harder and less glamorous. They live in workflow ownership, domain-specific execution, enterprise embedment, independent distribution, infrastructure flexibility, governance architecture, switching friction, and the accumulation of context that does not transfer cleanly when a user tries the native alternative.

That is why the next phase of AI competition will be less about who can make the model feel magical and more about who owns the environment in which important work actually happens. The more AI moves from experimentation into operations, the more buyers care about where the system runs, what it can touch, how it can be audited, who controls approvals, how failures are contained, and whether the product is merely convenient or actually governable. In other words, the moat is shifting from generation to control.

This is the real significance of the recent panic cycles. The market is beginning to understand that the threat is not just another startup building faster. The threat is that the platform beneath the market has decided to move upward into whichever categories now look commercially obvious.

Why buyers and investors should care

It would be easy to treat this as a founder problem. It is not. This is also an enterprise procurement problem and an investor judgment problem.

Buyers selecting AI vendors now need to evaluate more than present functionality. They need to evaluate upstream dependency. If a vendor sits on top of a lab that may later bundle similar features, alter economics, redirect distribution, or tighten the relationship between model and interface, then the buyer is not just buying a product. The buyer is inheriting a strategic exposure.

That changes the due diligence questions. How portable is the workflow? What remains valuable if the platform ships a native imitation? How dependent is the vendor on a single model provider? Where do governance and control actually live? What can be self-hosted, audited, or moved? Does the vendor own real execution, or mostly presentation?

Investors should ask the same questions from the opposite angle. Revenue growth in AI can be a sign of durable advantage. It can also be a sign that a company has identified a category the platform underneath it may decide to internalize. That does not make the app layer uninvestable. It makes lazy underwriting dangerous.

The most fragile companies in the next phase will be the ones that confuse proximity to model capability with ownership of customer value. The more durable ones will be the companies that own trust, workflow, embedment, and execution strongly enough that a native feature is annoying rather than fatal.

The closing stack

The most important thing happening here is not that every application company is doomed. It is that the stack is beginning to close.

The open narrative around AI suggested a broad ecosystem in which frontier labs would provide intelligence and a vast market of independent applications would flourish above them. Some of that remains true. There will be many winners above the model. But the power structure is clarifying. The companies closest to the model are moving outward, not inward. They are using capability advances, interface ownership, enterprise contracts, marketplaces, partner ecosystems, routines, and adjacent product launches to widen their strategic perimeter. As that perimeter expands, the application layer will stratify.

Some companies will remain vulnerable because they are too close to generic capability and too far from owned workflow. Some will survive because they become the trusted layer through which real work is governed and executed. Some will be forced into narrower specialization. Some will be acquired because their strongest future is inside a bigger platform. Some will build enough independent distribution to negotiate from strength. And some will discover too late that what looked like a tailwind from the foundation model was really a temporary lease on borrowed power. That is the real meaning of the panic around Lovable and n8n, and the sharper market reaction to Figma.

The tools are not dead. But the innocence of the application layer may be.