I started thinking about this as a future problem. That was already too late. By late October, China’s AI-heavy STAR 50 had fallen roughly 40 percent from its July high in less than four months, wiping out about $480 billion in market value. The broader CSI 300 had fallen about 20 percent from its June high, while some of the most celebrated names in robotics, semiconductors, and optical infrastructure had been cut in half. Even after the selloff, the STAR 50 was still trading around one hundred times trailing earnings.
A few days later, Beijing was already reaching for stabilization tools. More than 100 billion yuan in central-bank swap liquidity was being lined up to support equity purchases, financing for corporate share buybacks was made cheaper, and regulators were encouraging major technology companies and funds to support the market. So the question is no longer whether an AI boom can outrun itself. One already has, and that changes the argument from a distant warning into an immediate design problem.
The Warning Is Not That AI Failed
It would be easy to look at that and declare that the AI bubble is bursting everywhere. That would be lazy.
Artificial intelligence didn’t stop working because a technology index lost forty percent. Models didn’t forget how to write code, reason through documents, operate tools, generate media, or automate workflows. The infrastructure being built around them didn’t suddenly become useless. The long-term capability curve didn’t reverse because public markets decided they had paid too much for the near-term story.
The warning is more specific. Capital can outrun utility. Valuation can outrun revenue. Infrastructure can outrun demand and a technology can be genuinely transformative while an enormous amount of money is still allocated badly around it.
Those things aren’t contradictory. They are usually how technology booms end.
China Just Demonstrated the Mechanism
The Chinese market is useful because the mechanics became visible very quickly. AI-related stocks had grown to roughly 42 percent of Chinese market capitalization and 57 percent of turnover across mainland and Hong Kong-listed shares. Capital became heavily concentrated in a narrow technology narrative, retail participation became aggressive, and companies associated with robotics, chips, optical networking, and AI infrastructure were rewarded as though the future had already arrived on schedule.
Then the assumptions changed. Concerns about demand increased. Export restrictions became more threatening. Investors started asking whether valuations could be justified by actual earnings. Some of the most aggressively priced companies fell by half or more.
The technology didn’t need to become worthless for the trade to collapse. The price only needed to stop making sense. That distinction should matter to anyone building software right now, because the same behavior exists outside public markets. Companies are committing workflows, infrastructure, staffing assumptions, product roadmaps, and entire operating models to an AI ecosystem that is still rearranging itself underneath them.
The debt isn’t only financial. It’s architectural.
The United States Is Not Watching From Somewhere Safe
China may be further through the correction, but it isn’t operating in isolation. In October, more than $2 trillion was erased from the Nasdaq 100 as investors became less tolerant of enormous AI spending without equally enormous evidence of return. Technology companies still represent an extraordinary share of the S&P 500, and the largest hyperscalers are preparing to spend amounts on AI infrastructure that would have sounded absurd only a few years ago.
That spending may ultimately be justified. Some of it almost certainly will be.
The uncomfortable part is that nobody can reliably identify all of the durable value yet, because we’re still inside the expansion phase. Infrastructure is being financed before demand is fully understood. Products are being funded before defensibility is clear. Companies are being valued around future margins that depend on compute pricing, energy, model economics, customer behavior, and competitive structures that are all still moving.
China has shown what happens when confidence changes faster than the underlying technology does. The United States shouldn’t assume that geography alters that mechanism.
This Is the Part We Have Seen Before
The shape is familiar even when the technology is different. The dot-com collapse didn’t prove that the internet was useless. It proved that “the internet will change everything” wasn’t sufficient justification for every company with a domain name, a pitch deck, and access to speculative capital. The infrastructure remained. The network remained. The talent remained. Many of the businesses built around the initial frenzy didn’t.
The social-web era followed a different trajectory, but the consolidation pattern was recognizable. A period of extraordinary experimentation produced networks for every imaginable niche, behavior, audience, and identity. Capital flowed because growth itself was treated as evidence of durability. Eventually the surface area contracted, platforms consolidated, acquisitions absorbed competitors, and a much smaller number of companies controlled far more of the social graph than the early period suggested would ever be possible.
The useful technology survived both transitions. The surrounding assumptions didn’t. Artificial intelligence won’t replay either era exactly, but human incentives haven’t evolved enough to make the underlying pattern mysterious.
Money discovers possibility, possibility becomes narrative. Narrative attracts more money, money rewards duplication.Duplication produces saturation, saturation exposes weak economics, then gravity returns.
The Post-Boom Problem Is Already an Engineering Problem
The important question isn’t whether today’s AI companies survive. It is whether the systems being built on top of them survive. That is a much more immediate concern.
A business that tightly couples its internal workflows to one model vendor is making a financial prediction whether it admits it or not. It is predicting that the vendor remains available, that pricing remains acceptable, that the API remains compatible, that regulatory requirements remain manageable, that performance remains competitive, and that the provider’s business survives whatever consolidation follows. An organization that gives agents broad credentials is making another prediction. It’s assuming today’s capability boundary is a reasonable proxy for tomorrow’s authority boundary.
A company that allows generated internal tools to become production infrastructure without ownership, observability, documentation, or maintenance plans is making yet another prediction. It’s assuming the cost of creation is somehow related to the cost of stewardship. These are architectural bets disguised as convenience. The boom makes them feel cheap. The correction makes them visible.
Technical Debt Is Deferred Belief
Technical debt is usually described as engineering work postponed for later. The current cycle is producing another kind. Deferred belief. We believe the vendor will still be there. We believe token prices will continue falling. We believe the model will behave similarly six months from now. We believe the orchestration framework will remain maintained. We believe the provider won’t change its terms. We believe the generated application is simple enough that somebody else can understand it later. We believe the person who created the workflow will still be available when it fails. We believe the agent’s current permissions will remain appropriate as its capabilities grow.
Individually, any one of those assumptions may be reasonable. A system becomes fragile when too many of them must remain true simultaneously. That’s where the post-boom damage accumulates.
Not because artificial intelligence disappears, but because the organizations using it discover how much of their operating model depended on conditions that only existed during the expansion.
Build for Consolidation, Not Permanence
A durable AI system should assume churn. Model vendors will change. Pricing will change. Capabilities will converge. Frameworks will disappear. Standards will compete and some will lose. Acquisitions will eliminate products that looked permanent.
Regulation will change what can move across borders and what can be processed where. Security expectations will tighten. Procurement decisions will force migrations for reasons completely unrelated to benchmark performance. None of this should be surprising.
It should be a requirement. The goal isn’t to predict which model company wins. It is to build systems where being wrong about that prediction is inexpensive.
- Own the intent.
- Own the data.
- Own the control plane.
- Own the boundaries.
Treat the models, vendors, frameworks, interfaces, and APIs as machinery.
- Useful machinery.
- Powerful machinery.
- Replaceable machinery.
Authority Is More Durable Than Capability
This matters most with agents. The durable question isn’t what an agent can do, it’s who decided it was allowed to do it.
Capability is changing too quickly to be the foundation of authority. A model that can’t perform a task today may perform it trivially after the next release. A tool that currently exposes three safe operations may expose thirty later. An agent that once needed a human to coordinate several steps may collapse the entire workflow into a single action.
If authority is implicitly defined by capability, every improvement expands power. That’s backwards. Authority should remain stable enough that capability can evolve underneath it.
A better model should be able to perform twice as much work without automatically receiving twice as much permission. A new tool shouldn’t inherit every credential available to its host merely because integration was convenient. Automation should be able to increase without accountability decreasing at the same rate.
That’s what control looks like when the intelligence underneath the system refuses to stay still. The objective isn’t to put a human approval dialog in front of every mundane action. That would simply automate the work and preserve the burden. Mundane execution should disappear wherever it safely can. Ownership shouldn’t disappear with it.
Durable Does Not Mean Slow
There is a reason engineers flinch when someone says “future-proof.” It often means overbuilding. It means solving problems nobody has yet, creating abstractions around imagined requirements, and slowing down useful work in the name of a future that may never arrive. That isn’t durability. Durability means knowing which parts of the system should be replaceable and which parts express your actual authority.
A model provider is replaceable. Your ownership of the underlying data isn’t. A vector store is replaceable. The provenance of the information inside it isn’t. An orchestration framework is replaceable.
The policy defining what an agent may do on behalf of a person isn’t. A user interface is replaceable. The intent of the workflow behind it isn’t. Good boundaries make technological change boring.
That’s the goal. If a vendor becomes too expensive, replacing it should be annoying rather than existential. If a service shuts down, the organization should lose an integration rather than institutional memory. If regulation changes, the system should have somewhere coherent to express the new constraint. If a more capable agent arrives, it should encounter existing authority boundaries rather than quietly inventing new ones.
That isn’t moving slowly. It is refusing to pay the same migration tax every time the industry discovers something new.
The Boom Rewards the Opposite
The problem is that durable systems are difficult to sell during speculative periods. Visible motion wins. A company announcing a new agent product gets attention. A company making its data portable, reducing credential scope, separating its model layer, documenting provenance, and building clean operational boundaries has accomplished something far less exciting to put in a press release.
Markets are terrible at valuing disasters that didn’t happen. Nobody celebrates the rewrite that was unnecessary. Nobody sees the migration that took two days because the architecture was clean.
Nobody counts the vendor lock-in that never occurred, the compromised credential whose blast radius was already constrained, or the project that continued operating because its core logic didn’t belong to a provider that disappeared. Durability often expresses itself as an absence of drama. During a boom, that looks boring. After one, boring starts looking a lot like competence.
The Correction Changes What Companies Value
This is the part that matters commercially. During expansion, organizations reward speed, novelty, growth, and proximity to whatever technology currently signals relevance. They tolerate duplication because capital is available. They tolerate weak economics because adoption is growing. They tolerate architectural shortcuts because the next funding round, product milestone, or competitive announcement feels more urgent than the eventual maintenance bill.
Corrections reverse those incentives. Suddenly operating cost matters. Portability matters.
Security matters. Maintainability matters. Vendor exposure matters.
Institutional knowledge matters. The ability to replace one dependency without rebuilding five surrounding systems matters. The skills that look unnecessarily conservative during a boom become extremely valuable once companies are forced to distinguish capability from expense.
That’s why preparing for the correction doesn’t mean sitting outside the boom. It means using the boom to build the foundation the correction will eventually reward.
Sell the Present. Architect for the Future.
Nobody is going to approve a budget because someone promises “resilience against unknown post-bubble market conditions.” They will approve a useful capability they need now. So build that.
Use the models. Build the agents. Automate the repetitive work.
Reduce the amount of human effort wasted on execution that no longer requires it. Create products that would once have needed teams, vendors, and months of coordination. But build those capabilities through boundaries that survive the implementation currently providing them.
Ship today’s AI feature without making today’s AI vendor part of your identity. Use current demand to pay for infrastructure that remains useful when the terminology changes. Solve the present problem.
Preserve the ability to change your mind later. Sell the present, but architect for the future.
What China Changed About the Argument
A week ago, this argument could have sounded like cautious speculation. China removed that luxury. A forty-percent drawdown in one of the world’s most concentrated AI trades doesn’t prove the global AI expansion is ending. It doesn’t prove American technology stocks must follow the same path, and it certainly doesn’t prove the underlying technology is failing.
It proves something simpler and more useful. The relationship between technological importance and financial durability isn’t automatic. A technology can matter enormously and still produce terrible investments.
A product category can be inevitable and still contain companies that shouldn’t exist. Infrastructure can be necessary and still be financed at prices its eventual demand can’t justify. A market can correctly identify the future and still wildly misprice the path required to reach it.
That should be enough to change how we build. The question is no longer whether there will eventually be a period when companies discover which AI assumptions were durable. That process has started somewhere already.
Build for What Remains
The eventual winners of this cycle won’t necessarily be the companies that attached themselves most aggressively to the current wave. Some will be. Others will simply have been standing on durable ground when the water moved somewhere else.
The useful parts of this technology will remain. Models will remain. Automation will remain. Agentic systems will remain. Small teams operating with extraordinary leverage will remain. The ability to compress execution cost across entire categories of work will remain.
A great deal of everything surrounding those capabilities will change. That isn’t a reason to retreat. It’s a reason to own the parts that matter.
Build for the boom if you want attention. Build for what survives the boom if you want leverage.
The objective isn’t to become invulnerable to change. Nothing in technology deserves that confidence. The objective is to make change ordinary.
If the next model is better, use it. If the vendor changes, replace it. If the framework dies, remove it.
If regulation changes, express the new boundary. If a category collapses because it never deserved the capital thrown at it, let it collapse. The durable system isn’t the one that predicted the future correctly. It’s the one that never required the future to cooperate in the first place.
Sources / Working References
- Reuters, October 26–27, 2026 — “In China, the AI bubble has already burst”
- Reuters, November 2, 2026 — China market support measures following the October selloff
- Reuters, November 3, 2026 — global market risks from concentrated AI investment
- Reuters, November 3, 2026 — China National Financial Regulatory Administration warning on AI investment concentration