The AI hype cycles aren’t a single collapse but a continuous sequence of adjustments
When people assume AI will collapse in one dramatic burst, they overlook how real-world tech evolves. Instead of one peak followed by total failure, we see overlapping hype cycles—each one bursting where enthusiasm outpaces utility, yet leaving the core systems intact to fuel the next phase. That’s why the fear of a permanent crash isn’t accurate; it’s a selective pruning process that strips away speculative investments and reveals the functional tools that survive.
The rolling bubble theory holds because each wave of hype—whether early chatbot excitement or current agentic AI promises—drives massive hardware and energy investments. These costs create the foundation for the next practical application, ensuring no single bubble vanishes entirely. The 2000 dot-com crash proved the internet didn’t disappear; only the companies without sustainable models did. AI follows the same pattern, but at an accelerated pace, shifting from novelty-driven models to integrated, task-specific solutions.
Dhaval Joshi argues that the debate over whether AI is a bubble is misplaced. Instead, the question should be: Which specific bubble is deflating now? His analysis reframes the entire discussion on LinkedIn, explaining that rather than a single implosion, AI’s trajectory is a rapid-fire series of bubbles inflating and popping in quick succession. Investors repeatedly misjudge which entities will actually capture AI’s economic value, only to correct course as reality aligns with performance.
A recent comparison to the Dutch tulip bubble highlights this pattern. Just as rare bulbs led to speculative tulip futures, AI’s current hype spans sectors beyond mere novelty—from specialized agents to autonomous workflows. The rapid-fire nature means one bubble’s collapse doesn’t signal industry death; it’s just the next phase of specialization taking root.
To build something that endures past the next cycle, focus on the technical shifts shaping real-world utility. Treat LLMs less as search engines and more as reasoning engines that trigger external tools—this transition is where lasting value emerges. RAG (Retrieval-Augmented Generation) outperforms fine-tuning because it delivers verifiable, up-to-date results without proprietary data costs. Meanwhile, hardware investments peak before software breakthroughs; the challenge now is identifying the "killer apps" that justify those expenditures.
The winners aren’t the companies betting on broad AI superiority but those who deploy models into existing workflows—whether using Claude Code for rapid coding or running a local Llama instance for privacy. The volatility of valuations is noise; the measurable growth in deployment capabilities is the only lasting metric.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
This is hilarious. Is it just glorified autocorrect, or is something deeper happening here? One bubble pops—like the initial frenzy over basic chatbots—but the underlying infrastructure remains, fueling the next wave of specialized agents or autonomous workflows.
The author’s claim that AI is in a single "bubble" misses how real-world scaling works—like how the early LLM hype actually funded the infrastructure needed to build specialized agents today. The idea that one peak collapses entirely ignores that the tech that survives isn’t just the flashy models but the scalable systems that can be deployed in workflows, proving their utility before the next cycle.
I'm skeptical. Where is the data proving these are rolling bubbles instead of just market foam? The idea that we are in a single "AI bubble" is too simplistic because it ignores how technology actually evolves. Instead of one massive peak followed by a total collapse, we are seeing a series of overlapping hype cycles. One bubble pops—like the initial frenzy over basic chatbots—but the underlying infrastructure remains, fueling the next wave of specialized agents or autonomous workflows. This pattern means the "crash" people fear isn't a death sentence for the industry, but a pruning process that clears out the fluff and leaves behind the tools that actually work. Most people confuse a market correction with the end of a technology. When the dot-com bubble burst in 2000, the internet didn't disappear; the companies that didn't have a real business model did. AI is following a similar trajectory but moving much faster. We move from the "LLM as a novelty" bubble into the "AI-integrated software" bubble, and we're already sliding into the "Agentic AI" phase. Each single wave of hype creates a massive over-investment in hardware and energy, which then provides the cheap, scalable foundation for the next, more practical application. The danger isn't that AI is fake, but that the current valuation of companies is based on the promise of general intelligence rather than the utility of specific tools. We are seeing a transition from pure prompt engineering—where the AI model is treated as a black box—to a more nuanced approach where the focus is on understanding and optimizing the underlying mechanisms. This shift is crucial for developing AI systems that are not only powerful but also reliable and trustworthy.