After the drop in storage stocks, is AI still the main theme of the market?

On July 7, Samsung estimated its Q2 operating profit would reach 19 times the figure from the same period last year, yet its share price fell by 6.9%, followed by Micron and Western Digital. Also in early July, TrendForce projected that Q3 contract prices for DRAM (excluding HBM) and NAND would still rise by 13%-18% and 10%-15% respectively, albeit at a narrowed growth rate.

A combination of skyrocketing profits, sustained price increases, and falling stock prices signals that the market's previous enthusiasm for the memory sector is starting to wane. Soon, the drop in memory stocks was interpreted as a sign of weakening AI demand, sparking widespread market concerns over whether the core market trend would shift. But whether the market is truly moving away from AI cannot be determined solely by the ups and downs of the memory segment.

The critical judgment to make is whether the current adjustment remains a rotation within the AI sector, or if the market's main narrative is shifting away from AI. As long as the volume of tasks delegated to AI keeps growing, the AI thematic still has a solid demand foundation; which specific segment ultimately captures the profits becomes a relatively secondary concern.

The core challenge is that there is no real-time price metric to measure the volume of tasks assigned to AI. A 2026 report from the International Energy Agency (IEA) estimates that the power consumption of a standard AI task has dropped by at least 90% annually in recent years, yet power usage by AI-dedicated data centers will grow by 50% in 2025. Google disclosed in June that its monthly processed Token volume has surged roughly 330-fold over two years. Power consumption reflects physical computing load, while Token volume reflects usage scale, and both metrics are expanding rapidly.

When such a combination emerges, the classic explanatory framework is the Jevons Paradox: Improved efficiency makes resources cheaper, which in turn drives greater usage, ultimately leading to a rise in total consumption despite lower unit costs.

In December 2025, the National Bureau of Economic Research (NBER) published a working paper by Mert Demirer and three co-authors, which used real API data from OpenRouter and Microsoft Azure to observe usage changes following price reductions. A 10% price cut correlated with roughly an 11% increase in usage. For context, spending 1000 currency units previously bought 100 units of service; after the unit price fell from 10 to 9, usage rose to approximately 111 units, keeping total spending around 1000 units.

Such studies only capture the short-term price elasticity of usage, and cannot distinguish whether the growth stems from a greater number of requests, or from individual requests becoming more computationally intensive. Monthly active users and request counts only track how many people are using the service and how many calls they initiate, but cannot reveal how many discrete steps a single task is split into behind the scenes. From a user interface perspective, translating a sentence, organizing meeting recordings, verifying figures, drafting a to-do list, or writing a client email might all appear as a single instruction, yet the underlying workload varies dramatically.

A genuine reversal in AI demand would require evidence that the number of AI-powered steps per task is decreasing, that the adoption of AI into professional fields has stalled, and that work previously delegated to AI is being reclaimed by human workers.

01

AI Will Take Over More Work Steps

Enabling AI to handle more discrete steps in human workflows has become an explicit competitive focus among leading model developers.

When OpenAI introduced GPT-5.3-Codex in February 2026, it highlighted long-form tasks combining research, tool calling, and complex execution as a core priority. In June, Zhipu AI released GLM-5.2, explicitly labeling it "built for long-duration tasks" and testing the model with open-ended projects spanning hours or even dozens of hours. Competition no longer centers solely on who can deliver a better single response, but on which model can sustain continuous work toward a shared goal for longer periods.

In June 2026, Anthropic analyzed approximately 400,000 interactive Claude Code sessions, covering around 235,000 users. For users unfamiliar with a given task, each human prompt triggered an average of 5 execution actions and roughly 600 words of output. For users deeply familiar with the task, that figure rose to 12 actions and 3200 words of output.

To isolate the impact of other variables, researchers controlled for work style, task value, calendar month, user occupation, and model version. Conventional wisdom suggests that users more familiar with a task would require less AI assistance, but the data showed the opposite: for each level of increased user familiarity, AI-executed actions rose by 9% and output volume by 13%.

The more clearly a user understands how a task should be completed, the more effectively they can delegate its specific steps to AI. The real shift here is not that humans are asking AI more questions, but that between two points of human intervention, AI can now autonomously complete far more rounds of research, execution, inspection, and correction. Results from prior execution steps directly inform the AI's subsequent decision-making, allowing it to adjust plans and re-invoke tools without waiting for a new human instruction.

Once tasks shift to an Agent-based workflow, backend resource consumption rises; if tasks are further distributed across multiple AI agents, the division of labor will expand even more. In Anthropic's own research system, a single AI agent consumes roughly 4 times the Token volume of a standard chat interaction, and multi-agent collaborative workflows consume approximately 15 times as many Tokens. When completing a research task, different agents independently retrieve relevant materials, before a lead agent aggregates and cross-validates the results.

In such systems, even though users ultimately receive a single consolidated output, numerous parallel task streams have already run in the background. These multipliers, sourced from Anthropic's internal data (and not controlled experiments comparing identical tasks across different architectures), indicate that multi-agent research systems can have far higher Token consumption than basic chat applications.

Even frequent, experienced AI users share a common intuitive observation: when delegating identical categories of tasks to the same AI, the execution path and Token consumption for each completed task are inconsistent, with discrepancies sometimes being extremely large.

A working paper published by Stanford's Digital Economy Lab in April 2026 compared the execution trajectories of 8 state-of-the-art models on the same coding benchmark, finding that Token consumption across different runs of the identical task could vary by up to 30 times. Spending more Tokens does not guarantee better performance, as accuracy often plateaus at mid-level cost thresholds.

This evidence suggests that the current seemingly exponential Token growth does include inefficiencies. Model developers and users are still experimenting with how to split tasks into appropriate steps, when to retry operations, and what level of validation is sufficient; a 30x increase in runs for the same task does not mean all that excess consumption will persist long-term.

As methodologies mature, redundant retries and invalid rework will decrease, and the Token requirement for a given task will likely decline. However, as long as work still requires step sequencing, tool invocation, and result validation, these processes will not disappear simply due to improvements in model efficiency. The more complete the tasks people delegate to AI, the more intermediate steps the AI needs to handle; as more professional fields adopt this same workflow, it will drive the next wave of growth.

02

AI Is Taking Over More Than Just Coding

Coding is inherently suited for AI agents because code can be run, debugged, tested, and iteratively modified, with clear feedback available at every step. Other professional domains have their own unique datasets, tools, and validation workflows, and life sciences is one of the key verticals that leading companies are prioritizing.

On June 17, 2026, OpenAI released LifeSciBench, a benchmark of 750 real-world scientific tasks curated by 173 scientists with PhD training and biotech/pharmaceutical industry experience, 79% of which require multi-step reasoning or decision-making. These tasks include processing evidence, analyzing datasets, designing experiments, and validating results, translating the complex daily work of researchers into evaluatable AI tasks.

In April 2026, OpenAI launched the life sciences specialized model GPT-Rosalind, alongside a Codex Life Sciences plugin that supports connections to over 50 scientific tools and data sources. On June 30, Anthropic released Claude Science, which integrates with lab computing resources, more than 60 scientific databases, and dedicated AI agents for result verification. Both companies are reframing their general-purpose models into dedicated workbenches for the life sciences sector.

In February 2026, OpenAI partnered with Ginkgo Bioworks to integrate GPT-5 into cloud-automated laboratories, where the model designs experiments under human supervision, robots execute the procedures, and results are fed back to the model. Six rounds of testing evaluated over 36,000 reaction combinations, reducing production costs for a target protein and a cell-free protein synthesis system by 40% compared to the previous best baseline. AI is now making iterative decisions based on physical experiment feedback, embedding itself into real-world experimental cycles.

Coding relies on test-based validation, life sciences on experimental feedback, and financial reconciliation, contract review, and industrial simulation each have their own rules, parameters, and acceptance criteria. When AI can access relevant domain data and tools, and submit outputs for validation by professionals or real-world feedback, it can form its own end-to-end workflows. The more professional domains AI penetrates, the greater the volume of work that can be delegated to it.

According to real usage data disclosed by Anthropic, between October 2025 and April 2026, the share of Claude Code sessions dedicated to debugging and error fixing dropped from 33% to 19%, while the share for software deployment and execution rose from 14% to 21%, and the combined share for writing and data analysis increased from roughly 10% to 20%. The scope of Claude Code is expanding beyond pure coding tasks to encompass software operation and non-code knowledge work.

I am a direct personal example: I have no programming background, yet I already use Codex and Claude Code to organize research materials, verify data, and advance investment research, gradually delegating the entire end-to-end investment research workflow to AI. The execution paradigm first popularized by programmers is equally applicable to research, marketing, finance, and legal work. The product logic and AI capabilities iterated in the coding use case are now extending to other business functions within enterprises.

On July 9, 2026, OpenAI consolidated Chat, ChatGPT Work, and Codex into a single ChatGPT desktop application: the Work module handles research, analysis, and deliverable generation, while Codex focuses on software development. Both modules support built-in browser access, unifying chat, knowledge work, and software development into one desktop environment.

For most investors, even if they are aware of these product evolutions, it is difficult to fully grasp how advanced current AI capabilities have become. Many people already use ChatGPT, but their usage patterns remain stuck in the 2022-2023 era of simple one-turn question-and-answer interactions: opening a mobile app, entering a stock ticker, asking "what is your analysis of this company," reading the response, and ending the session.

When usage stops at this level, users only see a final text answer; the processes of retrieving materials, cross-verifying data, invoking tools, and iteratively refining outputs—tasks previously done manually—have not truly been delegated to AI. As a result, users cannot perceive the volume of continuous work AI can complete after a single instruction.

The shift from "asking a question" to "assigning a task" represents far more than a change in usage pattern. The former can be abandoned at any time, but once the latter can reliably deliver consistent outputs, stopping AI usage means reverting the entire workflow back to human labor. Once people grow accustomed to delegating repetitive, tedious yet necessary tasks to AI, it becomes extremely difficult to take those workflows back into manual control.

03

Conclusion

The current market dilemma is that both bullish and bearish arguments about AI have reasonable supporting evidence. Fluctuating memory prices and stock valuations justify caution, while the continuous expansion of Token volumes, power consumption, and AI capability boundaries supports optimistic outlooks.

The market prices memory and AI companies on a daily basis, but the actual volume of work being delegated to AI is difficult to directly measure, which is exactly where AI demand is most prone to misjudgment.

However, there is no need to force the pursuit of a misleadingly precise single number. A more practical approach is to track what new types of work we are delegating to AI: whether a task that previously required a simple question now involves autonomous information retrieval, tool invocation, inspection, and revision by AI; whether beyond coding, fields like investment research, life sciences, finance, and legal work are adopting these same workflows. As long as these transformations continue, the total volume of tasks assigned to AI will keep growing.

The real risk worth monitoring is not a few days of memory stock declines, or a minor earnings miss by a single company. The actual turning point would be when model capabilities are still improving and usage costs are still falling, yet people stop delegating new work to AI, and even begin to reclaim tasks that were previously automated. Only at that point would AI demand truly reverse. Until then, the market is essentially re-evaluating which participants in the AI value chain will ultimately capture the profits from this trend.

Disclaimer: This article is for educational and informational purposes only and does not constitute investment advice.

This article is sourced from the WeChat public account "Tangping Index", authored by Tangjie, and published with authorization from 36Kr.