Graphic | Sister Tang

On July 7, Samsung estimated its Q2 operating profit to be 19 times that of the same period last year, yet its stock price fell by 6.9%, with Micron and SanDisk following suit. Also in early July, TrendForce predicted that Q3 DRAM (excluding HBM) and NAND contract prices would still rise by 13%—18% and 10%—15%, respectively, albeit with narrower margins.

The combination of surging profits, rising prices, and falling stock prices suggests that market enthusiasm for memory stocks is waning. Soon, the decline in memory stocks was interpreted as weakening AI demand, sparking concerns about a potential shift in the market's main theme. However, whether the market is truly moving away from AI cannot be determined solely by the fluctuations in the memory sector.

The real question is whether adjustments are still rotating within the AI sector or if the main theme is indeed shifting away from AI. As long as the tasks assigned to AI continue to increase, the demand foundation for AI as the main theme remains intact; which specific segment reaps the profits is a relatively secondary issue.

The challenge lies in the lack of real-time pricing for the volume of tasks assigned to AI. A 2026 report by the International Energy Agency (IEA) estimated that the electricity consumption of the same AI task has decreased by at least 90% annually in recent years, yet electricity usage by AI-specific data centers is expected to grow by 50% in 2025. Google disclosed in June that its monthly token processing volume has increased about 330-fold over the past two years. Electricity usage reflects physical load, while tokens reflect usage scale—both are expanding.

When such a combination occurs, the classic explanation is Jevons' Paradox: as efficiency improves and resources become cheaper, usage increases, leading to a rise in total consumption.

In December 2025, the National Bureau of Economic Research (NBER) released a working paper by Mert Demirer and three others, using real API data from OpenRouter and Microsoft Azure to observe changes in usage after price reductions. A 10% price drop led to an approximately 11% increase in usage. For example, originally, 100 units cost 1,000 yuan; after the unit price dropped from 10 yuan to 9 yuan, usage increased to about 111 units, with spending still around 1,000 yuan.

Such studies focus on short-term price elasticity of demand, making it difficult to distinguish whether growth comes from more requests or heavier tasks behind each request. Monthly active users and request counts only record how many people use the service and how many calls are made, not how many steps a single task is broken into behind the scenes. From a user interface perspective, translating a sentence and organizing meeting recordings, verifying data, listing to-dos, and writing client emails may all appear as a single instruction, yet the backend workload varies greatly.

A true reversal in AI demand would require a reduction in AI steps within single tasks, a halt in AI's expansion into professional fields, and a reversion of previously outsourced work to human labor.

01 AI Will Take Over More Steps

AI taking over more steps in human work has become a Public competition (publicly competitive) direction for model companies.

OpenAI highlighted long tasks composed of research, tool invocation, and complex execution as a focus when introducing GPT-5.3-Codex in February 2026. Zhipu AI, when releasing GLM-5.2 in June, directly included "built for long-cycle tasks" in the title and tested the model with open projects lasting hours or even tens of hours. The competition now extends beyond who answers better in a single instance to who can work continuously toward the same goal for longer.

Anthropic analyzed about 400,000 Claude Code interactive sessions covering approximately 235,000 users in June 2026. For users unfamiliar with a task, Claude executed an average of about 5 actions and output about 600 words per interaction; for familiar users, it was 12 actions and 3,200 words.

To minimize the impact of other factors, researchers controlled for work style, task value, month, occupation, and model version. Logically, the more familiar someone is with a task, the less AI assistance they should need; however, the data showed the opposite. For each level of increased familiarity, AI actions increased by 9% and output by 13%.

The clearer one is about how a task should be done, the more specific steps can be delegated to AI. What's truly changing here is not that people are asking AI more questions but that between human interventions, AI can continuously complete more retrieval, execution, inspection, and correction steps. The results of one step directly inform AI's decision for the next; it can adjust plans and invoke tools again without waiting for a new human instruction.

Once tasks enter an agent-based workflow, backend consumption increases; if further divided among multiple AI agents, the division of labor expands further. In Anthropic's research system, a single AI agent consumes about 4 times the tokens of ordinary chat, while collaboration among multiple agents increases consumption about 15-fold. When completing a research task, different agents retrieve data separately, with a lead agent summarizing and verifying the results.

In such systems, although users ultimately receive a single result, multiple task lines have run in parallel behind the scenes. These multiples come from Anthropic's research and are not controlled experiments comparing the same task across different architectures; they merely illustrate that token consumption in multi-agent research systems can be far higher than in ordinary chat.

Even among frequent AI users who are relatively skilled, a direct observation is that when the same type of task is assigned to the same AI, the path it takes to complete the task and the tokens consumed vary each time, sometimes significantly.

A April 2026 working paper from Stanford's Digital Economy Lab compared the execution trajectories of 8 cutting-edge models on the same coding benchmark, finding that token consumption for the same task varied by up to 30 times across different runs; spending more tokens did not necessarily yield better results, with accuracy often peaking at moderate costs.

Such evidence suggests that some of today's seemingly excessive token growth is indeed wasteful. Model companies and users are still exploring how to break down a task, when to retry, and to what extent to check; running the same task dozens of times more does not mean all that consumption will persist.

As methods mature, redundant attempts and ineffective rework will decrease, and the tokens required for the same task may also decline. However, as long as work still requires step-by-step planning, tool invocation, and result inspection, these segment (links) will not disappear due to improved model efficiency. The more complete the tasks delegated to AI, the more steps it must complete; as more professional fields adopt similar approaches, another layer of growth will emerge.

02 AI Is Taking Over More Than Just Programming

Programming became the first major battleground for AI agents because code can be run, debugged, tested, and iteratively modified, providing clear feedback at each step. Other professional fields also have their own data, tools, and methods for check and acceptance (acceptance methods); life sciences is one of the key directions Top companies (leading companies) are actively pursuing.

On June 17, 2026, OpenAI released LifeSciBench, with 750 tasks closely resembling real scientific research work written by 173 scientists with doctoral training and experience in biotechnology or pharmaceuticals. 79% of these tasks require multi-step reasoning or decision-making, including processing evidence, analyzing data, designing experiments, and verifying results—turning the complex work researchers face daily into assessable AI tasks.

OpenAI launched the life sciences model GPT-Rosalind in April 2026, along with a Codex life sciences plugin connecting to over 50 scientific tools and data sources. Anthropic released Claude Science on June 30, capable of connecting to laboratory computing resources, over 60 scientific databases, and AI agents responsible for verifying results. Both companies are transforming general-purpose models into life sciences workstations.

In February 2026, OpenAI and Ginkgo Bioworks connected GPT-5 to a cloud-based automated lab, with the model designing experiments and robots executing them under human supervision. Experiment results were fed back to the model. After six rounds of testing over 36,000 reaction combinations, they reduced production costs by 40% compared to the previous best baseline for a protein and a cell-free protein synthesis system. AI also began deciding next steps based on physical experiment feedback, entering real experimental cycles.

While code relies on testing for validation and life sciences on experimental feedback, financial reconciliation, contract review, and industrial simulation each have their own rules, parameters, and acceptance methods. If AI can access relevant data and tools and submit results for professional or real-world feedback, it can form its own workflow. The more professional fields AI enters, the more work can be delegated to it.

In real usage data disclosed by Anthropic, from October 2025 to April 2026, the proportion of Claude Code sessions used for error detection and correction dropped from 33% to 19%, while those for deployment and software operation rose from 14% to 21%, and writing and data analysis combined increased from about 10% to 20%. Claude Code's scope is expanding from handling code issues to software operation and non-code knowledge work.

I am a direct example. I cannot program but have used Codex and Claude Code to organize data, verify information, and advance investment research, gradually delegating entire research processes to AI. The execution approach first adopted by programmers is equally applicable to research, marketing, finance, and legal work; product logic and AI capabilities iterated in programming scenarios are now expanding to other business functions within enterprises.

On July 9, 2026, OpenAI integrated Chat, ChatGPT Work, and Codex into a single ChatGPT desktop app: Work handles research, analysis, and delivery, while Codex handles software development; both modes include a built-in browser, bringing chat, knowledge work, and software development into one desktop application.

For most investors, even if they know these products have changed, it is hard to truly perception (perceive) how far AI capabilities have advanced. Many use ChatGPT but remain stuck in the 2022-2023 question-and-answer mode, such as opening the app, entering a stock code, asking, "What do you think of this company?" and reading the analysis before stopping.

Stuck at this level, users only see a single answer; the processes of retrieving data, verifying information, invoking tools, and iterative revisions—originally done by humans—have not truly been delegated to AI, so they cannot perception (perceive) how much work AI can now complete continuously after a single instruction.

Shifting from "asking a question" to "delegating a task" changes more than just usage patterns. The former can stop anytime, while the latter, once reliably delivering results, means discontinuing AI use requires returning entire workflows to humans. Once people adapt to delegating repetitive, tedious yet necessary tasks to AI, it becomes difficult to take them back piece by piece.

03 Conclusion

The market's current dilemma is that both bullish and bearish views on AI can find plausible evidence. Memory prices and stock performance warrant caution, while token consumption, electricity usage, and expanding capability boundaries provide reasons for optimism.

The market prices memory and AI companies daily, but how much work is truly being delegated to AI remains hard to observe directly—this is where AI demand is most easily misjudged.

However, there is no need to force a seemingly precise number. A more practical approach is to observe which tasks we are delegating to AI: has the same task evolved from a single question to having AI retrieve data, invoke tools, inspect, and revise on its own? Beyond programming, are investment research, life sciences, finance, and legal work beginning to use the same approach? As long as these changes continue, the work delegated to AI is still increasing.

The real concern is not a few days of decline in memory stocks or a slight profit drop at one company, but a scenario where models continue to improve, usage costs keep falling, yet people stop delegating new work to AI or even start reclaiming previously outsourced tasks. Only at that point would AI demand truly shift. Until then, the market is merely reselecting which players will profit within the AI theme.

Disclaimer: This article is for learning and communication purposes only and does not constitute investment advice.

Like, share, and repost! Your support is what keeps us updating!