Code up 300%, results up only 30%: The AI dividend faces an awkward reality

```

AI is creating an illusion of efficiency. Code output has surged, but software releases have only modestly increased; cumulative enterprise AI spending has surpassed $1 trillion, yet actual cost savings are generally disappointing. Multiple recent studies point to the same unsettling conclusion for the market: The productivity dividend of AI is already sharply diminished before it translates into real business value.

According to the Financial Times, MIT researchers tracked the performance of software developers before and after using AI tools and found that AI brought remarkable gains to basic tasks—the number of code files created or edited by developers increased by nearly 300%. However, this growth shrank dramatically after human steps such as code review and integration, with the ultimate increase in fully released software sitting at only around 30%.

Meanwhile, Bain’s survey of 951 large global enterprises shows that among those able to quantify AI’s cost savings, the largest group (40%) achieved cost reductions of only 10% or less, far below initial expectations.

The combination of these two studies puts pressure on the valuation logic of AI investments. The current high valuations of AI concept stocks are based on forecasts of future productivity rather than actual realized returns. Once the market starts seriously calculating ROI, the risk of valuation restructuring cannot be ignored.

Funnel effect: AI gains dissipate step-by-step in the workflow

MIT researcher Mert Demirer and co-authors provide the most systematic path map yet of AI productivity decay. The research team tracked developers' outputs at multiple levels: from the number of code files, independently edited files, work units submitted for review, to the final software release versions.

The results show a clear funnel structure: AI brings nearly 300% explosive growth at the upstream code generation stage, but by the submission for review stage, the increase is halved to about 150%; once in the complete software release stage, the gain shrinks fivefold again, ending at around 30%.

Researchers further examined whether the boost in software output under AI assistance led to actual consumer demand, and the conclusion is equally sobering. Over the past year, the number of mobile app releases grew significantly, but downloads did not rise correspondingly—most new apps failed to attract even a limited user base. This means that content produced faster by AI does not necessarily create the value the market needs.

Trillion-dollar investment, returns in doubt: Enterprises caught in a "cyclical bet"

Bain’s survey, from the perspective of corporate finance, confirms the core findings of the above research. This study—covering nine major global industries and enterprises with annual revenues over $100 million—shows that after global AI spending accumulated to over $1 trillion, the actual cost savings from automation generally fell far short of expectations.

More concerning is the decision-making logic of enterprises themselves. Bain found that 44% of large enterprises are using "cost savings from the previous round of AI, not yet realized" as the basis for funding the next round of AI investments, which Bain describes as "a structurally flawed cyclical bet." Meanwhile, Gartner predicts that over 40% of intelligent agent AI projects will be terminated before the end of 2027.

Specific cases at the enterprise level also confirm this trend. Uber’s CEO Dara Khosrowshahi recently revealed that the company exhausted its annual AI budget within a single quarter and plans to switch most AI usage to lower-cost models, reserving cutting-edge tools only for specific scenarios. Other recent studies on legal work show that pairing low-cost open-source AI with top-tier models can achieve better outcomes while greatly reducing costs. Bain summed it up: “The technology works, but the value hasn’t arrived.”

Historical lens: Structural change lags behind technological breakthroughs

Faced with the above data, the MIT team is not inclined to interpret it as evidence that AI’s capabilities are overrated. They believe a more likely explanation is that current organizational and market structures are not yet ready to absorb the true potential of AI.

History provides a strong reference. In the electrification wave from the late 19th to early 20th century, factories that simply swapped steam engines for electric motors and kept the original mechanical layout had very limited productivity gains. The real leap in efficiency came decades later—when engineers equipped each workstation with an independent small motor, completely restructuring production processes.

Something similar may be unfolding in the current AI sector. Traditional software and knowledge-work enterprises integrating AI into existing workflows gain limited productivity; but companies like Anthropic and OpenAI, built natively around AI, see explosive growth in usage, revenue, and productivity. This contrast suggests that true AI dividends may depend on the emergence of new organizational forms and business processes, rather than simply layering AI onto existing structures. For investors betting on AI valuations now, this means the waiting period may be far longer than expected.

Risk warning and disclaimerThe market involves risk; investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article fit their specific circumstances. If acting on this information, you do so at your own risk. ```