Marginal Analysis: How to Determine If Continued Investment Is Worthwhile?

Marginal Analysis: How to Determine If Continued Investment Is Worthwhile? #

In this article, we will delve into a concept widely regarded by economists as the core wisdom of economics—“marginal analysis.”

It helps you analyze any endeavor: whether to persevere, give up, or increase investment.

Take the “growth flywheel” often mentioned by entrepreneurs, for instance; there is no eternal positive feedback in the world. So, how do you determine if this flywheel is just starting or already nearing its end?

Marginal analysis can serve as a feedback-driven decision-making method, allowing you to adjust the next round of investment based on changes in output. Its advantage is that you can obtain clear decision signals without needing to master all system details beforehand.

I believe marginal analysis is a severely underestimated thinking tool. It should be as widely known as idioms like “don’t overdo it” and “an arrogant dragon will have cause to repent.”

Core Insight of Marginal Analysis #

When non-professionals observe things, their instinctive reaction is often to focus on their size: this company has strong capital, that organization has many members—impressive! However, large doesn’t equate to healthy. The company might be poorly managed, on the brink of insolvency; the organization might be demoralized. Stakeholders, at a minimum, pay attention to averages, such as how the company’s efficiency and profit margins perform.

However, these are merely past and present performances, whereas decisions must be forward-looking.

To judge whether something has development potential and is worth additional investment, economists focus on the “margin”: how much new output the next unit of input can bring.

For instance, you open a new factory. Initial heavy investment in factory buildings and equipment constitutes fixed costs. So, how many workers should you hire next? The first worker might only be able to operate one machine alone; after the fifth worker joins, division of labor becomes possible, for example, one person loading materials, another responsible for quality inspection; by the time the tenth worker arrives, the production line can operate truly efficiently—the more hands, the finer the division of labor, and the greater the additional output each new hire brings. This is increasing marginal product. With increased output, the fixed costs allocated to each product also decrease. These are all positive signals, indicating that you should continue hiring.

However, this process is not infinite. Workshop space is limited, and equipment quantity is fixed. When the 20th worker is hired, daily output still grows, but the growth rate has significantly slowed; by the 40th worker, laborers start waiting for equipment and hindering each other, at which point the additional output brought by each new person is zero. If you continue hiring, the workshop will become as crowded as a spring festival waiting hall, and output will actually decrease.

Without significant factory renovation, you will transition from increasing marginal product to decreasing marginal product, eventually even experiencing negative marginal product.

Evidently, more workers are not always better: the 5th, 20th, and 40th workers have entirely different implications for you. This is the profound insight of marginal analysis.

Four Scenarios of Marginal Benefit #

Next, we will elaborate on the concept more formally. Regardless of your endeavor, it will involve marginal revenue and marginal cost. Let’s define:

Marginal Benefit = Marginal Revenue - Marginal Cost

Marginal benefit can be positive or negative; its trend can be increasing or decreasing. Combining these two dimensions, you will face four basic scenarios—

Marginal benefit is positive and continuously increasing. This is the optimal period; be sure to significantly increase investment! Every additional unit of your input will yield more returns than before.

Marginal benefit is positive but decreasing. While still worth continuing—each additional unit still yields profit—the returns are gradually diminishing. At this point, be vigilant, as the future prospects of this endeavor are limited, and it may soon cease to be profitable.

If the current marginal benefit is temporarily negative, but the trend is improving, this might be the “cold start” phase we mentioned earlier: current investment may carry risks, but as long as this improving trend is genuine and reliable, it is worth investing, and profitability will eventually be achieved.

If marginal benefit is negative and continuously worsening, this is a bottomless pit of losses, and you should not hesitate to exit immediately.

The Magic of Increasing Marginal Benefit #

“Increasing marginal benefit” is arguably the area most worth significant investment in the world: each time you add investment, it not only brings new output, but the output quantity is greater than before! You are riding a positive feedback flywheel, unstoppable, and you might even feel that sleeping is a waste of time.

Why did we previously emphasize that “platforms” are the most powerful business model in the modern world? Precisely because they have an increasing marginal benefit structure: the more users, the more valuable the platform is to each user, the easier it is to attract the next user, and its value also increases. Platforms also benefit from a data flywheel effect: the larger the user base, the richer the data, making the product smarter, thereby attracting more users.

Increasing marginal benefit is also the secret to industrial profitability: with fixed initial costs, the more products produced, the less cost is allocated to each, and the greater the profit; furthermore, the more products produced, the higher the proficiency, allowing costs to fall further and profits to increase further.

In fact, as long as marginal benefit does not decrease, this endeavor is not only worth continuous investment but should also be constantly amplified using business leverage. You just need to reinvest the profits, forming compound interest, and growth itself will bring new growth—you will achieve the exponential growth people dream of…

In investor jargon, this is “scalability,” meaning the endeavor can be continuously amplified.

The Inevitability of Diminishing Marginal Benefit and the S-Curve #

However, we must admit that there is no eternal increasing marginal benefit in the world.

Every reinforcing loop exists within a larger system, and within that system, there are always balancing loops waiting to impose limits.

If your product sells well, competitors will enter the market, vying for users and suppliers, and launching price wars. Even if Messi’s skills are superb and his fan base enormous, the number of games he can play each year is limited, and he will eventually face aging. More often, the celebrity himself is still enthusiastic, but fans have grown aesthetically fatigued… Even if you avoid all these, there is still the hardest ceiling: total market size. When your market share reaches 50%, it’s impossible to double your growth again.

You will eventually encounter decreasing marginal benefit.

Every few years, the capital market relearns this lesson with real money. Recall the shared bicycle wars of yesteryear, where capital mistakenly believed that deployment volume was the flywheel: the more bikes, the more convenient, and the more convenient, the more users. The first half of the story was indeed true; however, once effective density was exceeded, the convenience brought by new bikes approached zero, while the costs of dispatch, maintenance, and loss continued to climb… The projects ultimately became unsustainable, turning into bicycle graveyards.

Therefore, even the best products in the world cannot deliver infinite exponential growth; their development trajectory typically presents as an “S-curve”—

With cumulative investment, the total outcome’s trend changes as follows: an initial cold start with huge investment but weak results; a mid-stage where the flywheel effect appears, marginal benefit increases, and the endeavor takes off; and a late stage where the market saturates, marginal benefit decreases, and the curve gradually flattens.

If you want to understand a product’s current life cycle stage, simply investigate its position on the S-curve.

In fact, the Chinese have long had a profound understanding of this situation where “the total is still high, but the margin has shown fatigue.” The Qian Hexagram in the I Ching progresses from “hidden dragon, do not act,” to “dragon appearing in the field,” all the way to “flying dragon in the sky,” but its final line is “an arrogant dragon will have cause to repent.” Is this not a portrayal of the S-curve? The Xiang Zhuan explains it even more directly: “Fullness cannot last long.”

So, if the product I relied on for growth has reached the peak of its S-curve, with marginal benefit at zero, facing the situation of “an arrogant dragon will have cause to repent,” how should I respond? Does it simply fade away? Far from it. You can fully develop a new product.

In 1994, British management thinker Charles Handy proposed the concept of “the second curve” [1]: to break through the ceiling of an old curve, a new S-curve must be drawn before the first curve reaches its peak. The continuous succession of S-curves can lead to sustained growth.

It is crucial to note the timing of activation—the second curve must be launched before the old curve reaches its peak. This is because pioneering a new curve requires funds, talent, and morale, all of which are only abundant when the old curve is still in its ascending phase. Once the new returns from the old curve are no longer sufficient to support new explorations, and funds, talent, and morale begin to dwindle, it will be too late to seek transformation.

Apple’s decision to release the iPhone while the iPod was still a bestseller, actively allowing a new product to cannibalize its star business, is a classic example of developing a second curve. I look forward to seeing what Apple’s next S-curve will be.

The Secret to Sustained Growth: Superposition of S-Curves #

Over the past few decades, humanity has witnessed two miraculous sustained growths.

One is Moore’s Law, which states that the number of transistors that can be accommodated on an integrated circuit roughly doubles every two years [2].

From Intel’s first microprocessor, the Intel 4004, with 2,300 transistors in 1971, to NVIDIA’s Blackwell GPU with 208 billion transistors in 2024 [3][4], the number of transistors has grown by approximately ninety million times over fifty years.

The second is the “Scaling Law” in the field of AI in recent years. In 2020, OpenAI researcher Jared Kaplan and others discovered that the cross-entropy loss (i.e., the accuracy of predicting the next word) of large language models continuously decreases with increases in parameters, data, and training compute [5]. This essentially means that the more you invest in a model, the smarter it becomes.

Fortunately, the scaling law remains effective to this day. Although there are occasional conjectures about whether it has reached its limit, it still works. It is precisely because of the effectiveness of the scaling law that we can expect AI to become increasingly intelligent, eventually achieving AGI (Artificial General Intelligence) and ASI (Artificial Superintelligence).

So, why can chips and AI continue to grow without falling into diminishing marginal benefit? The answer lies in the fact that the macro growth you observe is actually composed of multiple overlaid S-curves.

Moore’s Law is not the evolution of a single chip technology from beginning to end, but rather has undergone multiple technological iterations: from planar transistors to FinFET, and then to GAA; when one technological path approaches its limit, another takes over [6].

In the AI field, the scaling law is also continuously iterating. Initially, researchers mainly focused on increasing pre-training investment to make models larger and larger; subsequently, they found that many models “had huge brains but read too few books,” so they readjusted the ratio of model size to training data [7]. Later, investment shifted to post-training, reinforcement learning, and inference computation: not only making models read more data but also teaching them how to solve problems and making them think more deeply before answering; o1 and DeepSeek-R1, among others, belong to this new curve [8][9].

This reveals a truth: there’s no such thing as a “one-size-fits-all” solution in the world that will benefit you indefinitely. You must continuously invent new growth points to sustain an endeavor long-term.

Marginal Analysis and Strategic Transformation #

Marginal analysis, S-curves, and second curves collectively reveal that even the most excellent strategy must be adjusted at a specific time. Changing an old strategy is not because it is inherently bad, but simply because it no longer adapts to a new situation.

Taking China’s economy as an example, the rapid growth over the past four decades was largely driven by an investment-led curve: building roads, factories, and buildings—capital investment, followed by growth. In an era of severe infrastructure shortages, many projects had extremely high marginal benefits: for instance, the first expressway or the first modern factory could immediately unleash immense productivity.

However, investment will eventually encounter diminishing marginal returns. When Chong-En Bai and Qiong Zhang estimated China’s return on capital using various metrics in 2014, they found a clear downward trend after 2008 [10]. Indeed, the construction of the first expressway between two cities could immediately boost local economic development; but when the expressway network between major cities is already well-developed, extending roads further into remote mountainous areas might not recoup the investment for decades.

However, because infrastructure investment can immediately translate into GDP and be seen as a government achievement, local governments eagerly pursue such investments. Little do they know, these investments have now evolved into a heavy debt burden. The World Bank’s 2026 China Economic Update precisely points out: public investment faces diminishing returns “in many areas” and spending should be reallocated to uses with higher returns [11].

For example, by channeling funds towards public consumption. By 2025, China’s household consumption expenditure as a percentage of GDP is still only 40.0% [12]. To understand the low level of this figure, refer to the World Bank’s latest comparable 2024 data: China is at 40.0%, while the average for OECD member countries reaches 60.1%, a full twenty percentage points difference [13]. Chinese public consumption demand is robust, and driving growth through consumption currently offers significant marginal benefits.

Marginal Analysis of Personal Effort #

Diminishing marginal benefit is like a curse; any strategy used to its extreme will inevitably encounter it. Therefore, we must remain sensitive to this, so that we can take timely action to find new methods before the S-curve peaks.

For instance, the universally praised “effort” is worth subjecting to marginal analysis.

A student studies late into the night; the improvement from one more hour of study has long become negligible, while the negative impacts are increasingly growing. What is the point of continuing to study at this time? The choice with the highest marginal benefit for the next hour is to sleep.

When encountering a bottleneck, most people’s first reaction is to “increase effort”—unaware that what is needed at this time is actually to “change tactics.”

For instance, if your home is always messy, your first reaction might be to clean for one more hour on the weekend. This might initially be effective, but once items accumulate to a certain degree, no matter how diligent you are, you’ll just be moving clutter from one corner to another. Diminishing returns are reminding you: the bottleneck is no longer about cleaning speed but rather too many items. The correct approach is not daily tidying, but changing the variables—buy less, discard, redesign storage solutions.

You must shift your efforts from areas where marginal benefit approaches zero to areas where marginal benefit is increasing.

Beyond “Modest Satisfaction”: Embracing the Second Curve #

In recent years, a folk wisdom has become popular, called “modest satisfaction is better than perfect fulfillment,” meaning that if things develop to absolute perfection, they will, like the moon waning after full, turn to their opposite. This is essentially a sentiment about diminishing marginal benefit and the S-curve.

Since brilliance is inevitably followed by dimness, why shouldn’t I just remain in the state before the peak of the S-curve? Why not simply stick to “modest satisfaction” instead of pursuing “absolute fulfillment”? Some would say to retire successfully, not seeking great wealth or status, even quoting remarks like “those with abundance behind them forget to withdraw, those with no road ahead wish to turn back.”

Having understood the marginal analysis in this article, you will immediately see the superficiality of this view.

Elon Musk is already the richest person in the world, yet he continues to achieve great things one after another. Would you say he embodies “modest satisfaction” or “absolute fulfillment”? “Modest satisfaction is better than perfect fulfillment” is always a post-factum lament—you only regretfully cherish the state of “modest satisfaction” after experiencing the decline following “absolute fulfillment,” but what good is that then?

The true solution is not to reject “absolute fulfillment,” but to initiate a second curve before “absolute fulfillment” is reached.

Self-improvement and “an arrogant dragon will have cause to repent” are not contradictory.

【Poetic Evidence】

Spring is deep, yet blossoms not fully fallen, Deep in thick shade, autumn’s notes discerned. East wind no longer favors the flowering tree, But bestows its green shade upon new shoots.

Notes

[1] Handy, Charles B. The Empty Raincoat: Making Sense of the Future. London: Hutchinson, 1994.

[2] Intel Corporation. “Moore’s Law: Fun Facts.” Accessed July 20, 2026. https://www.intel.com/content/www/us/en/history/history-moores-law-fun-facts-factsheet.html.

[3] Intel Corporation. “Intel Marks 30th Anniversary of the Microprocessor.” November 15, 2001. https://www.intel.com/pressroom/archive/releases/2001/20011115corp_a.htm.

[4] NVIDIA. “NVIDIA Blackwell Platform Arrives to Power a New Era of Computing.” March 18, 2024. https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing.

[5] Kaplan, Jared, Sam McCandlish, Tom Henighan, et al. “Scaling Laws for Neural Language Models.” arXiv:2001.08361, 2020.

[6] Samsung Electronics. “From GAA to 3D Stacked FET: Expanding the Transistor into the Third Dimension.” Accessed July 20, 2026. https://semiconductor.samsung.com/news-events/tech-blog/from-gaa-to-3d-stacked-fet-expanding-the-transistor-into-the-third-dimension/.

[7] Hoffmann, Jordan, Sebastian Borgeaud, Arthur Mensch, et al. “An Empirical Analysis of Compute-Optimal Large Language Model Training.” Advances in Neural Information Processing Systems 35 (2022): 30016–30030.

[8] OpenAI. “Learning to Reason with LLMs.” September 12, 2024. https://openai.com/index/learning-to-reason-with-llms/.

[9] Guo, Daya, Dejian Yang, Haowei Zhang, et al. “DeepSeek-R1 Incentivizes Reasoning in LLMs through Reinforcement Learning.” Nature 645 (2025): 633–638.

[10] Bai, Chong-En, and Zhang, Qiong. “China’s Return on Capital and Its Influencing Factors Analysis.” World Economy, No. 10, 2014, pp. 3–30.

[11] World Bank. China Economic Update, July 2026: Rebalancing Growth. Washington, DC: World Bank, 2026. https://thedocs.worldbank.org/en/doc/0cb2fc6dd88d4db3816dc0433b5cb49b-0070012026/china-economic-update-july-2026.

[12] National Bureau of Statistics. Expanding and Upgrading the Consumer Market, Renewing and Improving Commercial Circulation—Series Report No. 12 on Economic and Social Development Achievements during the 14th Five-Year Plan Period. June 4, 2026. https://www.stats.gov.cn/zt_18555/ztfx/sswjjshfzcjbg/202606/t20260601_1963844.html.

[13] World Bank. “Households and NPISHs Final Consumption Expenditure (% of GDP).” World Development Indicators. Indicator NE.CON.PRVT.ZS. Accessed July 20, 2026. https://data.worldbank.org/indicator/NE.CON.PRVT.ZS?locations=CN-OE.