Exclusive Interview with Alexandr Wang, Founder of Scale AI: The Underlying Logic of AI Computing Power, High-Quality Data Annotation, and the Evolution of Human Work

From dropping out at 19 to a $29 billion AI unicorn: Alexandr Wang reveals the wealth, power, and future behind artificial intelligence

As the world competes for the smartest AI, some have built hundreds of billions of dollars in business empires by providing data for AI. The experience of Scale AI founder Alexandr Wang is not just a Silicon Valley startup story; it reveals the underlying logic of the redistribution of global capital, national competition, corporate organization, and the value of human labor in the next decade.

Artificial intelligence is undergoing a historic turning point.

In the past, humans relied on computers to improve work efficiency. Today, computers are beginning to understand language, write code, analyze information, solve complex problems, and even execute tasks autonomously.

This means that AI is gradually evolving from an auxiliary tool to a production factor capable of undertaking more and more cognitive labor.

Historically, the steam engine expanded human capacity to use mechanical power, electricity restructured industrial production systems, and the internet reduced global information dissemination and transaction costs.

Now, artificial intelligence is attempting to lower the cost of another important resource: knowledge labor.

As the marginal cost of acquiring some knowledge, analyzing information, and executing tasks continues to decline, the entire economic system's production methods, competitive barriers, and profit distribution may change.

In this round of transformation, one of the most noteworthy figures to study is Alexandr Wang, founder of Scale AI.

In 2016, he left the Massachusetts Institute of Technology (MIT) at the age of 19 to start Scale AI.

By 2024, Scale AI was valued at approximately $14 billion.

In June 2025, Meta announced an investment of about $14.3 billion in Scale AI, acquiring about 49% equity, with the company's valuation exceeding $29 billion. Wang subsequently joined Meta to help lead its AI strategy.

The significance of this deal goes far beyond creating a young billionaire.

Because Scale AI is neither the world's largest GPU manufacturer nor the most famous developer of general large language models, nor does it have a consumer social platform as large as Meta's.

It initially entered a seemingly insignificant field: data annotation.

Why could a company that helps AI process data achieve a valuation of hundreds of billions of dollars?

Why was Meta willing to invest such a huge amount of money?

As AI becomes smarter, will it make more people wealthy or concentrate wealth further?

The answers to these questions need to be sought from industrial economics, financial capital, technological revolutions, and the historical evolution of human society.

1. The true entrepreneurial wisdom of Alexandr Wang: not predicting winners, but finding what all winners need

Alexandr Wang grew up in Los Alamos, New Mexico.

This place is famous for the Los Alamos National Laboratory and was an important research center for the Manhattan Project during World War II.

Both of his parents are physicists.

Growing up in such an environment, mathematics, scientific research, and complex problem-solving were not distant concepts but part of daily life.

Wang participated in math competitions from a young age and later studied computer science at MIT.

In March 2016, AlphaGo defeated South Korean Go champion Lee Sedol.

This event profoundly changed the global tech industry's understanding of artificial intelligence.

Go has long been considered an activity that heavily relies on human intuition, experience, and complex strategic judgment. AlphaGo's victory demonstrated to the world that deep learning could surpass top human experts in extremely complex tasks.

Many tech entrepreneurs began searching for new AI applications.

Some studied robotics, others developed autonomous driving, and some focused on creating more advanced neural network algorithms.

But Wang noticed another issue.

No matter how advanced the AI model is, it needs effectively processed data for training.

For example, an autonomous driving company can collect vast amounts of road video through car cameras.

However, having the video does not mean the car can understand it.

The system must know which are pedestrians, traffic signs, bicycles, vehicles, and which road behaviors may pose dangers.

This information needs to be annotated, classified, verified, and quality-managed to become valuable machine learning data.

Wang saw this often-overlooked foundational link.

In 2016, he left MIT and co-founded Scale AI with co-founder Lucy Guo.

He did not directly participate in the crowded competition for AI end products but chose to sell critical production services to the entire AI industry.

This business model has repeatedly appeared in human economic history.

In 1848, the discovery of gold in California sparked the American Gold Rush.

Countless prospectors hoped to find gold and become rich overnight.

But prospecting is a highly uncertain activity. The location of deposits, mining costs, the number of competitors, and gold prices can all determine the final returns.

Meanwhile, businesses providing transportation, tools, finance, and living supplies also found commercial opportunities during the Gold Rush.

Levi Strauss's later clothing company is one representative case in the history of Western commercial development.

The economic logic here is not simply that "the Gold Rush should sell shovels."

The deeper meaning is:

When the ultimate winner of a new industry is still uncertain, providing common infrastructure for many potential winners can reduce dependence on a single terminal winner.

Scale AI's early advantage was just that.

No matter which autonomous driving company wins, they all need high-quality data.

No matter which large model company makes breakthroughs, they all need training, testing, and evaluation systems.

This horizontal service capability created market opportunities for Scale AI.

But there is another important question:

Infrastructure companies do not inherently possess permanent high profits.

As a service gradually standardizes and competitors increase, prices may decline.

Therefore, what Scale AI truly deserves to be studied for is not just discovering the demand for data annotation but continuously seeking higher value and stronger competitive barriers in the ever-changing AI technology landscape.

This is the key to whether a tech company can grow sustainably in the long term.

2. Why does AI need humans? Because data is not knowledge, and knowledge does not equal judgment

People often refer to data as the "new oil" of the artificial intelligence era.

This metaphor has some truth but is not complete.

Oil must be extracted, transported, and refined to become gasoline, diesel, and other industrial products.

Raw data also needs to be filtered, organized, cleaned, and processed to become effective model training material.

But there is a fundamental difference between data and oil:

The energy of oil mainly depends on its physical and chemical properties, while the value of data highly relies on information content, task requirements, accuracy, and usage methods.

One million pieces of repetitive, erroneous internet information may not be more valuable than a thousand high-quality medical expert reasoning cases.

The development of artificial intelligence is also changing the structure of the data industry.

1. First stage: enabling machines to recognize the world

Early machine learning data annotation primarily served computer vision.

For example, recognizing cats, dogs, pedestrians, cars, or traffic lights in an image.

This type of work often requires significant human involvement.

It helps machines establish basic recognition capabilities of the real world.

2. Second stage: enabling machines to understand human language and preferences

With the development of large language models, models not only need to understand text but also learn how to provide helpful, accurate, and task-relevant answers.

For example, facing a financial question:

"Why does a company's net profit increase while its operating cash flow continues to decline?"

A poor-quality answer might simply say:

"Because the company's financial situation has problems."

A higher-quality answer would explain:

Net profit follows the accrual basis of accounting, while operating cash flow reflects actual cash receipts and payments.

If a company's accounts receivable grow significantly faster than revenue growth, or if inventory continues to pile up, even if reported profits continue to rise, operating cash flow may deteriorate.

Further analysis should also examine the cash flow statement, changes in working capital, revenue recognition policies, and customer payment situations.

The value of the two answers shows a clear difference.

To train models to provide better answers, developers need high-quality feedback and evaluation mechanisms.

RLHF, or Reinforcement Learning from Human Feedback, is one important technical method.

Humans can compare different answers from the model and assess their accuracy, logic, helpfulness, and risks.

This feedback can be used to improve model behavior.

But RLHF is not the only way to train modern AI after training. Supervised fine-tuning, reward models, verifiable reward reinforcement learning, and synthetic data methods are also jointly promoting model capability development.

3. Third stage: enabling AI to learn professional judgment

As AI models become more powerful, the value of simple data annotation begins to change.

More complex and specialized tasks become important.

For example:

Medical experts can evaluate whether AI has missed dangerous symptoms.

Programmers can identify security issues in AI-generated code.

Mathematicians can check whether complex proofs hold.

Financial analysts can identify erroneous assumptions in financial reasoning.

Legal professionals can check whether models incorrectly cite legal rules or cases.

This means that the AI data industry is evolving from labor-intensive information processing to professional knowledge services.

Scale AI's Outlier platform is a representative of this model.

In an interview provided, Wang stated that Outlier's knowledge contributors earned about $500 million in total compensation over the previous year, covering over 9,000 towns in the United States.

These two figures belong to historical information at the time of the interview and should not be directly taken as the latest statistics for 2026.

But they reveal an important trend:

Artificial intelligence has not immediately eliminated all knowledge work; instead, it has created new demands for knowledge production, evaluation, and model training.

Of course, this does not mean that all participants have stable incomes.

Platform-based work may face fluctuations in task supply, changes in compensation, quality audits, and unstable job security.

The real long-term opportunity lies not just in providing manual labor for AI but in possessing irreplaceable professional capabilities.

In the future, the unit price of ordinary annotation work may decline.

But high-difficulty medical judgments, professional code reviews, scientific reasoning, and complex system evaluations will still require high-level knowledge.

It is also worth noting that even if AI models can automatically generate vast amounts of training data, they cannot indefinitely rely on unverified outputs to improve their capabilities.

Training data needs to provide effective new information, and evaluation data especially needs independence.

Otherwise, models may continuously repeat their own mistakes.

In this sense, one of the important roles of human experts will shift from directly providing answers to designing questions, verifying answers, and establishing reliable evaluation standards.

3. What truly controls the AI industry is not just algorithms, but five key production factors

Traditional views suggest that AI primarily relies on three factors for development:

Computing power, data, and algorithms.

This division is useful for understanding model training, but if viewed from the perspectives of global finance, national competition, and industrial chains, the framework needs to be further expanded.

The complete production system of the AI industry includes at least five key components:

Energy and capital, chips and computing infrastructure, data and expertise, algorithms and models, and commercial applications and customer channels.

Each link has the potential to generate enormous commercial value.

But their profit distributions are not the same.

1. Computing power: What is Nvidia's true competitive barrier?

Many people believe that its advantage lies solely in the ability to produce high-performance GPUs.

In reality, GPU performance is only part of the competitiveness.

The CUDA software ecosystem, computing tools, developer community, network interconnect technology, and system integration capabilities that Nvidia has accumulated over the long term together form a strong commercial barrier.

Suppose an AI company has already invested hundreds of millions of dollars in developing model training systems around a specific GPU architecture.

If it wants to switch to another computing platform, it may need to re-optimize code, adjust system architecture, test compatibility, and bear the risk of training interruptions.

This is what economists refer to as switching costs.

When the cost of migrating platforms is high, even if competitors offer lower-priced hardware, they may not be able to quickly acquire customers.

Nvidia's business advantage is the result of the mutual reinforcement of hardware capabilities and software ecosystems.

However, this advantage is not unshakeable.

Large cloud computing companies developing their own chips, new software frameworks reducing hardware dependency, and continuously improving model training efficiency could all change the long-term competitive landscape.

2. Manufacturing: Why does TSMC have strategic value?

Global advanced chip manufacturing heavily relies on a complex semiconductor supply chain.

TSMC's important position in advanced wafer manufacturing makes it a crucial infrastructure company for the global AI industry.

However, advanced chips cannot be produced by just one company.

Chip design, advanced lithography equipment, semiconductor materials, wafer manufacturing, advanced packaging, and high-speed network technology are distributed across multiple countries and regions.

For example, the United States has significant advantages in high-end chip design and related software ecosystems.

The Netherlands' ASML provides critical advanced lithography equipment.

Japan holds an important position in certain semiconductor materials and equipment fields.

Taiwan's TSMC possesses world-leading advanced wafer manufacturing capabilities.

South Korea has significant competitiveness in areas such as memory chips.

This means that the global AI industry actually relies on a high-precision industrial chain of multinational collaboration.

Any major disruption in a key link could affect the entire industry chain.

From a financial perspective, this represents supply chain concentration risk.

From a national security perspective, it makes advanced semiconductor manufacturing a strategically significant industry.

3. Energy: The more advanced AI becomes, the more it cannot escape the physical world

Software can spread globally through the internet, but running AI models requires real physical resources.

Large computing clusters rely on stable power supply, cooling systems, network facilities, data center buildings, and long-term capital investment.

The Colossus computing cluster built by Musk's xAI is a representative of this trend.

The expansion of the AI industry will not only stimulate semiconductor demand but may also change the markets for electricity, natural gas, transmission equipment, energy storage, industrial cooling, and data center construction.

This is significantly different from the software industry as commonly understood.

The marginal distribution cost of traditional internet applications can be very low.

But large-scale AI inference requires continuous consumption of computing resources.

Therefore, the business model of AI companies must not only focus on user growth but also measure the cost of each inference.

Suppose an AI service charges users $20 per month.

If the average cost of model inference and direct service per user is $15, the remaining margin for the company is very limited.

If, through technical optimization, the direct service cost is reduced to $3, the business model could fundamentally change.

This does not yet account for other expenses such as R&D, marketing, and management.

In AI business competition, reducing the unit cost of intelligent services is as important as improving model capabilities.

4. Data: Data scale does not equal competitive advantage

In the past, internet companies often relied on user scale to accumulate a large amount of behavioral data.

But in the AI era, it cannot be simply assumed that whoever has the most data will have the strongest model.

The reason is that different data contributes differently to model capabilities.

A large amount of low-quality, redundant information may provide limited additional training value.

In contrast, a professionally validated dataset that can reveal the weaknesses of a model may have a greater impact.

Therefore, the future commercial value of data will increasingly depend on:

Whether the data is scarce, whether it is obtained legally, whether it has high accuracy, whether it can be continuously updated, and whether it can truly improve performance on specific tasks.

For example, the clinical experience accumulated by medical institutions over the long term may be more suitable for certain medical AI tasks than ordinary internet articles.

The equipment failure records, production processes, and maintenance data owned by manufacturing companies may also become important resources for industrial AI.

However, when it involves personal privacy, trade secrets, or regulated information, the realization of data value must be premised on legal authorization, protective measures, and appropriate governance.

5. Commercial channels: Technological leadership does not necessarily mean the highest profits

Looking back at the history of the computer industry, one can find a recurring phenomenon:

Creating technology and obtaining the largest profits in the industry are often not the same thing.

In the personal computer era, hardware, operating systems, and application software companies controlled different value segments.

In the mobile internet era, smartphone manufacturers, operating system platforms, app stores, and internet service providers also formed different commercial positions.

The AI industry is likely to repeat this process.

Even if a company has a very advanced model, it may not be able to achieve the highest profits.

Companies with large customer bases, payment systems, corporate procurement channels, and mature business scenarios may have different types of advantages.

In the future, it is essential to focus not only on which companies have the strongest models.

But also on which companies can continuously convert model capabilities into user payments, corporate revenue, and free cash flow.

This involves the most fundamental distinction in financial markets:

Creating value does not equal capturing value, and revenue growth does not equal shareholder returns.

4. Why is Meta willing to invest $14.3 billion in Scale AI?

In June 2025, Meta announced a significant investment in Scale AI.

According to public information, Meta invested approximately $14.3 billion to acquire about 49% of Scale AI's equity, corresponding to a company valuation of about $29 billion.

At the same time, Alexandr Wang joined Meta to help drive the company's AI development.

This transaction is an important case for understanding capital competition in AI.

Because it involves technology, talent, supply chain, and strategic control.

First, Meta not only needs computing power but also organizational capability.

The development of large AI models is a highly complex system engineering task.

It requires not only researchers but also long-term collaboration between computing infrastructure, training data, model evaluation, engineering management, and product teams.

Even if a large tech company has ample funds, it does not mean it can immediately assemble a world-leading AI team.

Because funds can buy equipment but cannot directly purchase the established organizational synergy and technical experience.

Alexandr Wang has long operated Scale AI and has been exposed to significant demands in various AI research fields.

This gives him unique industry experience and organizational perspective.

Meta's investment in Scale AI and the introduction of its founder have clear strategic significance for talent and organizational capability building.

Second, top tech companies are using capital to buy competitive time.

The AI industry has very obvious technological competition and time pressure.

If a large platform falls behind in a generation of key technologies, it may face risks such as user loss, weakened ecosystems, and declining future profits.

Therefore, an investment cannot be evaluated solely based on the profits of the invested company.

It is also necessary to analyze whether it helps the investor enhance its core competitiveness.

For example, suppose a technology investment itself only creates limited direct financial returns but significantly improves a large platform's user retention, advertising efficiency, or product competitiveness.

Then, its overall value to strategic investors may exceed independent financial returns.

This is precisely the important distinction between strategic investment and traditional financial investment.

Financial investment primarily pursues capital returns.

Strategic investment may also pursue technological capabilities, market positions, supply chain security, or competitive advantages.

But strategic value is not a reason to pay any price.

If it ultimately cannot produce actual business improvements, no grand strategic narrative can replace investment returns.

Third, Meta's investment also brings new risks to Scale AI.

Scale AI originally served multiple large AI companies.

But Meta itself is also a competitor of these companies.

When a previously relatively independent supplier establishes a deep capital relationship with a large customer, other customers may have concerns.

For example:

Will their own R&D plans remain confidential?

Will key data be adequately protected?

Will the supplier prioritize serving a particular strategic shareholder in the future?

These concerns, even if they do not evolve into actual information leaks, may affect customer choices.

This reveals an important business rule for platform-type infrastructure companies:

Neutrality itself may also be an economically valuable asset.

In 2026, Scale AI has entered a new stage of development.

The company appointed Jason Droege as interim CEO in June 2025 and announced the appointment of Francis deSouza as the new CEO in July 2026, effective August 10.

Scale AI is no longer positioning itself solely as a data labeling service provider but is placing greater emphasis on model evaluation, enterprise applications, government systems, and reliable AI infrastructure.

According to information disclosed by the company in January 2026, its new business scale exceeded $1 billion in 2025, with corporate clients including institutions like Mayo Clinic, BP, and Allianz.

It is important to distinguish that the new business contract amount does not equal the actual revenue recognized that year, let alone profits.

This change indicates:

Scale AI is attempting to evolve from a supplier of model training materials to a technology service company that helps large organizations deploy and manage AI systems.

Whether this step can succeed will determine whether its future commercial value can continue to grow.

5. From a financial perspective, where is the biggest opportunity in the AI era?

One mistake investors often make is equating technological revolutions directly with investment returns.

For example:

Artificial intelligence will undoubtedly become increasingly important in the future.

So all AI-related stocks will rise.

This reasoning is flawed.

Historically, many great technological revolutions have indeed created enormous economic value, but investors involved do not necessarily achieve ideal returns.

1. The railroad revolution: An industry that changed the world could also lead to shareholder losses

In the 19th century, railroads completely transformed transportation, trade, and regional economies in the United States.

Railroads significantly reduced the costs of moving goods and people across regions.

They facilitated the formation of a national market and promoted the development of steel, coal, finance, and cities.

However, railroad construction required massive capital.

The industry experienced over-expansion, debt burdens, and fierce competition at certain stages.

Many railroad companies went through bankruptcy and restructuring.

The social value created by railroads is enormous, but not all railroad company shareholders received equal returns.

This illustrates that:

The more important a technology is to society, it does not mean that the stocks of related companies are worth buying at any price.

2. The internet bubble: Correctly predicting the future does not equal correct investment

In the late 1990s, the internet rapidly developed.

At that time, people believed that the internet would change shopping, advertising, finance, communication, and entertainment.

This judgment was fundamentally correct.

However, many internet companies lacked sustainable profit models.

Some companies relied on financing to maintain operations rather than achieving healthy growth through customer payments.

Around 2000, the internet bubble burst, leading to the collapse of many companies and significant losses for investors.

Meanwhile, the internet continued to change the global economy thereafter.

The most important lesson here is:

Investors may correctly predict the future of an industry, but ultimately incur losses due to overpaying for valuations, choosing the wrong companies, or underestimating competitive costs.

The AI industry faces the same issue.

3. What five financial metrics should AI investments analyze?

First, revenue quality.

Does an AI company's revenue come from recurring subscriptions or one-time projects?

Do customers renew their contracts?

Is revenue highly dependent on a few customers?

Second, gross margin.

What are the direct computing costs required to provide AI services?

Do model inference costs continue to rise with increased user usage?

Can the company reduce costs through technological optimization?

Third, capital expenditure.

How many GPUs does the company need to continuously purchase to maintain growth, and how many data centers must it build?

Is there a possibility that these assets will depreciate rapidly due to technological upgrades in the future?

Fourth, free cash flow.

How much cash does the company actually generate that can be used for debt repayment, buybacks, dividends, or reinvestment?

Profit growth does not necessarily mean that free cash flow will improve.

Fifth, competitive barriers.

Why can't customers easily switch suppliers?

Does the company possess unique data, patents, user networks, industry certifications, or is it highly dependent on third-party models?

If competitors can quickly replicate products, then revenue growth may not translate into long-term high profits.

4. The profits in the AI industry may continuously migrate.

The AI industry is not a static market.

In the early stages of infrastructure development, companies that possess scarce chips, computing resources, and equipment may gain strong bargaining power.

As computing power supply expands and model efficiency improves, some profits may migrate to model services and application layers.

As general model capabilities gradually become commoditized, companies that master customer relationships, industry workflows, and exclusive data may gain greater advantages.

However, this is not a certain linear process.

Infrastructure demand may continue to grow, and some leading chip companies may also maintain high profits through technology and ecosystems.

Therefore, research on AI investments must continuously answer:

What is the most scarce resource currently?

How long can this scarcity last?

Do market prices fully reflect this advantage?

Truly valuable investment research is not about discovering a popular trend that everyone knows, but about understanding whether the market's expectations for a company's future cash flow are reasonable.

6. US-China AI competition: Who holds the decision-making power over the next generation of technology and national strength?

AI competition is no longer just a commercial competition between companies.

It also involves national security, industrial foundations, financial capital, information systems, and international influence.

However, the US-China AI competition cannot be simply understood as one country having more data or the other having more advanced chips.

What determines long-term competitiveness is whether multiple systems can effectively collaborate.

1. America's biggest advantage is the combination of technological innovation and capital markets.

The US has significant technology companies, research universities, venture capital institutions, and capital markets.

These organizations can collectively support high-risk, long-cycle technology research and development.

According to Stanford University's "2026 AI Index Report," private AI investment in the US is expected to reach approximately $285.9 billion by 2025, while China's is about $12.4 billion.

However, it should be noted that China's AI investment also includes various financing channels such as government-guided funds, so the difference in private capital scale cannot be directly interpreted as a difference in total AI investment between the two countries.

This set of data truly reflects that:

The US private capital market has a strong capacity for risk-bearing and capital concentration.

It allows some technology companies that are not yet profitable but have enormous growth potential to obtain large-scale research and development resources through financing.

America's advantage comes not only from the amount of capital but also from the connections between capital, talent, technology, enterprise management, and commercialization capabilities.

2. China's competitiveness is not just data, but also engineering capabilities and manufacturing systems.

China has a large pool of engineering talent and a complex, complete manufacturing ecosystem.

Its internet market has long experienced intense product competition, and many companies have strong experience in cost control, rapid iteration, and large-scale operations.

The development of companies like DeepSeek shows that AI model competition does not solely depend on who can acquire the most GPUs.

Model architecture, algorithm optimization, system engineering, and resource utilization efficiency can also change the competitive landscape.

The Stanford 2026 AI Index points out that as of March 2026, the gap between top models in China and the US has significantly narrowed based on the leading model evaluation metrics they adopted.

However, this does not mean that both sides are completely identical in all aspects.

AI model performance, advanced chip manufacturing, software ecosystems, industrial robots, international markets, research papers, and the flow of high-end talent need to be evaluated separately.

3. Chip export controls are a strategic game surrounding production capacity.

The US has implemented export restrictions on certain advanced computing chips and related technologies, one of the key goals being to limit access to specific advanced AI computing capabilities.

The strategic logic behind this approach is not new.

Historically, industrial powers have often attempted to maintain competitive advantages by controlling key production technologies and equipment.

However, the actual effects of technological blockades are often very complex.

On one hand, it may increase the costs for the restricted party to obtain advanced technologies.

On the other hand, it may also stimulate domestic alternative research and development, resource reallocation, and improvements in computing efficiency.

Therefore, evaluating export control policies cannot solely look at the number of chips exported in a given year.

It is also necessary to observe long-term technological gaps, industrial substitution progress, supply chain shifts, and innovation costs.

4. AI is becoming an important component of national sovereignty.

Traditional national sovereignty usually involves territory, law, currency, and defense.

The digital age adds an important dimension:

Data infrastructure and information systems.

In the future, if a country heavily relies on external AI systems for critical public services, financial institutions, communication infrastructure, and government decision-making, it will need to consider risks such as technology supply interruptions, data security, system audits, and governance control.

As a result, more and more countries are beginning to discuss Sovereign AI.

Its core is not that every country must manufacture all chips from scratch or independently develop world-leading large models.

Rather, it is to ensure that critical systems have sufficient control, reliability, auditability, and operational continuity.

This also explains why AI infrastructure has begun to enter national strategic discussions.

7. How does AI affect history, culture, and politics? The future control of information deserves high vigilance.

Artificial intelligence has an important distinction from traditional search engines.

Search engines typically present multiple web pages, allowing users to compare different information sources.

Generative AI, on the other hand, tends to directly organize a complete answer.

This improves the efficiency of information retrieval but also brings new problems:

Users may only see the final conclusion without understanding the process of information filtering and answer formation.

If a large number of users rely on a few AI systems to understand historical, political, financial, and social issues, then the data sources, content policies, and fact-checking mechanisms of the models will have significant public impact.

Different countries and companies' AI models may be influenced by different legal environments, training data, system rules, and security policies.

In the face of certain sensitive historical events or political issues, some models may refuse to answer, some may provide restrictive answers, and some may offer more detailed information.

However, it must be emphasized:

Willingness to answer does not mean the answer is necessarily correct.

Refusal to answer does not necessarily mean the model lacks corresponding knowledge.

What is truly important is whether the information can be independently verified.

For example, studying a controversial historical event should not rely solely on the narrative of a single AI model.

A more reliable method is to cross-check historical archives, contemporary news, academic research, and original materials from different perspectives.

What is most concerning in the AI era is not just the increase in the amount of false information.

But rather, that false information may be disseminated in a highly fluent, structurally complete, and seemingly professional manner.

Why will information warfare become more complex?

Generative AI lowers the cost of mass-producing text, images, audio, and video.

This may make the creation of false identities, automated propaganda, and misleading content easier.

At the same time, AI can also help researchers, media, and security agencies identify abnormal dissemination behaviors, verify sources, and analyze data.

Therefore, AI is not inherently only for information control; it can also be used to improve transparency.

What determines the outcome is how the technology is deployed, and whether there are independent review, accountability mechanisms, and verification of information sources.

For the media industry, this means that future competitiveness will not be about who can produce the most news.

But rather, who can consistently provide accurate, timely, transparent, and verifiable information.

As content production costs tend to decline, credibility may become an increasingly scarce media asset.

8. AI and cyber warfare: Future national security competition will increasingly rely on machine speed.

Modern nations' financial systems, telecommunications networks, energy facilities, and military communications are highly dependent on digital infrastructure.

This makes cybersecurity an important component of national security.

The case of Salt Typhoon mentioned in the interview is worth noting.

US agencies have publicly disclosed that cyber attack activities linked to China have targeted telecommunications systems for infiltration.

China has rebutted some of the US's cybersecurity allegations.

Such incidents indicate that modern international competition occurs not only in traditional military domains but also in communication, networks, and information systems.

The development of AI may further change the landscape of cyber offense and defense.

In the past, large-scale vulnerability analysis, malware detection, and abnormal traffic investigation often required significant human involvement.

In the future, AI may improve the efficiency of code review, risk identification, threat analysis, and security response.

However, attackers may also use AI to find weak points and expand automated attack capabilities.

This creates a new security competition:

Defenders hope to detect problems earlier, while attackers seek to find opportunities faster.

However, it cannot be simply stated that AI has fully surpassed top human security experts.

The reliability of different tasks varies greatly, and real network environments still have complex contexts and risks.

The most likely trend is that human security experts will be responsible for strategy formulation and accountability, while AI systems will undertake more verifiable and automatable analytical tasks.

This will make cybersecurity capabilities increasingly dependent on the level of automation, system reliability, and human oversight mechanisms.

9. Will AI cause human unemployment, or will it give everyone their own digital employees?

This is the question that ordinary people are most concerned about.

Alexandr Wang proposed a future work model worth considering:

Employees may shift from personally completing all tasks to managing multiple specialized AI agents.

This perspective has significant economic implications.

However, it must be avoided to interpret it as all professions will automatically upgrade.

The impact of AI on different professions depends on specific tasks, not just job titles.

1. Why does AI first impact certain white-collar jobs?

In the past, the value of many knowledge-based professions was built on information processing capabilities.

For example:

Collecting data, drafting reports, organizing documents, translating texts, analyzing contracts, creating presentations, and writing basic code.

These tasks often share three common characteristics:

First, the input information can be digitized.

Second, the workflow has certain rules.

Third, the results can be checked and evaluated to some extent.

These characteristics make them more suitable for AI automation.

For example, a market analyst used to take three days to collect competitor information, organize product details, and create a preliminary report.

Now, they can use AI for data retrieval, preliminary classification, and report drafting.

But ultimately, they still need to verify information sources, assess whether market changes are significant, and make recommendations to management.

Therefore, labor value may gradually shift from mere information organization to judgment and decision-making.

2. Why do some blue-collar jobs have stronger protection in the short term?

Electricians, plumbers, maintenance workers, and construction workers often need to work in complex, non-standard physical environments.

For example, repairing a plumbing issue in an old building may require assessing the building's structure, the aging of the pipes, the actual installation space, and on-site safety risks.

These tasks not only require knowledge but also perception, operation, and on-site adaptability.

AI language models can explain maintenance methods, but that does not equate to having reliable physical operational capabilities.

Robotics technology is advancing and may change these professions in the future.

However, from digital software capabilities to stable physical execution capabilities, issues such as mechanical structure, sensors, costs, safety, and adaptation to complex environments still need to be addressed.

Therefore, the automation speed of some on-site service professions may be slower than that of purely digital tasks.

3. The most profound change brought by AI is the reduction of individual organizational costs.

In the past, if an entrepreneur wanted to complete a full business project, they might need to hire programmers, designers, market researchers, and copywriters.

Even if the project had not yet generated revenue, they might still have to bear high labor costs.

AI is changing this process.

Suppose an entrepreneur is preparing to develop an educational product.

They can use different AI tools for:

Market research, course design, product prototyping, code generation, promotional copywriting, and user feedback organization.

This does not mean that one person can replace an entire professional team without cost.

Complex products still require engineering reviews, real user testing, safety guarantees, and ongoing operations.

But the costs of some early exploratory work may indeed decrease.

Economically, this is equivalent to reducing some fixed costs of entrepreneurship and organizational production.

Certain tasks that previously required establishing a company can now be preliminarily validated by one person using AI tools.

In the future, the gap in individual competitiveness may increasingly depend on whether a person can effectively organize technological resources, rather than just how many specific tasks they can complete themselves.

4. Increased production efficiency does not necessarily mean individual income will increase.

There is an easily overlooked issue here.

Suppose a certain type of work becomes three times more efficient with AI.

Will employees' incomes also increase threefold?

Not necessarily.

If a large number of competitors can obtain the same AI capabilities, market competition may drive service prices down.

For example, in the past, producing a standard business report would cost $1,000.

After the widespread adoption of AI, many service providers can quickly generate similar reports.

If clients perceive no significant difference between different service providers, the market price of the reports may decrease.

In this case, AI has increased production efficiency but does not guarantee that service providers will earn higher profits.

Therefore, those who can truly gain more from the efficiency improvements brought by AI often need to possess other scarce resources:

Client relationships, brand reputation, unique data, professional judgment, industry experience, or intellectual property.

This is also a wealth distribution issue that needs to be understood in the AI era.

10. When everyone can use AI, what will be the truly scarce abilities?

In the early stages of AI adoption, being proficient in using the tools themselves was a competitive advantage.

But as technology matures, basic operational capabilities may quickly become widespread.

Just like today, being able to use search engines or office software is no longer a long-term competitive barrier.

In the future, the truly scarce abilities may be fivefold.

First, the ability to identify problems.

AI excels at providing answers based on tasks.

But determining which problems are worth solving often requires a deep understanding of the real world.

For example, developing a new financial software is not difficult.

The challenge lies in discovering what high-cost, unmet needs exist among a certain type of client.

Second, the ability to make independent judgments.

AI can generate multiple business plans.

But how to choose between different plans requires considering risks, funding, time, competitors, and long-term opportunity costs.

Decision-making not only means making choices but also entails bearing consequences.

Third, aesthetic and creative abilities.

When a large number of users utilize the same models and similar prompts, content may become homogenized.

Truly distinctive artistic styles, character settings, narrative abilities, and cultural insights still hold commercial value.

However, originality is not solely the domain of humans; AI can also produce novel combinations and valuable ideas.

What truly matters is whether the work can form a sustained audience recognition.

Fourth, the ability to build trust.

Technology can increase the speed of information production but cannot automatically endow a company with credibility.

In fields like finance, healthcare, law, and journalism, clients especially value reliability.

Therefore, transparent information sources, professional standards, and a long-term trustworthy brand record may become important competitive advantages.

Fifth, the ability to integrate resources.

The most competitive individuals in the future may not necessarily be the strongest in all professional fields.

Rather, they may be those who can clearly define goals, coordinate different AI systems, invite professionals to participate, and transform results into commercial value.

This is similar to the responsibilities of business managers.

The difference is that the resources being managed in the future may include human employees, software systems, AI agents, and external service platforms.

11. Five types of opportunities that entrepreneurs should focus on in the AI era.

Understanding the story of Scale AI ultimately cannot just stop at studying the success of one company.

More importantly, it is about identifying business opportunities suitable for one's own participation.

1. Industry-specific data and knowledge systems.

Many industries have a wealth of specialized knowledge that has not yet been effectively digitized.

For example, records of equipment failures in manufacturing, legal contract cases, financial research materials, and specialized maintenance processes.

Entrepreneurs can build searchable, verifiable, and continuously updated professional knowledge systems around legally obtained data.

The real commercial value comes from helping clients improve efficiency or reduce error costs, rather than just putting documents into a database.

2. Enterprise AI workflows.

Many enterprises do not need to develop world-leading large models themselves.

What they need is to make existing business operations run better.

For example, a logistics company may want to automatically organize transportation documents, identify abnormal orders, analyze reasons for delays, and generate customer notifications.

An insurance institution may want to improve the efficiency of claims document processing while maintaining manual review and compliance mechanisms.

This demand creates opportunities for vertical industry AI service providers.

However, enterprise-level products typically require longer sales cycles, as well as capabilities in data protection, system integration, and after-sales support.

3. AI model evaluation and safety services.

When AI enters the fields of finance, healthcare, and public services, companies cannot just focus on whether the model is smart.

They also need to know whether the model is reliable.

Therefore, evaluation services surrounding accuracy, hallucinations, biases, privacy, model safety, and professional compliance may form a sustained demand.

Scale AI's development towards model evaluation and reliable AI systems is a representation of this trend.

4. AI-driven content and intellectual property industries.

AI is lowering the production costs of images, animations, videos, and digital content.

This makes it easier for individual creators and small teams to attempt projects that previously required significant personnel involvement.

However, it is essential to distinguish between content production and commercial success.

Generating an animation at a low cost does not mean that the animation will attract an audience.

What truly determines long-term commercial value still includes character design, world-building, story quality, brand communication, fan relationships, and intellectual property management.

Disney's core competitiveness has never been just the ability to produce animations.

It also includes decades of accumulated character assets, narrative abilities, global distribution channels, licensing businesses, and emotional connections with consumers.

AI may lower content manufacturing costs but will not automatically create a brand that is loved by the market.

Therefore, for entrepreneurs looking to establish original IP, the most important factor is not how many videos they can generate daily, but whether they can build a character and story system with long-term recognition and commercial value.

5. Trustworthy information and professional media.

As AI makes content production increasingly easier, the internet may see more automated articles, videos, and comments.

An increase in the quantity of information does not mean that the quality of information improves simultaneously.

This environment may actually increase the value of trustworthy information services.

For example, a media outlet focused on AI, finance, and blockchain can build a long-term reliable content system around public documents, original interviews, regulatory information, and corporate announcements.

Its competitive advantage should not just be speed of publication.

It should also include factual accuracy, transparent sources, professional explanations, and ongoing tracking of complex events.

Further business models can include professional research subscriptions, industry databases, corporate information services, and compliance labeling business collaborations.

However, if paid content is confused with independent news, it may damage the media's most important asset: trust.

12. How should ordinary people face the AI revolution?

For ordinary people, the biggest risk may not be that AI directly replaces jobs.

Rather, it is continuing to rely on a skill that is rapidly becoming commoditized.

If a certain ability was once very scarce but can now be obtained at low cost by almost everyone through AI, the market price of that ability may decrease.

Therefore, individuals should build a new combination of abilities.

First, choose a professional field with long-term demand.

For example, finance, law, healthcare, industrial manufacturing, software engineering, education, or the content industry.

Second, learn how AI can improve workflows in that field.

Do not just learn the operation buttons of a certain software; understand the entire process from demand generation to result delivery.

Third, build your own professional knowledge and case accumulation.

AI can help summarize public information, but the experience, client understanding, and judgment ability formed from real projects over the long term are often harder to replicate.

Fourth, try to use AI to complete real projects.

Instead of learning a large number of unrelated tools, choose a specific goal:

Build a website, develop a small application, conduct business research, design an educational product, or set up an automated information processing workflow.

Finally, establish a habit of independent verification.

AI can quickly provide suggestions, but when it comes to funding, safety, healthcare, law, and important business decisions, it is essential to verify key facts and professional bases.

A 30-day experiment worth practicing.

You can choose a repetitive task you are most familiar with.

In the first week, record the complete workflow and the time required for each step.

In the second week, try using AI to handle tasks that can be digitized and easily verified.

In the third week, increase quality checks and compare the error rates of human and AI collaboration.

In the fourth week, calculate the time saved, tool costs, rework costs, and final quality changes.

For example, a task that originally required 10 hours per week.

After introducing AI, the initial processing only takes 3 hours, but human verification and modification still require 2 hours.

The actual savings are 5 hours, not the so-called tenfold efficiency advertised.

If the added tool costs and error risks exceed the saved costs, automation may even lack commercial value.

Therefore, truly meaningful AI productivity is not about generation speed but about the net benefits after quality and cost verification.

13. Will the AI revolution lead to greater wealth concentration?

From economic history, every major general technology revolution may simultaneously bring about wealth creation and wealth redistribution.

The industrial revolution increased social production capacity but was also accompanied by conflicts of interest between capital, land, and labor in its early stages.

The internet reduced the costs of information dissemination and transactions but also helped a few large platforms establish strong network effects.

Artificial intelligence is generating similar structural issues.

On one hand, AI tools lower the costs for individuals and small businesses to access certain knowledge work.

Individuals have the opportunity to complete tasks that previously required a team.

On the other hand, the most advanced AI models may still rely on massive computational investments, complex R&D teams, and large computational infrastructures.

This creates two simultaneous forces within the AI industry:

One force is the popularization of technology, lowering the barriers to entrepreneurship and production.

The other force is the concentration of infrastructure, giving large enterprises with capital, computing resources, and distribution channels an advantage.

Which force will ultimately dominate is still uncertain.

Future wealth distribution may be influenced by market competition, open-source ecosystems, technology costs, educational capabilities, government policies, and the bargaining power of workers.

For individuals, one cannot simply expect technological advancements to automatically increase income.

Attention should also be paid to whether one possesses assets that can share in the benefits of technological progress.

Here, assets are not necessarily just stocks.

They could also be equity in businesses, original intellectual property, professional brands, legally accumulated data assets, and business systems that can continuously serve customers.

Wages primarily come from selling labor time and skills.

Business equity may allow holders to share in the future residual profits of the company while also bearing the risk of operational failure.

Intellectual property can generate income through licensing, issuance, and commercial cooperation, but it also requires market demand, legal protection, and ongoing operation.

From the perspective of long-term wealth accumulation, one of the most important distinctions is whether individuals can gradually own productive assets with economic value beyond providing labor.

However, it must be emphasized that this does not encourage everyone to blindly start businesses, concentrate investments in AI stocks, or take on risks beyond their capabilities.

The value of any asset depends on real demand, cash flow, legal rights, competitive environment, and acquisition costs.

14. What is the most important variable in AI competition in the next decade?

If one only focuses on which AI company has launched the latest model, it is easy to be attracted by short-term news.

From a long-term perspective, it is more worthwhile to track six variables.

First, the unit cost of intelligent services.

How much computing resources and funds are needed to complete the same task?

A decrease in costs can expand the market but may also lower prices.

Second, the reliability of AI.

Can the model stably complete complex, continuous, multi-step tasks, rather than just performing well in demonstrations?

Third, the quality of professional data.

Which industries have specialized knowledge that is difficult to obtain publicly and can be legally used for model development?

Fourth, the true autonomous execution capability of AI Agents.

Can it continuously complete tasks under clear authority and supervision mechanisms and safely stop in case of anomalies?

Fifth, whether companies are willing to continue paying.

Technical demonstrations can attract attention, but the long-term payment of real customers is an important test of the business model.

Sixth, who can continue to make profits.

How is value distributed among chip companies, cloud computing platforms, model developers, application enterprises, and end customers?

These six variables together determine how AI technology will translate into economic growth and where the wealth generated by that growth will ultimately flow.

Conclusion: What is truly worth learning from Alexandr Wang's story is not that he dropped out at 19.

Looking back at Alexandr Wang's entrepreneurial journey, the most noticeable aspects are his age, education, and wealth.

He left MIT at 19.

Founded Scale AI.

At 24, he joined the ranks of young self-made billionaires.

The company's valuation grew from several billion dollars to over $29 billion.

These numbers are very compelling.

But they are not the most important part of this story.

What is truly worth learning is Wang's method of discovering business opportunities.

While the entire industry was chasing more advanced models, he chose to study the indispensable data production system behind the models.

When simple data labeling faced competition, Scale AI gradually entered expert feedback, model evaluation, and enterprise AI applications.

And when AI became the core technology for large tech companies competing for future competitiveness, Scale AI entered the center of global tech capital restructuring.

This shows that a successful entrepreneur does not necessarily have to create the most eye-catching end product.

Sometimes, identifying the key bottlenecks that exist across the entire industry and establishing a business system that can continuously solve this bottleneck can also create tremendous value.

But more importantly, bottlenecks will change.

Today, what is scarce is GPUs; in the future, it may be energy, professional data, model reliability, or AI systems that can truly complete complex tasks.

If companies rely solely on past advantages and cannot adapt to new technological structures, they may still be surpassed by competitors.

For countries, AI competition is not just about developing a leading model, but also about a comprehensive competition involving capital, talent, chips, energy, manufacturing, systems, and commercialization capabilities.

For enterprises, what truly matters is not whether they use AI, but whether they can create value that customers are willing to pay for through AI and establish a sustainable profit model.

For ordinary people, AI means that some old skills face price pressure, but it also means that the cost of acquiring new capabilities and establishing new businesses may decrease.

In the future, what determines an individual's economic status is not just how much knowledge they possess.

It also includes whether they can identify real needs, organize technology and talent, make reliable judgments, and own assets that can continuously create value.

Artificial intelligence may make acquiring knowledge and executing certain tasks increasingly cheaper, but discovering worthwhile problems to solve, building trust, taking responsibility, and creating real business value will still be the core of long-term competition.

In the future, what is most worth paying attention to is not just which machines are getting smarter, but who can turn that intelligence into real, reliable, and sustainable value.

Sources and Further Reading

1. Original interview with Alexandr Wang, This Past Weekend with Theo Von: youtube.com
2. Reuters, June 13, 2025, Meta invests in Scale AI and Alexandr Wang joins Meta: reuters.com
3. Scale AI, June 12, 2025, Announcement about Meta's investment and company management adjustments: scale.com
4. Scale AI, January 22, 2026, 2025 Business Review and 2026 Strategy: scale.com
5. Scale AI, July 30, 2026, Announcement of new CEO appointment: scale.com
6. Stanford University, 2026 AI Index Report, Global AI Technology, Capital, Economic and Social Impact: hai.stanford.edu

ABAB NEWS｜Opinion

This article is based on public interviews, corporate announcements, and research materials, containing industry analysis and trend judgments, and does not constitute any investment advice.