While the world remains fixated on the staggering fundraising rounds and valuations of AI model companies, a more fundamental question is emerging: just how big can this business ultimately become? A newly released research report from Sinolink Securities offers a disruptive perspective — stop measuring against traditional software market sizing. The true ceiling is hidden in human paychecks.
The report’s core logic is blunt and unsentimental: enterprises adopt AI not to chase trends, but driven by the fundamental imperative to replace human labor, boost efficiency, and compress costs. Therefore, AI’s real revenue ceiling is determined by the size of the human wage pool that AI can reprice. Sinolink defines this as the “AI-repriceable wage pool.”
This assessment stands in stark contrast to the current frenzy of global AI infrastructure investment. According to a report by 36Kr, just before the research was published, Nvidia quietly launched a new business that could be described as “compute financing.” Nvidia recognized that while AI startups worldwide are queuing up to train models, the prohibitive cost of GPUs and increasingly scarce venture capital funding have left these customers in a predicament — crowding at the storefront but unable to pull out their wallets. Nvidia’s solution: turn chips into financial products, accepting future revenue or profit shares from AI companies in exchange for present-day computing power. Australia’s Sharon AI and Singapore’s Firmus Technologies have already become the first partners, with plans to deploy a combined total of up to 210,000 GPUs.
This “using shovels to buy equity in gold mines” model precisely corroborates Sinolink’s observation: profits across the AI value chain are shifting violently from downstream to upstream, while revenue and profits at the application layer have lagged far behind. The industry’s cash is being siphoned upstream, leaving only a dry riverbed by the time it reaches downstream.
So how large a slice of the human wage pie can AI carve out?
Sinolink mapped the AI exposure of different occupations against 830 job categories in the U.S. Bureau of Labor Statistics’ 2025 Occupational Employment and Wage Statistics (OEWS 2025) survey. The results show that out of approximately $10.83 trillion in total U.S. annual wages, $1.45 trillion is already within AI’s disruption radius. This means AI can already perform or substantially assist with the work content of these roles.
Under a more aggressive theoretical exposure framework — referencing research by OpenAI and Eloundou — the potentially affected wage pool surges to roughly $5.68 trillion, representing over 52% of total U.S. wages.
The impact is even more striking when viewed through employment numbers. Of approximately 156 million employed persons in the U.S., 18.35 million — or 11.8% — are actually exposed to AI technology. The theoretical exposure figure reaches 68.3 million, or 43.9%.
Yet the money AI companies are currently earning from this is a mere drop in the bucket compared to these astronomical figures. The report cites leading model company Anthropic as an example: its annualized recurring revenue (ARR) stands at roughly $47 billion, representing just 3.2% of the $1.45 trillion actually exposed wage pool. Sinolink emphasizes that even if enterprises can achieve equivalent replacement of $100,000 in labor costs with just $10,000 in AI spending — causing this wage pool to face discounting when priced — the current tens-of-billions-of-dollars ARR of major model providers remains at extremely low penetration relative to a wage pool measured in the tens of trillions.
Unlike past waves of automation that primarily impacted manufacturing and repetitive physical labor, this AI wave is more directly targeting high-wage, knowledge-intensive, and service-sector jobs.
The report’s data shows that high-income groups face significantly higher AI exposure than middle- and low-income groups. Occupations at the lowest income percentiles — such as laundry workers and bakers — generally exhibit low AI exposure. In contrast, financial product managers (96.6% income percentile, 78.6% exposure), HR managers (95.3% income percentile, 76% exposure), and aerospace engineers (92.5% income percentile, 89.3% exposure) all face elevated substitution risk.
From an industry perspective, the three sectors with the highest theoretical exposure are Computer & Mathematical (87.6%), Business & Financial (78.2%), and Legal (78.0%). However, the ranking shifts when examining actual exposure. The highest actual exposure sectors are Computer & Mathematical (35.3%), Office & Administrative Support (33.2%), and Sales-related occupations (24.6%).
This gap reveals the complex reality of AI labor substitution: it is not solely determined by model capability, but is also constrained by job characteristics, liability attribution, and organizational processes. The legal profession involves interest mediation and lifelong liability, while financial services depend on client relationships and non-standardized information judgment — actual substitution in these fields has progressed more slowly. By contrast, programming roles, with their clearly defined work objects and short feedback loops, are seeing faster actual substitution. Among the top 20 occupations by actual exposure, eight belong to the Computer & Mathematical category, involving approximately 1.59 million employed persons. The report pointedly notes that for the computer industry, there is no necessary correlation between salary level and AI exposure — the entire industry faces near-uniform vulnerability, highlighting its systemic fragility amid technological iteration.
The financial industry presents a starkly different picture of internal divergence. Because some roles require accountability and have varying degrees of work output standardization, the financial sector’s overall actual exposure is relatively low, but internal polarization is pronounced. Market research analysts show 64.8% actual exposure, and financial and investment analysts 57.2%, while other roles requiring client relationship maintenance and non-standard judgment exhibit relatively low exposure.
In terms of total salary exposure, the $1.45 trillion in actually exposed wages is concentrated in five major sectors: Office & Administrative Support ($289.6 billion), Business & Financial ($247.4 billion), Management ($221.7 billion), Computer & Mathematical ($215.2 billion), and Sales-related ($199.5 billion). The report suggests this provides directional guidance for enterprise services from specialized large models: those seeking certainty can deepen their focus on administrative, computer, and financial sectors where clear substitution has already emerged; those pursuing “zero-to-one business breakthroughs” may find greater potential in education and medical diagnostics.
Sinolink’s report provides a macro benchmark for understanding the AI industry’s revenue ceiling, while Nvidia’s “compute financing” demonstrates, at the micro level, the profound restructuring of power dynamics across the value chain.
36Kr’s analysis identifies three motivations behind Nvidia’s move. First, customers are genuinely running out of money. Global AI infrastructure investment has expanded exponentially over the past two years, but end-application revenue has lagged far behind. OpenAI and Anthropic remain deeply loss-making, and cloud providers need to amortize massive capital expenditures over many years just to approach breakeven. Nvidia itself announced plans in June to issue at least $20 billion in debt — it needs downstream customers to be capable of absorbing the expensive chips it produces.
Second, competition is intensifying. AMD is catching up, Intel is transforming, and both Google’s TPUs and Amazon’s Trainium chips are attempting to reduce the industry’s dependence on Nvidia GPUs. Nvidia needs to use financial tools like “compute financing” to buy time and capture market share before alternatives mature, blanketing every data center on the planet with its GPUs.
The most subtle motivation lies in the revenue structure. Nvidia’s choice of profit-sharing over simple installment payments signals that it sees through the ceiling of hardware’s one-time-sale business model. By turning GPUs into equity stakes in “money-printing factories,” Nvidia is transforming from a hardware company into an infrastructure platform — from a seller of shovels to someone who uses shovels to buy equity in gold mines and share in every ounce of gold produced.
This is quietly reshaping the relations of production in the AI industry. Computing power is no longer merely a means of production to be purchased or leased, but has become an input — akin to equity — that can be obtained in exchange for future profits. Nvidia is becoming a super-hub, controlling the allocation of computing power and holding claims on the profits of future AI winners. It doesn’t need to predict which company will win; it only needs to ensure that whoever wins owes it money.
Sinolink’s report concludes by emphasizing the critical distinction between “exposure” and “substitution.” Exposure means tasks may be assisted, automated, or reorganized by AI, but does not imply that these wages will disappear proportionally. What truly determines AI’s economic impact remains enterprise adoption speed, the boundaries of model capability, organizational process redesign, and regulatory constraints.
AI’s macro impact will not manifest simply as a linear decline in employment numbers. A more likely trajectory is this: some single-function roles will be replaced, while many multi-function roles will be restructured; some wage costs will be compressed, while more labor processes will be repriced. Particularly notable is that AI Agents exhibit a “higher wage, higher substitution rate” characteristic, making AI’s potential impact on the income and consumption side potentially more far-reaching.
For investors, both reports point toward a sobering conclusion: the revenue potential of the AI industry should not be understood solely through the lens of software market size, but should be anchored against the far larger pool of labor costs. Current revenue penetration by major model providers remains at extremely low levels — but the flip side of this coin is that human compensation structures are facing a systemic restructuring that capital markets have yet to fully price in. And Nvidia’s “compute financing” represents an early move by upstream hardware suppliers to secure a ticket to future profits amid this restructuring.
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
Follow the three steps to link a messaging app like Telegram or LINE, and get schedule notifications pushed to you automatically
The news and data on this website are for reference only and do not constitute investment advice or an offer to buy or sell. Information is sourced from exchanges and public sources, and may be delayed, interrupted, or updated. While we strive for accuracy, we do not guarantee timeliness, correctness, or completeness. Content may include external links for which we are not responsible. By using this site, you agree that we and our partners are not liable for any losses. Investment carries full responsibility; please carefully assess risks and consult professionals. If there are errors in the content, please contact us for correction.
Leave a Reply