The $5 Trillion USA Bet Meets the Open-Weight World: America is building an AI economy around scarcity. China may be betting on abundance
The American artificial intelligence boom is usually described as a competition among technology companies: OpenAI versus Anthropic, Microsoft versus Google, Nvidia versus the chipmakers hoping to challenge it. Yet the scale of the investment has grown far beyond a contest among a few corporations. Goldman Sachs Research estimates that the large technology companies leading the AI buildout could spend approximately $5.3 trillion between 2025 and 2030. That figure represents projected capital expenditure rather than money already spent, and it includes technology infrastructure that cannot always be separated neatly into an “AI-only” category. Even with those qualifications, it reveals the magnitude of the wager now being made on the future of machine intelligence.
The United States is not merely developing better software. It is constructing an industrial system around artificial intelligence, including semiconductor fabrication, data centers, power generation, electrical transmission, cooling infrastructure, fiber networks, cloud platforms and the financial machinery required to fund them. Federal Reserve researchers estimate that annualized American data-center investment could reach approximately $370 billion by the second quarter of 2026, while the Bank for International Settlements reports that the five largest hyperscalers are on track to spend more than $1 trillion on AI-related capital investment across 2025 and 2026 alone. The BIS also notes that these commitments are growing faster than the earnings and free cash flow of the companies making them, pushing some firms toward greater reliance on debt and outside financing.
Into this extraordinarily expensive American buildout comes a very different strategy from China. Instead of keeping every advanced model locked behind a proprietary subscription or application programming interface, Chinese laboratories are increasingly releasing powerful open-weight systems that other organizations can download, examine, modify and deploy. Kimi K3, developed by Moonshot AI, is a particularly dramatic example. Moonshot describes it as a 2.8-trillion-parameter model with native multimodal capabilities, a one-million-token context window and strong performance across coding, reasoning and knowledge work. The company has announced that the full model weights will be released by July 27, 2026.
Kimi K3 does not establish that China has surpassed the United States in every dimension of artificial intelligence, and its enormous size means that most individuals and small organizations could not operate it locally. Nevertheless, it demonstrates that a Chinese company can now build near-frontier intelligence and pursue a distribution strategy fundamentally different from the one supporting the American investment boom. The collision between those strategies could become one of the defining economic stories of the AI era.
America’s wager on scarce intelligence
The business model of the leading American frontier laboratories depends on more than producing capable systems. It also depends on maintaining control over those systems. Users do not receive the weights behind the strongest models offered by OpenAI or Anthropic. Instead, they rent access through subscriptions, hosted applications and metered services. The provider controls the model, sets the price, establishes the usage rules and decides when a product will be updated, restricted or discontinued.
That arrangement can support enormous revenues as long as frontier-level intelligence remains scarce. Businesses may tolerate premium prices and dependence on a single vendor because only a small number of companies can provide the required capability. The cost of switching may also grow as companies build workflows, internal applications, customer records and institutional knowledge around one provider’s ecosystem. In that world, the model company does not merely sell computation. It operates a tollbooth through which customers must repeatedly pass.
Much of America’s infrastructure boom rests on the expectation that demand for artificial intelligence will become enormous while the providers of the most capable models retain enough pricing power to recover their investments. Data centers can remain useful for decades, but investors generally do not finance them merely because they will perform socially valuable computation. They expect future revenue streams large enough to justify the cost of land, chips, electricity, cooling, financing and construction. The American wager therefore depends not only on AI becoming widely used, but also on advanced intelligence remaining profitable to sell.
That wager may prove correct. Artificial intelligence could become embedded in nearly every company, public institution, profession and device. Yet another possibility is emerging beside it: AI use could become enormous while the market price of the underlying intelligence falls sharply. If highly capable models become widely available through open weights and competing hosting providers, the economic value of controlling one proprietary foundation model could decline even as the technology itself becomes more important.
Open weights change who holds the leverage
An open-weight model is not automatically free. A system as large as Kimi K3 requires expensive accelerators, high-speed networking, substantial electricity, specialized engineering and enormous amounts of memory. Moonshot itself recommends deployment on supernode configurations containing at least 64 accelerators, making the full model impractical for an ordinary workstation or small company server. The company’s hosted service also charges for usage, which means “open” should not be confused with costless.
Open weights nevertheless change the structure of the market because they separate the model from the company that created it. A proprietary frontier model can generally be accessed only through its owner or an authorized platform. An open-weight model can potentially be hosted by multiple cloud companies, specialized inference providers, universities, governments, cooperatives and large enterprises. Organizations may be able to operate it inside their own security environment, adapt it to specialized work, study its internal behavior or move among infrastructure providers without abandoning the underlying model.
That portability changes the balance of power between customers and AI companies. A business using a proprietary model may face higher prices, changing product rules or discontinued features with few realistic alternatives. A business using an open-weight model can, at least in principle, change hosts, negotiate lower prices or bring parts of the system in-house. The practical difficulty of operating a model may remain substantial, but the owner of the original weights no longer possesses exclusive control over every future use.
China appears to recognize that controlling the single best model may be less strategically important than making Chinese models part of the world’s default AI infrastructure. A 2026 report from the U.S.-China Economic and Security Review Commission describes China as going “all in” on open AI, with many laboratories publishing weights while charging considerably less for hosted access than global competitors. The report argues that this strategy creates a reinforcing loop: widespread adoption encourages experimentation and derivative models, which improve the ecosystem and encourage still more adoption. It notes that Alibaba’s Qwen family had already produced more than 100,000 derivatives on Hugging Face by the time of publication.
This is not merely a contest over which chatbot provides the most polished response. It is a competition between two theories of technological power. The dominant American approach concentrates the strongest capability inside a small number of companies and sells controlled access. China’s emerging strategy distributes more of that capability, encourages other organizations to build upon it and then gains influence through adoption, infrastructure, services and technical standards.
Your retirement account may already be part of the bet
Most Americans have not knowingly invested in an AI data center or a frontier-model laboratory, but many are financially connected to the buildout through their retirement savings. At the end of March 2026, Americans held approximately $9.9 trillion in 401(k) plans, with mutual funds managing $5.7 trillion of those assets. Equity funds accounted for $3.3 trillion, while hybrid funds such as target-date funds held another $1.6 trillion. These portfolios often own broad collections of public companies rather than narrowly identified “AI investments,” but the largest technology companies, chipmakers, utilities and infrastructure suppliers are woven throughout the market.
The S&P 500 alone contains 500 leading American companies and represents approximately 80 percent of available U.S. market capitalization. Because it is weighted by company size, the largest corporations play an especially important role in funds that track or benchmark themselves against it. When technology companies rise in value because investors expect enormous AI profits, retirement savers can benefit even if they never intentionally choose an AI stock. The reverse is also true: if those expectations are repriced, losses can spread through ordinary retirement accounts alongside the portfolios of wealthy investors and institutions.
The exposure extends beyond publicly traded shares. Bond funds may hold debt issued by technology companies, utilities or infrastructure operators financing the buildout. Pension systems, insurance companies and private-credit funds may invest in data centers, electrical generation, transmission projects or real estate partnerships. Local governments may offer tax incentives, road improvements or utility commitments to attract major facilities, while communities may allocate land, power and water based on projected economic benefits.
This means the AI boom is no longer a private gamble confined to Silicon Valley. It is becoming embedded in retirement savings, credit markets, utility planning, construction, regional development and public policy. Ordinary Americans may be participating indirectly through their pensions, index funds, electricity bills, local tax systems and employment markets. The gains could be broadly distributed if the investment produces sustained economic growth, but so could the consequences if the expected profits fail to materialize.
What happens if intelligence becomes cheap?
The arrival of cheaper open-weight models would not necessarily make America’s AI infrastructure useless. Every model still requires machines, electricity, storage and networking, and falling prices can cause consumption to rise rather than decline. A company that can afford only ten expensive AI agents today might operate hundreds or thousands if inference becomes dramatically cheaper. Schools, hospitals, nonprofits, local governments and small businesses could adopt systems that were previously beyond their budgets. Under that scenario, demand for computing capacity could continue rising even as the price of each unit of intelligence falls.
The economic danger lies in the difference between widespread usefulness and investor profitability. Infrastructure can remain socially valuable while producing much lower returns than the people who financed it expected. If open competition compresses the prices charged for model access, customers and developers may benefit while model laboratories, cloud platforms and infrastructure owners face thinner margins. The technology could succeed magnificently as a public capability while disappointing investors who assumed that a small number of companies would continue collecting premium rents.
History offers many examples of transformative technologies that destroyed substantial investor capital. Railroads reshaped commerce and settlement even though numerous railroad companies failed. Fiber-optic overbuilding during the telecommunications boom helped create the modern internet, yet many of the firms that financed those networks collapsed. Society inherited valuable infrastructure, but the original investors often paid dearly for it. An AI investment correction could follow a comparable pattern in which the machines remain useful, demand continues and intelligence spreads, while valuations built around permanent scarcity are dismantled.
If two models produce similar business outcomes, customers may not pay several times more simply because one was developed in California rather than China. They may choose the system that is cheaper, easier to customize, more private or less restrictive. Foundation models could gradually become commodities, pushing value toward applications, proprietary data, workflow integration, customer relationships, security, support and efficient hosting. OpenAI and Anthropic might remain highly successful, but they would need to compete through complete products and trusted services rather than assuming that exclusive control of raw model intelligence will provide an enduring moat.
American laboratories already appear aware of this pressure. OpenAI has released the open-weight gpt-oss-120b and gpt-oss-20b models under the Apache 2.0 license, allowing organizations to operate and customize them on their own infrastructure. These are not OpenAI’s most advanced hosted systems, but their release suggests that the company recognizes the strategic danger of allowing Chinese developers to dominate the open-model ecosystem uncontested.
The political temptation
This raises an uncomfortable political question: what will the United States government do if Chinese open-weight systems begin seriously threatening the economics of the American AI buildout? There are legitimate security concerns surrounding foreign-hosted services, especially when personal, corporate or government information is transmitted to servers controlled by a company operating under another country’s laws. Models may also contain vulnerabilities, hidden behaviors, ideological biases or capabilities that could be exploited for cyberattacks and espionage.
A downloadable model operating entirely on trusted American infrastructure, however, is not identical to a Chinese-controlled online service. Policy can distinguish between foreign data transmission and local model operation, just as it can distinguish between government systems, critical infrastructure, consumer applications and academic research. Security testing, procurement standards and restrictions on sensitive data can address specific risks without automatically prohibiting access to the underlying weights.
The danger is that national-security language could be used to erase those distinctions once economic pressure grows. Restrictions would not necessarily appear as a dramatic nationwide law declaring Chinese model weights illegal. They could emerge gradually through federal procurement rules, government contracting requirements, cloud-hosting restrictions, cybersecurity certifications, intellectual-property litigation, liability standards and special obligations imposed on regulated industries. The result could be a soft prohibition in which possession remains technically legal but commercial use becomes too risky or complicated for most organizations.
The economic motive for such restrictions would extend well beyond protecting OpenAI, Anthropic or a small group of technology executives. Policymakers could view restrictions as necessary to defend stock-market valuations, retirement savings, corporate debt, utilities, domestic employment, data-center projects and a national strategy built around American AI leadership. The Federal Reserve’s May 2026 Financial Stability Report noted that market participants were already concerned that AI valuation issues could help trigger a correction in risk assets, while the BIS has questioned whether the current scale of AI investment can be sustained if expected productivity gains and revenues do not arrive.
A comprehensive ban is not inevitable. In fact, the Trump administration’s current AI Action Plan explicitly recognizes that open-source and open-weight models possess economic and geopolitical value, and the administration has publicly supported the development of American open models. That stated policy could make a broad attack on open weights politically contradictory. Yet Washington could still draw a boundary around Chinese-origin models while continuing to promote American ones, presenting the distinction as a matter of national security rather than economic protection.
The toll road and the open route
America may be building the most sophisticated computational infrastructure in history. That infrastructure could support scientific discovery, medical research, industrial automation, education, public services and forms of intelligence that remain difficult to imagine. The physical investment may prove extremely valuable even if some of the financial assumptions surrounding it turn out to be wrong.
The unresolved question is who will control access to the intelligence running through those machines. The dominant American business model resembles an enormous toll road, financed on the assumption that users will continue paying premium prices to pass through a limited number of corporate gates. China’s open-weight strategy offers another route, one that is not truly free because hardware, electricity and expertise still cost money, but that does not necessarily leave a single model company controlling every entrance and collecting every toll.
If Chinese laboratories continue releasing open-weight models within striking distance of the American frontier, the United States will face a defining choice. It can compete by producing better models, opening more of its technology, lowering prices and building services that customers voluntarily prefer. Alternatively, it can protect existing investments by restricting access to foreign models and making the American market less open than the technology itself.