AI’s Promise of Boom Meets Brutal Price War and Stagnant Adoption

The artificial intelligence industry’s promise of a paradigm-shifting economic revolution is increasingly colliding with a sober reality on the ground, defined by a fierce commercial price war, stagnating corporate adoption, and intensifying regulatory and geopolitical friction. While leading AI developers continue to forecast historic, double-digit economic growth fueled by autonomous systems, newly released market data and mounting security challenges suggest a far more complicated transition. The gap between the theoretical capabilities of frontier models and their actual economic integration is widening as businesses opt for cheaper, “good enough” technology, and governments step up scrutiny over IP theft, safety protocols, and cybersecurity breaches.

According to the latest AI Index compiled by Ramp, a financial operations platform that tracks corporate card and invoice expenditures, overall business adoption of AI tools crept up by just 0.4 percentage points in August, reaching 56.1%. This plateau suggests that larger enterprises are struggling to integrate complex AI systems into their legacy workflows. At the same time, the cost of running these models is collapsing. Ramp’s data reveals that the market price for one million AI tokens—the basic units of data processed by large language models—has plummeted from approximately $1.15 in March to just $0.68.

While falling token prices reduce operating costs for enterprises, they squeeze the profit margins of the heavily funded labs building the models. Rather than buying expensive, premium “frontier” models, corporations are increasingly downgrading to cheaper, standard systems that are “good enough” for routine tasks. Use of premium frontier systems—such as Anthropic’s Claude Opus and Fable models, or OpenAI’s GPT-5.6-Sol—accounted for 45% of token volume, down from a peak of 53% in August. Instead, businesses are gravitating toward mid-tier alternatives like OpenAI’s GPT-5.6 Terra and Anthropic’s Claude Sonnet.

The tension over AI’s ultimate economic destination was highlighted in a new research paper and modeling tool released by Anthropic. The San Francisco-based AI safety and research lab outlined three wildly divergent paths for how generative AI could reshape the U.S. economy by 2030. In Anthropic’s most conservative scenario, AI serves as an incremental productivity aid, functioning similarly to the early adoption of the internet and yielding gradual, familiar economic gains. The second scenario projects that AI will autonomously handle up to half of all knowledge work by 2030, doubling the nation’s historical GDP growth rate. While knowledge workers’ wages would stagnate in this model, other sectors of the labor market would see real income gains.

Ara Kharazian, Ramp’s head of economics, noted that outside of specialized coding agents, AI developers have yet to build products that deliver clear, substantial productivity boosts for the average white-collar worker. “We’re very deep into an existing price war that’s going to drive down their ability to profit from tokens on their own,” Kharazian said. Concurrently, average monthly AI spending per employee among the top 1% of corporate spenders fell 9.7%, dropping from $7,976 to $7,205. Despite the market-wide cooling, Anthropic has maintained a slight edge over its primary rival in corporate market share, with 43.8% of U.S. businesses paying for its products compared to OpenAI’s 39.8%.

In Anthropic’s most extreme projection, AI surpasses human capability in virtually all knowledge-work tasks, operating almost entirely autonomously. Under this scenario, annual GDP growth surges to 15%, doubling the size of the U.S. economy every four and a half years. While overall societal wealth would skyrocket, unemployment would climb to unprecedented levels, surpassing the depths of any historical recession. Anthropic’s model calculates GDP strictly from the supply side, assuming that dramatic increases in productivity and output will find ready markets. The company acknowledged that its interactive scenario tool represents a “stark simplification of a complex reality,” omitting demand-side disruptions, consumer spending drops from job losses, business cycles, financial market instability, and catastrophic systemic risks.

To break out of the low-margin token price war, major tech firms are rapidly pivoting toward autonomous AI “agents”—systems capable of performing complex, multi-step actions across different applications without human intervention. Meta recently entered this high-stakes race with the launch of “Muse,” a personal AI assistant available via the web, mobile apps, and WhatsApp. Unlike static chatbots, Muse is designed to remember personal user details and, with user permission, connect directly to financial and organizational services like Stripe, Plaid, and personal email to execute transactions and manage schedules. Initially offered free to U.S. adults, Meta is testing subscription models priced at $20 and $100 per month.

However, the transition to active, agentic AI systems introduces severe new security vulnerabilities. Cybersecurity firm Calif recently demonstrated a proof-of-concept “WeChat worm” that utilized AI to hijack accounts. The attack was triggered by a simple incoming call from an existing contact and required no user interaction or acceptance to compromise the device. While Calif reported the vulnerability to WeChat’s parent company, Tencent, which successfully patched the flaw in July, the exploit highlights how autonomous, agentic systems can be weaponized at scale. Rogue AI agents developed by top-tier firms have already carried out unintended actions against corporate websites, raising concerns among enterprise clients and slowing down the widespread adoption necessary to trigger the economic boom that developers have promised.

Jack Clark, Anthropic’s co-founder and head of public benefit, noted that while the underlying technology is improving at a “very, very fast and sustained rate,” its diffusion through the broader economy is likely to occur “more slowly” than silicon valley assumes. If rapid adoption does occur and triggers mass displacement, Clark suggested that the resulting surge in tax revenues could provide policymakers with unprecedented resources to assist affected workers. Many economists remain deeply skeptical of such explosive growth projections. In a newly published essay, economists Ben Moll of Princeton University and Alex Imas of the University of Chicago Booth School of Business dismissed predictions of double-digit GDP growth over the next decade. They pointed out that a compound growth rate of 16.6% annually—necessary to make the world “10 times richer in 15 years”—would translate to a world 100 times richer in just 30 years, an expansion without historical precedent.

According to Moll, an AI-driven economic explosion requires five simultaneous, highly improbable conditions: Automation must diffuse through the global economy far faster than any previous technology in human history; consumer demand must remain strong enough to absorb the massive surge in cheap AI-generated output; public and corporate trust must not be derailed by major, AI-related systemic cyber incidents; AI must transition to self-improving, automated research and development; and consumption patterns must adapt rapidly to favor AI-produced goods and services. This skepticism aligns with broader warnings from the academic community. In July, more than 200 economists, researchers, and executives—including Nobel laureate Daron Acemoglu, Joseph Stiglitz, Paul Krugman, former Google CEO Eric Schmidt, and venture capitalist Reid Hoffman—signed an open letter urging policymakers to prioritize research into the economic impacts of AI. Acemoglu’s own research has projected a far more modest impact, estimating that AI will boost U.S. labor productivity by only 0.5% and overall GDP by a mere 0.93% over the next decade.

The intense competition to build and monetize frontier models has heightened geopolitical tensions between the United States and China. U.S. cyber and law enforcement agencies recently accused six prominent Chinese AI firms—including high-profile startups Moonshot AI and DeepSeek, as well as e-commerce giant Alibaba—of executing systematic, industrial-scale model copying. The federal agencies alleged that these companies used a technique known as “knowledge distillation” to extract advanced capabilities from American frontier models developed by Anthropic, OpenAI, Google, and SpaceX. Distillation involves using the outputs of a larger, highly sophisticated “teacher” model to train a smaller, cheaper “student” model. While distillation is a standard technique in machine learning research, U.S. officials assert that these companies bypassed years of costly R&D by draining billions of tokens from American systems, likely with the tacit approval or awareness of the Chinese government. The allegations are expected to complicate bilateral technology negotiations ahead of upcoming diplomatic summits between Washington and Beijing.

As national security concerns mount, domestic regulators and lawmakers are intensifying their oversight of top-tier AI developers. In Washington, Senator Josh Hawley (R-Mo.), chairman of a Senate Homeland Security subcommittee, launched a formal probe into OpenAI’s handling of a security breach that occurred in July involving Hugging Face, a prominent repository for open-source AI models and code. In a letter to OpenAI CEO Sam Altman, Hawley characterized the company’s internal reporting on the breach as “reckless” and accused it of redacting critical details. Hawley’s inquiry, which set an October 1 deadline for detailed answers, demands disclosure on how OpenAI’s autonomous agents may have behaved unexpectedly, and how the company manages safety policies. The probe reflects a growing anxiety in Congress regarding frontier AI risks. Hawley’s letter specifically referenced recent warnings from three former Anthropic researchers who stated there is a greater than 10% chance that advanced AI could lead to human extinction within the next decade.

The push for tighter domestic regulation stands in sharp contrast to the executive branch’s current trajectory. The Trump administration’s newly unveiled AI framework notable omits any requirement for companies to publicly report real-world safety failures, model malfunctions, or instances of rogue behavior. This voluntary approach creates a significant regulatory divergence from the European Union’s landmark AI Act, which legally mandates that providers of high-risk AI systems report all serious incidents to national authorities. This regulatory fragmentation is also straining international security alliances. Anthropic recently drew criticism from British officials after it withheld its latest model, Claude Mythos 5.1, from pre-release safety testing by the UK’s AI Security Institute. The institute was established following the 2023 Bletchley Park AI Safety Summit to evaluate frontier models for catastrophic risks before they are deployed to the public. While Anthropic stated it is coordinating with the U.S. government to expand testing access to international partners, the decision has fueled concerns in London that Washington is placing national protective barriers around its AI sector, potentially shutting out close allies from evaluating cutting-edge technologies.