Anthropic’s $30 Trillion Fantasy – Marcus On AI | Substack

Cognitive scientist and AI critic Gary Marcus has published an essay on his Substack newsletter, Marcus on AI, disputing a widely circulated Anthropic projection that artificial intelligence could deliver on the order of $30 trillion in economic gains. Marcus argues the figure rests on optimistic assumptions rather than credible evidence, and his critique has added to an ongoing debate over whether the AI industry’s economic projections are supported by what the technology can currently deliver.

The dispute centers on estimates attributed to Anthropic suggesting that AI, if deployed broadly across the economy, could generate tens of trillions of dollars in value over the coming years and decades. Such projections have become a common feature of AI industry commentary, often cited to justify large capital spending on data centers, chips, and energy infrastructure by major technology firms.

Marcus, writing on his Substack, contends that the $30 trillion figure rests on assumptions about AI capabilities that current systems do not possess. His essay argues that today’s large language models — including those built by Anthropic and its competitors — remain prone to errors, hallucinations, and reliability problems that limit their usefulness in high-stakes, high-value economic domains. He argues that extrapolating from limited current deployments to a transformation of the entire global economy overstates the available evidence.

Anthropic, for its part, has positioned itself as a builder of increasingly capable AI systems and has argued that the technology’s economic potential is substantial. The company is one of the best-funded AI labs in the world, backed by billions in investment from partners including Amazon and Google. Its economic forecasts, like those of competitors such as OpenAI, rest on the expectation that AI systems will continue improving rapidly and will be adopted across industries at scale. Marcus’s essay questions both premises: that capability gains will continue at the assumed pace, and that adoption will translate into measured productivity growth on the scale claimed.

Stakes in the AI Investment Debate

The exchange is notable because trillion-dollar projections are influencing economic decisions. Investors, policymakers, and energy planners are making choices today based on forecasts of AI-driven growth, and the major AI labs are spending large sums on compute infrastructure in expectation of substantial future returns. If those forecasts prove inflated, analysts have warned of a potential misallocation of capital into data centers and chips that may not generate the returns their backers expect.

Marcus’s critique also points to a measurable question in the economic data: despite rapid corporate adoption of AI tools, aggregate productivity statistics have so far shown only modest gains. Economists remain divided over whether AI’s impact is delayed, concentrated in specific sectors, or smaller than industry projections suggest. The debate over Anthropic’s figure reflects that larger question.

For Anthropic specifically, the dispute touches on credibility. The company has cultivated a reputation for AI safety research and careful communication, and a prominent public challenge to its economic forecasts places pressure on how it substantiates them.

Marcus’s Long Running Skepticism

Gary Marcus, a cognitive scientist at New York University and author of books including Rebooting AI, has been one of the most consistent public skeptics of the large language model paradigm. For years he has argued that current AI systems lack robust reasoning and world knowledge, and that industry timelines for transformative AI are overconfident. His Substack newsletter has become a prominent venue for critiques of AI hype.

The $30 trillion figure belongs to a broader set of AI economic forecasts. Other labs and consultancies have published estimates of AI adding trillions annually to global GDP, while figures like Sam Altman of OpenAI have spoken of AI eventually driving growth comparable to the Industrial Revolution. Anthropic’s projection sits within that same category of forecast, which skeptics like Marcus argue understates deployment frictions, error rates, regulatory constraints, and the difficulty of integrating AI into complex real-world workflows.

“The $30 trillion figure rests on assumptions that current AI systems cannot support.”

— Gary Marcus, Marcus on AI (Substack)

Numbers Nobody Can Verify Yet

The central difficulty with any multi-trillion-dollar forecast is that it cannot be verified in the present. It is not yet clear exactly which inputs and assumptions underlie the $30 trillion estimate as Anthropic presents it, over what time horizon it is meant to apply, or whether the figure refers to cumulative gains, annual output, or market value. Readers should treat it as a projection, not a measured fact.

It is also unclear how Anthropic has responded, if at all, to Marcus’s specific critique. Because the essay was published on an independent newsletter rather than in a peer-reviewed venue, the exchange so far consists largely of claims and counter-claims rather than settled analysis. Whether measured productivity data over the next several years supports the optimists or the skeptics remains an open empirical question.

Watching the Evidence Accumulate

The practical test of competing claims will come from data. Over the next several quarters, economists will be watching national productivity statistics, corporate earnings disclosures, and AI revenue figures from the major labs for evidence of whether AI adoption is translating into the growth its backers project. Anthropic and its peers are expected to continue publishing economic research supporting their investment cases, while critics like Marcus will likely continue pressing for methodological transparency.

A direct public response from Anthropic to Marcus’s essay, or a detailed disclosure of the assumptions behind the $30 trillion figure, would sharpen the debate. In the absence of that, the disagreement between AI optimists and skeptics will be resolved gradually — by revenues, productivity data, and eventually returns on the hundreds of billions being invested in AI infrastructure.

Source: Anthropic

Key Questions

What is the $30 trillion claim?

It is a projection associated with Anthropic suggesting AI could generate roughly $30 trillion in economic gains. The exact time horizon, methodology, and definition of the figure have not been fully detailed publicly, which is one of Marcus’s main criticisms.

Who is Gary Marcus?

A cognitive scientist at New York University and longtime critic of AI hype. He writes the Marcus on AI newsletter on Substack, where he argues that current large language models are overpromised relative to their actual capabilities.

Why does this debate matter?

Because large investment decisions — in data centers, chips, and energy — are being made on the basis of AI growth forecasts. If projections like the $30 trillion figure prove inflated, analysts warn of a significant capital misallocation risk.

Has Anthropic responded to the critique?

It is not yet clear whether Anthropic has issued a specific response to Marcus’s essay. The debate currently consists of the original projection and Marcus’s published counter-argument.

Can either side be proven right?

Not immediately. The dispute will be settled over time by observable evidence: productivity statistics, AI-related revenues, and returns on AI infrastructure investment over the coming years.

Source: Anthropic