Washington and Silicon Valley have found a new fight to pick over artificial intelligence in “pacing,” or the deliberate throttling of frontier model development until safety, alignment, and society at large can catch up. To its detractors, pacing is unilateral disarmament in the race with China. To its champions, pacing is the only responsible path for a technology whose own creators warn of catastrophic risk. 

Both camps have fallen prey to the “Compute-to-GDP Fallacy”—the mistaken belief that every incremental leap in AI model performance immediately translates into macroeconomic output. Every prior general-purpose technology took decades to diffuse into measurable productivity. AI is following the same curve at an accelerated pace, but everyone seems to buy the hype that the laws of history or of economics do not apply this time.

In reality, Corporate America is already years behind the AI frontier, and the labs’ commercial fortunes will be decided by trust and adoption, not raw capability. Pacing would cost the economy remarkably little. Here’s why we—whether out of arrogance or misdiagnosis—are simply having the wrong argument.

The pacing skeptics’ suspicions are not frivolous. Is pacing real, or a savvy marketing gambit by frontier labs and cybersecurity companies polishing their financials ahead of IPOs? Would pacing cede the U.S. lead in AI to China, or would Beijing reciprocate and pace in its own manner?

The Frontier Problem

AI has plainly reached a critical capability milestone. Warnings of catastrophic or existential risk can no longer be dismissed outright, even if the near-term probability remains modest. Yet by focusing almost exclusively on cutting-edge models, frontier labs have mismanaged both their messaging and the public trust. More than 100 recent conversations with CEOs, policy leaders, and AI scientists for our coming book, When Machines Act, have convinced us that the pacing debate has lost sight of first-principles thinking.

Lost in the noise is the distinction between the cutting-edge research the labs conduct behind closed doors and the products they release to the public. The real question may be whether the labs need to slow down at all or simply do a better job of ensuring their products are safe for consumption.

The Alignment Problem

Since the release of ChatGPT in 2022, corporate leadership has scrambled with a speed unmatched in modern commercial history. Even so, while executive suites have mobilized with unprecedented urgency, the structural physics of enterprise architecture—fragmented data silos, legacy ERPs, strict compliance regimes, and basic data hygiene—make true economic absorption an inherently slow slog. As corporate budget shocks from runaway “tokenmaxxing” demonstrated, many daily enterprise workflows require far simpler models, and precious few tasks at the average Fortune 500 company demand a frontier system at all. Pacing, therefore, will neither harm economic output nor choke off the labs’ commercial revenues, because enterprises need time simply to assimilate the capabilities already on the table. 

Among high-performing companies, more than two-thirds identify data as the primary barrier to implementing AI, a figure that has proven stubborn even as the models themselves have leaped forward. Only 7% describe their data as “completely ready” for AI; fewer than a quarter have a data strategy at all; and 63% either lack AI-suitable data management or are unsure whether they have it. 

The Fallacy Problem

As McKinsey Senior Partner Asutosh Padhi emphasized on air with Fareed Zakaria, technical availability is fundamentally different from economic transformation. General-purpose technologies have historically required decades to reorganize workflows and generate broad-based productivity gains. Electricity took 75 years to lift productivity economy-wide. Computers required 50 years, and the Internet and mobile devices demanded 25. The underlying models may be ready, but the systemic organizational restructuring they demand will take substantial time. When McKinsey surveyed the business community, the firm found that only 6 percent of companies reported a “significant” impact and modest earnings attribution.

Companies are concentrating on the high-reward, low-risk automation tasks that models one or two generations old can already solve. As one highly respected former Wall Street CEO told us, these systems will run in parallel with legacy systems for years to confirm they operate correctly and that no regulatory risk is unknowingly absorbed.

Despite advances in frontier labs, corporate America will set the pace itself, ensuring a secure rollout regardless of what the labs decide. No company in any industry should release a product it believes to be dangerous, and AI is no exception.

A parallel dynamic has emerged in the economics of silicon. Older-generation chips, initially cast aside in the scramble for cutting-edge accelerators, are finding a second life as workhorses for the practical inference tasks that dominate enterprise demand. As Growth Protocol founder and CEO Miro Dimitrov noted at last week’s Yale CEO Caucus, deploying neuro-symbolic architectures has allowed his enterprise reasoning platform to slash inference costs by roughly 80-fold in live client deployments, largely by shifting workloads off ultra-expensive GPUs and onto everyday enterprise CPUs.

The Three Phases of AI Adoption

Corporate AI adoption is best understood in three phases, distinguished by how much work a company can responsibly hand over, which is gated by data readiness and the trust systems have earned. The first phase, assistance, consists of off-the-shelf copilots that ride atop enterprise platforms