Anthropic chief executive Dario Amodei published an extensive essay calling on the artificial intelligence industry to voluntarily pace capabilities development so safety measures can catch up. Semiconductor equities took a brief dip on the headline, but equity analysts quickly brushed the warning aside. D.A. Davidson analyst Gil Luria pointed out that nobody is actually slowing anything down. Across Alphabet, Amazon, Meta, and Microsoft, capital expenditure on data centers and silicon expanded past $293 billion in the first half of 2026 alone, with Meta increasing infrastructure spending by more than 50 percent quarter over quarter.
The contrast between lab rhetoric and financial reality is striking, but dismissing Amodei's warning as pure theater misses the technical substance. Frontier labs have direct telemetry on internal training runs that equity research desks cannot see. If capabilities in autonomous multi-agent swarms, recursive tool use, and automated vulnerability exploitation outpace defensive evaluations, the resulting tail risks are real. When a research team that has raised billions urges caution on capability scaling, that signal reflects empirical observation of internal model behavior.
The breakdown occurs when moral caution meets the game theory of competitive markets. In economics, a voluntary multilateral agreement to restrict production is notoriously unstable without external enforcement. If one lab deliberately caps cluster utilization or delays its next generation run, its rivals capture the frontier lead, attract the top researchers, and lock in enterprise software contracts. Open-weights research groups and sovereign computing projects have zero incentive to join a voluntary pause negotiated in San Francisco. A firm that slows down unilaterally does not eliminate capability risks for the world. It merely transfers the frontier advantage to competitors who chose not to pause.
Beyond competitive game theory, capital allocation operates under physical inertia that cannot be redirected on short notice. Hyperscaler capital expenditure is not a software flag that product managers can toggle. Building a multi-hundred-megawatt compute campus requires three-year lead times for substation transformers, bespoke power purchase agreements with regional utilities, and non-cancellable wafer commitments with foundries. The capital deployed during the first half of 2026 was committed eighteen to thirty-six months ago, when current frontier models were still in early planning stages.
Depreciation schedules do not pause because an industry executive issues a public plea. The multi-billion-dollar depreciation charges on server racks and cooling infrastructure hit corporate income statements every quarter regardless of whether those clusters run frontier pre-training or idle in standby mode. If a cloud provider leaves thousands of liquid-cooled accelerators underutilized out of caution, fixed depreciation continues to erode operating margins without generating inference revenue to cover the carry cost.
Institutional investors view this spending through an options framework. For a hyperscaler balance sheet, the distribution of outcomes from overbuilding compute is heavily asymmetric compared to the distribution from underbuilding.
If a hyperscaler overbuilds infrastructure, the downside is temporary margin compression, depressed server utilization, and lower pricing on spot inference tokens. As Wall Street analysts noted, an operator that overbuilds can eventually trim forward capital expenditure, sweat its existing asset base over extended useful lives, and convert deferred depreciation into free cash flow.
If a hyperscaler underbuilds, the outcome is existential. Failing to provision sufficient capacity means losing the foundational platform of enterprise software for the next decade. Workloads migrate to competing cloud ecosystems, developer loyalty evaporates, and the enterprise value of the core franchise faces permanent impairment.
Overbuilding compute represents an expensive long call option on future technological rent. Underbuilding is equivalent to writing naked puts on a cloud monopoly. Under that payoff distribution, rational corporate finance officers will choose to overbuild every single quarter.
There is also an incumbent incentive embedded in calls for collective moderation. When established frontier labs suggest regulatory standards or synchronized safety gates, they naturally propose criteria anchored near the frontier they already dominate. A mandated development pause or complex pre-deployment certification establishes high capital barriers that deter younger startups from challenging the leading incumbents. What sounds like industry-wide caution often functions as regulatory moat construction.
Equity markets did not ignore Amodei's essay out of disbelief in his technical acumen. They ignored it because capital cycles, depreciation accounting, and competitive payoffs govern corporate behavior far more reliably than public appeals for restraint. Until the economic loss function penalizes unaligned deployment more severely than losing the frontier race, capital will continue funding compute at maximum speed.
Originally published at https://deanlee.info/essays/anthropic-slowdown-wall-street-capex/.
Top comments (0)