Back to News
Market Impact: 0.62

The AI economy could crash on mounting chip costs — and those token costs won’t help

NVDATSLAMSFTUBERITGOOGL
Artificial IntelligenceTechnology & InnovationInflationCorporate Guidance & OutlookCredit & Bond MarketsTrade Policy & Supply ChainRegulation & LegislationCompany Fundamentals

AI chip spending is surging into the billions per data center, while Goldman Sachs forecasts token consumption will rise 24-fold to 120 quadrillion tokens per month by 2030. The article argues that constrained chip supply, higher manufacturing costs, and rapid depreciation are pushing up prices, feeding inflation and pressuring AI company profitability, with added risk from debt-funded chip purchases and circular investment deals. It also warns that even a 90% drop in inference costs may not lower enterprise AI spend because agentic systems consume far more tokens per task.

Analysis

The market is still treating AI capex as a linear growth story, but the more important second-order effect is margin compression from the consumption side: if inference becomes the dominant workload, unit economics may worsen even as model costs fall. That means the bottleneck is shifting from semiconductor supply alone to enterprise willingness to pay, which creates a fragile bridge between hyperscaler capex and future revenue conversion. In that regime, the businesses with the most exposure to token monetization—not just chip demand—become the highest-beta equities. This is net positive for upstream compute vendors in the very near term, but the durability is less clear. If customers start rationing usage, the market could move from “more chips required” to “capex discipline required” within 2-4 quarters, especially for companies already leaning on circular financing or aggressive AI-forward guidance. The risk is not a collapse in demand outright; it is a slower and more painful reset where growth remains intact but returns on capital deteriorate. The more interesting competitive dynamic is that scarcity favors incumbents with balance-sheet capacity and procurement scale, while smaller software and hardware buyers get squeezed out. That widens the moat for the largest platform players, but it also raises the odds of regulator scrutiny around concentrated supply, debt-backed financing, and opaque partner commitments. In other words, the immediate winners may be the only entities able to absorb the cost shock, but the medium-term losers are the rest of the ecosystem that depends on cheap, abundant compute. Consensus is likely underestimating how quickly enterprise budgets can cap usage once AI stops being a novelty and becomes a line item. The market narrative assumes cost declines translate into adoption gains; the article suggests the opposite can happen if agentic workflows simply consume the savings. That makes efficiency breakthroughs a bearish catalyst for chip demand in the medium term, even if they look superficially bullish for adoption.