The Horrible Economics of AI Are Starting to Come Crashing Down

AI companies are facing rising infrastructure and compute costs, with Anthropic moving users to pay-as-you-go API billing after capacity constraints and usage spikes strained its systems. Gartner estimates the industry may need nearly $2T in annual revenue by 2029, and token consumption would need to rise 50,000 to 100,000 times by 2030 to support current economics. The article warns that price hikes or ads may be needed to offset data center capex, but could slow adoption and growth.

Analysis

The important shift here is not “AI is expensive” — it is that marginal demand for inference is starting to outrun the industry’s ability to monetize it without destroying usage growth. That matters most for the pick-and-shovel layer: GPU cloud, interconnect, power, and data-center infrastructure should keep outperforming on backlog visibility, while model-layer companies face a much harder optimization problem because every pricing action risks lowering token velocity just as capex intensity is peaking.

The second-order effect is that enterprise buyers will likely rationalize consumption before they abandon AI. That creates a near-term squeeze on low-ROI agentic workflows, internal copilots, and workflow automation vendors whose value prop depends on high call volumes; the usage-based model becomes a tax on experimentation. The winners inside software will be those that can compress tokens per task or bundle AI into subscription pricing, effectively subsidizing usage and preserving seat expansion.

The market is probably underestimating how quickly procurement teams will enforce AI budgets once token invoices become visible in P&Ls. That should slow the “AI everywhere” narrative over the next 2-4 quarters and pressure names whose growth assumptions embed unconstrained usage. The contrarian point is that this is not necessarily bearish for the ecosystem as a whole: higher prices can improve unit economics and reduce waste, which may ultimately favor the strongest platforms and infrastructure providers with pricing power and scale.

Tail risk is a capex-air-pocket if hyperscalers and model providers simultaneously slow spending to protect margins; that would hit semiconductor lead times and data-center supply chains with a lag. But the more likely path is a bifurcation: premium customers keep paying, while marginal users get rate-limited, downgraded, or pushed to smaller models. That makes this more of a dispersion trade than a blanket AI short.

Key Decisions for Investors

• Long infrastructure over application-layer AI: buy a basket of NVDA/MSFT/AMZN on any 3-5% pullback, with a 3-6 month horizon; thesis is that pricing pressure shifts spend toward the scarce layer with the most pricing power and backlog visibility.
• Short or underweight AI-native software names with usage-based monetization and weak gross-margin protection for a 1-2 quarter window; focus on companies where AI is a cost center disguised as growth, as customer throttling can hit revenue before management guides to it.
• Pair trade long AMT/CCI-equivalent data-center-adjacent infrastructure exposure versus short a basket of compute-heavy SaaS names; risk/reward favors the infrastructure side if enterprise token budgets tighten and capex remains locked in.
• Buy call spreads on NVIDIA into any post-earnings or post-capex selloff; the risk is multiple compression if hyperscalers pause orders, but the upside remains intact if the market continues to re-rate AI compute as a utility bottleneck.
• Avoid chasing AI agent beneficiaries until we see evidence of enterprise willingness to pay; if tokens are being rationed, the first products cut are discretionary automation layers, so wait for confirmed budget line items before adding exposure.