what if ai fails to deliver ? | 1 | **Capex/ROI mismatch** The master risk. The industry is spending as if AI revenue will arrive quickly and at high margins. Goldman’s baseline model implies **$765B annual AI capex in 2026**, rising to **$1.6T in 2031**. If enterprise ROI is slow, the denominator of the ROI equation explodes before the numerator arrives. | 2 | **Hyperscaler capex reversal** The most important near-term trigger. If Microsoft, Google, Amazon, Meta, Oracle, or CoreWeave slows AI capex guidance, the market immediately questions Nvidia demand, data-center growth, cloud growth, and AI multiples. Microsoft disclosed **$37.5B Q2 FY26 capex**, with roughly **two-thirds on short-lived GPUs/CPUs**. | 3 | **Token price compression / open-source commoditization** Cheap models do not need to beat frontier labs everywhere. They only need to be good enough for ordinary enterprise workflows. OpenAI’s API page lists high-end text pricing in dollars per million tokens, while DeepSeek lists V4 pricing far lower, including extremely cheap cache-hit pricing. | 4 | **GPU depreciation and impairment risk** ai chips are short-lived assets. If useful economic life is shorter than accounting life, reported earnings overstate true economics. Microsoft explicitly said roughly two-thirds of capex was for short-lived GPUs/CPUs, and CoreWeave’s Q1 2026 results already show **$1.147B depreciation and amortization** in one quarter. | 5 | **Utilization gap: capacity before workloads** Capacity is being built before enterprise workflows are fully proven. If clusters sit underutilized, providers discount pricing, resale values fall, and debt/capex ROI collapses. This is the classic infrastructure-bubble failure mode: demand is real, but timing and scale are wrong. | 6 | **Nvidia + Mag-7 concentration risk** The S&P 500 is highly exposed to AI-premium names. Nvidia’s own filing showed four direct customers accounted for **22%, 17%, 14%, and 12%** of accounts receivable as of Oct. 26, 2025, showing concentrated payment channels even if end demand is broader. ([SEC][4]) | 7 | **Infrastructure credit / GPU-backed debt risk** The AI boom is becoming a credit story. CoreWeave reported **$99.4B backlog**, but also **$536M interest expense**, **$740M net loss**, and heavy depreciation in Q1 2026. If utilization or customer demand weakens, GPU-backed debt and data-center finance become fragile. | 8 | **Enterprise ROI evidence remains weak** Adoption is not the same as ROI. NBER found most firms report little measurable impact so far, including **89% reporting no impact on labor productivity** over the past three years. PwC’s 2026 CEO Survey found many CEOs still do not see AI financial benefits. | 9 | **Software gross-margin leakage** AI may make software better but margins worse. SaaS companies historically enjoyed very high gross margins, but AI introduces inference costs, model costs, monitoring, and support. If customers expect AI features to be included, vendors absorb cost without proportional pricing power. | 10 | **Agentic AI cost explosion** Agents use many more tokens than chatbots. GitHub is moving Copilot to usage-based billing from **June 1, 2026**, with usage calculated from input, output, and cached tokens. That is a clear signal that agentic workflows are too compute-intensive for simple flat-rate economics. | 11 | **Power/grid/social-license backlash** AI infrastructure needs electricity, land, water, substations, and local political permission. IEA projects data-center electricity consumption will reach around **945 TWh by 2030**, growing about **15% per year**. Data Center Watch says **$64B** of U.S. data-center projects have been blocked or delayed. | 12 | **Model routing destroys premium inference mix** Enterprises will not use premium frontier models for everything. They will route easy tasks to cheaper models and reserve premium models for hard tasks. That means usage can rise while premium revenue mix falls. This is especially dangerous if capex was justified by high-priced frontier inference. | | 13 | **Data cleanup + integration debt** AI demos are easy; enterprise integration is hard. Messy permissions, legacy systems, poor metadata, siloed data, and bad process design mean companies need data engineering and workflow redesign before AI creates real ROI. This cost is often ignored in capex optimism. | 14 | **Verification/rework tax** Workday found nearly **40% of AI time savings are lost to rework**, including correcting, rewriting, and verifying outputs; only **14%** of employees consistently see clear positive net outcomes. That directly reduces net productivity. | 15 | **AI features become table stakes** Customers may expect AI inside every product without paying much extra. That converts AI from a pricing-power tool into a defensive cost of doing business. This is dangerous because the capex was spent assuming incremental monetization, not merely product maintenance. | 16 | **Backlog quality / circular demand risk** Backlog only matters if it is durable, profitable, collectible, and not circularly financed. If AI labs, cloud providers, chip vendors, and neoclouds are funding each other’s demand, headline revenue may overstate true end-customer economics. | 17 | **Legal/IP/data-licensing tolls** Anthropic’s proposed **$1.5B author settlement** and California’s training-data transparency rules show the free-data era may become more expensive and legally constrained. That raises model-development cost and can slow scaling. | 18 | **On-prem/private AI drift** Enterprises may still adopt AI, but through private, hybrid, open-weight, or on-prem systems. That can weaken centralized API margins and reduce cloud vendor pricing power, even while total AI usage rises. | 19 | **Customer concentration** AI supply-chain demand is more concentrated than it appears. Nvidia’s receivable concentration shows that a small number of direct customers matter heavily. If a few hyperscalers or integrators slow orders, the revenue impact can be fast. ([SEC][4]) | 20 | **Procurement discipline and usage caps** As AI bills become visible, CFOs will demand usage caps, cheaper models, vendor competition, and outcome-based pricing. GitHub’s move to usage-based billing makes token consumption financially explicit, which will train customers to optimize costs. | 21 | **Search/ad cannibalization** AI search may reduce traffic to websites and pressure the content ecosystem. Pew found Google users clicked traditional links in **8%** of visits with an AI summary versus **15%** without one. This can hurt publishers, content supply, and search monetization. | 22 | **Physical supply-chain bottlenecks** AI buildout needs more than GPUs: fiber, optics, transformers, substations, cooling, construction labor, and power generation. Bottlenecks raise cost and delay monetization, reducing returns on already-committed capex. | 23 | **Human accountability limits labor substitution** In legal, finance, healthcare, compliance, and high-risk operations, humans still need to review and approve outputs. That caps labor savings and turns AI into augmentation rather than full substitution. | 24 | **Synthetic-data / scaling limits** Epoch AI estimates the effective stock of quality-adjusted public human text at around **300T tokens**, with full utilization possible between **2026 and 2032** if trends continue. If public data becomes scarce, future model improvements may cost more. ([Epoch AI][12]) | 25 | **Consumer surplus captured by users, not providers** AI may be enormously useful but economically deflationary. Users and customers may capture the benefits through cheaper intelligence, while infrastructure owners compete returns away. This is the telecom-fiber analogy: society wins, overfunded capital loses. 1. AI vendor subsidizes usage during early adoption 2. Users build habits around cheap/flat-rate tools 3. Agentic workflows multiply token consumption 4. Vendor discovers gross margin pressure 5. Vendor shifts to credits/usage billing 6. Customer discovers true cost 7. CFO imposes budget caps 8. Teams route to cheaper models 9. Premium AI revenue growth slows 10. AI labs either cut prices or lose volume 11. Hyperscaler capex ROI weakens