The stack, not the magic

Dario Amodei (Anthropic CEO) told Axios in May 2025 that AI could eliminate roughly 50% of entry-level white-collar jobs within 1-5 years and push unemployment to 10-20%. It is the starkest public prediction from a frontier-lab executive with direct commercial incentive to be optimistic. Whether the forecast is right or directional, the fact that the CEO of an AI company is making it is itself a data point about where the builders think this lands.

Challenger, Gray & Christmas: AI-attributed layoffs accounted for ~4.5% of total 2025 US layoffs. An HBR January 2026 survey of 1,006 executives found most AI-linked layoffs were "in anticipation of" rather than "in response to" measurable AI-driven productivity gains. The gap between stated rationale (AI) and measured productivity (limited) is the "AI washing" phenomenon: layoffs that were going to happen anyway, given an AI-shaped cover story.

Job postings for entry-level writing and editing roles fell sharply after ChatGPT launched in late 2022. Junior software-engineer postings have grown far slower than overall tech employment. Customer-service rep postings declined in industries that deployed chatbots aggressively. Aggregate unemployment for college-educated workers stayed low; the displacement is showing up as slower entry-level hiring and elevated output expectations per remaining employee, not mass firings. The pipeline into senior roles is narrowing.

Goldman Sachs / OpenAI estimated ~30% of US occupations have tasks substantially exposed to LLM automation; legal, finance, software engineering, and administrative support rank highest. Acemoglu (MIT) and other labor economists have produced lower estimates (5-10% realistic impact over a decade), but all credible estimates are non-zero and concentrated in the same white-collar mid-skill band where wage premiums have historically been highest.

NVIDIA holds approximately 80% of the AI accelerator market by revenue (higher by training-specific measures, lower for inference where alternatives like Google TPU and AWS Trainium take share). The company's data-center segment revenue grew from ~$15B in FY23 to >$100B in FY25. NVIDIA's CUDA software lock-in is the most durable piece of the moat; hardware alternatives exist but porting costs mean even well-funded buyers (Meta, Microsoft) run mostly NVIDIA.

TSMC manufactures roughly 90% of the world's leading-edge semiconductors (≤5nm nodes), including essentially all advanced AI chips (NVIDIA H100/H200/B100, Apple Silicon, AMD MI300). Samsung has a minority share; Intel 18A is not yet at competitive volumes. The Taiwan-strait geopolitical concentration of the most strategically critical chip supply chain is unmatched in any industry.

OpenAI, Anthropic, Google DeepMind, and (distant fourth) xAI + Meta together account for essentially all frontier-model releases as of 2025. Training a frontier model now costs $100M-$1B+ and requires access to tens of thousands of H100-class GPUs, which NVIDIA allocates by relationship. Open-weight models (Llama, Mistral, Qwen) lag the frontier by 6-12 months on benchmarks. The 4-lab concentration replicates the pre-antitrust Big Tech shape at a higher capital intensity.

Synergy Research: AWS (~32%), Microsoft Azure (~22%), and Google Cloud (~11%) together take ~65% of global cloud infrastructure spend. AI workloads intensify the concentration because training and serving frontier models requires data-center footprints only the top three can deliver at scale. OpenAI runs on Azure; Anthropic on AWS + GCP; Meta and Google run their own. The 'AI economy' is functionally the hyperscaler economy with a new product layer on top.

EPRI and Lawrence Berkeley National Lab: US data-center electricity consumption is projected to reach 9-12% of national total by 2028, up from ~4% in 2023. The growth is almost entirely AI training and inference. Utilities are deferring coal retirements and signing multi-billion-dollar long-term contracts with hyperscalers; in some grids (Northern Virginia, Georgia, Ohio) residential rates are rising to cover infrastructure build-outs driven by data-center demand.

Epoch AI: estimated training costs for frontier models passed $100M with GPT-4 (2023), crossed $500M with Gemini Ultra and GPT-5-class models (2024), and are projected to reach $1B+ per training run by 2026. The cost is dominated by GPU-hours; a single training run uses 20,000-100,000 H100-equivalent GPUs for months. Doubling cost per generation cannot continue indefinitely; the scaling-law extrapolation that justifies the spend has weaker empirical footing past ~2× current frontier.

Independent benchmarks: a typical LLM chat response consumes roughly 10 Wh of end-to-end compute energy, compared to ~0.3 Wh for a Google web search. For large models and long responses the gap is 50-100×. At the hundreds-of-millions-of-daily-queries volume ChatGPT and competitors now serve, the aggregate electricity and water-cooling demand is meaningful at regional grid scale.

Northern Virginia and Phoenix-area AI data centers report water consumption (mostly evaporative cooling) of 3-6 million gallons per facility per day during peak summer operation. Microsoft's 2023 sustainability report disclosed 34% year-over-year global water-withdrawal growth tied to AI workloads. Water-stressed regions (AZ, NV, TX) are seeing local disputes as data-center siting competes with residential and agricultural use.

PJM Interconnection's 2025/26 capacity auction cleared at $269.92/MW-day, up from $28.92/MW-day the prior year, a 9× jump driven primarily by data-center load growth colliding with deferred coal retirements. PJM's auction pays generators a fixed amount for being available to meet peak demand; the cost flows through to retail bills across the grid's 13-state, 65-million-person footprint (Mid-Atlantic + much of the Midwest). The aggregate uplift on the 2025/26 auction is roughly $14.7 billion. Monitoring Analytics, PJM's independent market monitor, attributed the spike to capacity scarcity created by AI-driven demand growth outpacing new generation interconnection.

PJM has approved or accommodated retirement deferrals on roughly 3.6 GW of coal capacity since 2023, with hyperscaler load growth cited as the proximate cause. Indianapolis Light & Power, AEP Ohio, and Dominion all pushed planned coal exits into the 2030s. The deferrals lock in emissions that the affected utilities had previously committed to retire and complicate state-level decarbonization targets. The trade is explicit: keep coal running so that the AI capex cycle has firm-power headroom.

Sherwood News (Apr 29 2026) reports 14 states are actively considering moratoriums on new data-center builds, with residential electricity prices projected to rise 15-40% over the next 5 years across affected grids. Local pushback has run ahead of federal policy: Indianapolis IN extended its 2024 moratorium through 2025; Prince William County VA paused approvals on the 23,000-acre 'Digital Gateway' corridor and Fairfax, Loudoun, and Stafford counties tightened zoning rules; Atlanta-metro suburbs Dunwoody, Peachtree Corners, and Norcross GA passed full moratoriums while Marietta GA imposed a study pause; Tucson AZ effectively blocked water-intensive builds via a 2024 water ordinance, with Chandler and Mesa AZ raising water-use thresholds; Memphis TN has open public-health complaints over xAI's Colossus gas-turbine emissions. State consumer advocates in MD, NJ, and OH have filed FERC complaints arguing the PJM auction cost-shift to residential ratepayers is unjust. The externality is land + water + grid, all governed by the city, county, and state authorities federal AI policy doesn't touch.

Sherwood News reporting (Apr 29 2026) cites projected 15-40% residential electricity price increases over the next 5 years driven by data-center load growth. Hyperscalers (Alphabet, Amazon, Meta, Microsoft) are spending >$650B on data centers in 2026 alone; data-center electricity demand is expected to double by 2030 to roughly the combined draw of France + Germany. ~2,300 GW of generation and storage is stuck in interconnection queues, more than the entire installed US power-grid capacity, so most of the new demand has to be served by either deferred fossil retirements or capacity-auction premiums. Both routes flow through to retail bills.

Stanford HAI 2024 study "Hallucinating Law": tested GPT-4, Claude, and PaLM on ~200,000 legal queries. Hallucination rates ranged from 58% to 82% depending on model and query type, even for purpose-built legal assistants. The authors concluded that "current legal LLMs are substantially incapable of serving as reliable research assistants." Courts have sanctioned multiple law firms for filings that contained fabricated case citations.

NewsGuard and independent fact-checking groups: approximately 40% of news content produced by unsupervised LLM pipelines (auto-summarization, auto-rewrites) contains factual errors or unverifiable claims. The rate is higher for time-sensitive content (financial news, politics). The error rate has not meaningfully improved over 2023-25 despite model upgrades; the failure modes shift but the aggregate unreliable-output rate is roughly stable.

Obermeyer et al. (Science 2019) showed that widely-deployed clinical risk-score tools systematically under-prioritized Black patients for care. Follow-up audits through 2024 have found similar patterns in FDA-cleared AI diagnostic tools: training data skewed toward specific demographics produces quieter failures at worse rates for under-represented populations. The FDA's 2024 AI/ML guidance requires reporting of demographic performance gaps; compliance and auditing remain uneven.

Independent reviews: fewer than 3% of US firms have publicly reported measurable ROI from generative-AI deployments as of 2025; most cite productivity improvements without quantified bottom-line impact. MIT's Sloan Management Review 2024 survey: ~75% of enterprises report ongoing AI pilots, ~15% report production deployments, and a minority report measurable efficiency gains. The gap between investment levels (record enterprise spending) and realized returns is structurally similar to the dot-com era 1998-2000.