Training frontier AI models now costs tens to hundreds of millions of dollars and requires compute that grows roughly 4-5x per year, concentrated in three to four companies. Labor market evidence is early but shows measurable productivity gains for knowledge workers alongside documented displacement in lower-skill tasks. Data centers consumed 1-2% of global electricity in 2023 and Goldman Sachs Research estimates this rises to 3-4% by 2030, with water consumption a largely untracked second externality.
AI economics covers the cost structure of building and operating AI systems, who captures the value, and how compute investment is distributed. The jobs question asks whether AI augments human workers or displaces them, and which workers are most affected. The environmental question asks how much energy, water, and carbon AI consumes across the full lifecycle: chip fabrication, data center construction, model training, and inference at scale.
These are not independent questions. Compute concentration determines who can build frontier models, which determines market structure. Energy demand from AI is growing fast enough to reverse decades of flat electricity demand in the US and Europe, forcing utility investment decisions that will shape the grid for decades. Labor displacement creates political pressure that shapes the regulatory environment. And the costs of training directly determine which organizations can participate in frontier AI development.
Training compute for notable AI models has grown approximately 4-5x per year since 2010, according to Epoch AI's analysis of 333 models published through May 2024. GPT-3 (2020) used roughly 3 x 10^23 FLOP. GPT-4 (2023) is estimated at 2 x 10^25 FLOP. Gemini Ultra (2023) at approximately 5 x 10^25 FLOP. Each generation is roughly 4-5x larger than the previous.
In dollar terms, training GPT-4 is widely estimated to have cost in the range of $50-100 million in compute alone, based on reported hardware and cloud pricing. Frontier models training on 2026 hardware (H100/H200 clusters) at scale cost more. The critical dynamic is that only organizations with access to large GPU clusters can participate. As of 2026, this means primarily OpenAI (backed by Microsoft), Google DeepMind, Anthropic (backed by Amazon and Google), Meta AI, and xAI. A handful of Chinese labs. No meaningful academic participation at the true frontier.
Hardware concentration mirrors model concentration. NVIDIA dominates AI accelerator supply with roughly 80%+ market share in training chips. The H100 and H200 GPUs (and successor Blackwell architecture) are produced by TSMC in Taiwan on 4nm and 3nm processes. Export controls imposed by the US in 2022 and expanded in 2023 restrict sales of advanced AI chips to China, which has responded by accelerating domestic chip development through Huawei and others.
Inference costs are lower per query but aggregate to significant sums at scale. A ChatGPT query requires approximately 2.9 watt-hours of electricity versus 0.3 watt-hours for a Google search, according to the International Energy Agency, roughly a 10x difference. At billions of queries per day, this adds up.
The labor economics of AI is contested. Early evidence divides roughly as follows:
Productivity augmentation evidence:
GitHub Copilot studies showed developers completing tasks roughly 55% faster. Experimental studies of AI writing tools show measurable output gains for knowledge workers. A 2023 study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond examining a customer-service AI tool found a 14% average productivity increase, with the largest gains for lower-skilled workers. This suggests AI can reduce skill differentials within a job category.
Displacement evidence:
The McKinsey Global Institute estimated in 2023 that generative AI could automate 60-70% of work activities in some occupations. The most exposed roles are those involving routine information processing, document drafting, and data entry. Translation, paralegal support, and entry-level coding are already experiencing headcount constraints. Fiverr and Upwork reported reduced demand for certain freelance categories post-ChatGPT.
The augmentation vs. displacement divide by job type. Higher-skill cognitive work (research, strategy, engineering) tends toward augmentation: AI tools raise productivity without eliminating the judgment component. Lower-skill information tasks (transcription, basic data entry, template drafting) face more displacement risk. Manual labor requiring physical dexterity is currently less exposed. Healthcare, education, and trades are considered more resilient.
Historical context:
Prior technology transitions (electrification, computing) displaced some occupations while creating others. AI may follow the same pattern over decades while causing short-term disruption. Whether AI is categorically different because it targets cognitive work specifically is a live research question without resolved consensus.
Electricity:
Global data centers consumed approximately 200-250 terawatt-hours per year through roughly 2020, roughly 1% of global electricity, largely flat due to efficiency improvements. Since 2020, AI workloads have broken this plateau. Goldman Sachs Research estimates data center power demand will grow 160% by 2030, rising from 1-2% of global electricity to 3-4%. In the US, data centers used about 3% of national electricity in 2022 and Goldman expects this to reach 8% by 2030. A single ChatGPT-scale query requires nearly 10x the energy of a Google search. By 2028, Goldman analysts project AI will represent about 19% of data center power demand.
The CO2 implications are significant. Goldman Sachs estimates data center carbon emissions may more than double between 2022 and 2030 even accounting for renewable energy investments, at a present-value "social cost" of $125-140 billion.
Water:
Less visible than electricity but measurable. Data center cooling systems use large amounts of water. Training GPT-3 in Microsoft's US data centers directly evaporated approximately 700,000 liters of freshwater, according to research published by Li et al. (arXiv:2304.03271, accepted Communications of the ACM). Globally, AI demand is projected to account for 4.2-6.6 billion cubic meters of water withdrawal in 2027, more than the total annual withdrawal of 4-6 countries the size of Denmark, by the same research. Water stress is location-dependent: data centers in arid areas draw from scarcer supplies.
Training vs. inference. Training a large model is a one-time (though repeatedly iterated) compute spike. Inference, serving millions of queries daily, accumulates steadily. For deployed models with large user bases, cumulative inference energy likely exceeds training energy within months to years of launch.
Carbon intensity:
Varies dramatically by grid. A data center powered by Icelandic geothermal or Norwegian hydro has near-zero operational carbon. One powered by coal-heavy grids in parts of the US Midwest or Southeast has much higher emissions. Tech companies are investing in renewables and entering power purchase agreements, but they are also increasing total consumption faster than they are decarbonizing.
FLOP (floating-point operation):
The standard unit for measuring training compute. A model trained on 10^23 FLOP used roughly that many arithmetic operations. Compute is the most legible proxy for model size and cost.
Inference:
Running a trained model to produce outputs (answers, images, recommendations). As opposed to training, which produces the model itself.
Power Usage Effectiveness (PUE):
The ratio of total data center energy to IT equipment energy. A PUE of 1.0 is perfect; most modern data centers run 1.1-1.5. Hyperscalers (Google, Microsoft, Meta) achieve close to 1.1.
Compute concentration:
The degree to which training compute is held by a small number of actors. Highly concentrated as of 2026, with barriers to entry including capital cost, chip access, and engineering talent.
Export controls:
US Commerce Department restrictions (October 2022, updated 2023) limiting export of advanced AI chips (NVIDIA H100 and above, AMD MI300 class) to China, Russia, and other designated countries.
Epoch AI's 2024 analysis of 333 notable models found consistent 4-5x per year compute growth since 2010, with the largest current models (GPT-4, Gemini Ultra) landing where this trend would predict.
Goldman Sachs's 2024 research reports documented the trajectory from flat data center power consumption through 2020 to the current surge, projecting 200 TWh of incremental annual demand by 2030 attributable to AI.
Li et al. (2023) provided the first systematic methodology for estimating AI water footprints, finding that GPT-3 training consumed 700,000 liters of freshwater and projecting global AI water withdrawal to exceed 4 billion cubic meters annually by 2027.
NVIDIA's revenue grew from roughly $27 billion in FY2023 to over $130 billion in FY2025, driven almost entirely by data center AI chip sales, illustrating how AI compute spending concentrates in the hardware supply chain.
"AI will eliminate most jobs in five years."
No credible research supports this specific timeline. Evidence supports productivity shifts and displacement in specific task categories, not wholesale elimination of employment at five-year horizons.
"AI will create as many jobs as it destroys."
Historically true for prior technology transitions, but the pace of this transition may not allow for smooth reabsorption. Distribution matters: the people displaced may not be the same people hired for new roles, and geographic and education mismatches are real.
"Training cost is the main environmental impact."
For widely deployed models, cumulative inference energy quickly surpasses training energy. The conversation has focused disproportionately on training because it is a visible, dateable event.
"Renewable energy commitments mean AI is carbon neutral."
Additionality matters. Buying renewable energy certificates from existing capacity does not reduce grid emissions. Real decarbonization requires building new clean generation capacity that would not otherwise exist.
"Efficiency improvements will solve the energy problem."
Data center efficiency has improved consistently for decades, but absolute consumption has still risen because deployment scale grows faster than efficiency gains. This is a standard rebound effect.
Estimates of AI's electricity demand have firmed up since the Goldman Sachs figures above. The International Energy Agency's 2025 Energy and AI report put global data-center electricity use at about 415 terawatt-hours in 2024, roughly 1.5 percent of world supply, and projected it to more than double to around 945 terawatt-hours by 2030, with AI as the main driver. The IEA's follow-up Electricity 2026 report found data-center demand grew 17 percent in 2025 alone, and power use at AI-focused sites rose about 50 percent that year. The United States and China together account for nearly 80 percent of the projected growth.
For the US, a December 2024 Lawrence Berkeley National Laboratory report (Shehabi et al.) estimated data centers used 176 terawatt-hours in 2023, about 4.4 percent of national electricity, and could reach 325 to 580 terawatt-hours by 2028, between 6.7 and 12 percent of US supply.
The harder limit is the grid itself. The IEA warns that without major transmission investment, up to 20 percent of planned data-center projects risk delay. Interconnection waits in hotspots like Northern Virginia reportedly run several years, and long-lead equipment such as transformers has become the bottleneck rather than chips. Announced capacity far exceeds what utilities can actually connect.
The open question is whether efficiency gains, like DeepSeek's cheaper 2025 models, curb demand or feed it. Many analysts point to the Jevons paradox, where cheaper compute raises total use rather than lowering it (arXiv:2501.16548). Treat every 2030 figure here as a projection, not a forecast.
Every claim traces to a cited source below.