The AI Economy: Capex, Power, Bubble Indicators, GDP Scenarios
This content was developed with AI assistance and is regularly reviewed for accuracy. It is not investment advice.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology. Most figures on this page come from analyst estimates and press coverage rather than company filings; they are marked "reported" and dated.
The honest one-line summary of the AI economy in October 2026 is: an investment boom, with productivity gains that are still mostly measured inside individual firms. Hundreds of billions of dollars a year are going into data centers, chips, and power. Official forecasters count that spending in their growth numbers, but their baselines do not yet assume AI makes the economy more productive. Whether that gap closes, and how fast, is the real question behind every "AI bubble" headline.
What you'll get here:
- How much is being spent on AI infrastructure, and who is spending it
- Why electricity, not chips, became the binding constraint
- A set of indicators for reading the bubble debate yourself, instead of a verdict
- The range of GDP scenarios for 2030, and what they imply for wages and jobs
The spending: a physical-infrastructure story
AI is often described as software. Economically, in 2026, it is construction. The largest cloud companies plus Oracle were projected to spend $635 billion to $690 billion on capital expenditure in 2026, an increase of roughly 67% to 74% over 2025. Goldman Sachs analysis puts cumulative spending at about $1.15 trillion for 2025 through 2027, and Dell'Oro projects worldwide data-center capex of $1.7 trillion a year by 2030 (all reported via analyst summaries; example).
The scale shows up in individual companies. Amazon was reported to spend $44.2 billion on capex in a single quarter, an annualized pace near $175 billion, while its trailing free cash flow fell to about $1.2 billion. On the private side, the Stanford AI Index 2026 counted $285.9 billion in US private AI investment (reported).
Meanwhile the price of using AI keeps falling. When Anthropic released its September 2026 flagship, it cut the per-token price by 20% against the previous generation and said typical workloads cost about 40% less (Anthropic). Spending rises because demand for compute rises faster than its unit cost falls. Current prices live on the AI Model Comparison page.
The binding constraint: power
In 2024 the bottleneck was chips. By 2026 it was electricity.
Prices. PJM, the grid operator covering much of the US mid-Atlantic and Midwest, cleared its capacity auction for 2026 to 2027 at $329.17 per megawatt-day, against about $28 two years earlier, a more than tenfold rise (the price hit a cap of about $325). The consultancy ICF estimates residential electricity rates could rise 30% to 60% by 2030, though measured average retail prices were up a more modest 2.3% year over year (reported; Bisnow).
Capacity. Wood Mackenzie reported in March 2026 that US data-center development slowed because the grid is reaching its limit: about 25 gigawatts were added to project pipelines in the last quarter of 2025, half the previous quarter's pace. One utility, AEP, reported 190 gigawatts of requests in its connection queue against a 37-gigawatt system. EPRI projects data centers could use up to 9% of US electricity generation by 2030, up from 4% in 2023, and Gartner warns that power shortages could restrict 40% of existing AI data centers by 2027 (all reported).
Power links the AI economy to your household bill and to the AGI debate. It is the most frequently cited brake on how fast capability can scale, which is why AGI: Definitions and the Timeline Debate treats it as part of the timeline question.
Reading the bubble debate as indicators
"Is AI a bubble?" has no single answer, and both sides have 2026 data. A more useful habit is to watch a handful of indicators and notice which way they move.
| Indicator | Bearish reading | Bullish reading |
|---|---|---|
| Capex versus cash flow | Spending is outrunning free cash flow at some hyperscalers (Amazon's trailing figure near $1.2B, reported) | Supporters argue the biggest spenders can fund it from profitable core businesses and slow down if demand disappoints |
| How it is financed | Bank of America flagged a possible "air pocket" in 2026 data-center debt (December 2025) | Supporters argue the debt risk is concentrated in specific projects rather than the whole build-out |
| Valuations | Critics such as Peter Schiff argue roughly $1 trillion a year is going into assets that may be obsolete in five to six years | Nvidia's forward price-to-earnings ratio near 22 is far below dot-com peaks (reported, March 2026) |
| Utilization | Capacity built for demand that never arrives would sit idle | Cloud and GPU capacity at Microsoft, Google, Amazon, and specialist providers was reported sold out |
| Power | Grid limits and rising prices could strand projects and provoke political backlash | Constraints slow the build-out, which can prevent overbuilding |
None of these settles the question alone. What would change the picture is a combination: utilization falling while debt financing rises and AI revenue growth slows. If you see those three together in the news, the bear case is strengthening. If revenue keeps growing into the capacity, the bull case is.
What official forecasters say
The International Monetary Fund raised its 2026 forecasts in January 2026 (global growth 3.3%, US growth 2.4%) "in part" because of AI infrastructure investment, while its chief economist warned that unmet productivity expectations "could spark a correction in high market valuations" (Reuters via Kitco).
The October 2026 World Economic Outlook is reported to hold global growth at 3.1%, with AI investment propping it up, and to model an adverse scenario in which unwinding AI investment cuts global output by up to 0.4% by 2028, or 0.6% with tighter financial conditions. Its baseline assumes no acceleration in productivity from AI (secondary summary; not verified against the IMF text). In other words, the official view counts the spending and waits for the productivity.
GDP scenarios for 2030
Because point forecasts are so uncertain, the most useful 2026 work presents ranges. Anthropic's Economic Scenario Explorer (September 2026) models US GDP in 2030 at 2025 prices under three scenarios (Anthropic):
| Scenario | GDP versus no-AI path | Labor's share of income | What it means for workers |
|---|---|---|---|
| Modest | +1.6% | 59.4% | Small changes to wages and jobs |
| Substantial | +8.3% | 56.1% | Knowledge-worker wages roughly flat |
| Extreme | +32.4% (about 15% annual growth) | 45.2% | Knowledge-worker wages down more than 10%; unemployment could spike |
The key uncertainty separating the substantial and extreme paths is whether AI becomes able to improve itself. Notice the pattern in the right-hand columns: higher growth comes with a smaller share of income going to workers. Faster growth and broadly shared gains are not the same thing, and which one happens is partly a policy choice.
People disagree sharply about which scenario is likely. In Anthropic's August-September 2026 polling, the median US adult's answers implied about 5.4% extra growth by 2030, people who took its online quiz implied 12.5%, and the median Anthropic employee implied 17.1% (Anthropic, The AI expectations gap). Workers who expect AI to do most of their work also expect much higher unemployment in 2030 than workers who do not.
Productivity: large inside firms, small in the statistics
At the level of people and companies, the gains are visible. In Anthropic's June 2026 survey of its users (not representative), 86% reported faster work, 82% broader scope, and 69% higher quality. Microsoft's 2026 Work Trend Index found 66% of AI users spend more time on high-value work. Inside Anthropic, code commits per engineer rose about fivefold in six months. Most of these figures are self-reported or confined to technology firms, and Goldman Sachs analysts have noted that measurable effects so far are mostly within the technology sector.
The gap between firm-level gains and national statistics is normal early in a technology's spread; electricity and computers both took years to show up in productivity data. It is also why the jobs question is better answered at the level of roles and career stages than with a single GDP number. AI and Your Career in 2026 covers that personal picture, from entry-level hiring to the AI-skills wage premium.
Key takeaways
- The AI economy in 2026 is an investment boom: $600 billion-plus of reported hyperscaler capex this year, with productivity gains still mostly firm-level
- Power is the binding constraint: capacity prices on the largest US grid rose more than tenfold in two years, and data-center pipelines slowed (reported)
- Read the bubble debate through indicators (cash flow, financing, valuations, utilization, power) rather than verdicts
- Official forecasters count AI spending but do not yet assume AI productivity gains; the IMF models a correction scenario
- 2030 GDP scenarios range from +1.6% to +32.4%, and the higher-growth paths come with a smaller share of income for workers
Next steps
- Continue to: AI and Your Career in 2026 for what these numbers mean for your job
- Continue to: AGI: Definitions and the Timeline Debate for the self-improvement question behind the scenarios
- Continue to: The Global AI Regulatory Landscape for how governments are responding