$1.5 Trillion in "Shadow Output": SemiAnalysis Says AI Is Creating Wealth Invisible to GDP

```

Investment in AI is already reflected in electricity bills, GPU orders, and data centers, but much of the value it generates may not yet be reflected in GDP.

According to a recent report by the well-known AI industry research institute SemiAnalysis, "AI Shadow Output: The Visible Costs of Invisible Output," AI is creating a macroeconomic statistical challenge: costs are visible, but the output may be invisible.

SemiAnalysis calls this phenomenon "AI Shadow Output." It is defined as: The economic value brought by AI already exists, but is invisible or seriously distorted in GDP, price indices, labor statistics, or industry accounts. Such "AI shadow output" may be as high as $1.5 trillion.

In other words: "AI output exists first in reality, and only then is it observed/measured."

The significance to the market is direct. AI capital expenditure, data center construction, electricity demand, GPU procurement, and token consumption all enter company accounts and are seen by investors. But if the work AI replaces or adds is not sold at a clear price, traditional macro data may only see the cost but not the corresponding output.

As AI growth leans more towards capital market financing, "any indicator that fails to show results from AI will be used to examine whether there are signs of a bubble."

To use the computer revolution as a reference: Nobel laureate Robert Solow once said, "You can see the computer age everywhere except in productivity statistics."

In 2013, a seemingly dull revision to U.S. GDP accounting methods included R&D and intellectual property investment, increasing total output by about $3.6 trillion in the 1990s, equivalent to nearly 30% of 2000's full-year GDP. Thus, the measurement problems brought by AI may be even greater.

Despite the bursting of the Internet bubble, the market value of the "Magnificent 7" is now 1.8 times that of Europe. The market's pricing of technological revolutions may anticipate macro-confirmed productivity growth.

What exactly does "$1.5 Trillion" mean?

The most watched number in the report is $1.5 trillion.

But SemiAnalysis makes clear: this is not "$1.5 trillion worth of labor has already disappeared," nor is it "GDP is undercounted by $1.5 trillion." It means: With the current generation of AI capability, there are credible market signals that tasks corresponding to about $1.5 trillion in labor costs have significant potential to be enhanced or automated by AI.

In other words, this is the "labor cost exposed to AI impact," not the scale of layoffs that have occurred, nor confirmed missing output.

This estimate is based on level 4 and above signals on the "evidence ladder." Weaker evidence includes benchmarks and product marketing; stronger evidence includes public corporate claims of AI in production, courts recognizing related AI work, and insurance companies willing to underwrite AI error risks.

SemiAnalysis also emphasizes that no level 5 or level 6 evidence has been observed so far. This "should be taken as a warning against overly optimistic AI narratives."

So far, most evidence points to "AI enhancement," not "AI full replacement." This is also a key limitation for interpreting the $1.5 trillion number.

Why would AI output "disappear" from GDP?

AI shadow output is divided into two categories.

The first type is "substitutional shadow output." That is, work previously done by humans, now done by AI.

Take legal documents as an example. For a simple legal document, whether drafted by a lawyer or by AI, the actual value to users may be similar. But in GDP statistics, the two paths differ.

If lawyers charge fees, that income enters the services statistics. If AI completes it with tokens, the lawyer's bill disappears, and there might only be a few dollars or even cents of token spending left in GDP, perhaps recorded under software or AI services, not legal services.

From a GDP standpoint, "this transaction essentially disappears, leaving only a few dollars of tokens in an unrelated sector."

SemiAnalysis gives a price change example: the drafting cost for a basic will has fallen long-term with technology, from $400 to $150 over about 30 years; dropping from $150 to $0.50 took just one year, a fall of over 99%. The former creates statistical bias, the latter may simply disappear from the data set.

Also, the legal services price index was not included in the CPI until 1987; as of September 2024, the index has risen 4.6 times. This index largely resembles the employment cost index, as it does not fully capture productivity improvements in legal services.

The second type is "incremental shadow output." That is, work no one did before because it was too expensive, but now is done because AI makes it affordable.

For example, when the cost of a literature review drops from $2,000 to $2, people might not just save $1,998, but might do a literature review for every project. Similarly, compiling six months of past emails, generating research material before interviews, or reviewing a contact's background ahead of meetings all create real value, but aside from tokens, API calls, cloud costs, or subscription fees, leave little clear economic record.

There are signs that a significant portion of current token spending is for new work, not just replacement of existing work. But the exact scale is not transparent, as these activities "are hidden behind the token's anonymous curtain."

Why are some AI outputs captured by statistics?

SemiAnalysis notes a third situation: captured AI output.

Example: A company previously paid $10,000 for outsourced HR services. Now, it pays $10,000 for AI HR services. The transaction price hasn't changed; GDP still records $10,000 spent, but with possibly fewer jobs.

But if the same HR task is done internally with $10 worth of tokens, the external $10,000 transaction becomes a $10 expense. GDP, thus, falls by $9,990, even though the same work is done.

Here's the crux: Whether AI is captured by GDP depends not only on whether it creates value, but whether it does so through visible prices, visible transactions, or recognized industry classifications.

Services are the biggest problem area

Manufacturing automation is easier to capture, since items are countable.

If factories produce screws more cheaply, statisticians can count the screws, measure lower costs, and see increased output. Over six centuries, the real price of screws fell over 99%, with global output rising dramatically; real GDP can capture such growth and productivity improvements fairly well.

But services are different.

SemiAnalysis writes: "We lack a usable language to describe units of service and cognitive labor." There is no "one ton of literature review," and no "barrel of consulting." Service industry statistics usually rely on receipts, wages, hours, and price indices, from which they derive "quantity."

This may work when prices change slowly. But if AI causes service costs to collapse by 99%, the system may read lower income as lower output, not higher productivity.

For example, in the Industrial Revolution, "horsepower" let people compare machine, animal, and human output; tokens cannot do this. One million tokens may produce junk, or a useful email summary, a legal document, or a strategic decision. Economic value depends on output, not token count.

An observable signal: Fewer jobs, but higher average wages

AI shadow output can't be directly observed, but may leave "fingerprints" in other data.

For example: Entry-level employees are most affected.

If AI first replaces routine jobs and low-paid entry-level positions decrease, the average wage for remaining employees may rise. "The cheapest workers disappear from the data; nobody gets a raise, but average wages go up."

SemiAnalysis has observed that sectors highly exposed to AI employment underperform the overall economy, but these underperforming sub-sectors actually see relative wage increases. The report sees this mismatch between jobs and wages as one of the statistical traces of AI substitution.

Another signal comes from token usage.

Anthropic's March 2026 Economic Index shows 37% of tokens are used in computer and mathematics-related fields.

Yet software investment's GDP contribution has not broken free of its pre-AI trend, and hasn't set new highs.

This may be an early sign of incremental shadow output: extensive AI use is happening, but not yet clearly reflected in traditional GDP software investment.

Why the market should care about this statistical problem

Macroeconomic data is still the best tool for observing the economy. Investors use it to judge real prosperity, policymakers to weigh unemployment and inflation, businesses for hiring, automating, or expanding production.

But if AI breaks the "labor—output—price—industry" data chain, decisions may get worse.

New Fed chair Walsh said in December 2025: "If you're looking at the data, you're looking in the rearview mirror. You'll be late. You won't realize the country can grow faster without inflation. So you have to bet."

If AI's output shows up before statistical confirmation, investors and policymakers may, for some period, only see two types of data: on one side, rising data center, GPU, electricity, water and token spending; on the other, employment shifts, wage structure changes, and changes in service income.

In this case, AI may look like a cost increase, or even falling output, in data; but in firms and user experience, some work is already being completed more cheaply and faster.

Where might the statistics break?

SemiAnalysis lists four main points of breakage.

First, boundary shifts. Services once purchased on the market become handled internally by businesses or households using AI. Paid research briefs become internal AI workflows; outsourced tasks become prompts for employees. The value may still exist, but the transaction that made it visible disappears.

Second, price collapse. Services have no standalone units of quantity or quality. If statistics only see fewer receipts, higher average wages, they may read "higher inflation, lower productivity and output."

Third, industry mismatch. AI might create value in one industry, but income is recorded in another. For example, a hospital uses AI for administrative work, which manifests as efficiency in medical services, but the transaction shows only as revenue for an AI or software company. This, the report says, makes AI vendors appear to be the value source, while AI-adopting industries look stagnant.

Fourth, new work is hard to see. For example, preparing meeting profiles with AI costs just a few tokens, but may help you prepare better. The report says such value leaves no clear receipt, and is hard to show up in traditional accounts.

This isn't the first time statistics have missed economic activity

Research shows the UK's Office for National Statistics once estimated household production was equal to 63.1% of measured GDP; the ILO estimates there are 16.4 billion hours of unpaid care work globally each day, worth $11 trillion annually, triple the global tech industry.

GDP boundaries have long missed productive activity without market transactions. AI may shift more information work from the "priced and measurable" area to the "output exists but no clear price" area.

However, SemiAnalysis's current framework only measures market labor with wages, employment, and task classification—not unpaid care, household production, or informal-economy AI usage. This is due to data auditability limits, not absence of the phenomenon.

What the monitor can and can’t say

SemiAnalysis emphasizes its "shadow output monitor" is a stress map, not a layoff or shadow output forecast.

It tracks tasks, professions, wages, evidence level, token costs, and the possible full-time position replacement scale. These are labor- and input-side indicators: enough to show where substitution economics have appeared.

But high-exposure sectors doesn't mean those jobs are already gone. You should interpret this as: tasks are identifiable, wage pools are large, market evidence is strong, token costs are low—the economic conditions for substitution are visible.

As for whether there will be large-scale substitution, or whether expanded demand, AI enhancement, and new tasks absorb shocks, it still needs time to validate, says the report.

SemiAnalysis concludes: "Shadow output is not a reason to deny AI's costs—it's a reason to measure the other side of the ledger."

In other words, AI's impact on land, electricity, water, chips, jobs, and token spending is already seen by the market. But the question is: if AI is creating changes on an industrial revolution scale, does economic data also need to observe output without clear invoices, traditional units, or stable industry categories?

Risk Disclaimer and Limitation of LiabilityThe market has risks; investment must be cautious. This article does not constitute personal investment advice and does not consider the unique investment objectives, financial situation, or needs of any individual user. Users should consider whether any views, opinions, or conclusions in this article are appropriate for their individual situation. Acting on this content is at your own risk. ```