McKinsey says applying AI to existing work could be worth $2.52 trillion a year to the global economy, and the same analysis shows how quickly that prize erodes. In a modeled scenario, a 30% rise in effective AI costs cuts the global value pool to $1.67 trillion, a loss of roughly $850 billion, or 34%, as AI Affairs summarizes the findings. The report is titled The AI value curve: The economics of applying AI to existing work.
Headlines round the number to $2.5 trillion. The report itself says $2.52 trillion, and that gap matters less than the story underneath it. The Economic Times and NDTV Profit both carried the figure, and the angle that travels is the one buyers care about: AI’s economic upside depends on what it costs to run, check and fix.
That framing lands at an awkward moment. Model makers are shipping new releases at a pace we tracked in our look at the AI infrastructure race, and the bills for compute are climbing alongside them. McKinsey’s contribution is to put a dollar value on what happens if the cost side of the ledger moves the wrong way.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the McKinsey AI Value Curve Report Actually Says
The McKinsey AI value curve report estimates a $2.52 trillion annual global economic opportunity from applying AI to work that people already do. The mechanism is twofold. AI can automate some tasks outright, and it can augment workers on others, so the same job gets done faster or at higher quality. The estimate covers certain sectors, regions and occupations rather than the whole economy.
The word “existing” in the title carries weight. This is not a forecast of brand-new AI-native industries or products nobody has invented yet. It is an accounting of value locked inside current workflows, from drafting and analysis to customer handling and coding, if AI absorbs part of the labor.
The report’s second move is the one that made headlines. McKinsey does not treat AI cost as a fixed input. It defines cost broadly and then asks what happens to the value pool when that cost rises. The answer is a steep curve, which is where the report gets its name. Value falls faster than most executives assume once the full cost stack is counted.
The 30% Cost Shock: Numbers Behind the $850 Billion Loss
The headline scenario is simple to state. Raise effective AI costs by 30%, hold everything else constant, and the global annual value pool drops from $2.52 trillion to $1.67 trillion. The gap is approximately $850 billion, which McKinsey puts at 34% of the original opportunity.
Look at the ratio. A 30% cost increase produces a 34% value loss, so each point of cost growth removes about 1.13 points of value. That is my arithmetic on McKinsey’s published figures, not a number from the report. It suggests the value pool is slightly more than proportionally sensitive to cost, because the marginal use cases at the edge of viability drop out first and take their entire contribution with them.
| Metric | Baseline | After 30% cost rise | Change |
|---|---|---|---|
| Global annual value pool | $2.52 trillion | $1.67 trillion | About -$850 billion (-34%) |
| US annual value pool | $960 billion | $640 billion | About -$320 billion (-33%) |
| Rest of world (derived) | $1.56 trillion | $1.03 trillion | About -$530 billion (-34%) |
| US share of global pool (derived) | 38.1% | 38.3% | Roughly unchanged |
| Effective AI cost index | 100 | 130 | +30% |
The rows marked derived come from subtracting McKinsey’s US figures from its global figures. They are simple subtraction, but they are mine, so treat them as an illustration of proportion rather than as McKinsey regional estimates.
Why the US Loses $320 Billion in the Same Scenario
For the United States, the same modeled increase cuts the annual value pool by approximately $320 billion, from $960 billion to $640 billion. By my arithmetic on McKinsey’s figures, the US holds roughly 38% of the global pool, a share that barely moves under stress.
That stability tells you something. The cost shock is not hitting one region harder because of local wages or adoption rates. In McKinsey’s modeling, it hits everywhere at close to the same proportion. The US simply has more value to lose in absolute terms, since its pool is the largest single-country slice of the total.
What the rest of the world loses
Subtracting the US from the global totals leaves roughly $1.56 trillion of baseline value outside America, falling to about $1.03 trillion. That is a drop near $530 billion. For readers in Europe, India or Southeast Asia, the practical reading is that cheap, reliable AI is not an American concern alone. Anyone whose business case rests on a fixed per-task cost is exposed to the same curve.
A Modeled Scenario, Not a Forecast
McKinsey is explicit that these figures are a modeled scenario, not a forecast. It did not predict that costs will rise 30%. It asked what a 30% rise would do, and reported the arithmetic. Reports on the study repeated that caveat, and it deserves repeating here, because stress-test numbers have a habit of being quoted as predictions a week later.
The distinction shapes how to read the $850 billion. It is not money that will vanish from the economy. It is a measure of how much of the opportunity depends on cost staying in a certain range. Costs could fall, hold steady or rise, and the scenario just gives each outcome a price tag.
Policy economists are working on the same problem from a different direction. The Federal Reserve Bank of Chicago has published a 2026 working paper titled Forecasting the Economic Effects of AI, an example of how central bank researchers now treat AI’s economic impact as a modeling exercise. Its findings are separate from McKinsey’s, and this article does not draw on them.
What Counts as an AI Cost: Tokens Are Only Part of It
The most useful part of the report may be its definition of cost. McKinsey counts model and token costs, then adds human review, validation, verification, and rework or risk activities. That is a wide net. A per-token price cut at a model vendor addresses only the first item.
| Cost component | What it covers | Who mostly controls it | Typical lever |
|---|---|---|---|
| Model and token costs | Inference spend per request | Model vendor and buyer | Cheaper models, caching, routing |
| Human review | Staff reading AI output | Buyer | Narrower scope, better prompts |
| Validation | Checking output meets a standard | Buyer | Automated test suites and rubrics |
| Verification | Confirming facts and sources | Buyer and vendor | Grounding, citations, audit logs |
| Rework and risk | Fixing errors, absorbing failures | Buyer | Guardrails, escalation paths |
The last two columns are my analysis of how each component tends to work in practice. They are not McKinsey’s classification. The point stands either way: four of the five cost lines sit with the buyer, not the model vendor.
Token spend itself is a moving target. Our comparison of Gemini, ChatGPT and Claude API cache pricing shows how much the effective price of a request can swing depending on how a team structures its calls. Two companies using the same model can pay very different effective rates.
Human Review, Verification and Rework: The Hidden Cost Layer
Most AI business cases start with a token calculation. A task costs a fraction of a cent to run, a human takes ten minutes, and the savings look enormous. The McKinsey framing breaks that math. If a person has to read the output, check it against a source, and redo it when it is wrong, the real cost per finished task is the model fee plus all of that labor.
This is where the 30% scenario bites. Effective cost can rise without any model vendor touching its price list. If error rates climb on a harder task mix, review time grows. If a regulator tightens documentation rules, verification gets heavier. If a bad output reaches a customer, rework and risk costs spike. Each of those moves counts toward the 30%.
Shadow usage hides the true bill
Cost accounting gets harder when employees adopt tools on their own. Our report on shadow AI use among 45% of workers covers how much usage sits outside formal budgets. Review and rework time in those cases rarely shows up on any AI line item, which means leaders may underestimate effective cost by a wide margin.
Software teams know the pattern well. AI can speed up writing code, but reviewers then absorb the load. We looked at that tradeoff in our piece on how AI cuts coding time while flooding open source with low-quality submissions. Time saved at the keyboard becomes time spent in review.
The $7 Trillion Infrastructure Bet and the $760 Billion Profit Pool
The cost story has a supply-side twin. McKinsey estimates that approximately $7 trillion could be invested in AI infrastructure by 2030. To justify that spending, it says roughly $760 billion in annual profit pools would be needed.
Run the numbers and the pressure is plain. $760 billion a year against $7 trillion is an annual return of about 10.9% on the capital. That is my calculation, and it ignores timing, depreciation and how fast hardware ages. Still, it shows the scale of revenue the industry must find to make the buildout pay.
Now connect it to the value pool. $760 billion is about 30% of the $2.52 trillion baseline and roughly 45% of the $1.67 trillion stress case. The comparison is imperfect because a value pool measures economic benefit to users, and a profit pool measures what vendors keep. Even so, the direction is clear: the same cost shock that shrinks user value also shrinks the room for infrastructure profit.
Spending commitments keep growing in the meantime. Our coverage of Meta lifting AI capex to $145 billion is one example of a hyperscaler leaning in while a peer trims. Whether that bet pays off depends on whether customers see enough net value after paying for review and rework.
Market Impact: Who Gains and Who Loses if Costs Rise
A cost increase does not land evenly. Model vendors with pricing power can pass costs along. Cloud providers with spare capacity can discount. The buyers in the middle, enterprises wiring AI into thin-margin processes, feel the squeeze first because their business case was built on a fixed unit cost.
Consider three buyer profiles. A high-margin software firm can absorb a 30% rise in AI cost with barely a ripple, since AI is a small slice of cost of goods sold. A call center running AI on every interaction faces a real hit, because AI spend is large relative to margin. A government agency under audit rules may see verification costs balloon even if token prices stay flat. These profiles are illustrative, not drawn from the report.
For investors, the report reinforces a question already circulating: how much of the AI revenue forecast is durable demand and how much is subsidized experimentation. OpenAI’s reported pursuit of fresh capital, which we covered in its talks around a $1.25 trillion valuation, depends on the first answer. A value pool that shrinks by a third under modest cost pressure makes that answer harder to give with confidence.
McKinsey 2023 vs 2026: How the Estimates Compare
McKinsey’s earlier work on generative AI produced a bigger headline range. In 2023, the firm estimated generative AI could create $2.6 trillion to $4.4 trillion in annual economic benefits across 63 use cases and 16 business functions. The 2026 figure of $2.52 trillion looks smaller, but the two are not like for like.
| Attribute | 2023 McKinsey report | 2026 McKinsey report |
|---|---|---|
| Title | The economic potential of generative AI: The next productivity frontier | The AI value curve: The economics of applying AI to existing work |
| Headline estimate | $2.6 trillion to $4.4 trillion a year | $2.52 trillion a year |
| Scope | 63 use cases, 16 business functions | Existing work across certain sectors, regions and occupations |
| Mechanism | Generative AI benefits | Automation and augmentation |
| Cost stress test | Not part of the figure cited here | 30% cost rise cuts pool to $1.67 trillion |
| Cost definition | Not part of the figure cited here | Tokens, review, validation, verification, rework and risk |
The shift in emphasis is the story. Three years ago the question was how big the upside could get. In 2026 the question is how much of it survives contact with real operating costs. That is a maturity signal for the field, and it mirrors how enterprise buyers now talk.
What the Cited Research Says, in McKinsey’s Own Framing
The verified citations behind this story come from McKinsey itself, as relayed in the research record. Each line below is attributed to the firm, with the source linked on first use.
On the headline estimate: “McKinsey estimated that AI could generate $2.52 trillion in annual global economic value by automating and supporting existing work.” (McKinsey and Company, The AI value curve, 2026)
On the stress test: “McKinsey modeled that a 30% increase in effective AI costs could reduce the global annual value pool from $2.52 trillion to $1.67 trillion.” (McKinsey and Company, The AI value curve, 2026)
On the size of the loss: “McKinsey calculated that the modeled cost increase would eliminate approximately $850 billion, or 34%, of the global AI value opportunity.” (McKinsey and Company, The AI value curve, 2026)
On the American slice: “McKinsey estimated that the United States annual AI value pool could decline from $960 billion to $640 billion if effective AI costs rose by 30%.” (McKinsey and Company, The AI value curve, 2026)
And for historical contrast: “McKinsey estimated that generative AI could create $2.6 trillion to $4.4 trillion in annual economic benefits across 63 use cases and 16 business functions.” (McKinsey and Company, The economic potential of generative AI, 2023)
Coverage by NewsBytes repeats the same core figure, which is a useful cross-check. Multiple outlets reading the same report arrive at $2.52 trillion, so the rounding to $2.5 trillion in headlines is cosmetic.
Competitive Comparison: Vendors, Clouds and Enterprise Buyers
The report does not name winners, and neither will I. What it does is change the competitive logic. If effective cost drives value, then the vendor that cuts total cost per finished task wins share, not necessarily the one with the top benchmark score. A model that is slightly less capable but produces fewer errors can beat a stronger one on net economics.
Price competition at the model layer
Price gaps between Western and Chinese labs are already wide. Our breakdown of DeepSeek versus Claude pricing shows how far apart per-token rates can sit. McKinsey’s framework adds a caution: a low sticker price helps only if review and rework do not eat the difference. Cheap output that needs heavy checking is not cheap.
Cloud providers face a different test. They must keep capacity full enough to justify the multi-trillion-dollar infrastructure outlook, while customers push for lower unit prices. Those two pressures pull in opposite directions, and the $760 billion profit-pool requirement sits right where they meet.
Enterprise buyers hold the most leverage and know the least. Few can say what a finished task costs once review time is counted. Whoever builds that measurement first gains negotiating power with every vendor they use.
Historical Context: Cloud Bills Taught This Lesson Before
Technology buyers have seen this movie. When cloud computing spread, early business cases leaned on low unit prices and ignored the labor of managing usage. Bills grew, finance teams objected, and a discipline called FinOps emerged to tie spending to business value. The pattern repeated with SaaS sprawl and, later, with data warehouse costs.
AI is now at the stage cloud was in when finance teams first asked hard questions. The difference is that AI adds a human-in-the-loop cost that classic infrastructure never carried. A virtual machine does not need someone to proofread its output. A language model often does, and McKinsey’s definition treats that proofreading as a first-class cost.
The earlier McKinsey estimate also set expectations high. Back then, headline benefit ranges drove boardroom enthusiasm. The 2026 version supplies the missing counterweight, a number showing how much of the upside can slip away through ordinary operating friction.
Predictions: 5 Things to Expect From Here
These are my read of where the economics point. They are analysis, not reported facts, and none comes from McKinsey.
- Cost per completed task becomes the standard metric. Buyers will move from quoting token prices to quoting the full cost of a finished, checked output, because that is the number that decides whether a project survives budget review.
- Review tooling turns into its own spending category. Automated evaluation, test suites and audit trails exist to shrink the human-review line, and vendors will sell them as cost reducers.
- Pricing shifts toward bundles and outcomes. Vendors that want to shield customers from variable cost will offer flat or per-outcome plans, which moves risk back to the seller.
- Model routing goes mainstream. Teams will send routine work to cheaper models and reserve top-tier models for hard cases, trimming effective cost without giving up quality where it matters.
- Infrastructure investors face tougher questions. The $760 billion profit-pool requirement will draw more scrutiny as buyers push back on unit prices, and expect analysts to model the stress case McKinsey published.
Timing is the open variable. None of these shifts requires a crisis, only the steady arrival of finance teams asking the same question McKinsey asked.
What Engineering and Finance Teams Should Do Now
Start by measuring. Pick two or three AI-assisted workflows and log four numbers for each: model spend, minutes of human review, share of outputs reworked, and incidents that reached a customer. Those map directly onto McKinsey’s cost definition and give you a baseline effective cost.
Next, run your own stress test. Add 30% to the effective cost of each workflow and see which ones still clear their hurdle rate. The ones that fail are your fragile use cases. The ones that pass are where to invest. This is the same logic McKinsey applied at global scale, just applied to a single budget.
Then attack the big lines. If review dominates, narrow the task scope or add automated checks. If tokens dominate, test caching and cheaper models. If rework dominates, fix the prompt or retrieval step that produces the errors. Teams that want a wider view of the field can browse more in our AI and machine learning coverage.
Limits of the Analysis and What Remains Unconfirmed
A few boundaries are worth stating plainly. The $2.52 trillion is an estimate built on McKinsey’s assumptions, and it covers certain sectors, regions and occupations rather than all work. The 30% cost rise is a hypothetical input. The $7 trillion infrastructure figure and the $760 billion profit-pool figure are also McKinsey estimates, not audited commitments.
The available reporting also does not establish which specific products, vendors or named executives the analysis refers to, so I have not attributed any claim to individuals. The $2.5 trillion in headlines is a rounded form of $2.52 trillion. Where I did my own arithmetic, such as the rest-of-world figures and the return on infrastructure capital, I said so in the text.
The takeaway: McKinsey’s number is less a prediction than a sensitivity test. It shows that AI’s economic case is real and large, and also that it hinges on a cost definition most buyers have not yet learned to measure.
FAQ: McKinsey’s AI Value Curve and the $850 Billion Risk
How much could AI add to the global economy according to McKinsey?
McKinsey estimates a $2.52 trillion annual global opportunity from applying AI to existing work through automation and augmentation. Headlines often round this to $2.5 trillion.
Where does the $850 billion figure come from?
It is the modeled loss if effective AI costs rise 30%. The global value pool falls from $2.52 trillion to $1.67 trillion, a drop of about $850 billion, or 34%.
Is the McKinsey figure a forecast?
No. McKinsey describes it as a modeled scenario, not a forecast. It shows how sensitive the value pool is to cost and does not predict that costs will rise.
What counts as an AI cost in the report?
Model and token costs, plus human review, validation, verification, and rework or risk activities. Token prices are only one part of the total.
How does the scenario affect the United States?
The same 30% cost increase would reduce the US annual value pool by approximately $320 billion, from $960 billion to $640 billion.
How much AI infrastructure spending does McKinsey expect?
McKinsey estimates about $7 trillion could be invested in AI infrastructure by 2030. It says roughly $760 billion in annual profit pools would be needed to justify that spending.
How can a company test its own exposure?
Measure model spend, review time, rework rate and incidents for each AI workflow, then add 30% to the total and check which workflows still make financial sense.
