AI revenue gap forecasts compared across Gartner, Goldman, Bain and JPMorgan

The AI revenue gap exceeds $6 trillion through 2032 in Gartner’s September consolidation model. Infrastructure owners need more than $14 trillion in cumulative revenue to cover capital, financing and a 12% return. Gartner forecasts no more than $8 trillion attributable to those assets. The difference is my calculation.

That shortfall is not a write-down forecast. Gartner separately estimates more than $4 trillion in asset value exposed to write-downs under that return assumption.

Meanwhile, Goldman’s $417 billion annual requirement and Bain’s $6 trillion annual 2031 market use different financial tests. My infrastructure comparison followed spending into the buildout. This analysis follows the returns.

• Who earns the revenue matters. Application sales support infrastructure returns only insofar as they contribute to the owner’s economics.
• The accounting clock changes the hurdle. Equipment lives, replacement cycles and financing affect how much revenue is needed, and when.
• Higher demand can widen the funding gap. More sales may require more capacity, so spending growth alone does not establish returns.

AI revenue gap forecasts measure four different financial tests

Gartner’s model follows the infrastructure owner across multiple years. Goldman asks what annual revenue would support a specific annual investment base, while Bain estimates the market size needed to sustain a future spending rate. J.P. Morgan provides another return requirement, but the publicly available citation preserves an older research vintage. (Goldman Sachs, Bain, J.P. Morgan Asset Management)

Table 1 compares financial hurdles, not interchangeable market forecasts. Goldman covers six companies, Bain covers a global market and JPM’s January 2026 article retains the November 2025 research vintage. (Goldman, Bain, J.P. Morgan)

Gartner’s row comes from Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032 (G00861169), September 14, 2026. The comparison keeps the report’s September assumptions separate from Gartner’s October forecast update.

For that reason, a common percentage would conceal those differences. Dividing current cloud revenue above a historical trend by a future revenue requirement is not equivalent to dividing a projected 2031 market opportunity by a 2031 market requirement. The percentages would look comparable while answering different questions.

The assumptions crosswalk

Two models can use the same return target and still require different revenue because they apply it to different capital, companies and periods. The crosswalk makes those boundaries explicit before testing which assumption could change the conclusion.

Evidence behind the model comparison

Gartner’s column uses Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032 (G00861169), September 14, 2026. Goldman’s thresholds come from its September 24 report, with asset lives from the separately identified companion-framework reproduction. The other columns follow Bain’s public model and J.P. Morgan’s January 2026 discussion. Analytical-risk judgments are mine.

• A 10% target does not make Goldman and JPMorgan directly comparable. Matching the percentage leaves the invested-capital base, revenue boundary and research vintage unresolved. The $233 billion difference between their headline requirements is arithmetic, not a measured disagreement about the same market.
• The models disagree about the unit of analysis before they disagree about demand. Gartner follows revenue attributable to assets, Goldman tests annual requirements, and Bain sizes a market against an investment rate. A research note should preserve all three perspectives rather than average them into an unsupported consensus.

Where the models can and cannot be reconciled

Gartner’s cumulative model is useful for studying recovery across asset generations. Goldman’s return ladder is useful for testing a higher hurdle within one investment base. (Goldman) Bain’s calculation tests whether a global revenue pool can sustain an annual spending rate. (Bain)

• Do not rank the firms from most optimistic to most pessimistic using the headline alone. A larger number can reflect broader industry coverage, a later endpoint or cumulative rather than annual revenue. Those are scope differences before they are disagreements about demand.
• Return targets and spending ratios need separate columns. ROIC compares operating profit after tax with invested capital. A capex-to-revenue ratio compares spending with sales and does not establish the profit earned on that spending.
• A defensible reconciliation starts with a common asset base. Choose the owners, construction and equipment cohorts, revenue recipients and period first. Only then can useful lives, utilization, pricing and margins be varied without changing the question mid-calculation.

The practical disagreement is not simply how much customers will spend. Gartner tests how much of that spending reaches infrastructure owners over the recovery period. Goldman isolates the annual revenue required at different returns, while Bain asks how large the industry must become to support a future investment rate. (Goldman, Bain) Successful application adoption can strengthen Bain’s demand argument without resolving Gartner’s revenue-allocation problem.

Customer savings are not infrastructure-owner revenue

For example, consider an enterprise that saves $1 million of labor expense after buying a $100,000 AI service. Those hypothetical savings establish customer value, but the provider earns $100,000, not $1 million. If that provider pays $20,000 for infrastructure, the infrastructure owner receives that amount rather than the full subscription. In other words, these invented amounts illustrate the accounting boundary, not an estimate of industry margins. Adding savings, subscription revenue and the infrastructure payment would count different economic concepts as one revenue pool.

This is why enterprise productivity is necessary evidence but insufficient proof of infrastructure payback. Buyers must retain enough value to renew, application providers need enough margin to serve them and infrastructure owners need enough contribution to recover capital. A model that improves one participant’s economics by squeezing another has redistributed the return problem rather than eliminated it. My earlier analysis of Gartner’s $5.95 trillion AI forecast examines the broader market definition, a useful starting point before assigning revenue to infrastructure owners.

Who gets paid determines Gartner’s AI revenue gap

Gartner’s September report projects $7.5 trillion of cumulative global AI capital spending between 2025 and 2030. Annual capital spending approaches $2 trillion by 2030, but the model attributes no more than 30% of revenue from AI-powered products and services to underlying capital investments during that period.

Revenue allocation drives the difference. Gartner assigns all AI-optimized infrastructure-as-a-service revenue directly to data centers, while allocating only part of software and service revenue to those assets. A model provider’s sales do not become its infrastructure supplier’s revenue dollar for dollar.

Selected assumptions from the report show how that allocation changes over time. These are model inputs for 2025 and 2035, not measured market shares or a table of 2030 forecasts.

These are selected inputs from Gartner’s September model. Changes are my calculations in percentage points, not percentage growth.

Gartner’s attribution also accounts for pass-through revenue and seeks to avoid counting the same infrastructure contribution twice. For consumer AI, the report specifically notes that some attribution is already included in AI-optimized IaaS.

This makes the model sensitive to where AI providers retain their margins. An application company can grow quickly while its infrastructure supplier earns too little to recover equipment costs and fund replacement capacity. As a result, industry revenue growth alone does not resolve the owner’s return problem.

Revenue allocation across the eight selected segments

Where revenue attribution rises most

The largest changes in the selected attribution assumptions are not in IaaS, which remains fully allocated. Embedded agentic AI rises 35 percentage points, from 25% to 60%, while AI models rise 35 points, from 35% to 70%. For the application-market context, my agentic AI and chatbot spending comparison separates adoption forecasts from the infrastructure-owner return question.

• The model already assumes greater infrastructure participation in future AI sales. Its shortfall persists despite rising attribution in those categories. A rebuttal therefore needs to explain why revenue, attributable share or capital efficiency should exceed the modeled path.
• Attribution is not the same as profitability. Booking more revenue against a data center does not eliminate electricity, staffing, networking, financing or replacement costs. A higher attributed share helps only within the rest of the model’s cost structure.
• Double counting can create a false solution. Counting an application subscription and the compute purchased to deliver it as two independent sources of final demand overstates the economic base. Separate final customer spending from transactions within the supply chain.

Attribution creates a second growth lever

A fixed revenue pool can support more attributed infrastructure revenue when the model assigns a larger share to the assets. For example, hold a hypothetical segment’s annual sales at $100 billion to isolate that effect. Moving embedded agentic AI attribution from 25% to 60% raises the amount assigned to infrastructure from $25 billion to $60 billion, a 2.4-fold change without any increase in sales. The AI-model assumption moves from $35 billion to $70 billion on the same illustrative sales base, a doubling.

Those amounts are my illustration of the two endpoint assumptions, not new Gartner market forecasts or evidence of a contractual transfer. Thus, they show why an infrastructure-return model needs both a sales forecast and an allocation method. Applying the later attribution share to the earlier revenue base changes the scenario even if the headline market forecast stays fixed.

• A larger attributed share needs an economic explanation. Ask whether compute use per sale rises, the revenue mix changes or the accounting treatment captures more pass-through payments. These mechanisms have different implications for costs and margins.
• Segment growth and attribution growth can compound. If both increase, attributed revenue can grow faster than final-customer spending. That does not make the added revenue an independent pool of demand, and summing supplier and application sales still risks double counting.

Depreciation and the forecast boundary matter

Gartner depreciates semiconductors over five years and other infrastructure, including cooling, power and shells, over 20 years. The model excludes complex arrangements such as exchanging compute for equity and multitier leasing structures. The report argues that accounting and financing changes alter the timing and allocation of costs more than the underlying revenue requirement.

After 2030, forecast inputs give way to extrapolation. Gartner extends its IT spending forecasts into projections, assuming capital spending grows only with inflation and extending revenue with a line of best fit adjusted for AI intensity. The result is a conditional path to returns, not an independently forecast annual spending series through 2032.

Under those assumptions, Gartner says the aggregate buildout does not achieve its 12% return hurdle before 2032 at the earliest under the baseline. Faster replacement cycles, different financing terms or a different allocation of revenue would change the result.

The return crossing does not eliminate impairment risk

Gartner’s original figure qualifies the recovery date. It places the 12% return crossing in 2033, while the report’s prose describes 2032 as the earliest possible recovery point. Therefore, the defensible conclusion is a conditional early-2030s crossing, not a precisely dated promise of profitability. The chart’s post-2030 lines also depend on the extrapolation described above.

• An improving annual revenue flow does not erase earlier capital damage. The figure still shows about $4.2 trillion of infrastructure exposed to write-downs in 2033, near the marked return crossing, and $3.5 trillion in 2035. Annual revenue and the recoverable value of the installed assets answer different questions.
• Successive write-down exposure estimates must not be added together. Each describes infrastructure value exposed to write-downs at that point in the modeled path. Summing successive years would repeatedly count exposure rather than calculate a cumulative revenue shortfall or a forecast of booked annual impairments.

How equipment life changes the annual cost burden

The table below is a deliberately simple sensitivity illustration, not a reconstruction of any firm’s model. It assumes $100 of equipment cost, straight-line depreciation, no residual value and no financing, operating cost or tax effects.

Dividing equipment cost by useful life gives the annual depreciation charge. Extending the accounting life lowers the annual depreciation charge in this illustration, but it does not change the original cash payment or guarantee that the equipment remains competitive.

• Physical life and economic life can diverge. A functioning accelerator can lose commercial value if a newer system delivers the required workload at a much lower cost. The model needs evidence about usable output and realized pricing, not only whether the equipment still runs.
• Depreciation relief is not free cash for replacement. An owner that depreciates equipment over six years but replaces it earlier still needs to fund the replacement. Compare accounting schedules with the operating plan.
• Construction delays can consume the recovery window. If a completed asset waits for power or customers, revenue arrives later while capital remains committed. Utilization timing belongs beside asset life in a return assessment.

Write-down exposure is a separate calculation

Gartner estimates more than $4 trillion of asset value exposed to write-downs through 2032 at the 12% return target. Raising the target toward 25%, the return level the report associates with hyperscaler capital investment before the AI buildout, takes the modeled exposure above $6 trillion.

Neither figure is the same as the greater-than-$6 trillion revenue shortfall calculated from the report’s cumulative revenue bounds. Revenue is earned over a period. In contrast, impairment concerns the recoverable value of assets.

Diversified cash flows give established providers a funding option that compute-only businesses lack in Gartner’s analysis. Gartner forecasts that more than 80% of neocloud providers will cease independent operations by 2030. That is a consolidation forecast, not a statement that 80% will enter bankruptcy.

Scarce compute can still produce weak investment returns

Gartner’s September report explicitly expects compute constraints through at least the end of the decade while questioning sustainable returns. Its explanation includes expensive power and hardware, competition supported by outside capital and falling token costs that subsidize greater consumption rather than expand margins. Scarcity is therefore not enough to establish who captures the economic benefit.

My reading is that three prices have to be separated. An equipment supplier charges the infrastructure owner for capacity. The owner sells compute to a model or application provider, which then charges a final customer. If equipment and energy costs rise faster than the owner can raise realized compute prices, heavy demand can coincide with an inadequate return. In short, a long queue of customers does not reveal the margin on serving them.

• A diversified hyperscaler and a compute-only provider do not face the same funding test. Gartner identifies established cash flows as an advantage for diversified providers. That advantage can extend the period over which a company finances the buildout, but financing an asset longer does not establish that the asset earns its hurdle rate.
• A write-down changes the investor’s loss, not necessarily the asset’s usefulness. In an illustrative restructuring, an asset sold below its original cost can support a viable business for the buyer while leaving the original investor with a loss. Continuing to operate the data center would not disprove a forecast that its original capital was impaired.
• Customer continuity is a separate diligence question. Gartner’s consolidation forecast makes portability, contractual protections and fallback capacity relevant even to buyers that own no infrastructure. A provider acquisition, service withdrawal or pricing change can affect a product without causing a physical shortage of functioning equipment.

October’s higher spending forecast does not automatically close September’s gap

Gartner’s October 7 AI spending forecast raises AI-optimized IaaS spending for 2026 and 2027 by 13% and 14%, respectively. The 3Q26 data places that market at $47.7 billion in 2026 and $199.7 billion in 2030.

October’s update also raises the cost side of the buildout. Its accompanying report increases the 2030 AI-optimized server outlook by $701 billion and the AI processing semiconductor outlook by $312 billion. Gartner attributes the changes to memory pricing, infrastructure demand and expanding inference workloads.

Annual spending and the AI revenue requirement

Selected annual spending estimates from the October forecast show the scale without treating different markets as interchangeable. For the broader equipment market, my analysis of Gartner’s $9.76 trillion IT spending forecast separates servers from other IT categories.

All three series come from the same 3Q26 forecast workbook and are annual market spending in current US dollars. They are separate market categories, not components to sum into a new capex figure.

Using one forecast edition keeps the growth comparison on a consistent basis. The following calculations use 2026 and 2030 annual market spending, with compound growth measured over four years.

Growth rates do not establish investment returns

• IaaS grows faster in percentage terms than the hardware categories shown. That is consistent with an expanding service-revenue opportunity. It does not prove that service margins are sufficient to cover the entire installed and replacement asset base.
• Server spending adds the largest dollar amount. The forecast increases annual server spending by about $1.16 trillion between the endpoints. This is an increase in the annual spending rate, not cumulative investment over those four years.
• These rows cannot produce a payback ratio. Server buyers include owners whose monetization extends beyond the IaaS category. Dividing one market by another would mix economic boundaries and omit operating costs.

These are selected rows from Gartner’s 3Q26 AI Spending Forecast, rounded from millions of dollars. They are not an independently additive capex total, and this article does not subtract the IaaS row from server spending to manufacture a new return estimate.

The growth comparison changes after 2028

Indexing each category to 2026 = 100 exposes a timing difference that endpoint growth rates miss. Server spending initially expands faster than IaaS. By 2028, the server index reaches about 247 against 228 for IaaS. The ordering reverses in 2029, when IaaS reaches about 312 and servers about 291, calculated from the same Gartner workbook.

That sequence is consistent with equipment purchases leading subsequent service expansion, but it does not prove that the same assets generated the later IaaS revenue. The categories have different customer and supplier coverage. Matching equipment purchases to service revenue requires asset-generation data and a consistent set of infrastructure owners.

• Timing can improve while the dollar burden still grows. IaaS wins the four-year percentage-growth comparison, yet servers add roughly 7.6 times as many annual spending dollars between the endpoints. This is a comparison of increases in two markets, not a funding ratio.
• Late-period catch-up is an assumption worth monitoring. If service growth slows before installed capacity becomes productive, the revenue path can miss the forecast even when the equipment spending has already occurred. Check recognized service revenue and paid utilization alongside each spending update.

How the forecast translates into a revenue shortfall

The September consolidation report says achieving a 12% return by 2030 would require AI-optimized IaaS revenue above $1.2 trillion, holding its other assumptions constant. October’s $199.7 billion forecast provides newer market context, but inserting it into one part of September’s model would leave the investment base, attribution and financing assumptions unreconciled.

Gartner raised both the revenue opportunity and infrastructure spending expectations. The two reports do not include a rerun of the September consolidation model using all of October’s inputs.

The percentage bounds use 8 divided by 14 and its complement. Required revenue exceeds $14 trillion while attributable revenue does not exceed $8 trillion, so the exact upper bound is 8/14. The displayed 57.1% is a rounded description of that bound, not an exact threshold.

• Even the most favorable reading leaves more than two-fifths of required revenue missing. That describes the September model’s required cumulative revenue, not a share of assets that must be written off. It also does not measure current revenue collected.
• The shortfall and impairment figures cannot be added. They describe a revenue stream and an asset-value consequence of the same modeled economics. Adding them would double count different expressions of the problem.

Goldman’s $417 billion requirement starts with a different investment base

Goldman’s September 24 US Equity Views report estimates that annual AI revenue of $308 billion would cover depreciation and operating expenses in its model. At a 10% return on invested capital, the requirement rises to $417 billion, based on average annual AI capex in 2026 and 2027. (Goldman, pp. 5–6)

Goldman’s required-revenue model covers Amazon, Alphabet, Meta, Microsoft, Oracle and SPCX, identified as SpaceX in the companion framework. Its company universe is therefore broader than the three firms in Goldman’s cloud-backlog chart. (Goldman, reproduced companion framework)

Required revenue comes from Goldman’s Exhibit 8. Differences and percentage increases are my calculations from its rounded figures, which explains the $110 billion final increment rather than $109 billion. (Goldman)

• Break-even leaves a substantial step to an acceptable return. Goldman’s table requires another $109 billion of annual revenue to move from break-even to 10% ROIC, a 35.4% increase over $308 billion by my calculation. (Goldman, Exhibit 8) Covering operating costs and depreciation is therefore a different milestone from earning the modeled return on capital.
• Higher return targets increase the required revenue within the same investment base. The requirement rises from $417 billion at 10% to $526 billion at 20%, another $109 billion. (Goldman, Exhibit 8) This is a cleaner sensitivity comparison than placing Goldman’s annual number beside Gartner’s cumulative figure.

Depreciation and backlog require separate checks

Goldman’s companion framework uses five-year useful lives for compute and 15 years for data-center shells. That methodology is available in a third-party reproduction, while the revenue thresholds above were checked directly in Goldman’s own PDF. (Companion framework, Goldman)

Goldman also estimates roughly $70 billion of annualized cloud revenue above a pre-AI growth trend across Amazon, Alphabet, Microsoft and Oracle. That is a proxy for incremental revenue, not a disclosure of all AI revenue earned by every company in its investment model. (Goldman, pp. 6–7)

Its backlog evidence is stronger when described precisely. Amazon, Alphabet and Microsoft have about $1.7 trillion of combined total cloud revenue backlog in the report, while the cover uses the broader wording “exceed $1.5 trillion.” Both figures appear in Goldman’s original document. (Goldman, pp. 1, 6–7)

Backlog supports the demand case, but it is neither annual revenue nor an AI-only measure. The useful follow-up is how quickly those commitments convert into recognized revenue and what costs the providers incur to fulfill them. (Goldman, Exhibit 10)

Backlog reduces one uncertainty while leaving three others

Goldman’s three-company total-cloud backlog and its four-company above-trend cloud-revenue proxy are not the same population as the six-company investment model. (Goldman, Exhibits 8–10) Subtracting the proxy from required revenue would create an apparent shortfall without reconciling company coverage or revenue definitions.

For investment analysis, a signed commitment still needs a delivery schedule, a contribution margin and a collection path. A commitment recognized over several years cannot be treated as cash available to fund this year’s equipment purchase. Capacity that must be built before recognition can increase funding needs even while the order book strengthens. These are the tests I would apply to backlog, rather than assuming the entire balance is either guaranteed profit or worthless demand.

The same distinction changes how to read Goldman’s return ladder. Its $308 billion break-even figure covers the modeled annual depreciation and operating costs, whereas $417 billion includes the 10% return requirement. (Goldman, Exhibit 8) Crossing the first threshold is progress, but it does not establish the cash funding available for a larger next-generation buildout.

Bain’s $6 trillion is a market-size requirement

Bain projects annual global AI infrastructure spending reaching $1.5 trillion by 2031, including new data centers, compute capacity and upgrades to installed equipment. Assuming capex settles at about 25% of industry revenue implies an AI market approaching $6 trillion annually. (Bain)

That 25% is a capex-to-revenue ratio, not a return on invested capital. Bain is estimating the revenue base needed to sustain an investment rate rather than applying Goldman’s depreciation-and-return calculation to the same companies. (Bain, Goldman)

Consumer subscriptions and advertising contribute a projected $200 billion to $400 billion by 2031 in Bain’s model. Enterprise adoption adds $1 trillion to $1.4 trillion in provider opportunity, leaving a combined $1.2 trillion to $1.8 trillion against the $6 trillion requirement. (Bain)

Its headline $4.2 trillion innovation requirement uses the high end of that existing opportunity range. Using the low end would leave $4.8 trillion, an arithmetic extension rather than Bain’s headline estimate. Neither amount is a forecast of asset impairment. (Bain)

Bain identifies physical AI as a $900 billion opportunity and “autonomous everything” as another roughly $400 billion, covering vehicles, trucks, drones and industrial automation. Those opportunities sit inside its argument for new revenue creation, not proof that infrastructure owners will receive every dollar generated. (Bain)

Reconstructing Bain’s revenue bridge

Bain’s market requirement and its opportunity ranges answer separate questions. Keeping both visible shows how much of the required market depends on additional innovation rather than the categories in its existing consumer-and-enterprise estimate. (Bain)

Coverage and capital intensity

• Identified opportunities cover only part of Bain’s required market. The $1.2 trillion to $1.8 trillion range equals 20% to 30% of the $6 trillion requirement by my calculation from Bain’s inputs. (Bain) That is a within-model comparison of future opportunities, not a measurement of revenue already earned.
• The 25% spending assumption deserves its own sensitivity test. Holding $1.5 trillion of annual investment constant, a 20% capex-to-revenue ratio implies $7.5 trillion of revenue, while 30% implies $5 trillion. Those are my illustrative alternatives to Bain’s 25% assumption, not additional Bain forecasts. (Bain’s baseline)
• Innovation opportunities should not be added without checking overlap. Physical AI and “autonomous everything” describe potential revenue pools in Bain’s expansion argument. (Bain) A credible bridge must specify which sales are incremental and which replace or overlap with already-counted enterprise spending.

Two opportunity cases leave a large innovation requirement

Both ends of Bain’s consumer and enterprise ranges leave most of its $6 trillion market requirement unmet. These are arithmetic endpoints, not probability-weighted scenarios or additional Bain forecasts. (Bain’s inputs)

The upper endpoint supplies 30% of the required market, leaving 70% to additional revenue creation. At the lower endpoint, the split is 20% supplied and 80% still needed, calculated from the same Bain inputs. (Bain)

How much the spending ratio changes the answer

Bain’s $6 trillion result follows from dividing $1.5 trillion of annual infrastructure spending by 25%. (Bain) The sensitivity matrix below varies both inputs to show how strongly they determine required revenue.

Only the $1.5 trillion / 25% cell reproduces Bain’s baseline. Every other cell is my sensitivity illustration, calculated as annual capex divided by the assumed capex-to-sales ratio, with no return-on-capital or operating-margin claim.

• The revenue hurdle moves in direct proportion to spending at a fixed ratio. Raising annual capex from $1.5 trillion to $2 trillion at 25% raises required revenue from $6 trillion to $8 trillion. That extra $2 trillion comes from the assumed spending rate, before any deterioration in profit margins.
• A lower sustainable capex share requires a larger revenue pool. At $1.5 trillion of annual capex, moving from 25% to 20% raises required revenue by $1.5 trillion. A business that can sustain less capital spending per dollar of sales needs more sales to fund the same buildout.

A higher capex share makes the revenue requirement look smaller, but it does not make the business less capital intensive. In the 30% illustration, more of each sales dollar is committed to capital spending before considering the other demands on cash. Therefore, the smaller denominator-driven revenue hurdle is not evidence of a better return.

How spending assumptions compound

The matrix also exposes how two apparently modest changes can compound. Holding the 25% ratio fixed, 10% more capex requires 10% more revenue. Reducing the ratio by 10% in relative terms, from 25% to 22.5%, raises required revenue by 11.1% at fixed capex. Combining 10% more capex with that lower ratio raises the baseline requirement by 22.2%, from $6 trillion to about $7.33 trillion. These are my algebraic sensitivities, not scenarios assigned probabilities by Bain.

A stronger demand forecast therefore needs a corresponding investment plan. If every additional dollar of demand triggers enough new equipment purchases, market growth can preserve rather than close the financing problem. The useful counterargument to Bain is a demonstrated change in capital intensity or monetizable demand, not simply a larger AI market headline.

Productivity gains and the older JPMorgan comparison

Bain’s identified enterprise opportunity is provider revenue, not the full value of productivity that customers obtain. (Bain) The commercial test is whether new applications create repeatable paid demand beyond existing subscriptions, not simply whether they perform useful work. Substituting an AI product for another software purchase can still support an individual vendor, but it does not automatically create an equally large addition to industry-wide customer spending.

J.P. Morgan adds a useful historical comparison rather than a fresh 2026 forecast. Its January 9 article cites a $650 billion annual revenue requirement for a 10% return and explicitly attributes that figure to November 10, 2025 research, AI Capex – Financing the Investment Cycle. The public article does not provide enough methodology to reconcile that estimate fully with Goldman or Gartner. (J.P. Morgan Asset Management)

The next test is revenue per useful workload

Gartner’s October 22 report on scaling inference describes prefill as compute-intensive and decode as bandwidth-bound. Separating those stages lets operators match different hardware to each, but the report warns that moving the key-value cache between pools requires adequate networking. Buying faster accelerators does not resolve every inference bottleneck.

Gartner recommends prioritizing GPU orchestration, model serving and basic observability before advanced optimization. Those operating choices give buyers a way to examine costs at the workload level instead of assuming that every task requires frontier hardware.

For an enterprise procuring AI, the immediate decision is whether a useful result remains economical when the provider’s subsidy, hardware mix or price changes. For an infrastructure owner, the question is whether the margin on that result contributes enough to recover today’s assets and support replacement capacity.

Utilization and margin have to work together

An owner cannot recover a fixed annual capital burden from unused capacity. To isolate that relationship, assume a hypothetical asset needs $30 million of annual contribution after variable delivery costs to meet a separately specified fixed-cost and capital-recovery target. Assume prices and workload mix remain constant, and paid utilization scales revenue proportionally. This is my simplified operating illustration, not a reconstruction of Goldman, Gartner or JPM’s ROIC models.

Each cell is the annual sales potential at 100% utilization needed to produce $30 million of contribution at the stated actual utilization and margin. The formula is $30 million divided by utilization and contribution margin. It is not actual recognized revenue. For example, the 70%-utilization, 60%-margin cell implies $50 million of actual sales, which produces $30 million of contribution.

At a fixed 50% contribution margin, improving utilization from 50% to 90% lowers the required full-capacity sales potential from $120 million to $66.7 million. The improvement works because more of the same asset earns revenue, not because its purchase price changes. If higher utilization requires steep discounts, the margin column must change too. Presenting a utilization gain without its realized price can hide that offset.

Cheaper inference does not guarantee more contribution

Suppose a hypothetical completed task sells for $1.00 and costs $0.60 in variable delivery expense. Contribution is $0.40. A 30% reduction in that expense lowers cost to $0.42, but a simultaneous 25% price cut lowers revenue to $0.75 and contribution to $0.33. As a result, the provider now needs 21.2% more completed, paid tasks just to preserve the previous contribution dollars.

If the selling price instead falls 40%, to $0.60, contribution becomes $0.18. Restoring the original contribution then requires 122.2% more task volume, before any extra capacity investment or fixed operating expense. These are calculated illustrations, not price forecasts. They show why lower inference costs and higher consumption can coexist with a difficult capital-recovery problem.

• Count completed paid work rather than raw activity. A token, retry or abandoned reasoning step is not automatically a monetized customer result. The workload definition needs to match what the customer pays for.
• Measure each equipment generation separately. New assets can add revenue while masking declining utilization or prices on older hardware. Contribution by equipment generation, remaining investment and replacement timing reveal whether the installed base is paying for itself.
• Test the capacity limit of every volume-based recovery plan. A plan requiring twice as much work is incomplete unless the existing assets can deliver it at the assumed cost and service level. Buying more capacity to recover the first investment raises the hurdle again.

The price cut can outweigh the engineering gain

In the illustrated $1.00 task, a 30% reduction in variable cost saves $0.18. A 25% selling-price cut gives the customer $0.25, so the provider’s contribution falls by $0.07, or 17.5%. The engineering improvement is real, but the customer captures more than the cost saving.

This distinction separates three beneficiaries of inference efficiency. Customers may receive cheaper output, application providers may sell more useful work and infrastructure owners may deliver more tasks per asset. However, none of those outcomes establishes the other two. To support an investment case, the analysis needs the realized price, the cost per successfully completed task and the available capacity for the additional volume.

• Lower prices are not necessarily bad economics for every participant. A customer can gain while a supplier’s margin contracts. Assess the return at the entity that owns the capital, rather than treating a consumer saving as an infrastructure payment.
• Volume recovery needs a capacity check. The 2.22-fold task requirement in the steep-price-cut case assumes unchanged unit economics. If serving that volume requires new capital or a worse workload mix, the calculation must include those costs instead of treating volume as free.

An operating checklist for the AI revenue gap

The following checks turn the forecasts into a review agenda. They are my synthesis of the models, not an additional analyst forecast.

Require the investment period, revenue recipient and return target beside every AI revenue-gap number. A supplier that cannot provide those three items has not given the buyer a usable economics model.

What would make the economics improve or deteriorate

Paid utilization, realized prices and replacement timing can change these models before the next forecast arrives. My framework below identifies the evidence that would strengthen or weaken the return case.

What to monitor between forecast updates

• Efficiency gains help only if the owner retains part of the benefit. A lower cost per token is insufficient evidence when selling prices fall just as quickly. Track contribution per useful task and the amount of paid work delivered by each equipment generation.
• Demand growth and return improvement are separate milestones. Higher usage can coincide with lower margins, replacement spending or delayed collections. The return case needs cash generation as well as a rising consumption chart.
• The strongest counterargument to a revenue-gap forecast is an operating bridge. Show the specific changes in utilization, pricing, costs, asset lives and attributable revenue that move the result. A larger total-addressable-market estimate alone does not reconcile the model.

Software Strategies Blog AI infrastructure projections through 2033

All values after 2030 in this section are Software Strategies Blog projections, with analysis by Louis Columbus, not Gartner forecasts. I extend the published series in two ways to test how strongly the result depends on retaining earlier growth versus the slower pace at the end of the source forecast. Neither path has an assigned probability.

CAGRs calculated from the forecast

I calculate compound annual growth as (ending value / starting value) raised to 1/n, minus one, where n is the number of annual intervals. The 2025–2030 calculation has five intervals, while 2026–2030 has four. These are growth rates calculated from Gartner’s forecast series, not realized investment returns or measured historical revenue growth.

For each projection, I multiply the 2030 source value by (1 + the selected annual rate) for each year after 2030. I use the supplied CSV values at their published precision and retain the full calculated growth rate before rounding the displayed result. A separate check against the workbook’s additional decimal places leaves every displayed CAGR and projection unchanged. All amounts remain annual market spending on the source’s current-US-dollar basis, without a separate inflation or exchange-rate model.

Software Strategies Blog projection if the 2026–2030 CAGR continues

Software Strategies Blog projection if only the final-year growth rate continues

Gartner’s Forecast: AI Spending, Worldwide, 2025–2030, 3Q26 (G00862061), October 7, 2026, and its workbook supply the values through 2030. Every 2031–2033 amount is a Software Strategies Blog projection, abbreviated SSB in the tables. I calculated the growth rates and scenario comparisons.

Server spending is the most sensitive to retaining the earlier growth rate

My CAGR-continuation case puts annual AI-optimized server spending at about $4.02 trillion in 2033, compared with $2.21 trillion when only the 9.5% final-year growth rate continues. That $1.81 trillion spread is generated by the growth assumption, not by a disagreement between research firms. The ratio between the two results is about 1.82.

For IaaS, my two paths produce $584.0 billion and $482.7 billion in 2033, a smaller 1.21-fold spread. Network fabric reaches $137.5 billion or $85.7 billion, a 1.60-fold spread. The broader implication is that extending one average growth rate can disproportionately enlarge the equipment requirement when the source forecast already anticipates a slowdown.

• The longer CAGR contains the fastest expansion years. Every category’s 2025–2030 CAGR exceeds its 2026–2030 CAGR, and both exceed its final-year growth rate. Extending the earlier average therefore carries more of the initial surge into a later period.
• The slower case is not a downside boundary. Even the final-year-continuation case assumes positive growth every year. A capacity pause, falling equipment prices or weaker demand could produce results below both paths, while a new investment cycle could exceed them.
• These projections do not recalculate the AI revenue gap. They do not update asset lives, operating costs, financing, revenue attribution or the installed investment base in Gartner’s consolidation model. Subtracting projected IaaS from projected server spending would still mix different markets and would not measure a funding shortfall.

In short, the projection is useful as a sensitivity test, not a substitute for a new forecast. Its next check is whether observed company guidance, service revenue and equipment demand support the assumed continuation. It does not turn a spending series into evidence that the assets earn an adequate return.

What would change the AI revenue-gap verdict

This comparison does not establish that AI demand is weak. Instead, it shows why demand growth alone cannot establish whether infrastructure owners earn their target return. My conclusions differ by the decision being made.

• Investors need an operating explanation, not another market-size headline. Ask which changes in paid utilization, realized prices, costs and equipment lives would close the modeled shortfall. Then check whether those changes require more capital than the model assumes.
• Enterprise buyers need a workload that survives a provider change. Test whether the application remains economical under different prices, model choices and migration costs. A useful application can still depend on a supplier’s temporary subsidy.
• Forecast readers need an explicit test of what would change the conclusion. Keep the revenue recipient, investment period and return target fixed when comparing revisions. Otherwise, a smaller headline gap may reflect a new definition rather than better economics.

The next convincing signal is not another upward spending revision. It is evidence that paid work covers delivery, financing and replacement costs while leaving a return for the infrastructure owner.

Frequently asked questions

Does Gartner forecast a $6 trillion AI write-down?

Gartner’s September report does not make that claim at its 12% return target. It forecasts more than $14 trillion of required cumulative revenue against no more than $8 trillion attributable through 2032, while separately estimating more than $4 trillion of asset value exposed to write-downs.

Why are Goldman’s and Bain’s revenue requirements so different?

Goldman models annual revenue needed to support average annual AI capex in 2026–2027 at specified returns. Bain estimates the annual global market needed in 2031 if infrastructure spending equals 25% of revenue, a different financial test and market scope. (Goldman, Bain)

Does higher AI spending prove the infrastructure will earn its target return?

Higher spending does not establish who retains the revenue or the cost of earning it. Gartner’s October update raises both IaaS revenue expectations and infrastructure spending, without supplying a fully updated version of the September return model.

How I built this comparison

The Gartner analysis uses original research PDFs and their accompanying forecast files. Report links identify the research behind each calculation. The principal report is Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032 (G00861169), September 14, 2026. Its revenue bounds are on page 3, investment and attribution discussion on page 5, impairment analysis on page 7, and methodology on pages 8–9.

In addition, the October spending figures come from Forecast: AI Spending, Worldwide, 2025–2030, 3Q26 (G00862061), dated October 7, 2026, and its accompanying workbook and CSV. For clarity, dollar amounts in the selected spending table are rounded to one decimal billion from the source’s millions. The inference discussion uses Scaling AI Inference Systems Software (G00863660), October 22, 2026.

Goldman’s thresholds and backlog were checked in its September 24 original PDF, while its detailed depreciation framework is identified as a third-party reproduction. Bain and J.P. Morgan are linked to their first-party publications, with JPM’s underlying 2025 vintage retained. The greater-than-$6 trillion Gartner revenue difference is my subtraction of the report’s bounds, not a separately named Gartner forecast. (Goldman, companion framework, Bain, J.P. Morgan)

The customer-value example, utilization matrix and inference-price examples are hypothetical teaching calculations. They do not introduce additional market forecasts or claim to reproduce an analyst’s proprietary model. Contribution margin in the utilization illustration excludes the fixed-cost and capital-recovery target to prevent counting that burden twice.

Original charts and source graphics

Gartner’s original Figure 1 is reproduced with attribution and its original labels retained. The Bain revenue-opportunity comparison is an original Software Strategies Blog chart based on Bain’s published inputs, not a reproduction of Bain’s Figure 2. The four-firm comparison, Goldman return ladder, spending charts, attribution comparison, indexed-growth chart, Bain sensitivity matrix and inference-margin illustration are my graphics. Calculated values and hypothetical scenarios are identified separately from the publishers’ inputs. Each image links to its full-size file.

More analysis from this blog

• AI infrastructure spending forecasts compared. Compare the spending estimates behind the buildout, with each firm’s scope kept separate.
• Gartner’s broader IT spending outlook. Read the server and IT-services comparison through 2030 for the equipment-market context.
• Agentic AI market forecasts compared. Review the major agentic AI market estimates before treating application-market opportunity as infrastructure-owner revenue.

This is my personal analysis. It does not represent my employer.