AI Investment Is Driving Growth While Exposing Economic Risks

TD Securities has identified a sharp economic imbalance: AI-related projects produced nearly three-quarters of US business investment growth during 2025. The Google News headline captures the boom, but not the dependence forming beneath it. AI represented roughly 8% of capital expenditures while generating an outsized share of their growth.

That difference changes the economic debate. Spending on data centers, chips, power facilities, software, and research now supports growth before AI delivers broad productivity gains. The economy is receiving the investment first, while the operational payoff remains uncertain.

The central conflict is therefore investment-led growth versus productivity-led growth. Alphabet, Amazon, Meta, Microsoft, and other large technology companies can keep the first engine running through capital spending. The second engine requires thousands of ordinary businesses to reorganize work around AI and produce measurable results.

What the Google News Headline Leaves Out

TD Securities is describing an economy increasingly dependent on one narrow investment cycle, not simply another strong year for technology spending.

TD Economics estimates that AI-related investment contributed about $137 billion beyond its pre-boom trend during 2025. Its analysis combines equipment, structures, software, and research spending connected with the AI buildout.

Data centers and semiconductor facilities accounted for a combined $125 billion in investment that year. Software and research added an estimated $68 billion of AI-related spending under TD's methodology.

The resulting concentration is the most important finding in the business investment outlook. AI-related activity generated nearly three-quarters of business investment growth despite representing approximately 8% of total capital expenditures.

That pattern matters because spending outside the AI complex remained weak. Fewer than one-third of more than 50 investment categories expanded during 2025, according to TD's assessment.

Outside the pandemic, that breadth was the weakest since 2010. The comparison does not mean the economy has returned to post-financial-crisis conditions. It shows how little support the investment expansion received from unrelated industries.

Trade uncertainty and elevated borrowing costs offer part of the explanation. Interest-sensitive businesses faced tighter financing conditions while large technology companies continued funding data centers from substantial cash flows.

The imbalance creates two very different readings of the same data. One reading says private companies are building the infrastructure for a major productivity advance. Another says headline growth has become unusually exposed to continued spending by a small corporate group.

Both readings can be true for a time. Construction creates demand, equipment orders support manufacturing, and software development counts as investment. Those effects appear before businesses prove that AI can raise output across the wider economy.

The original story also requires a naming distinction. TD Securities is the capital-markets business within TD, while the detailed estimates come from TD Economics research. Treating the headline as a single corporate forecast would blur that distinction.

What changed is not merely investor enthusiasm. AI infrastructure has become large enough to influence national investment statistics, industrial supply chains, power planning, and expectations for economic growth.

AI Investment Is Carrying More of the Economy

The immediate economic benefit comes from building AI capacity, while the durable benefit depends on what companies eventually do with that capacity.

A data center activates a long supply chain. Developers need land, construction labor, electrical equipment, cooling systems, servers, networking hardware, storage, chips, and grid connections.

Those purchases raise investment during the buildout. They also support revenue for semiconductor manufacturers, utilities, engineering firms, cloud providers, and equipment suppliers.

The Bureau of Economic Analysis explains this relationship through capital, labor, energy, materials, and services. Its growth accounting framework traces how those inputs contribute to industry output and national growth.

Information-industry capital was among the leading contributors to real US GDP growth between 2021 and 2024. That result is consistent with substantial spending on AI and related services.

However, official statistics do not contain a clean category called AI investment. AI software crosses multiple industries, and computer purchases include equipment unrelated to machine learning.

The Federal Reserve faces the same measurement problem. Its July 2026 AI buildout analysis combines software, computing equipment, data centers, and power infrastructure to approximate AI's contribution.

That analysis found meaningful contributions to quarterly GDP growth from 2025 through the first quarter of 2026. Software and computer equipment supplied the largest positive components.

Yet imports complicate the calculation. A company can make a major domestic investment while purchasing much of its computing hardware from foreign suppliers.

The purchase raises equipment investment, but the import offsets part of its direct contribution to GDP. This explains why capital spending and domestic economic growth can move by different amounts.

The measurement gap also runs in the other direction. Investment accounting captures construction and equipment more easily than downstream improvements in decision-making, product quality, or worker output.

A customer-service team using AI to resolve cases faster might produce more value without buying a new server. A software team might release more features using the same headcount.

Those gains matter economically, but separating AI's contribution from other changes remains difficult. Firms also differ in how they measure saved time, error rates, output quality, and displaced work.

The current cycle therefore has a visible side and an uncertain side. The visible side includes cranes, chips, substations, and capital budgets. The uncertain side includes organizational adoption and sustained productivity.

This is why the Google News framing deserves scrutiny. A spending boom already shaping GDP is not proof that AI has transformed production across the economy.

Big Tech Spending Puts the Rest of Business Under Pressure

The AI investment economy rewards companies that can fund infrastructure now, while pressuring smaller firms to justify adoption without comparable resources.

The largest technology groups entered the cycle with cash, cloud businesses, established customers, and access to global capital markets. Those advantages allowed investment to continue despite higher interest rates.

TD research estimates that quarterly capital spending among major hyperscalers rose from roughly $40 billion in 2023 to about $130 billion in 2025. Hyperscalers operate computing platforms large enough to serve extensive customer demand.

That spending supports demand for advanced chips and data-center capacity. It also raises the competitive threshold for companies building models, cloud services, and AI applications.

A smaller developer does not need to construct a data center. However, it usually depends on infrastructure controlled by larger providers and pays for computing as usage grows.

Enterprise buyers face another pressure. Executives have heard that AI will raise productivity, but many still lack dependable processes for evaluating output quality and economic returns.

Buying access to a model is only an early step. Companies must connect data, redesign workflows, establish review rules, train employees, and decide where human approval remains necessary.

Those complementary investments rarely appear in a product demonstration. They determine whether AI becomes operational infrastructure or remains an underused subscription.

TD's Canadian productivity research makes this distinction clearly. Its productivity assessment says early US gains primarily reflect capital deepening rather than broad organizational transformation.

Capital deepening occurs when workers receive more or better equipment. It can raise measured output, but it differs from reorganizing production around a new general-purpose technology.

Real US investment in information-processing equipment rose about 28% year over year during the fourth quarter of 2025. Computer and peripheral investment increased approximately 72%.

US data-center construction starts also rose about 190% during 2025, according to figures cited by TD. These numbers show extraordinary capacity expansion.

They do not show that a hospital reduced administrative delays or that a manufacturer improved its defect rate. Those outcomes depend on implementation inside each organization.

A survey cited by TD covered roughly 6,000 executives across the United States, United Kingdom, Germany, and Australia. It found broad adoption but limited realized productivity improvements attributable to AI.

That gap pressures technology vendors as well. They must move customers from experimentation into repeatable deployment while controlling errors, security exposure, and oversight costs.

For knowledge workers, the lesson is practical. Better models help, but reliable value often depends on organizing the information supplied to those models.

A maintained AI knowledge base can reduce the time spent locating project context. Its business value still depends on accuracy, access controls, and consistent use.

The forced response is different for each participant. Big technology companies must keep infrastructure utilization high. Enterprise buyers must prove returns, while smaller vendors must build differentiated products on rented computing capacity.

This pressure will persist beyond the current spending cycle. Infrastructure can be financed and built relatively quickly, but organizational change moves through budgets, policies, and employee behavior.

The Mechanism Runs From Data Centers to Productivity

AI spending becomes durable economic growth only when infrastructure supports cheaper computing, broader adoption, and higher output across many industries.

The cycle begins with expected demand. Technology companies anticipate greater use of training and inference, which means operating an AI model after it has been trained.

They respond by ordering chips, constructing facilities, and securing electricity. Suppliers expand capacity when those orders appear durable enough to justify their own investments.

More capacity can lower the cost of computing over time. Lower costs allow developers to add AI features and let enterprises run more tasks economically.

Broader adoption then creates opportunities for productivity gains. Workers can accelerate research, summarize records, generate drafts, inspect data, or automate defined portions of a process.

However, each connection in this mechanism can weaken. Grid delays can leave buildings unfinished or prevent completed facilities from reaching full operation.

Chip shortages can raise costs and extend delivery schedules. TD observed that semiconductor prices accelerated during the second half of 2025 as demand collided with constrained supply.

Software economics create another uncertainty. Lower computing costs do not guarantee profitable applications if customers resist higher spending or require extensive human verification.

The final step, organizational transformation, is the hardest. Earlier general-purpose technologies needed complementary investments in skills, management, infrastructure, and redesigned processes.

Electricity did not immediately transform factory productivity when businesses first connected motors. Many gains arrived after managers redesigned production around distributed electric power.

Computers followed a similar path. Companies bought hardware before learning how to restructure information flows, supply chains, and administrative work.

AI has the same timing problem, but with added reliability concerns. A model can generate fluent output that contains incorrect facts, missing context, or insecure recommendations.

This creates what economists often call a productivity J-curve. Measured productivity can initially disappoint while organizations invest in intangible assets and redesign work.

The later payoff appears only if those changes produce more output than their ongoing costs. Training, governance, data preparation, and human review all belong in that calculation.

The mechanism also spreads benefits unevenly. Construction markets near large data-center clusters receive direct investment, while other regions experience few immediate gains.

Workers in AI-adjacent industries can benefit from higher demand. Other employees face workflow changes or uncertainty about whether employers will automate parts of their roles.

The result is an economy shaped by two clocks. The infrastructure clock moves through announced capital budgets and construction schedules.

The productivity clock moves through experimentation, trust, process redesign, and measured results. It does not automatically accelerate because another data center opens.

This distinction makes AI investment economy data more useful. It separates money spent to create capacity from value produced through sustained use.

If utilization rises and enterprise results improve, today's construction becomes the foundation of long-term growth. If adoption stalls, the same infrastructure becomes excess capacity with expensive operating requirements.

What the Investment Numbers Still Cannot Prove

The strongest evidence supports a large AI construction cycle, but it does not yet establish broad productivity transformation or adequate investment returns.

The first uncertainty is definitional. Government accounts do not isolate AI, so analysts construct estimates from categories containing both AI and non-AI activity.

Computers and peripheral equipment include ordinary workplace devices. Power infrastructure supports homes, factories, transportation, and conventional data services alongside AI facilities.

The Fed warns that extrapolating a pre-2023 baseline becomes less reliable as unrelated trends affect the same categories. Every estimate therefore depends on assumptions about inclusion and attribution.

The second uncertainty concerns concentration. An investment expansion driven by a small number of cash-rich companies can remain strong, but it carries limited breadth.

TD found that fewer than one-third of business investment categories grew during 2025. This makes aggregate investment more sensitive to hyperscaler budget decisions.

The third uncertainty is monetization. Infrastructure operators need enough paying demand to cover chips, power, cooling, maintenance, networking, and financing.

Consumer enthusiasm does not necessarily produce enterprise revenue. Business customers evaluate security, accuracy, integration costs, and measurable returns before expanding deployments.

The fourth uncertainty concerns financial markets. Higher technology valuations can support consumption through household wealth, but falling prices can reverse that channel.

TD modeled this risk in its market stress analysis. A hypothetical 30% S&P 500 correction would reduce annual growth by more than one percentage point under its assumptions.

The same scenario would push unemployment above 5% by year-end. TD described an adjustment of that size as a tail risk, not its central forecast.

A 10% correction produced much smaller effects in the model. That contrast shows why the scale of any repricing matters more than the simple arrival of market volatility.

TD also identified circular financing and greater reliance on debt as potential vulnerabilities. Circular arrangements occur when participants fund customers or partners that purchase their products.

Such agreements can accelerate infrastructure deployment. They can also obscure independent demand when capital, computing contracts, and supplier revenue move within the same commercial network.

Power presents another constraint. Global data-center capacity requires generation, transmission, substations, and lengthy connection processes.

TD estimates that global installed data-center power demand rose from about 60 gigawatts in 2020 to approximately 100 gigawatts in 2024. The United States and China led that expansion.

Its grid constraints report argues that regions must connect projects without shifting unreasonable costs onto ordinary electricity customers.

That problem can turn investment into a political conflict. Communities weigh tax revenue and construction work against land use, water consumption, transmission upgrades, and electricity bills.

None of these risks invalidates the AI investment thesis. They define the conditions required for the thesis to succeed.

The evidence supports strong infrastructure demand and meaningful effects on business investment. Claims about economy-wide productivity require more direct operational evidence.

Readers should therefore distinguish three statements. AI spending is large, AI spending supports growth, and AI adoption has transformed productivity are separate claims.

The first two have substantial evidence. The third remains an open test that cannot be settled through capital expenditure totals.

Three Signals That Will Decide What Comes Next

The next phase will be judged by investment breadth, infrastructure utilization, and verified productivity gains rather than larger spending announcements alone.

The first signal is the breadth of business investment. Analysts should watch whether more non-AI equipment, structures, and intellectual-property categories begin expanding.

Broader growth would strengthen TD's view that the economy can move beyond a concentrated buildout. Continued weakness would show that AI remains an unusually isolated support.

This measure also provides context for future Google News headlines about capital spending. A rising total looks healthier when participation spreads beyond several technology companies.

The second signal is the relationship between data-center construction, equipment orders, power connections, and actual utilization. These measures track different stages of the same infrastructure cycle.

Continued construction accompanied by semiconductor output and active computing demand would support the investment case. Construction without utilization would raise the risk of excess capacity.

Grid access belongs in this signal because a planned facility does not become productive merely when construction begins. Delayed connections can strand capital and distort announced capacity.

The Fed specifically recommends comparing physical construction with computing equipment and semiconductor production. Divergence can reveal supply constraints or a changing spending mix.

The third signal is verified productivity outside the technology sector. Companies should report output, quality, cycle time, or cost improvements that survive normal operational scrutiny.

A retailer shortening inventory planning, a hospital reducing administrative work, or a manufacturer improving maintenance can provide stronger evidence than adoption surveys alone.

The measurement must include implementation costs. Saved employee time has limited value when organizations cannot redirect it toward additional output or better service.

This third signal decides whether AI becomes a general-purpose technology across the economy. It also determines whether infrastructure owners receive sustained demand after the initial buildout.

The challenge for developers is to create systems that work reliably within real business constraints. Accuracy, security, integration, and human oversight will matter alongside model capability.

Enterprise buyers should demand baselines before deployment. They need to know which measure will change, how it will be verified, and which new costs accompany the improvement.

Knowledge workers can apply the same discipline on a smaller scale. Track whether an AI workflow reduces retrieval time, improves completeness, or accelerates a recurring deliverable.

Investors face a broader version of that test. They must separate revenue supported by continuing capital deployment from revenue supported by profitable end-user demand.

TD Securities has helped move the economic conversation from model performance to capital allocation. That shift is valuable because investment now affects growth, power markets, supply chains, and financial risk.

The question is no longer whether AI attracts substantial spending. It is whether that spending creates productive capacity that businesses will use at sustainable rates.

Watch the next investment reports, infrastructure utilization disclosures, and enterprise productivity evidence together. If all three improve, the current concentration becomes an early stage rather than a structural weakness.

If only spending rises, the Google News narrative will remain incomplete. Readers should ask which industries are producing more, who captures the gains, and whether those gains outlast construction.