2026 hyperscaler capex hits $697B as AI race turns to debt

The binding constraint on the AI buildout has moved from compute to capital structure. J.P. Morgan's latest estimates put 2026 hyperscaler capex at $697 billion across five US companies, a figure $173 billion above what the bank projected at the start of the year. The financing behind that spending now carries far more debt than the sector has used before.

The scale has no modern precedent. PIMCO credit strategist Lotfi Karoui draws a comparison to the 19th century: after adjusting for inflation, the current AI investment wave is set to become the largest capital cycle since the railway boom. Consensus forecasts put hyperscaler capital spending above $1 trillion per year from 2027 onward, with no clear sign of moderation. Karoui has also cautioned that the ultimate scale remains deeply uncertain, even though the direction, higher, is not in dispute.

The 2026 numbers keep climbing

J.P. Morgan's $697 billion figure covers only the five largest US hyperscalers and has been revised upward by $173 billion within roughly seven months. The same bank's asset-management unit had framed the five-company total as the upper end of a range running from $650 billion to $697 billion for the year. Estimates have kept drifting higher as earnings-season guidance arrived: the projected 2026 outlay for those five companies is now roughly $760 billion, with about $211 billion in depreciation expected against that spending.

Amazon, Microsoft, Alphabet, and Meta are the dominant contributors, with their combined spending projected to rise about 71% year over year. Consensus projections across the market put total AI capex near $900 billion by the end of 2026, an 85% increase over 2025, and above $1.2 trillion by the end of 2027; Evercore and Bank of America analysts see the figure exceeding $1 trillion in 2027 alone. The spread across forecasts, from $650 billion to $760 billion depending on how many companies and which accounting year is counted, is itself a sign of how fast the ground is moving. J.P. Morgan's own tracking shows the 2026 hyperscaler capex estimate nearly doubling since mid-2025 as each quarterly guidance cycle has added to the total. J.P. Morgan's longer-horizon work puts cumulative AI and data center infrastructure spending between $5 trillion and $7 trillion by 2030.

The depreciation line is the part worth watching. Recognizing $211 billion in depreciation against roughly $760 billion of spending means about one dollar of cost is recognized for every 3.6 dollars invested, a ratio that frames the returns debate: at this pace, it takes years of compounding revenue for the accounting to catch up. J.P. Morgan has responded by raising its S&P 500 target to 8,000, arguing that profits will now have to carry the market.

Why hyperscaler capex now runs on debt

The funding mix sets this cycle apart. Instead of relying mainly on balance-sheet cash, AI infrastructure is now financed through corporate debt, project-level debt, and equity stacked in multiple layers; recent high-performance computing data center bond deals have reached loan-to-cost ratios of 95%.

Concrete deals show the pattern: $4.25 billion raised for Hut 8's Beacon Point facility and $5.25 billion for CoreWeave, both structured in a market where the average facility spans roughly 100,000 square feet and houses 2,000 to 5,000 servers. J.P. Morgan projects that about $150 billion of the next five years of buildout will come from leveraged finance, with up to $200 billion more in data center securitizations.

The structural reason is simple: spending is rising faster than operating cash flow at the largest operators, which is why external financing has become a permanent feature of the cycle rather than a temporary bridge. The corollary is that hyperscaler capex now behaves like a credit cycle as much as a technology cycle. High-yield data center debt has performed well enough to keep issuance flowing, but that performance has also normalized structures, like 95% loan-to-cost terms, that leave lenders with a thin equity cushion. At a 95% loan-to-cost ratio, lenders are funding nearly the entire construction bill, and equity covers little more than a rounding error if asset values fall.

Power, not compute, is the real bottleneck

Capital structure is not the only constraint tightening. J.P. Morgan points to power as the hardest constraint on AI data centers: grid capacity cannot keep up with demand, which pushes project timelines out. The bank tracks 122 gigawatts of new data center capacity planned between 2026 and 2030, all of it landing on a grid that already struggles to connect the facilities under construction. In this market, power contracts rather than chip deliveries set the delivery dates.

Federal funding is part of that picture. Project Stargate, the US government's flagship effort, commits $500 billion over four years to AI and energy infrastructure, an acknowledgment that compute expansion and power expansion now move together. For operators, site selection has effectively become a power-procurement exercise: securing firm energy supply matters as much to project economics as the cost of the GPUs inside the building.

The trade-off: balance sheets versus credit markets

The trade-offs sort the field into two camps. The largest hyperscalers enter this phase from a position of balance-sheet strength, funding a large share of capex from operating cash flow and using debt to close the remainder. Smaller operators and AI infrastructure companies depend almost entirely on the credit markets, which makes their cost of capital and the depth of the high-yield market existential variables.

The risk profile is hard to track from any single vantage point. As leverage rises, operators become more exposed to interest-rate moves and credit-market disruptions, and the layered structure of the financing, equity stacks above project debt above corporate debt, obscures aggregate exposure. Data center securitizations add another layer: they package lease cash flows into rated notes whose performance depends on tenants that are themselves running on borrowed money. Investors have responded by demanding rigorous credit analysis even as they deploy at record scale, and corporate event data from Wall Street Horizon, covering more than 11,000 global companies, shows the second quarter produced the most secondary equity offerings in five years as operators reached for additional capital.

The railway comparison PIMCO draws is not flattering: that cycle ended in overcapacity and consolidation. What differs this time is that actual compute demand is real and growing, which gives the largest balance sheets a credible path to returns. The returns question cuts both ways: if AI revenue grows as projected, the debt was cheap relative to the asset base; if it does not, 95% structures convert a slowdown into a solvency event. The open question is whether the debt-funded middle of the market survives the period before depreciation and interest costs catch up with revenue.

The verdict, for now, favors the leveraged but not the debt-dependent. J.P. Morgan's own framing is that investors are maintaining discipline with rigorous credit analysis despite the scale of deployment, which keeps the door open for well-structured deals while pricing weaker ones harshly.

Why this matters

For decision-makers, the practical takeaway is that who participates in the AI buildout now comes down to access to capital and access to power. The hyperscaler capex wave of 2026 is financed on terms that reward balance-sheet strength and punish dependence on cheap credit, and the leverage accumulated this year is the number to watch as depreciation and interest costs compound through 2027.

Related Articles

• Global AI Infrastructure Spending to Drive $2.59 Trillion Market by 2026
• AI Chip Rally Hits Turbulence as Investors Eye Hyperscaler Spending Slowdown
• US Data Center Electricity Demand Rises as AI Infrastructure Spending Hits $725 Billion

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.