Artificial intelligence is reshaping how the world's largest tech companies finance themselves. This article examines a fundamental shift underway among the five major hyperscalers (Alphabet, Amazon, Meta, Microsoft, and Oracle) as their AI-driven capital expenditure programs outpace the cash flows that once funded them.
With aggregate capex expected to exceed $690 billion in FY26 and free cash flows under mounting pressure, these companies are increasingly turning to external financing through debt issuance, equity raises, and large-scale leasing arrangements to sustain their AI infrastructure build-out.
We explore what is driving this funding shift, how each player is responding, and what the implications are for credit ratings, capital markets, and the longer-term race to monetize artificial intelligence.
In the last 18 months, hyperscalers Alphabet, Amazon, Meta, Microsoft, and Oracle have moved from almost fully self-funded capex to raising external capital at scale. Incremental annual debt rose from 9% of capex in FY24 to 32% LTM by mid-2026, and equity has now returned to the funding mix: Alphabet priced an $84.75bn equity raise in June 2026, and Oracle plans $40bn of combined debt and equity for FY27 (ending in May 2027).
A race to raise capital is underway ahead of anticipated IPOs from Anthropic and OpenAI in 2026 and 2027, respectively. Alphabet’s move confirms the urgency of securing capital before markets are asked to absorb even larger AI issuances.
Surging AI capex and declining free cash flows (AI costs are front-loaded with returns expected over a longer horizon) are the main drivers of the funding shift. Aggregate cash capex of five hyperscalers is expected to exceed $690bn across their respective FY26 periods (>80% YoY vs. ~70% YoY in FY25). Calendar 2026 capex guidance points to close to $800bn (including finance leases and customer pre-payments), reflecting higher component prices, especially memory chips. As a result, FY26 free cash flows are expected to move close to zero or turn negative for all except Alphabet and Microsoft.
Recent moves by SpaceX and, reportedly, Meta to lease data center capacity highlight strong demand from AI foundation model developers and an opportunity to generate near-term returns on massive data center investments. While this adds financial flexibility, it is not yet clear how easily leased data center/GPU could be reclaimed for internal operations if internal AI demand gains momentum, or whether these leases will in practice remain in place long term despite short exit clauses.
Rising capex requirements are starting to differentiate the players by credit strength: S&P downgraded Oracle on July 9 to BBB- (still investment grade) from BBB, citing surging capex, negative free cash flow and customer concentration. A matching move from Moody’s (Baa2 Negative, one notch above S&P’s new rating) would likely pressure bond prices and widen credit default swaps just as Oracle plans to raise further funding.
While the other hyperscalers retain substantial headroom before any downgrade, the underlying themes such as customer concentration (OpenAI and Anthropic) and declining cash flows are evident across the group.
Over the long-term, hyperscaler capex will largely depend on how quickly AI monetization catches up with investment. In the short- to medium-term period, it is evident that external financing will play a prominent role, alongside leasing and off-balance-sheet JVs for data center construction.
Aggregate capex for the five hyperscalers has risen steeply over the past few years, from $95bn investing cashflows in FY20 to around $490bn in the LTM (to May 2026), directed primarily at AI compute, data centers, and relevant infrastructure. Estimates point to more than $690bn in FY26 (>80% growth YoY, the largest annual increase of the cycle) and more than $900bn by FY28. In percentage terms, FY26 marks the peak growth year while absolute spending keeps rising thereafter.
Hyperscaler Capex ($bn)
Estimates sit close to company-announced capex guidance for 2026 (or FY27 in Oracle’s case), though guidance for Microsoft is higher as it includes items such as finance leases.
While 2026 represents the peak growth year for now, estimates have been revised up repeatedly (in line with company guidance), and the timeline for peak capex growth keeps extending. For example, Alphabet raised 2026 capex guidance to $180bn-190bn in Q1-26 from the $175bn-185bn announced at Q4-25 results and flagged a further increase for 2027. Meta also lifted its 2026 capex guidance to $125bn-145bn from $115bn-135bn.
The AI build-out is expected to settle a lower growth rate within the next few years, as data center investments are inherently front-loaded and the majority of build-out is expected to be complete. Yet successive waves of larger and more capable AI models and agentic AI workloads have so far pushed that peak further out.
Upcoming hyperscaler earnings calls should be scrutinized for any further increase in capex guidance, together with cloud growth, remaining performance obligations (RPO), and the share price reaction.
Historical Capex, Guidance and Estimates
Data center build-out (both IT and physical infrastructure) is driving the capex, fuelled by expected demand from artificial intelligence. Compute (processors and servers) represents a majority of the investments, with a significant step-up reaching $380bn in 2026 (roughly doubling from 2025) driven partly by component price inflation, particularly memory as demand outstrips supply of high-bandwidth memory (HBM) chips in AI accelerators. The HBM shortage is also lifting prices of other memory types, feeding into storage and networking (NAND in storage, DRAM in networking gear), though memory is a smaller share of cost in those categories, which dampens the impact of the price hikes.
The capex mix is shifting towards short-lived assets such as compute, which is expected to represent ~60% of capex in 2026, up from ~43% in 2022, whereas the cost portion of physical long-lived (10+ years) infrastructure, buildings, and land is declining. The assumed useful life of compute assets (GPUs/CPUs) is therefore increasingly critical for hyperscaler earnings.
Shorter lifetimes (3-4 years instead of the 5-6 years currently assumed) would imply significantly higher depreciation charges, as investors such as Michael Burry have argued. For now, the market evidence supports the longer schedules as contracted GPU leasing rates and second-hand prices are holding up well even for models launched 3-6 years ago (Nvidia’s H100 and A100), indicating the chips remain economically useful within that range, partly reflecting today’s compute scarcity.
A further implication of the mix shift is that a growing share of capex becomes recurring replacement spend rather than one-off construction. That would make capex plans harder to wind down while raising technological disruption risk if, for example, a new chip architecture displaces GPUs as the main base for running AI workloads. Hyperscalers are already developing their own silicon, with Google leading the way with its tensor processing units (TPUs). It has secured its first large external customer in the last 12 months, turning in-house chips into a revenue stream as well as a cost hedge.
Hyperscaler Data Center Capex Split by Category (%)
In the past, hyperscalers relied on operating cash flows to fund their capital expenditure. The unprecedented increase in AI investments has pushed them toward other avenues of financing. Since FY24, free cash flow has trended downward for most players and is expected to decline further in FY26 (except for Microsoft, which depends on finance leases more than other firms).
Free Cash Flow (in $bn)
Despite the pressure on FCF, most hyperscalers have strong balance sheets and low debt levels. Even with recent debt issuance, gross leverage remains low and is likely to remain so. However, shrinking FCF may limit share buy-backs and non-AI R&D, and could eventually impose more general caution on capex if AI returns disappoint.
In July 2026, Meta was reported to be planning a cloud business to sell excess AI compute to external customers. Management has publicly hinted at the possibility, with Mark Zuckerberg saying that a cloud business was “definitely on the table”. However, Meta has yet to confirm its plans, though the firm’s Q2-26 earnings should shed more light on the near-term roadmap.
The implications of a Meta cloud unit are: 1) Meta’s near-term internal AI demand is not absorbing all the capacity it is building; and 2) external demand for AI compute remains strong amid scarce capacity. Meta’s shares are up +17% between July 1 and July 16, suggesting investors welcome a second monetization route for the capex that would ease pressure on FCF.
Naturally, all hyperscalers are racing to maintain leadership in the AI cloud market, and FCF will be supported with growing AI demand over time. However, the timeline for an FCF recovery is uncertain, and it depends on the inflection point at which AI profits overtake the capex build-out.
Debt
Incremental debt as a share of capex rose from 9% in FY24 to 32% for the LTM before June 2026, taking aggregate total debt to ~$700bn. Continued issuance to bridge the funding shortfall is likely. Amazon recently placed a $25bn bond (2.5x oversubscribed, but lower than 3.2x on its March issue), while Oracle plans to raise ~$20bn of debt in FY27 (the implied debt portion of its $40bn funding plan).
Total Debt and Debt Increase as a % of Capex
Despite increasing debt levels, total debt/EBITDA remains around 1x or below for four of the five hyperscalers, and lower still on a net basis given large cash reserve. Net leverage for the four larger names sits well below typical downgrade thresholds of 1.0x-1.5x for their respective investment-grade ratings, making downgrades unlikely in the short- to medium-term period.
Oracle is the outlier, with significantly higher and growing leverage driven by contractual obligations to build capacity for OpenAI. It has been downgraded by S&P from BBB Stable to BBB- Stable in July 2026, one notch above speculative grade, on concerns about increasing capex, rising business risk, and weaker cash flow. Any downgrade move from Moody’s is likely to pressure bond prices.
Total Debt/ EBITDA
Equity
Equity, long absent from big-tech funding, has returned. In June 2026 Alphabet announced an $80bn equity raise, upsized and priced a day later at $84.75bn (the largest equity capital transaction for a listed corporate). Of that, $44.75bn (including a $10bn private placement with Berkshire Hathaway) is directed at general corporate purposes such as AI capex, while the remaining $40bn at-the-market program primarily covers employees’ equity-award tax obligations (an administrative shift that nonetheless preserves corporate cash for investment).
Oracle raised $43bn of debt and $5bn of equity in FY26 and plans roughly $40bn of combined debt and equity in FY27, with no additional debt issuance expected in calendar 2026. With the anticipated IPOs of Anthropic and OpenAI, capital markets will absorb substantial new AI-related supply. Hence, hyperscalers raising equity early protects cash flows and rating headroom and arguably secures capacity in markets that may become more crowded.
Leasing and Alternative Structures
A third funding route sits more hidden in financials: leasing wholesale capacity at third-party owned data centers (usually funded by private credit) with the hyperscaler holding no stake or a minority stake and committing to long-term leases or capacity offtakes. For hyperscalers, such leases replace upfront funding with multi-year operating expense, reducing the need for debt or equity issuance or self-funded capex.
Future lease obligations (leases signed but not yet commenced) predominantly relate to data centers and represent sizeable commitments not yet recognized on the balance sheet. Lease obligations are, however, only one component of a broader set of commitments that include purchase obligations, which are generally tied to compute, other equipment, and energy supply.
Leasing is not new, but its scale is a direct consequence of the funding needs of AI infrastructure. Aggregate lease-related commitments, not recognized as balance sheet liabilities, are material across the five hyperscalers at ~$820bn. Oracle is the most exposed, committing to lease obligations with 15- to 19-year terms commencing between FY27 and FY29 that total almost 3x its FY27 capex guidance.
Historical Capex, Guidance, and Estimates
Over the 12 months to July 2026, the S&P 500 has outperformed most of the five hyperscalers. That is a notable reversal for a group that led the index through the first phase of the AI rally. Alphabet has been the only outperformer, supported by its new Gemini models, a smaller but faster growing cloud unit than Amazon’s or Microsoft’s, and the first external deals for its TPU chips with Anthropic and Meta.
While geopolitical tensions weighed on the broader market, the year-long divergence from the S&P 500 reflects two structural forces: 1) fears around the sustainability of AI investment, and 2) shifts toward other areas of the AI supply chain.
On the first, investors have grown more cautious on capex, keeping the group’s share prices suppressed since late 2025. On the second, capital within the AI theme has rotated from the companies spending the capex (hyperscalers) toward those receiving it (semiconductors, especially memory), and the rotation has since extended into energy and power infrastructure assets.
Share Price of Five Hyperscalers (1 July 2025 = 100)
This article was co-written by Guniz Kama and Bina Rajput.
This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.