How are Alphabet, Microsoft and peers actually financing the AI infrastructure boom, and what happens to their balance sheets and to lessors if AI monetization lags?

Advocate A’s weakest point is treating bond-market appetite as evidence that the AI build-out is economically sound. A 10-tranche Alphabet investment-grade corporate bond — meaning a bond rated high enough that default risk is judged low — can be attractive because lenders have recourse to Alphabet’s whole advertising, search, YouTube and cloud cash-flow base. That does not prove the incremental compute capex, i.e. spending on servers, GPUs, networking, power and data-center capacity, will earn adequate returns. The telecom parallel is not that Alphabet equals Global Crossing; it is that creditors can fund a real infrastructure asset cycle whose unit economics later deteriorate faster than debt and lease commitments can adjust.

A’s framing is also incomplete on leases. A build-to-suit lease is a structure where a landlord or infrastructure vehicle finances and constructs a facility tailored to the tenant, then recovers capital through long-term rent or take-or-pay-style commitments. Calling that “risk allocation” misses Moody’s central warning: for the top five U.S. hyperscalers — the giant cloud platforms such as Amazon, Microsoft, Alphabet, Meta and Oracle — $969 billion of future data-center lease commitments existed as of year-end 2025, and $662 billion related to leases not yet commenced, equal to 113% of their most recent adjusted debt. Moody’s further notes that renewal options and residual-value guarantees can make reported liabilities understate likely cash outflows. That is precisely why off-balance-sheet risk matters: it is not illegal, but it can defer recognition until capacity comes online.

The telecom analogy is strongest on transmission channels, not on identical balance sheets. In 2000–2002, the problem was not that internet traffic vanished; it was overfinanced capacity, price compression and debt schedules. Contemporary reporting on the telecom crash described nearly $300 billion of debt tied to the build-out, while WorldCom bondholders and Global Crossing creditors discovered that real networks could still produce disastrous creditor outcomes when revenue assumptions failed. Today’s AI risk is similar: even modest disappointment in enterprise AI revenue, inference pricing, utilization rates, or hardware refresh assumptions can convert “optional” growth capex into fixed lease cash outflows.

A’s demand evidence is therefore backward-looking. GlobalFoundries forecasting above-estimate revenue on data-center deployments is evidence that suppliers are currently benefiting, but also that supplier revenues are levered to hyperscaler capex. The Philippines BPO sector shows the other side of the same trade: business process outsourcing — contracted customer-support, back-office and IT-enabled work — is already being repriced as AI handles customer-service tasks. Concentrix cut its fiscal 2026 revenue outlook midpoint by about $130 million, and Bloomberg Intelligence attributed the pressure to AI shrinking demand for core customer-experience outsourcing faster than higher-value AI services expand. That is not a contained fortress-balance-sheet story; it is a capital cycle with lessors, suppliers and labor markets all exposed to the same monetization bet.

The specific weakness in A’s rebuttal is that it concedes the most important point but then treats the concession as harmless because the contracts are “disclosed” and bought by sophisticated investors. A acknowledges that $662 billion of not-yet-commenced hyperscaler leases is real, that residual-value guarantees exist, and that headline lease figures can understate eventual cash exposure. That concession implies the debate is no longer about whether Alphabet or Microsoft are near-term default candidates; it is about whether AI infrastructure leverage is being transformed into long-duration, partly off-balance-sheet obligations whose economic risk appears only after facilities commence, hardware is installed, and revenue assumptions are tested. Disclosure does not eliminate cyclicality; it only tells investors where to look.

A’s “sophisticated investor” defense is also weaker than it looks. The telecom cycle was not dangerous merely because some companies committed fraud. Brookings’ post-crash analysis identified the core mechanism as capacity growth vastly outstripping demand, high fixed costs, low marginal costs, price compression, bankruptcies and layoffs. That mechanism can recur even with legal accounting and credible borrowers. AI data centers have similar fixed-cost traits: land, power, cooling, networking and chips are committed before demand and pricing are fully known. If inference prices fall, utilization disappoints, or customers renegotiate workloads across clouds, the cash obligations remain less flexible than the revenue.

A’s backlog argument is therefore not dispositive. Cloud backlog is a useful demand signal, but it is not the same as risk-adjusted profit on incremental AI compute. Backlog can include multi-year cloud, software and marketplace commitments, not necessarily high-margin AI inference revenue sufficient to justify the next dollar of GPUs, power and lease payments. Alphabet’s own Q2 commentary shows rapid AI adoption and paid token usage, which I accept; it does not prove that the capital intensity of supplying those tokens will produce durable returns after depreciation, energy, networking, model-serving costs and hardware refreshes.

A also mishandles the BPO evidence. If AI pressures Concentrix and Teleperformance because clients automate customer-service work, that shows substitution is real; it does not prove hyperscalers will capture enough of the labor savings to service a trillion-dollar infrastructure build-out. Enterprise customers may keep most savings, vendors may compete token prices down, and BPO-heavy labor markets like the Philippines can absorb the downside before cloud providers prove the upside. Moody’s states that leases yet to commence will materially increase adjusted debt and cash outflows and reduce flexibility if industry conditions change. That is the core issue A has not answered: the risk is not immediate Big Tech insolvency, but a financed capacity cycle whose losses would transmit through lessors, private credit, suppliers and exposed labor markets if monetization lags even modestly.

A’s strongest case is that the core hyperscalers are not weak telecom start-ups: Alphabet, Microsoft and Amazon have diversified cash flows, investment-grade access to bond markets, and current cloud demand signals that look stronger than 1999 fiber speculation. A is also right that some stress has appeared first in more leveraged peripheral AI infrastructure firms, while the largest platforms can absorb mistakes that would bankrupt smaller carriers.

I still think B’s position is stronger because A’s evidence proves resilience at the parent-company level, not safety of the financing cycle. The central question is not whether Alphabet can repay one bond tranche; it is whether a high-fixed-cost compute build-out is being funded through long-duration obligations whose cash consequences arrive after capacity is committed. Moody’s reported $969 billion of top-five U.S. hyperscaler future lease commitments at year-end 2025, including $662 billion for leases not yet commenced and therefore not yet recognized as balance-sheet lease liabilities; Moody’s also warned that residual-value guarantees and renewal economics can understate economic risk. That is not refuted by today’s margins, because the risk appears when facilities begin, utilization normalizes, hardware ages, and AI inference prices are competed down.

A’s margin/backlog argument is relevant but insufficient. Cloud backlog often bundles ordinary cloud, software, marketplace and multi-year enterprise commitments; it does not disclose project-level return on the marginal AI data center after GPUs, power, cooling, networking, model-serving costs, depreciation and refresh cycles. The telecom analogy is not “fraud equals Big Tech”; Brookings’ post-crash account identified the broader mechanism: capacity outstripped demand, high fixed costs met low marginal costs, prices fell, firms could not service debt, and bankruptcies and layoffs followed. AI compute has the same dangerous shape if capacity becomes abundant faster than paying use cases scale.

A’s “sophisticated investors price it” response also fails historically and logically. Meta’s Blue Owl Hyperion transaction publicly includes a capped 16-year residual-value guarantee, which shows disclosure, but disclosure does not transfer all economic risk away from the tenant or eliminate lessor refinancing risk. Sophisticated private-credit and infrastructure investors can still underwrite correlated assumptions about AI utilization, tenant renewals, power costs and residual values.

Finally, the downside is already visible outside hyperscaler income statements. Bloomberg-reported pressure on Concentrix and Teleperformance shows that AI substitution can hit BPO labor markets before cloud providers prove they can capture durable profits from the same automation. The strongest version of my claim is therefore narrower than “Big Tech will fail”: the AI build-out can be financially dangerous even if Big Tech survives, because losses can cascade through lessors, private credit, suppliers and exposed labor markets.