← Back to feed

A Financial Times investigation published August 4 uncovered the financing architecture that may be the most durable competitive advantage in the AI infrastructure race — one that has nothing to do with chip performance.

Google has quietly built a credit guarantee function around its Tensor Processing Unit sales. The result: data center operators building on Google’s TPUs borrow at a median rate of 7.1%. Operators building on Nvidia chips pay 9.3% for equivalent debt. The 2.2 percentage-point gap is not a market accident. It is engineered.

The Numbers

Alphabet’s Q2 2026 10-Q, filed July 24, discloses:

• $43.8B in data center lease guarantee exposure — contingent liabilities covering Anthropic-bound facilities if tenants default
• $24.1B in future guarantees still awaiting final terms
• $85.2B in signed but not-yet-commenced data center leases
• $7.6B in power-related guarantees
• Total off-balance-sheet AI infrastructure commitments exceeding $150B

That is separate from Alphabet’s $811B in binding purchase commitments for chips, energy, and data center construction disclosed in the same filing — a $479B increase from a single quarter earlier.

Jefferies analyst Jonathan Petersen called the borrowing-rate differential a “structural cost-of-capital disadvantage” for companies operating in Nvidia’s ecosystem. At $35B in initial hardware financing, the 2.2-point gap translates to roughly $770M per year in additional interest for Nvidia-backed operators — a compounding penalty that widens as AI infrastructure spending scales toward hundreds of billions.

How It Works

Google’s guarantee mechanism transforms the creditworthiness of data center projects. Without a Google backstop, a former crypto mining company building AI facilities — the typical developer — carries no investment-grade credit rating. Lenders price that as high-yield debt.

When Google guarantees the lease, lenders assess Google’s balance sheet instead. That substitution moves the instrument toward near-investment-grade quality, where institutional capital participates at meaningfully lower rates.

Five completed projects across 1.4 gigawatts of power collectively raised $15B in debt using this structure. Ten developments now underway represent 2.4GW of dedicated TPU power.

Google did not guarantee without compensation. In exchange for lease backstops on TeraWulf’s 360MW New York facility, Alphabet received penny warrants representing approximately 14% of TeraWulf equity. Cipher Digital (168MW in Texas) surrendered warrants representing approximately 5.4%. Hut 8’s $7B River Bend deal included no warrants.

The Broadcom Layer

The structural innovation that prices the whole network runs through Broadcom. Under agreements Broadcom signed, if Anthropic stops making lease payments and TPU hardware sells below loan value on the secondary market, Broadcom absorbs the shortfall on senior debt tranches — approximately $30B of the $35B initial structure.

That guarantee converts lender exposure from “will Anthropic pay?” to “will Broadcom honor?” Since Broadcom holds investment-grade credit, the risk transformation is what moves senior tranches toward near-investment-grade pricing. Broadcom’s incentive: the company has $128B in purchase commitments covering fiscal 2027 and 2028. The Anthropic relationship alone could generate $21B in Broadcom AI revenue in 2026 and $42B in 2027, per Mizuho estimates.

The Structural Risk

Aviation finance worked because aircraft hold secondary-market value. A repossessed Boeing 737 can be re-leased to another carrier. The secondary market for one million purpose-built Ironwood TPUs is untested at this scale.

Google has already previewed its 8th-generation TPU split — a training chip (Sunfish, Broadcom-designed) and an inference chip (Zebrafish, MediaTek-designed), both targeting TSMC’s 2nm node, expected in late 2027. If Ironwood has an 18-24 month competitive lifespan before next-gen chips arrive, any default scenario that unfolds post-transition could find Broadcom honoring a $30B guarantee against substantially degraded hardware.

The entire structure rests on Anthropic’s annualized revenue — currently at $30B-plus, led by Claude Code — sustaining growth to meet chip and data center lease obligations. Google’s Q2 operating cash flow exceeded $185B over the trailing twelve months, and Google Cloud’s 82% growth to $24.8B in a single quarter suggests the infrastructure spending is chasing signed demand. But the framework has never been stress-tested.

The Bigger Picture

What Google built is a manufacturer-finance model borrowed from aviation — GE Capital signed its first aircraft lease in 1967. Applied to AI chips, it turns raw compute sales into a capital markets operation where Google participates in the upside via equity warrants and wins market share by undercutting the cost of Nvidia-ecosystem financing by 220 basis points.

At $200B in off-balance-sheet commitments, the AI infrastructure market has its first structural lender. The question is whether Ironwood chips will hold residual value the way Boeing aircraft do — or whether the guarantee chain gets tested before anyone has had to answer that question at scale.