The Underwriter Week — NVIDIA Guaranteed $265B in Seven Days, Then Its Own Stock Sold the Risk

The Underwriter Week

In the week of July 21–27, NVIDIA committed capital as a lender — not as a chip supplier — in three consecutive deals. Monday: reports emerged that NVIDIA is considering a $250 billion loan backstop for OpenAI and SoftBank's planned 10GW AI data center in Ohio (DataCenterDynamics, 7/27). The same day, NVIDIA committed $5 billion to Ilya Sutskever's Safe Superintelligence (SSI) alongside supplying the Vera Rubin platform. Tuesday: NVIDIA joined Brookfield as a co-investor in Naver's $10 billion expansion of the Sejong data center to 200MW.

Combined: roughly $265 billion in one week. Realized equity is maybe $15 billion of that; the rest is loan guarantees and platform commitments. On a semiconductor company's balance sheet, this is not a normal week.

The Market Priced It First

NVDA closed roughly 5% down on Monday. Reuters and Barron's revived circular financing as the headline concern. The mechanism: NVIDIA backstops the loan that funds OpenAI's data center → OpenAI uses proceeds to buy NVIDIA GPUs → NVIDIA books the revenue. The revenue is real. The funding source is a contingent liability parked off-balance-sheet on NVIDIA's own book. Accounting-neutral; risk-cumulative.

This adds a new line item to the NVIDIA valuation. Until now, NVDA traded as a proxy for AI compute demand. This week it started trading as the lender of last resort for the AI infrastructure buildout. Those are different framings. Once the market starts computing capital ratios and commitment exposure the way it does for banks, a 60%-range gross margin stops being a premium and starts being a risk buffer.

Why Now

Bloomberg NEF projects US data center power demand at 118GW by 2030 and up to 207GW by 2033 (Utility Dive, 7/28). From today's ~45GW baseline that is a 5× expansion in eight years. Sell-side models keep concluding that the combined annual CapEx of MSFT + GOOGL + META (~$350B) is insufficient to fill this curve.

The gap gets closed from two wallets. First, private credit and infrastructure funds — Brookfield, KKR, Blackstone. Second, vendor financing. Wallet one has been open. What this week confirmed is wallet two.

Same Balance Sheet, Three Directions

The same week saw reports that NVIDIA halved SOCAMM memory allocation on Vera Rubin systems, with memory cost climbing to 29% of BOM. Margin defense requires either lower cost or lower spec; the former apparently failed in negotiations with SK Hynix and Micron. So the spec was cut. NVIDIA's balance sheet is now spread across three growing demands: (1) customer financing commitments, (2) internal R&D and fab-partner CapEx, and (3) memory BOM. Each is expanding; each pressures margin from a different side.

The Competitive Frame

AMD chose the same week to launch its Helios rack-scale data center GPUs, claiming 30% better tokens-per-dollar than NVIDIA (EE Times, 7/27). The timing is what matters. As NVIDIA leans on capital advantage (financing customers), AMD leans on unit-cost advantage (cheaper inference tokens). For hyperscalers, the choice becomes NVIDIA chips + NVIDIA-backed financing versus AMD chips + own financing + 30% lower token cost. Emerging hyperscalers with weaker credit (Oracle, CoreWeave, Nebius) prefer the first bundle. Established hyperscalers with cash flow (MSFT, GOOGL, META) can seriously weigh the second.

Supply Pipe vs. Demand Pipe

Stepping back one week, two capital pipes opened in opposite directions. On the supply side, CXMT raised $8.6B in its Shanghai IPO and surged 466% on debut (SiliconAngle, 7/27); SK Hynix fell 13%, SanDisk 12%, Micron 5%. On the demand side, NVIDIA underwrote hyperscaler CapEx with its own balance sheet. When those two pipes meet, residual margin compresses into two places: (1) financing spread and (2) HBM contract volumes.

Positioning

For NVDA in the next earnings call, ask for three numbers: (1) total off-balance-sheet financing commitments, (2) accounting treatment for circular-financing contingent liabilities, and (3) percentage of revenue funded by NVIDIA-backed capital. None are guaranteed to be disclosed, but non-disclosure widens the valuation discount.

Rerating on AMD's token-per-dollar claim depends on independent data center benchmarks (expected by year-end). The current trajectory of evidence points toward compression of NVIDIA's residual margin ceiling. On the memory side, watch whether CXMT's listing translates into DDR5 spot price movement in August. Today's DDR5 16Gb spot sits at $50.93 — still near cycle highs, but how it behaves once the supply-side capital pipe is wet-tested determines SK Hynix and Micron's next-quarter guidance.

Key Sources: - Nvidia considers $250bn backstop for OpenAI's 10GW Ohio data center (DataCenterDynamics, 2026-07-27) - BNEF chip-based model pegs US 2033 data center demand at up to 200 GW (Utility Dive, 2026-07-28) - Warning shots fired as AMD announces new data center GPUs at Advancing AI event (EE Times, 2026-07-27) - Chinese memory maker CXMT jumps 466%+ in first day of trading (SiliconAngle, 2026-07-27) - South Korea's Naver confirms it will expand Sejong data center to 200MW (DataCenterDynamics, 2026-07-27) - plus 55 more