NVIDIA Turns AI Demand Into Cash; Korea Raises Capital’s Price

The day’s strongest signal is a change in the AI financing question. NVIDIA’s August 26 results show that frontier-compute demand is no longer supported only by capital-expenditure plans or model usage claims: a major supplier is converting it into revenue, margin and cash at extraordinary scale. Its August 26 Form 10-Q also shows how much inventory, supplier capacity, credit support and long-duration commitment now sit behind that conversion. The Bank of Korea’s decision early on August 27 supplies the macro counterweight. The central bank raised its rate while explicitly describing robust AI investment and semiconductor strength as part of a faster-growth environment that also carries inflation and financial-stability pressure. These records are connected by capital intensity, not causality.

NVIDIA reported fiscal second-quarter revenue of $96.221 billion, 18% above the previous quarter and 106% above a year earlier. Data Center revenue reached $89.0 billion, up 18% sequentially and 117% year over year; Edge Computing contributed $7.2 billion. GAAP gross margin was 75.0%, GAAP operating income was $63.734 billion and GAAP net income was $59.688 billion. The company’s non-GAAP net income was lower, at $53.954 billion, because its reconciliation excludes net gains on equity securities and other items while, beginning this fiscal year, no longer excluding stock-based compensation. That qualification matters: both measures are disclosed, but neither should be silently substituted for the other.

Cash conversion was substantial. For the first six months of fiscal 2027, NVIDIA generated $74.421 billion of operating cash, compared with $42.779 billion a year earlier. The quarter’s release says the company returned about $26.0 billion through repurchases and dividends. Its third-quarter outlook calls for $108.0 billion of revenue, plus or minus 2%, with 74.0% GAAP and non-GAAP gross margins, plus or minus 50 basis points. The outlook assumes no Data Center compute revenue from China. That exclusion makes the forecast easier to interpret but also preserves export controls and market access as material sources of upside and downside.

The balance sheet turns the success story into a duration-and-risk story. Accounts receivable rose to $63.059 billion at July 26 from $38.466 billion at January 25, and inventory rose to $31.575 billion from $21.403 billion. Customer advances increased to $2.8 billion from $160 million. None of those movements proves deterioration: receivables and inventory normally expand with a business that has more than doubled revenue, while advances can improve funding visibility. They do show that more cash, components and customer credit are moving through the system before every transaction is finally settled.

Future commitments are larger still. NVIDIA disclosed $366 billion across future supply and capacity, cloud-service agreements, data-center leases not yet commenced, equity investments and capital expenditures. Supply-and-capacity commitments alone reached $279 billion, up from $119 billion one quarter earlier, primarily for memory and manufacturing facilities needed for current and future data-center systems. The company also disclosed land, power and shell guarantees for selected AI-cloud partners’ lease obligations; their fair values were not significant at quarter-end. Memorandums with large capital providers aim to mobilize more than $500 billion of third-party infrastructure capital over time, but the 10-Q says those preliminary arrangements may not become definitive agreements. NVIDIA may offer limited residual-value support on selected projects at its option. The $500 billion is therefore neither booked NVIDIA revenue nor a committed single fund.

This distinction is the central limitation on the phrase “compute is revenue,” which NVIDIA’s chief executive used in the earnings release. The quarter confirms revenue and cash for NVIDIA; it does not establish the returns of every AI-cloud operator, data-center landlord, power project or enterprise customer purchasing the capacity. NVIDIA can receive cash before the full economic life of a financed AI factory is known. Investors and operators should separate vendor revenue, customer utilization, project debt service and end-user productivity. They occur at different layers and on different clocks.

The US Census Bureau’s July durable-goods release prevents one supplier’s growth from becoming a claim about all capital spending. Total new orders rose 1.1% to $339.3 billion, but the gain was 0.4% excluding transportation. Nondefense capital-goods orders excluding aircraft—a common proxy for business-equipment demand—rose just 0.2% after a revised 1.7% increase in June. Machinery orders increased 1.2%, while computers and electronic products fell 1.1%; within that category, communications-equipment orders rose 0.4%. The series is seasonally adjusted but not adjusted for price changes, excludes semiconductor manufacturing from new and unfilled orders, and comes from a non-probability panel for which sampling error and statistical significance cannot be measured. It is a directional cross-check, not a clean measure of AI capex.

The Bank of Korea made the financing constraint explicit. By a six-to-one vote, its Monetary Policy Board raised the base rate 25 basis points, from 2.75% to 3.00%, the second consecutive increase. The statement expects the global economy to grow moderately with support from robust AI investment even as energy costs keep inflation elevated. Domestically, it says exports and investment remain strong because of the semiconductor cycle. The bank raised its 2026 growth forecast from 2.6% to 3.3% and its 2027 forecast from 2.1% to 2.9%. At the same time, July headline inflation was 2.8%, core inflation rose to 2.6%, Seoul-area housing prices remained on a strong upward path and household loans increased substantially.

That is not evidence that NVIDIA’s quarter caused Korean inflation or the rate decision. The bank cites accumulated cost pressure, stronger domestic demand, housing, household credit, oil, exchange rates and geopolitical uncertainty. The operational connection is narrower: Korea’s semiconductor exports and investment benefit from the AI cycle, while the same stronger growth reduces the central bank’s room to overlook inflation and leverage. The Board now says it will decide the timing and pace of additional increases. A technology boom can therefore strengthen income and investment while also raising the hurdle rate applied to the next factory, lease, cloud contract or startup round.

For infrastructure operators, the useful metric is no longer accelerator supply alone. Watch utilization, customer concentration, receivable days, inventory transitions, power readiness, lease guarantees, residual-value exposure and the match between long-lived project debt and shorter technology cycles. For enterprises—including Iranian firms paying for compute in foreign currency—the choice between owned hardware, reserved capacity and usage-based cloud should be tested against completed useful work, not nominal GPU access. A lower token price does not rescue an underused reservation, and a cheaper server does not solve power, cooling, maintenance or currency risk.

The next evidence is specific. NVIDIA must convert its $108 billion outlook without a China Data Center contribution, keep margins near guidance and turn rising receivables and inventory into cash while executing a $279 billion supply-and-capacity book. The proposed financing platforms need definitive agreements, independent underwriting and transparent risk allocation. Korea’s next inflation, housing-credit, export and semiconductor-production data will show whether stronger growth keeps the tightening bias alive. US manufacturing revisions will test whether July’s narrow equipment growth broadens. Today’s defensible conclusion is that AI demand has cleared the supplier-revenue test; the harder test is whether the financed physical system can earn more than its rising, increasingly visible cost of capital.