What Matters More Than a 5% Yield: Whether AI Profits Become Cash

The U.S. 10-year Treasury yield closed at 4.96% on September 11. From September 1 to September 11, the two-year yield rose 24 basis points while the 10-year rose 17 basis points. This looks less like a crisis of confidence isolated to long bonds and more like a repricing of the policy-rate path and real yields.¹²

AI supply-chain profits were also real. NVIDIA reported $89.0 billion of Data Center revenue, while Samsung Electronics reported KRW 89.2 trillion of operating profit in its Device Solutions division. A weak macro environment does not erase these results. Strong results, however, do not prove that every future AI capital project will earn an adequate return.³⁴

For the next six to 12 months, the investment focus should be on AI businesses whose valuation still works under conservatively reduced earnings and whose cash collection is visible, rather than on a blanket bet on macro deterioration. Within that group, memory ranks ahead of long-duration expansion stories that depend heavily on external financing. Samsung Electronics is the clearest Korean example in this review.

Market data in this article is measured through September 11, 2026. Company results refer to each issuer’s latest reported quarter. Different fiscal periods are not combined into a single quarter.

The Rate Move Looks More Like a Discount-Rate Shock Than a Credit Crisis

The U.S. Treasury curve shows that short yields rose more than long yields.

One basis point is 0.01 percentage point. This curve cannot identify the exact contribution of every cause. It does show why a simple story of indiscriminate panic selling across the Treasury market is incomplete.¹

The move becomes clearer when nominal and inflation-protected Treasury yields are compared.

The difference is a simple proxy for breakeven inflation. It contains risk and liquidity premia, so it is not identical to pure inflation expectations. Even so, 80% of the two-day increase in the nominal 10-year yield corresponded to a rise in the real yield.¹²

Higher real yields are especially difficult for growth stocks priced on distant profits. They are less damaging to companies that already earn substantial profits and trade at lower multiples.

Credit markets do not yet indicate a broad freeze. The ICE BofA U.S. High Yield Index option-adjusted spread was 270 basis points on September 10, only two basis points above September 4. Credit spreads can react late, but this move is not enough to establish that a credit crisis is already under way.⁵

U.S. core consumer prices rose 0.3% month on month in August, up from 0.2% in July. The year-on-year rate slowed from 2.5% to 2.4%. The monthly acceleration deserves attention, but it does not by itself prove a complete loss of inflation control.⁶

A 5% 10-year yield is not a universal corporate break-even point. Project economics depend on the spread between expected returns and the full cost of capital. Cash balances, debt maturities, customer contracts and payback periods differ across companies.

The task is not to sell every AI equity when the 10-year crosses 5%. It is to identify which business models lose their economics first when capital remains expensive.

Energy Risk Is Real, but That Does Not Prove Energy Stocks Are Cheap

U.S. refinery utilization was 98.0% in the week ended August 28 and 97.8% in the week ended September 4. With the system already operating at a high rate, there is limited room to solve product shortages quickly by simply running refineries harder.⁷

The U.S. Energy Information Administration forecast distillate inventories below 100 million barrels in September. Autumn refinery maintenance and harvest-season diesel demand add pressure. The sub-100-million figure is a forecast, not a confirmed inventory reading.⁸

The forecast’s assumptions matter. The September 9 outlook was completed on September 3. Its path toward lower crack spreads through mid-2027 assumes that normal tanker traffic through the Strait of Hormuz returns in the near term. The EIA also states that a constraint lasting beyond the end of 2026 would imply higher crack spreads than in its base forecast.⁸

Citing an official forecast and validating its assumptions are different jobs. The supply risk is real. That still does not establish that energy equities are cheaper than AI equities today.

Refiners require company-level work on crude access, operating continuity, product mix and valuation. This article does not underwrite normalized earnings and prices for individual energy companies, so it does not support shifting the portfolio center toward energy.

AI Results Are Real, but Cash Lands in Different Places

Reported revenue, profit and cash flow are stronger evidence than product announcements.

These results rebut the claim that all AI demand remains hypothetical. They do not establish that every future capital project will be economic.³⁴⁹¹⁰

Cloud capital expenditure becomes revenue for NVIDIA and then for memory suppliers. Those flows are not three independent sources of final demand. Investors must distinguish end-customer payments, cash generated by existing businesses and externally raised funding.

Oracle’s quarter ended August 31, 2026 illustrates the difference.

Subtracting the increase in these prepayments for analytical purposes produces -$16.759 billion. This is neither Oracle’s reported free cash flow nor a normalized cash-flow estimate. It is a comparison showing how much customers helped fund investment during the quarter.¹¹

Prepayments are positive evidence of customer commitment and burden sharing. They are also cash received before future service delivery, so they should not automatically be extrapolated as recurring cash generation.

NVIDIA also requires a cash check. Operating cash flow in the same quarter was $24.077 billion, down 52.2% from the prior quarter. The increase in accounts receivable reduced operating cash flow by $22.346 billion.³

Those figures do not establish collection failure or poor revenue quality. Rapid growth and payment timing can consume working capital. They do make the next quarter’s cash collection more important.

Microsoft generated $55.441 billion of operating cash flow in the quarter ended June 30. Subtracting $35.802 billion of additions to property and equipment leaves $19.639 billion. This does not include every lease and future commitment, but it disproves the claim that every AI infrastructure buyer has exhausted its internal cash.¹⁰

The weak category is not AI as a whole. It is the set of businesses whose dependence on financing grows faster than their revenue.

AI Growth and Macro Pressure Meet in the Same Causal Chain

The transmission path is straightforward:

Better AI performance and higher usage → more compute, memory and power demand → supply bottlenecks and higher investment cost → higher supplier profits and a heavier funding burden on buyers

AI may eventually lift productivity and reduce inflation. There is no guarantee that these benefits arrive before the investment-stage cost increases. Companies need time to change workflows and deploy the technology.

The same price move can create opposite effects inside one company. Samsung’s DS division earned KRW 89.2 trillion of operating profit, while its Mobile eXperience and Networks businesses recorded a KRW 0.7 trillion operating loss amid component-cost pressure. The economics of memory sellers and buyers are not the same.⁴

An investor can believe in AI progress without betting on a return to low rates. The expression is to prefer companies receiving higher costs as revenue and profit over companies only absorbing those costs.

Power bottlenecks are not uniformly positive for suppliers. Scarcity can raise the value of existing assets while delaying operation of the data centers that would use new equipment.

On August 3, Texas ordered a comprehensive audit of data centers in the ERCOT interconnection process. No project can move forward before the audit is completed, and non-compliant projects will be denied grid connection. This is a concrete delay risk, not evidence of a permanent nationwide stop to AI infrastructure construction.¹²

Announced investment budgets should receive less weight than projects with verified financing, permits, grid connection and operating schedules.

Samsung’s Price Can Absorb a Material Decline in Earnings

Ignoring dividends, taxes and transaction costs, a share price can be decomposed into earnings per share and the price-to-earnings multiple.

Future price ÷ current price = future EPS ÷ current EPS × future P/E ÷ current P/E

If earnings rise 50% while the P/E falls 20%, the price rises 20% arithmetically. If earnings rise 20% while the P/E falls 30%, the price falls 16%. Choosing the right industry and paying the right price are separate decisions.

Samsung Electronics common shares closed at KRW 259,500 on September 11. The company reported second-quarter EPS of KRW 10,849. Simply annualizing that quarter produces KRW 43,396. This is not a 2027 EPS estimate. It is a starting point for a stress test.⁴¹³

The table does not prove that Samsung is unconditionally cheap. It shows that the multiple remains reviewable after a substantial reduction in current earnings.

Using a 10x future P/E as an analytical assumption produces the following outcomes.

A 15% gain from the current price requires roughly KRW 29,843 of EPS at a 10x P/E, or about 69% of annualized second-quarter earnings.

The 10x multiple is not guaranteed. The 17% upside case requires both 70% earnings retention and a market willingness to assign 10x to those earnings. A low trailing multiple alone does not establish upside.

The downside must also be explicit. If only 50% of annualized earnings remains and the P/E falls to 8x, the implied value is KRW 173,584, about 33.1% below the September 11 close. This is not a maximum-loss estimate.

The central variable is not whether the 10-year yield touches 5%. It is whether Samsung can sustain annual EPS near KRW 30,000 under server-led demand and supply constraints. KRW 30,000 is not a verified 2027 forecast. It is a valuation threshold for testing what the current price requires.

Price and Cash Collection Rank Ahead of the Growth Story

Four groups produce a clearer priority order.

This is not a complete market ranking. SK hynix and all optical and power-equipment companies have not been compared under the same earnings and price framework. The ranking should not be extended to names that have not been underwritten.

Company preference and position size must also remain separate. Even an attractive issuer is subject to single-name concentration, correlated exposure to the same demand cycle and available settled cash. This article does not provide personal allocation or trading instructions.

The View Changes When Credit and Earnings Deteriorate Together

These thresholds are analytical monitoring proposals. They are not validated automated trading rules or economic laws.

The strongest counterargument is that today’s profits may represent the peak of a supply shortage. If memory supply expands while customer projects are delayed, earnings and valuation multiples can decline together. There is no proof that today’s low multiple absorbs all of that risk.

Even so, conservatively estimating durable earnings is more useful than trying to time the end of macro stress. The alpha does not come from predicting when the macro environment improves. It comes from the gap between profits that survive a bad environment and the price paid for those profits.

Disclaimer: For research and information purposes only. Not investment advice. Readers should perform their own due diligence and consult licensed advisers before making any investment decision.