AI Fundamentals Are Solid, the Problem Is Rates, the Trigger Is Oil: The Transmission Path from Brent at $94 to Korean Memory

Thu, 23 Jul 2026 15:00:00 +0900

Context: Alphabet’s Q2 results established that the demand debate was effectively over and the market’s scoring criterion had moved to cash. What sets the discount rate in that scoring is interest rates, and this summer’s trigger for rates is oil. This piece synthesizes two macro analysis notes that arrived on the same day, and adds on top of them a reading of the transmission path from a Korean semiconductor perspective. It is also an exercise in sharpening the macro axis that the 45/35/20 AI memory demand scenarios left as a premise behind the 20% downside case.

TL;DR

Strictly speaking, the trigger is not oil itself but the war premium. Of Brent at $94, physical supply-demand justifies $74-78 (the level realized at the June 24 ceasefire-hope moment), and the remaining $16-20 is a fear component. Inventories rose 1.4 million barrels, reversing the expected draw, and the Strait of Hormuz remains open. Because the fear component can round-trip on a single headline, it is the fastest and largest short-term variable in this cycle.
The transmission path runs on a timetable. Today’s oil price is shaping the July CPI due in mid-August, and that CPI will shape the September FOMC (roughly 68% hike odds currently priced by the market). By rule-of-thumb coefficients, a sustained $10/bbl move in oil is worth about +0.2pp on headline CPI and 20-30bp on the 10-year yield.
Oil is not the whole story behind rates stuck at 4.6-4.7%. The 30-year TIPS real yield of 2.95% is the highest since 2008, and behind it sit the fiscal deficit, AI-related bond supply, and the real investment demand of AI CAPEX. AI CAPEX has now entered a self-limiting loop in which it is both a cause of rising rates and a casualty of them.
A depreciation wave is coming for the cloud profit cycle. On an Alphabet basis, company-wide operating income growth is accounting-scheduled to decelerate from +28-32% in 2026 to +15-22% in 2027 and +10-18% in 2028, and the slope is set by Cloud margin. A multiple of around 27x is already demanding the upper end of that range.
Four channels reach Korean memory: valuation (multiple compression), flows (a strong dollar and foreign investors), cost (power and materials), and earnings. The first three are governed by rates, but 2026-2027 earnings are set by contracts, not by rates. Selling Samsung Electronics and SK Hynix on an oil-driven sell-off is a category error, and category errors have historically been opportunities.
The verdicts are clustered in time. In Korea time: the FOMC statement at 3:00 AM on July 30, Chair Kevin Warsh’s press conference at 3:30 AM, the Microsoft call at 6:30 AM, Samsung Electronics and SK Hynix earnings around the same day, and Amazon on July 31. Within 48 hours, the discount rate, big-tech demand, and Korean suppliers’ contracts will be judged in sequence.

Key Framing

Oil is the trigger, rates are the gun, and the AI multiple is the target. Korean memory stands next to the target, in the path of stray rounds. But a stray round and an aimed shot are not the same thing. Rates and oil strike the multiple and the flows, but 2026-2027 memory profits are set by contracts and volumes already signed. Telling this difference in time horizon apart is this summer’s alpha in Korean semiconductor investing, and confusing it will be the most common mistake.

1. Decomposing the Oil Price: $20 of the $94 Is Fear

The past eight weeks of round-trip price action tell the story of what oil actually is right now. Brent plunged from its May peak of $102.58 to $73.74 (the pre-war level) on June 24 ceasefire hopes, then rallied from $76.80 on July 10 to $94.13 on July 22, a $17 move in 12 days. That is a 22% gain in a month, a seven-week high. Over the same stretch, the US 10-year yield fell to 4.4% before snapping back to 4.63%, and oil erased the entire bond rally that had been built on the softer-than-expected CPI print earlier this month. That matches the market’s diagnosis that this rate move is, for the most part, a function of energy prices. [Fact: market data as of 2026-07-22]

Break down the components and an asymmetry appears. Supply risk is at its maximum. US strikes on Iran are now in their 11th day, the Houthis have declared a maritime blockade toward Saudi Arabia, a Caspian pipeline terminal was hit, and the US administration has dismissed negotiations and even floated warnings of strikes on nuclear facilities. Yet physical supply is intact. The Strait of Hormuz remains open (keeping it open is itself the US military’s operational objective), and weekly EIA crude inventories rose 1.4 million barrels, reversing an expected draw. [Fact: news and EIA] The price physical supply-demand justifies is around $74-78, the level realized on June 24, and the $16-20 above that is a war premium built up by geopolitical headlines. [Inference: decomposition]

The nature of this premium is what should determine strategy. The premium carries negative carry. Absent an actual escalation, it decays with time, and as June 24 demonstrated, a single day’s whiff of ceasefire can erase 4% of the oil price and 8bp of yield together. That is why time is an ally for risk assets here, and why chasing oil at $94 is arithmetically the disadvantaged side of the trade.

2. The Transmission Path and Timetable: Today’s Brent Is Writing September’s Rate Decision

The path breaks into three legs. From oil to prices: by rule of thumb, a sustained $10/bbl rise pushes headline CPI up about 0.2pp (gasoline pass-through takes 2-4 weeks, and energy is about 7% of CPI). If the +$20 move versus the June low holds, it puts roughly 0.4pp onto the July-August headline prints. From prices to rates: the observed beta is 20-30bp on the 10-year for every $10/bbl. Oil’s -4.3% on June 24 moved the 10-year -8bp, and the same relationship reappeared in reverse during the July rally. From rates to the Fed: the June minutes named Middle East conflict explicitly as a driver of the persistent-inflation scenario, and roughly half of the committee projected a hike this year. The market’s roughly 68% September hike probability is, in effect, a function of the oil price. [Fact: synthesis of the source notes]

Laid out in Korea time, the calendar reads as follows: the FOMC statement at 3:00 AM and Chair Warsh’s press conference at 3:30 AM on July 30 (the first read is whether he treats the oil move as transitory), weekly EIA inventories every Thursday pre-dawn, the July CPI in mid-August, and the FOMC on September 15-16. September is also the month OpenAI has floated as its IPO target. Today’s Brent is the first domino in this chain, and it is writing, right now, what kind of September the oil price will meet.

3. Dissecting a 4.7% Yield: Three Engines, and Oil Is Only One

The 10-year at 4.63% (July 21, a two-month high), the 2-year breaking above 4.20% technical resistance, the 30-year in the 5s. The whole curve is bear-steepening. Getting the response right requires decomposing this move into three engines.

Engine	Component	Character	Key Indicator
Oil / war premium	30-40bp round trip on a single headline	Volatile component, negative carry	Brent $78 / $100 / $110, EIA inventories
Policy repricing	Policy rate on hold at 3.50-3.75% vs. 2-year at 4.2-4.3%	About 3 hikes priced over 2 years	Dot plot, September hike odds 68%
Real yield / term premium	30-year TIPS at 2.95%, highest since 2008	Most structural	Fiscal deficit, AI bond supply, Treasury auctions

[Fact: market data and Fed publications]

The third engine carries the most weight. It is not expected inflation that is rising but the real price of money itself, and behind it sit a fiscal deficit of about $2tn a year, roughly $300bn a year of AI-related corporate bond supply, and the real investment demand of AI CAPEX, which from the big four alone already runs to 2.2% of GDP. It is telling that the June minutes named strong AI-related demand explicitly, alongside tariffs and the Middle East conflict, as a driver of the persistent-inflation scenario. The Fed has formally flagged the AI boom as an inflation factor. [Fact: FOMC minutes]

This is where the structure running through this whole series completes itself. AI CAPEX sits on both sides of the rate equation at once. Through bond supply, investment demand, and the inflation channel the Fed itself has flagged, it is a cause pushing rates up; through the discount rate and the cost of funding, it is a casualty rates wound in return. This self-limiting loop is a more powerful, exogenous brake on CAPEX than any CFO’s discipline, and it is also the ultimate arbiter of the 2028 FCF-turnaround debate covered in the Alphabet piece. [Inference: structural synthesis]

The disappearance of forward guidance should also be read as design, not accident. The newly appointed Chair Warsh has long been a critic of dot-plot-style forward guidance going back to his time as a governor, and in an environment where oil has round-tripped $28 in eight weeks and payrolls have cooled to 57K versus an expected 115K, any guidance goes stale the moment it is published. There is one practical implication. The Fed’s guidance is no longer the statement; it is the 2-year note. A 2-year at 4.3% is effectively the dot plot now, and the absence of guidance feeds through to a structural rise in the rate-volatility premium, which is to say a structural discount on the multiple.

4. The Cloud Profit Cycle: A Depreciation Wave, and an Upper Bound Already Priced In

The operating income growth rate one to two years out that big tech won’t tell you can be approximated with arithmetic. Alphabet’s actual Q2 incremental margin was 54% (an operating income increment of $6.0bn divided by a revenue increment of $11.15bn). On the other side of the ledger, once CAPEX of $205bn (2026) and $257bn (2027 consensus) lands on the income statement on a 5-6 year depreciation schedule, the annual increment to D&A is estimated at roughly $25bn in each of 2027 and 2028 (a D&A path of roughly $26bn in 2025 → $45bn in 2026 → $70bn in 2027 → $95bn in 2028). [Inference: own estimate]

The conclusion from the arithmetic is this. As long as Cloud revenue increments hold at $45-50bn a year (50%+ growth), incremental profit beats the incremental D&A and margin defense is achievable, and the 24-month conversion of the $514bn backlog contractually underwrites that growth rate through 2027. The honest answer is that the profit cycle’s runway is visible through 2027, and 2028 is contingent on how fast the backlog rebuilds.

Translated into a company-wide operating income growth path, it looks like this.

Year	Operating Income Growth (Est.)	Note
2026	+28-32%	H1 actual +30%
2027	+15-22%	Lower bound if Cloud margin is pushed down to 28%
2028	+10-18%	Upper bound if 35% margin holds plus TPU hardware margin adds on

[Inference: own estimate assuming +20%/yr revenue and 33-35% Cloud margin]

The sensitivity is simple. Every ±5pp in Cloud margin is ±5pp in company-wide operating income growth, and if 2027 CAPEX heads toward $300bn, the 2028 lower bound falls further into single digits. The deceleration itself is fixed by accounting; only the slope is unknown. The market isn’t getting a clean resolution not because management doesn’t know the slope, but because the moment they say it, it becomes guidance. And a multiple of around 27x is already demanding the upper end of that range. At a 10-year yield of 4.65%, a 22x earnings yield of 4.5% on large-cap tech falls below the risk-free rate, an inversion not seen before in this cycle. In this environment, the market paying for near-term cash flow rather than 2030 cash flow is not a matter of taste; it is the scoring standard rates are forcing on it.

5. The Scenario Matrix and the Invalidation Test

Scenario	Probability (Subjective)	Brent	US 10Y	Market Reaction
A. Ceasefire / de-escalation	~40%	Back to $74-80	4.3-4.5%	Relief rally in duration and growth stocks, September hike odds collapse
B. Sustained war of attrition	~40%	$85-95 box	4.5-4.8%	September hike goes through, multiples compress gradually, credit concessions widen
C. Escalation	~20%	$110+	5%+	Equities and credit damaged together, the only scenario where everyone loses

[Inference: scenario probabilities are subjective]

Where the asymmetry sits matters. Because A and C carry larger amplitude than B, a position that settles comfortably into the box (B) is the most vulnerable. And this matrix comes with an invalidation test attached. If oil falls below $80 but the 10-year holds above 4.5%, the culprit is not oil but the structural component (the fiscal deficit, AI bond supply, real yields). In that case the short-term-trigger hypothesis should be rejected and the intensity of de-risking should be raised, because a structural engine is far worse news than a temporary trigger.

The tactical frame here is exploiting the asymmetry, not making a directional bet. Under this regime’s correlation structure (oil up feeding into rates up, rates up feeding into AI multiples down), energy exposure functions as a natural hedge for an AI-centric portfolio. A ceasefire headline works as a pre-booked buying window for repurchasing beaten-down duration and growth names; conversely, chasing oil at $94 and panic-selling on a spike day both run against the nature of the premium (negative carry, capable of evaporating on a single headline). Both notes agree on the same conclusion: if selling is warranted, it should follow confirmation of Scenario C, not a price level.

6. The Transmission Path to Korean Semiconductors: Stray Rounds versus Aimed Fire

This is where this piece adds to the two source notes. The oil-driven rate chain reaching Korean memory decomposes into four channels.

Channel 1, valuation. When rising oil pushes rates up, growth-stock multiples compress, and KOSPI semiconductors get squeezed along with the rest of the global growth-stock basket. The July 13 and 16 sell-offs (SK Hynix -15.37% and -12.34%, Samsung Electronics -10.70% and -9.47%) layered this rate channel on top of earnings downgrades and fears over Chinese supply. There is, however, a difference in texture. Samsung Electronics’ and SK Hynix’s forward PER sits in the low-to-mid single digits, meaning the duration embedded in the multiple is short, unlike the 27x carried by US megacap tech. Rate beta is larger for assets whose earnings expectations sit further out, so low-multiple memory should, in theory, fall less under the same rate shock. If it actually fell more, that is a matter of flows and fear, not rates. [Inference: valuation structure]

Channel 2, currency and flows. A September hike combined with a strong dollar produces won weakness and a foreign-flow headwind. Won weakness itself is favorable for export earnings, but the pattern of large-cap semiconductor names with high foreign ownership becoming the sell window during emerging-market capital-flight episodes has repeated before. The longer Scenario B persists, the more pressure accumulates through this channel.

Channel 3, cost. Oil and power costs make up a single-digit share of fab cost, so the direct shock to memory manufacturing cost is limited. What matters more is data center operating cost. A rise in power prices raises the total cost of ownership of AI infrastructure, which collides head-on with the token-cost-cut goal that SK Hynix Chairman Chey Tae-won has talked about. If oil becomes structurally more expensive, the price elasticity of AI demand comes under a real test, which is a path that thickens the downside premise (demand destruction) in the 45/35/20 scenarios. That said, as the oil-price decomposition above shows, about half of this pressure is still fear rather than structure, so it is too early to call it structural.

Channel 4, earnings. And this is where the conclusion sits. 2026-2027 memory earnings are set by contracts, not by rates. The Q3 server DRAM contract price forecast of +13-18%, LTAs and HBM volumes, and the 2027-vintage HBM pricing negotiations set to begin in Q4 are the variables that determine earnings, and the demand backdrop for those negotiations was, if anything, strengthened by Alphabet’s results. Oil and rates cannot touch these contract variables. So selling Samsung Electronics and SK Hynix on an oil-driven sell-off is a category error, confusing the discount-rate variable with the earnings variable, and as long as the earnings path is undamaged, that category error leans toward being an opportunity. What warrants caution instead is Scenario C. Escalation calls up risk-off, a strong dollar, and demand-destruction fears all at once, so the low-multiple argument stops working as a shield.

Translated by scenario, memory looks like this. In A (ceasefire), the memory names that sold off alongside rate fear bounce back quickly, because the earnings variable was never actually damaged. In B (war of attrition), an earnings-differentiated market continues under multiple compression, with contract prices doing the defending. Stock selection over the index, and earnings revisions over the multiple, drive the outcome. In C (escalation), low multiples are no longer a shield, so hedging is the only response.

7. The Readout Table: A Triple Verdict in 48 Hours

This week is the proving ground for everything in this piece. In Korea time: the FOMC statement at 3:00 AM and Chair Warsh’s press conference at 3:30 AM on July 30, the Microsoft and Meta calls at 6:30 AM, Samsung Electronics and SK Hynix earnings around the same day, and the Amazon call at 6:00 AM on July 31. The discount rate (the Fed), big-tech demand (Microsoft, Amazon), and Korean suppliers’ contracts (Samsung Electronics, SK Hynix) will all be judged, in sequence, within 48 hours.

Warsh’s wording on energy. If he treats the oil move as transitory, scenario weight shifts from B toward A; if he leads with a response to persistence, it shifts from B toward C. Whether the 68% September hike probability gets reaffirmed is the only real news out of this press conference.
Three Brent levels. $78 (confirms the premium has dissolved), $100 (the forced pivot point toward Fed hawkishness), $110 (confirms escalation). Consecutive weekly increases in EIA inventories are a signal that the premium is melting.
The invalidation test. If the 10-year holds above 4.5% even as oil falls below $80, the structural component is the culprit, and the trigger hypothesis should be rejected.
Microsoft and Amazon. Alphabet proved demand, so it is now these two’s turn to prove margin and FCF. The day good earnings stop lifting the stock is the day liquidity has turned its back.
The Samsung Electronics and SK Hynix calls. Whether the Q3 contract price increase confirms the +13-18% range, and whether shareholder-return and partnership events show up in disclosures. The discriminator for memory is, once again, the DRAM contract price, not rates.

Stocks and assets mentioned in this piece are examples for analysis and are not a recommendation to buy or sell any specific security or asset. Responsibility for investment decisions and their outcomes rests with the investor. Scenario probabilities are not statistical values but subjective estimates reflecting public information, and the oil-CPI-rate coefficients are rules of thumb that vary by regime and range. The cloud depreciation path and the operating income growth ranges are our own estimates, not company guidance. The descriptions of hedging and level discipline are examples of a scenario-based response framework, not personalized advice. Market data is as of July 22-23, 2026, and does not reflect subsequent moves.

AI Multiples on Korea Invest Insights