Quick Cut 013｜Reuters朝刊4日間で見えたAI投資の選別——GPUは売れた。だが電力と信用は追いつかない｜AI Demand Is Real. Financing Is the Fault Line
Reuters Roundup: AI Demand Is Real. Financing Is the Fault Line Between October 17 and October 21, Reuters headlines appeared to tell conflicting stories. CoreWeave
Reuters Roundup: AI Demand Is Real. Financing Is the Fault Line
Between October 17 and October 21, Reuters headlines appeared to tell conflicting stories. CoreWeave debt traded at distressed yields. Anthropic reportedly offered double-digit returns for chip financing and later disclosed an enormous first-half net loss, much of it accounting-related. Microsoft withdrew from a UK data-center project where grid access might not arrive until the 2030s. Yet SpaceX reported a 161% jump in quarterly revenue, Super Micro raised its preliminary sales outlook, and ByteDance, Alibaba and Tencent placed more than USD 35 billion of orders for Nvidia products.
These are not contradictions. They describe different points in the same capital cycle.
Re-Cut Update｜21 Oct 2026, 12:19 JST
What changed: South Korea's exports in the first 20 days of October rose 105.4% year on year; the working-day average reached a record USD 4.74 billion, and semiconductor exports rose 222.5% to 44.7% of total exports. Japan's September exports increased 16.7%, below the 20.1% consensus, but Reuters linked the gain in semiconductor equipment to investment by South Korean chipmakers.
At the same time, Australia's NextDC fell 44% after a power-network delay pushed about A\$140 million of contracted FY2027 revenue into later periods. The company cut FY2027 net-revenue guidance to A\$475–500 million from A\$615–640 million, cut EBITDA guidance to A\$245–265 million from A\$385–410 million, reduced capex guidance and paused new developments. Additional power for its M3 Melbourne site is now expected in 2028.
Previous judgment: AI demand is real, while power, credit and execution determine which projects convert contracts into cash.
Change in view: The judgment is maintained, but the evidence is stronger on both sides. Korean chip exports confirm that near-term semiconductor demand is not merely a financing narrative. NextDC confirms that signed customer demand cannot become recognised revenue without usable grid capacity. The correct trade is therefore not “all AI up” or “AI is over,” but a separation between component suppliers that can ship now and infrastructure owners whose revenue depends on power-delivery dates.
Tokyo-market implication: Japanese equipment, testing and memory names gain a demand-side confirmation, but Japan's headline export miss and NextDC's guidance cut argue against treating the sector's morning rebound as a blanket risk-on signal. The next test is whether Advantest, Tokyo Electron, Disco, SCREEN and Kioxia hold VWAP and whether order backlogs convert without delivery delays.
The four supply lines
1. Compute: GPU, HBM, advanced packaging, testing and servers. Demand remains very strong. The evidence includes Chinese orders for more than three million Nvidia-compliant processors, SpaceX’s fast-growing compute-rental revenue and Super Micro’s higher server sales outlook.
2. Power and connectivity: A signed cloud contract cannot operate a GPU cluster without grid access, transformers, cooling, optical links and construction permits. Microsoft’s withdrawal from Nscale’s Essex project shows that demand does not create usable megawatts on command. Amazon’s expanded purchases of 3M optical-interconnect products show how AI spending is moving beyond chips into the physical network.
3. Credit: CoreWeave’s stressed bonds and the double-digit yield reportedly offered in Anthropic-related chip financing show that capital now has a price. Long customer contracts can support borrowing, but mismatches between debt maturity, contract duration and project completion create refinancing and renewal risk.
4. Monetization: SpaceX converted compute capacity into revenue, but AI absorbed USD 16.4 billion of its USD 19.6 billion quarterly capital expenditure and the unit still carried heavy operating losses. The correct question is not whether revenue exists. It is whether lifetime cash generation exceeds the cost of capital and continuing reinvestment.
TokyoBrainie’s cut
The AI boom is entering its second accounting phase. In the first phase, scarcity rewarded almost every supplier. In the second, investors must distinguish orders from paid utilization, revenue from free cash flow, and a prestigious customer from a bankable contract.
This matters for Japanese equities. Semiconductor equipment and HBM suppliers sell nearer to the front of the buildout and may receive revenue before the final AI service becomes profitable. Optical-fiber, cable, transformer and cooling suppliers benefit when bottlenecks shift from processors to the physical grid. Banks stand farther back: they are paid for bearing credit risk, but limited disclosure and special-purpose vehicles can obscure the final exposure. SoftBank Group is different again; its sensitivity comes through valuation, financing and portfolio concentration rather than the sale of physical components.
Market map
- Direct demand beneficiaries: Advantest, Tokyo Electron, Disco, SCREEN, Kioxia.
- Physical supply-line beneficiaries: Furukawa Electric, Fujikura, Sumitomo Electric, optical components, power equipment and cooling.
- Credit transmitters: MUFG and other lenders—systemic risk appears limited for now, but individual exposure and provisioning still matter.
- Narrative and valuation exposure: SoftBank Group and companies valued mainly on future AI optionality.
Counterpoint
The bear case can also be overstated. J.P. Morgan estimated that most global banks have limited CoreWeave exposure and described bank financing as secured and milestone-based. Strong server sales and compute-rental revenue show that the buildout is not merely fictitious circular financing. Falling yields, faster grid connections or rising utilization could repair the weakest parts of the thesis.
What changes the view
Watch paid utilization, customer concentration, free cash flow after capital expenditure, debt maturity versus contract length, grid-connection dates, transformer and optical-component lead times, and the share of AI spending funded by operating cash rather than new debt.
Verdict: AI demand is real. The next winners will be selected by the quality of their supply lines and the speed at which invested capital returns as cash—not by the presence of “AI” in a headline.
At 13:49 JST, Reuters reported that Samsung Electronics and SK Hynix are upgrading their Chinese NAND plants. The companies confirmed that upgrades are under way, but the specific layer counts, output targets and the extent of U.S. export-license approval rely partly on unnamed sources and remain incomplete. Samsung's Xian conversion reportedly targets 286-layer V9 NAND, while SK Hynix is seeking 240-layer-or-higher production at Dalian.
At 18:14 JST, Supermicro's preliminary update supplied a separate demand signal: estimated fiscal Q1 2027 revenue of \$18.5–19.5 billion versus prior guidance of \$14.5–15.5 billion, and gross margin of 12.0–12.5% versus 10.4–10.8%. The company attributes the upside mainly to AI-platform demand and customer/product mix. These figures are unaudited and may change when final results are released on 3 November U.S. time.
The earlier thesis is maintained, not reversed. AI demand is converting into server revenue and memory upgrades, but power, financing and export licences still decide where that demand can become usable capacity. For Tokyo, this is positive evidence for equipment, testing, server components and memory demand—but not a blanket buy signal. Kioxia faces both stronger end demand and potentially stronger NAND competition; Japanese equipment makers benefit only where licences and actual tool deliveries permit.
Wall Street confirmed that AI demand can still pull capital back into technology: on 21 October, the Nasdaq rose 1.05%, the Philadelphia Semiconductor Index gained 2%, Super Micro closed 9.6% higher, Dell added 5.3% and HPE rose 3.7%. This completed the first market test listed in the previous update: Super Micro's preliminary upside did spread to peers.
The post-close evidence was more selective. Meta reported quarterly revenue of \$64.60 billion versus the \$63.37 billion LSEG consensus, guided fourth-quarter revenue to \$73–76 billion, and raised the lower end of its 2026 capex range to \$135–145 billion. Its shares rose more than 4% in extended trading. However, third-quarter free cash flow was negative \$2.12 billion, versus positive \$10.63 billion a year earlier. This is stronger evidence of infrastructure demand, but not proof that rising AI expenditure is already converting into cash.
Lam Research reported record September-quarter revenue of \$8.22 billion and a 40.0% GAAP operating margin. Its December-quarter revenue guidance of \$8.6 billion, plus or minus \$0.4 billion, was close to market expectations rather than a large upside surprise; the shares fell roughly 5% after hours according to a market report. Texas Instruments provided the broader-cycle counterpoint: Q3 revenue was \$6.06 billion and its Q4 revenue range of \$5.8–6.3 billion was above Wall Street's estimate, indicating recovery in industrial and automotive analog demand alongside AI data-center growth.
Thesis update: the original thesis is maintained, but the Tokyo transmission becomes more selective. Demand is visible in servers, hyperscaler capex, analog chips and wafer-fabrication equipment. Yet equipment shares can fall after record results when valuation already assumes a larger guidance beat. The next Tokyo test is therefore not whether semiconductor shares open higher, but whether Advantest, Tokyo Electron, Disco, SCREEN and Kioxia hold VWAP and relative strength after the opening gap.
DISCO's 16:00 JST results converted the AI back-end demand story into audited operating numbers. Fiscal Q2 sales reached JPY 142.8 billion, up 36.4% year on year and 24.9% quarter on quarter; operating income was JPY 67.6 billion, up 52.3% year on year, with a 47.4% margin. Shipment value reached a record JPY 148.6 billion. First-half operating cash flow rose to JPY 111.7 billion and free cash flow to JPY 56.9 billion, although inventories increased by JPY 19.5 billion.
The next-quarter profile is deliberately more mixed. DISCO forecasts October–December sales of JPY 132.4 billion and operating income of JPY 58.1 billion, both below the September quarter, while shipment value is forecast at a fresh JPY 149.6 billion. Because equipment revenue is recorded on inspection and acceptance, shipment value is the cleaner near-term demand signal. The forecast assumes USD/JPY 151 and EUR/JPY 173.
At 18:39 JST, however, the market signal reversed from the morning's optimism: Nasdaq 100 futures were down 1.93%, Nvidia was down 2.4%, and major chip shares were 2.9–5.3% lower in U.S. premarket trading. Meta, which had risen more than 4% after hours, was down about 0.8% in premarket trade. Brent at USD 112 and renewed attacks in Saudi Arabia added an inflation and risk-premium shock.
Thesis update: demand is not the weak link in DISCO's numbers; valuation, macro risk and the conversion of record shipments into accepted revenue are tomorrow's tests. Strong fundamentals no longer guarantee a positive opening when oil, rates and AI capital discipline are repriced together.
The previous evening update recorded a premarket warning. The U.S. close has now converted that warning into an observable market test. On 22 October, the Nasdaq fell 2.60%, the S&P 500 lost 1.21%, and the Philadelphia Semiconductor Index dropped 6.7%—its largest one-day decline since 23 June. Meta fell 7.0%; Intel, AMD, Sandisk and Marvell each lost more than 8%.
The repricing spread from equity to credit. Meta's five-year CDS widened 7.6 basis points to a record 114.79 basis points after the company raised the floor of its 2026 capex outlook, reported a USD 2.12 billion quarterly free-cash-flow deficit and introduced non-voting shares. CoreWeave fell 8.4%, while other AI-infrastructure names also suffered heavy losses.
Safe-haven demand pulled the U.S. 10-year Treasury yield down 15.17 basis points to 4.994%, below 5% for the first time since 23 September. The yen strengthened 0.51% to JPY 157.33 per dollar. Yet Brent settled 4.0% higher at USD 113.17, its highest close since 4 May. Lower Treasury yields therefore did not signal an easier inflation environment; they reflected risk aversion while oil added a separate inflation shock.
Previous judgment: AI demand is real, but valuation, credit, power, execution and capital recovery determine the stock-price outcome.
Update: The judgment is maintained and strengthened. DISCO's operating results still validate back-end demand, but the U.S. close shows that strong demand cannot protect the sector when investors reprice financing, dilution and cash recovery at the same time. For Tokyo, the first test is no longer whether the U.S. premarket selloff persists—it did—but whether Japanese semiconductor leaders can absorb it without persistent VWAP failure.
The 23 October U.S. close provided the first counterevidence to Thursday’s AI-credit panic. The S&P 500 technology sector gained 0.35% and the semiconductor index rose about 0.4%, while the Nasdaq slipped only 0.21% after its 2.60% fall the previous day. Meta edged lower, SpaceX, Oracle and Dell gained 1%–2%, and Nvidia, Alphabet and Amazon fell by less than 1%. Reuters also reported a modest narrowing in CDS spreads for major AI companies.
This is stabilization, not resolution. The S&P 500 still fell 0.52% and the Dow 1.23%, while Brent crude settled at USD 116.72 and the U.S. 10-year yield returned to 4.99%. Higher energy and funding costs remain a direct challenge to data-centre economics and high-duration valuations.
Previous judgment: AI demand is real, but financing, dilution, capital recovery and valuation determine the market outcome.
Update: The warning is retained, but the “uncontrolled liquidation” branch weakened on Friday. For Tokyo on 26 October, semiconductor leaders have a better chance of a technical rebound, yet that rebound should be trusted only if price, VWAP, breadth and credit improve together.
Friday’s modest recovery in U.S. semiconductor shares reduced the probability of uncontrolled liquidation, but Tokyo did not confirm a sector rebound on Monday morning.
At 10:27 JST, the Nikkei 225 was down 0.43% at 62,425.47 while the broader Topix gained 0.48%. Market breadth was positive—166 Nikkei constituents rose and 59 fell—yet heavyweight technology shares dragged the index lower. Sumco fell 4.03%, Kokusai Electric 3.86% and SCREEN Holdings 3.78%.
This is not a broad risk-off signal. It is a relative-pricing signal: energy, trading companies and many non-technology shares advanced while semiconductor names remained under pressure. The earlier judgment that panic had paused still stands, but the expected technical rebound in Tokyo’s chip leaders has not yet been confirmed.
The next test moves to U.S. hyperscaler earnings and Asian memory makers. Microsoft, Alphabet and Amazon, along with SK Hynix and Samsung Electronics, report this week. Their capital spending, free cash flow and demand guidance will matter more than a one-session bounce.
Updated judgment: AI demand has not been disproved, and broad market breadth argues against systemic liquidation. However, Tokyo’s semiconductor sector remains the weak link. A durable rebound requires price, VWAP, breadth and credit to improve together.
