Chip Stock Selloff Deepens as China’s Lithography Breakthrough and Nvidia’s $750 Billion Deal Push Test the AI Trade

Last updated: July 28, 2026, 2:30 p.m. CEST (8:30 a.m. ET)

The global equity market split cleanly in two on Tuesday, and the fault line ran straight through the artificial-intelligence trade. Semiconductor and AI-linked shares were sold hard for a third consecutive session across Asia and Europe, while consumer staples, autos and other unglamorous cash-generative businesses attracted the money coming out of them. Japan’s Nikkei 225 closed roughly 4% lower, its weakest finish in more than two months. The pan-European STOXX 600 nevertheless edged higher, with the technology sub-index down about 0.8% and consumer names sharply up. In the United States, Nasdaq 100 futures pointed to a lower open even as Dow futures pointed higher — an unusually wide divergence between two indexes that normally move together.

This is not a market in broad retreat. It is a market changing its mind about one specific thing: whether the enormous capital being poured into AI infrastructure will earn a return, and who is actually paying for it.

Three events over the previous four trading days did the damage. First, reporting by The Information that a Chinese state-backed group has begun mass-producing domestically developed immersion deep-ultraviolet lithography machines knocked ASML shares to their lowest level since June and dragged the entire chip-equipment complex with them. Second, Chinese memory maker ChangXin Memory Technologies — CXMT — surged roughly 470% on its Shanghai debut on Monday in the largest mainland Chinese listing since 2010, an event that simultaneously validated the memory boom and reminded investors who might supply the next wave of it. Third, Bloomberg reported that Nvidia is working on a fresh round of AI agreements worth more than $750 billion, including a proposed backstop of as much as $250 billion tied to an OpenAI data-centre lease. That last item reopened the argument that has shadowed the AI rally for a year: how much of the demand for Nvidia’s chips is being financed by Nvidia itself.

Layered on top is a Federal Reserve meeting that began on Tuesday and concludes on Wednesday, July 29, with a policy statement at 2:00 p.m. ET. It is Chair Kevin Warsh’s second meeting since succeeding Jerome Powell, and unusually for a mid-cycle gathering, the risk is not framed around a cut. Nine of eighteen participants penciled in at least one rate increase this year in the June projections. Then Microsoft and Meta Platforms report after Wednesday’s close, and Apple and Amazon follow on Thursday. Between them, those four companies plus Alphabet are on course to spend something in the region of $700 billion on capital investment this year.

What follows is an account of how the pieces fit together, what has actually been verified, what remains rumour, and which of the competing interpretations the evidence currently supports.

Key Takeaways

• Main development: A three-day global selloff in semiconductor and AI-infrastructure shares intensified on July 28, driven by reports of Chinese progress in immersion DUV lithography, a blockbuster Shanghai listing by Chinese DRAM maker CXMT, and renewed scrutiny of Nvidia’s vendor-financing arrangements.
• Key figures: Nvidia closed at $196.51 on Monday, July 27, down 4.99% on the day. Japan’s Nikkei 225 fell about 4% on Tuesday to its lowest close since late May. ASML fell between roughly 4.6% and 7% on Monday depending on the listing and time of measurement.
• Where the money went instead: Unilever rose about 6% in London on Tuesday, its largest single-day gain in two years, after reporting 5.8% underlying sales growth in the second quarter. Mercedes-Benz gained more than 4% after beating consensus adjusted EBIT by roughly 39%.
• Macro backdrop: Oil extended a sharp retreat as the United States and Iran held a fragile pause in hostilities; gold traded below $4,100 an ounce; the dollar index sat near a one-month high at about 101.55 as traders priced a meaningful chance the Fed raises rates rather than cuts.
• Why it matters: The market is no longer treating AI capital expenditure as automatically value-accretive. Alphabet’s second-quarter free cash flow was negative $5.85 billion, the first negative quarter since its 2004 flotation, and the shares fell more than 7% the following session despite a 24% revenue increase.
• What comes next: The FOMC decision lands Wednesday, July 29 at 2:00 p.m. ET with no updated Summary of Economic Projections. Microsoft and Meta report the same afternoon; Apple and Amazon on Thursday, July 30.

The split tape: what actually happened on July 28

Start with the closing prices, because they are the only part of this story that is not open to interpretation.

On Monday, July 27, the Dow Jones Industrial Average rose 262.83 points, or 0.51%, to 52,210.08. The S&P 500 added 0.02% to close at 7,413.18. The Nasdaq Composite fell 0.18% to 24,932.08. That is a textbook rotation day: the industrial and consumer-heavy index up half a percent, the technology-heavy index down, the broad index almost exactly flat because the two cancelled each other out. Beneath the surface the dispersion was far larger than the headline numbers suggest. The VanEck Semiconductor ETF lost more than 2%, adding to losses from the previous Friday. Advanced Micro Devices fell about 5% and Teradyne about 4%. SanDisk dropped roughly 11%.

Nvidia, the single largest weight in most global equity benchmarks, fell 4.99% to close at $196.51.

Asia picked up the baton overnight. Japan’s Nikkei 225 fell roughly 4% to its lowest close in more than two months, with the broader Topix down a more modest amount — a gap that tells you the damage was concentrated in the index’s semiconductor-equipment heavyweights rather than spread across the market. Advantest, Tokyo Electron and Kioxia were among the worst affected. Korean chipmakers fell alongside them. Hong Kong’s Hang Seng managed a small gain, and mainland Chinese indices were mixed, which is itself informative: the news that hurt Tokyo and Amsterdam was, from Beijing’s perspective, good news.

Europe opened lower and then diverged. By mid-morning the STOXX 600 was up about 0.3% at 646.75, with the technology sector down roughly 0.8% after a nearly 2% fall the previous day, and consumer staples and consumer discretionary names leading the advance. Germany’s DAX, Britain’s FTSE 100 and France’s CAC 40 all posted modest gains. The Euro STOXX 50, which carries a heavier weighting in ASML and other large-cap technology, was close to unchanged.

The commodity and currency backdrop was doing its own work. Crude extended a multi-session decline on signs that the United States and Iran were sustaining a pause in hostilities, with West Texas Intermediate trading in the low $80s and Brent in the mid-$80s during European hours. Gold, which had carried a substantial war premium for months, fell more than 1% and traded below $4,100 an ounce. The dollar index held near a one-month high at roughly 101.55, with the euro at about $1.1366.

Put together, the picture is coherent: a violent, sector-specific de-rating of AI-linked hardware, offset by a bid for companies that sell soap, cars and consumer staples, all taking place under a monetary-policy cloud that has not lifted.

The trigger: China begins building its own immersion DUV scanners

The report that started this leg of the selloff came from The Information on Monday, July 27, and was quickly picked up across the trade and financial press. Shanghai Yuliangsheng, a lithography startup with ties to Huawei and to the state-backed equipment group SiCarrier, has reportedly begun mass production of an immersion deep-ultraviolet lithography scanner built largely from Chinese components. According to that reporting, roughly five systems are targeted for delivery in 2026 and about twenty in 2027, with the first units going to SMIC, Hua Hong and CXMT. SMIC has been trialling the tool since around September 2025 and could move it into production use as early as 2027.

ASML fell sharply on the news. Estimates of the size of the drop vary between roughly 4.6% and 7% depending on whether one measures the Amsterdam or Nasdaq listing, and at what point in the session — a reminder that a single headline percentage rarely captures a cross-listed stock’s day. What is not in dispute is direction: the shares closed at their lowest level since June, and the pressure continued into Tuesday’s European session.

Before assessing what this means, it is worth being precise about what a DUV immersion scanner is and is not, because the distinction is doing a lot of work in the bull case.

DUV, EUV, and the gap between them

Lithography is the step in chip manufacturing where circuit patterns are projected onto a silicon wafer coated in light-sensitive material. The shorter the wavelength of light, the finer the features that can be printed. Deep-ultraviolet systems use a 193-nanometre argon-fluoride laser. Immersion DUV improves resolution further by placing a layer of purified water between the final lens and the wafer, raising the effective numerical aperture. Extreme-ultraviolet systems use 13.5-nanometre light generated by vaporising tin droplets with a high-power laser inside a vacuum chamber, and represent a step change in complexity, cost and capability.

ASML is the only company in the world that makes EUV systems. It has held that position for more than a decade, and there is no credible evidence that any Chinese entity is close to matching it. Export controls agreed between the Netherlands, Japan and the United States have blocked EUV sales to China entirely since 2019, and have progressively tightened around the most advanced immersion DUV models as well.

The Yuliangsheng tool, as described in the reporting, is designed for chips at roughly the 28-nanometre node. With multi-patterning — the technique of exposing the same wafer multiple times to build up finer structures than a single exposure allows — it could reportedly be pushed toward 7-nanometre, or 5-nanometre at reduced yields. That is not a trivial capability. It is roughly the process class SMIC has used to produce Huawei’s advanced mobile processors. But multi-patterning is expensive: each additional exposure adds cost, cycle time and defect risk, which is precisely why the industry moved to EUV in the first place.

Two further caveats deserve emphasis. Key components in the Chinese systems are reportedly still imported from Japan, meaning the supply chain is not yet sovereign. And domestic supplier delays have already slowed production this year. A target of five machines in 2026 and twenty in 2027 is, in the context of ASML shipping hundreds of systems annually, a rounding error on near-term revenue.

So why did the market react so violently?

Because the market was not pricing near-term revenue. It was pricing permanence.

The investment case for ASML has rested on something close to a natural monopoly at the most technically demanding step in the most strategically important manufacturing process on earth. A monopoly of that kind supports valuation multiples that a merely excellent capital-goods company cannot. Any evidence that the moat is crossable — even slowly, even at a lower technology node, even with imported parts — changes the terminal-value assumptions embedded in the price. That is what a de-rating is: not a revision to next year’s earnings, but a revision to the confidence interval around the next fifteen years of them.

There is a second, more immediate channel. Mainland China accounted for about 19% of ASML’s first-quarter 2026 sales, down from roughly 33% for full-year 2025 and 36% in the fourth quarter of 2025. That decline was already under way, reflecting the digestion of a large pull-forward of Chinese orders placed ahead of tightening export restrictions. Domestic Chinese alternatives, however limited, accelerate the trend. Every mature-node fab in China that can be equipped with a domestic scanner is a fab that will not buy an ASML tool, and mature-node systems have historically been a meaningful profit contributor even if they are not where the technological glamour lies.

A more skeptical reading of the market’s reaction is also available, and several analysts made it. The reported production volumes are small, the technology is a generation behind, the components are not fully domestic, and the reporting is single-sourced. Independent confirmation from SMIC, Huawei, SiCarrier or Yuliangsheng has not been provided. On that view, Monday’s move was an emotional repricing of a structurally intact business, and the appropriate comparison is to previous China-breakthrough headlines that produced sharp one-day moves and little lasting damage. The counter-argument is that the same was said about SMIC’s 7-nanometre Kirin processor in 2023, and the industry consensus about what China could not do has moved steadily in one direction since.

CXMT’s Shanghai debut: the second China shock in twenty-four hours

While the lithography story was moving European equipment names, a separate event in Shanghai was doing damage to memory stocks.

ChangXin Memory Technologies, founded in Hefei in 2016 and now China’s largest DRAM producer, listed on the Shanghai Stock Exchange’s STAR Market on Monday, July 27. The shares were offered at 8.66 yuan and opened above 49 yuan, closing up roughly 466% to 470% depending on the measurement. The offering raised about 57.92 billion yuan, or roughly $8.6 billion, rising to as much as 66.61 billion yuan if the over-allotment option is exercised in full. It is the largest exchange listing in mainland China since Agricultural Bank of China’s flotation in 2010, and Asia’s biggest of the year.

The debut left CXMT with a market capitalisation in the region of 3.3 trillion yuan — approximately $460 billion to $490 billion depending on the exchange rate and the point of measurement — making it the most valuable company listed on a mainland Chinese exchange. Company disclosures cited in the prospectus put its global DRAM market share at 7.67% as of late 2025, which would place it fourth behind Samsung Electronics, SK Hynix and Micron Technology. Reported first-quarter 2026 revenue of about 50.8 billion yuan represented an extraordinary year-on-year increase; these figures come from prospectus-based reporting rather than from an independently audited filing available in English, and should be read with that in mind.

Memory stocks outside China fell on the news. SanDisk dropped roughly 11% to 12%, Micron about 5%, and SK Hynix around 8%. Part of that was mechanical — a high-beta group in a risk-off tape ahead of earnings — but the specific concern is straightforward. CXMT has told investors it intends to use most of the proceeds to expand production capacity and fund research. New DRAM supply, funded by an $8.6 billion equity raise and backed by the Chinese state, is the single most direct threat to the pricing environment that has made 2026 an extraordinarily profitable year for the memory incumbents.

There is an additional wrinkle. Reports that Apple has been testing CXMT’s DRAM chips, if borne out, would mark the point at which Chinese memory moves from domestic substitution into the global merchant market. Apple has not confirmed any such qualification, and readers should treat it as unverified.

The memory paradox: record prices, falling memory shares

It is worth pausing on how strange the memory sector’s position has become, because it explains why the same news can be read as bullish or bearish depending on the time horizon applied.

By any operating measure, 2026 has been a spectacular year for DRAM and NAND producers. Global memory sales reached a record $74.6 billion in July, up roughly 31.7% month on month, with NAND sales jumping about 40.7% to a record $25.8 billion. UBS has raised its DRAM price forecast to increases of 32% and 18% over the following two quarters, and expects NAND prices to climb around 30%. Samsung’s first-quarter operating profit rose more than sevenfold year on year. SK Hynix reported quarterly revenue above 50 trillion Korean won for the first time, with an operating margin reported as high as 72%.

The cause is structural rather than cyclical. Data centres now consume an estimated 70% of global memory output. Samsung, SK Hynix and Micron, which together control more than 95% of DRAM production, have reallocated capacity toward high-bandwidth memory for AI accelerators, leaving conventional DRAM and NAND in acute shortage. SK Hynix’s chief executive warned in July 2026 that the imbalance would probably persist well beyond 2030.

Yet memory shares have been falling. SanDisk lost roughly 14% in a single session earlier in July. Micron, SanDisk and Western Digital all dropped around 6% on July 13 when SK Hynix’s outlook disappointed. This is the classic behaviour of a cyclical sector at what the market suspects is peak margin: investors sell into record earnings because they are pricing the next turn, not the current one. CXMT’s arrival with fresh capital is exactly the kind of event that turns a suspicion into a thesis.

None of this makes the sellers right. Supply additions in memory take years to arrive, and CXMT’s capacity expansion will not meaningfully alter 2026 or 2027 pricing. The debate is about 2028 and beyond, and about whether the AI-driven demand base is durable enough to absorb whatever arrives. What is observable today is that the market has begun to answer that question pessimistically.

Nvidia’s $750 billion of deals and the circular-financing argument

The third strand of Tuesday’s selloff has nothing to do with China. It concerns the structure of the AI boom’s financing, and specifically Nvidia’s role in it.

Over the weekend and into Monday, two separate arrangements came into public view. Late on Friday, Nvidia announced a partnership with South Korea’s SK Group under which the two companies expect to do more than $500 billion in business with each other, tied to the construction of more than two gigawatts of AI data-centre capacity in South Korea. Then reporting first published by The Wall Street Journal and confirmed by other outlets described talks in which Nvidia would provide a backstop of as much as $250 billion supporting OpenAI’s lease of a proposed 10-gigawatt data-centre campus in southern Ohio, developed by SB Energy, a SoftBank subsidiary. The guarantee, as described, would cover the lease obligations and construction debt rather than the Nvidia hardware inside the building; financing for the chips themselves is reportedly a separate discussion involving a further sum in the region of $350 billion.

Two points of precision matter here. The SK Group arrangement is announced. The OpenAI backstop is reported to be under negotiation, at an early stage, and could change materially or collapse entirely. Neither Nvidia nor OpenAI has published binding terms. Bloomberg’s aggregate figure of “more than $750 billion” combines announced and reported items and should be understood as a characterisation of a deal pipeline, not a booked commitment.

The market’s objection is about the shape of these transactions rather than their size. In a conventional supply relationship, a customer earns money elsewhere and spends it with a vendor. In the arrangements now proliferating across the AI sector, the vendor supplies capital — as equity, as a guarantee, as a purchase commitment, or as an offtake agreement — that helps the customer acquire the vendor’s product. Revenue recognised by the vendor is therefore partly funded by the vendor’s own balance sheet. As long as the end demand is real and the customer eventually becomes self-financing, this is aggressive but defensible growth capital. If end demand disappoints, the vendor has simultaneously overstated its revenue quality and taken on the customer’s credit risk.

Axios reported that the OpenAI guarantee discussions reignited exactly this concern, and Michael Burry — a persistent critic of AI accounting — responded publicly with a characteristically terse reference to circularity. Nvidia’s own credit market has reportedly taken notice, with spreads on its debt widening as the scale of contingent commitments became clearer.

Nvidia’s defence, articulated repeatedly by Jensen Huang, is that these are not disguised sales but investments in building out an industry whose bottleneck is power and shell capacity rather than demand for accelerators. On that account, guaranteeing a lease on a data centre is closer to a utility underwriting a transmission line than to a chip vendor manufacturing its own order book. The company has also pointed out that its accounts receivable and revenue recognition are subject to standard audit and that no deferred or vendor-financed revenue has been recognised improperly.

The honest summary is that both descriptions can be true simultaneously. The physical build-out is real; the electricity contracts are real; the shortage of grid interconnection is real. And the financing structure genuinely does transfer risk from the customer’s balance sheet to Nvidia’s in a way that was not present in the 2023–2024 phase of the boom. What changed in July 2026 is that investors stopped granting the benefit of the doubt.

Alphabet’s negative cash flow and the moment the capex bargain broke

If one wants a single date on which the market’s attitude to AI spending changed, July 22, 2026 is a strong candidate.

That afternoon Alphabet reported second-quarter results that were, on almost every operating measure, excellent. Consolidated revenue rose 24% year on year to $119.8 billion, or 23% in constant currency, the twelfth consecutive quarter of double-digit growth. Google Services revenue increased 15% to $94.5 billion, with Search and other up 17% and YouTube advertising up 13%. Google Cloud revenue rose 82% to $24.8 billion. Operating income increased 30% and the operating margin expanded by two percentage points to 34%.

And the shares fell more than 7% the following session — their worst day in over a year.

The reason was in the cash flow statement rather than the income statement. Purchases of property and equipment roughly doubled to $44.9 billion in the quarter. Free cash flow came in at negative $5.85 billion. Alphabet has been a public company since August 2004 and had never before reported a negative free-cash-flow quarter. Management simultaneously raised full-year capital expenditure guidance to a range of $195 billion to $205 billion, up from $180 billion to $190 billion three months earlier — the second upward revision of the year.

Investors did the arithmetic. A company generating 34% operating margins on $119.8 billion of quarterly revenue was nonetheless consuming cash, because the infrastructure required to serve AI demand was being built faster than that demand converted into billings. Depreciation on those assets will land on the income statement over the following several years regardless of whether the revenue arrives on schedule.

The distinction matters and is frequently blurred in coverage. Capital expenditure does not reduce reported earnings in the period it is incurred; it appears on the balance sheet and is expensed through depreciation over the asset’s useful life. What it does immediately is consume cash. A company can therefore post record profits and negative free cash flow in the same quarter, which is precisely what Alphabet did. For a business whose equity valuation has historically rested on prodigious and reliable cash generation, that is a meaningful change in character.

What Microsoft, Meta, Apple and Amazon have to prove this week

Alphabet’s report set the terms for the reports that follow. Microsoft and Meta Platforms are scheduled to report after the close on Wednesday, July 29 — the same afternoon the Fed announces. Apple and Amazon follow on Thursday, July 30.

The scale of the collective commitment is worth stating plainly. Based on company guidance and analyst compilations, Microsoft has pointed toward roughly $190 billion of calendar-2026 capital spending, Alphabet to $195–205 billion, Amazon to approximately $200 billion, and Meta to $125–145 billion. Bloomberg’s compilation of analyst estimates puts combined 2026 capital spending for Alphabet, Microsoft, Amazon and Meta at about $724 billion, rising toward roughly $950 billion in 2027.

Apple’s position in that table is the most interesting. It has declined to build hyperscale AI infrastructure, opting instead to partner with external model developers to power its services. For most of 2024 and 2025 that looked like strategic timidity and the shares were treated accordingly. In July 2026 it has looked like discipline: Apple shares rose roughly 15% during the month while Meta fell about 9.8% and Amazon was close to flat for the year. A market that spends two years rewarding spending and then abruptly rewards its absence has not necessarily become wiser; it has changed the variable it is grading on.

What would satisfy investors this week is reasonably specific. For Microsoft: an Azure growth rate that has not decelerated, evidence that AI capacity is revenue-constrained rather than demand-constrained, and a credible bridge from Copilot seat counts to recognised revenue. For Meta: advertising growth sufficient to absorb a rising depreciation and headcount base without margin compression. For Amazon: AWS operating margin holding up while capital spending accelerates. For Apple: continued services growth and no sudden announcement of a $100 billion data-centre programme.

What would not satisfy them is another quarter in which revenue beats consensus and capital guidance rises again. Alphabet demonstrated that this combination is now punished rather than rewarded.

Where the money went: Unilever’s best quarter in more than a decade

Rotation requires a destination. On Tuesday it was consumer staples, and the single clearest illustration was Unilever.

The company reported first-half results on July 28 showing second-quarter underlying sales growth of 5.8%, of which underlying volume growth accounted for 5.5% and price for the remainder. That is the strongest quarterly volume performance the company has recorded in more than a decade, and a marked acceleration from 3.8% underlying sales growth in the first quarter. For the half, underlying sales growth was 4.8%, split 4.2% volume and 0.6% price. Management raised its full-year expectation, now guiding to growth within its multi-year 4% to 6% range rather than at the bottom of it.

The shares rose about 6% in London, the largest single-day gain in two years, taking the stock to its highest level since March.

The composition of that growth is what matters, and it is worth explaining why. For most of 2022 through 2024, consumer goods companies grew revenue almost entirely through price increases, passing through input-cost inflation while volumes stagnated or fell. That is a low-quality form of growth: it depends on pricing power that erodes as consumers trade down, and it cannot compound. Growth driven by volume, more units sold, indicates genuine demand and is far more durable. Unilever delivering 5.5 points of volume and 0.3 points of price is an almost complete inversion of the pattern that characterised the sector two years ago.

Whether that is company-specific execution or a broader consumer recovery is not yet clear from one result. Unilever has been through a substantial portfolio restructuring, including the separation of its ice cream business, and management has attributed the improvement to what it describes internally as a “desire at scale” strategy of concentrating investment behind fewer, larger brands. A single quarter of exceptional volume growth from a company midway through a restructuring should not be extrapolated without corroboration from peers.

What is not in doubt is the market’s reaction function. In a session where investors were selling companies that promise returns in 2029, a company delivering measurable volume growth today was rewarded disproportionately.

Mercedes-Benz: a large beat, a smaller business, and China still shrinking

The second European gainer of note was Mercedes-Benz Group, whose shares rose more than 4% after second-quarter results that beat expectations across every major division while confirming full-year earnings guidance.

Group revenue was €32.1 billion, down 3% year on year. Adjusted group EBIT was €2.30 billion, against a Jefferies estimate of €1.47 billion and a broader consensus around €1.65 billion — a beat of roughly 39% relative to consensus. Earnings per share came in at €1.14 against €0.95 in the comparable quarter, with net profit attributable to shareholders of €1.065 billion versus €915 million.

Two things should be held in mind before reading that as a straightforward recovery.

First, revenue declined. The beat came from margin and mix, from strong performance in Vans and Mercedes-Benz Mobility financial services, and from cost control, rather than from selling more cars at better prices. That is a legitimate way to earn money and it is what a well-run manufacturer is supposed to do in a soft market, but it is not the same as demand recovering.

Second, the company cut its unit sales outlook. Mercedes-Benz Cars is now expected to deliver unit sales slightly below the previous year’s level, which the company indicated could translate into a full-year decline in the range of 2% to 7.5%. Management pointed specifically to “the negative development of the Chinese market.” Chinese demand for German luxury vehicles has been eroding for several years as domestic electric-vehicle manufacturers have taken share at both the premium and mass-market ends, and there is no sign in this release that the trend has stabilised.

The share price context is also relevant. The stock has been trading near multi-year lows; a 4% gain from a depressed base after a 39% earnings beat is a modest response by historical standards, which suggests investors are treating the quarter as a well-managed decline rather than an inflection.

Philips: how a beat and a guidance raise produced a double-digit fall

The most instructive European result of the day was Philips, because it demonstrates how differently the market weighs backward-looking and forward-looking data.

The company’s second-quarter release showed comparable sales growth of 4% to €4.4 billion, ahead of the roughly 3.8% analysts expected. The adjusted EBITA margin reached 16.4%, far above a consensus near 12.1%. Management raised full-year adjusted EBITA margin guidance to 13.5%–14.0% from 12.5%–13.0%, confirmed comparable sales growth guidance of 3.0%–4.5%, and lifted the free cash flow forecast to €1.5–1.7 billion from €1.3–1.5 billion.

The shares fell roughly 9% to 11% in Amsterdam.

Two details explain the disconnect. The first is the quality of the margin. Of that 16.4% adjusted EBITA margin, approximately 4.2 percentage points came from a United States tariff refund — a one-off cash recovery rather than an improvement in the underlying business. Strip it out and the margin was around 12.2%, roughly in line with what analysts had modelled. The raised full-year guidance similarly includes about one percentage point of tariff-refund benefit. Put differently, the entire 4.3-percentage-point margin beat is accounted for by a 4.2-percentage-point one-off. That is not an earnings beat in any sense that affects valuation.

The second, and more important, detail is orders. Comparable order intake fell 1%, with low-single-digit growth in Diagnosis and Treatment offset by mid-single-digit declines in Connected Care, which management attributed to order phasing. For a capital-equipment business selling MRI scanners, ultrasound systems and hospital monitoring infrastructure, the order book is the forward indicator. Revenue in any given quarter reflects decisions hospitals made twelve to eighteen months earlier. Orders reflect what they are deciding now. A declining order book alongside a margin flattered by a tax refund is a combination that invites exactly the reaction the shares received.

The episode is a useful corrective to a habit of financial coverage: reporting whether a company “beat” or “missed” as though the two words carry information. They frequently do not. What carries information is which line beat, why, and whether the cause repeats.

Saipem, energy, and the cost of operating in a war zone

At the other end of Tuesday’s European movers list sat Saipem, the Italian energy engineering and drilling contractor, down more than 7%.

First-half results showed revenue up 1.9% to €7.35 billion and adjusted EBITDA up 9.4% to €836 million. Below that, the picture deteriorated: EBIT fell 21% to €240 million, adjusted net profit fell 6.4% to €131 million from €140 million, and reported net profit was €96 million after roughly €35 million of non-recurring charges linked to provisions for a voluntary early-exit programme agreed with trade unions in the fourth quarter of 2025. The company also disclosed approximately €70 million of additional costs incurred to manage logistical and operational difficulties in the Middle East and to ensure staff safety.

That last item is the most revealing line in the release, and it connects the corporate story to the macro one. Saipem’s business is concentrated in offshore and onshore energy construction, a substantial portion of it in the Gulf. The disruption to shipping and to regional operations caused by the 2026 conflict between the United States, Israel and Iran did not merely move the oil price; it imposed direct, quantifiable costs on the companies that build energy infrastructure in the region. Investors also noted a reduction in the company’s EBITDA outlook.

The broader European energy sector was among the day’s weakest performers, which is arithmetically unsurprising given crude’s decline. Falling oil prices help the index by relieving input costs across industrials, transport and consumer sectors; they hurt the energy sector directly. Both effects were visible on Tuesday.

Among the smaller European movers, Fiera Milano and SIG Group both rose more than 6%, while Philips and Saipem led the decliners — a distribution that reinforces rather than complicates the day’s theme.

Why Tokyo fell hardest

Japan’s roughly 4% Nikkei decline was the largest single-market move of the session, and the explanation is largely mechanical.

The Nikkei 225 is a price-weighted index, which means high-priced shares exert disproportionate influence regardless of company size. Several of Japan’s semiconductor-equipment and AI-linked names — Advantest and Tokyo Electron in particular — carry very high share prices and therefore very large index weights. When the global chip complex sells off, the Nikkei mechanically falls further than a capitalisation-weighted Japanese index would. That is precisely what happened: the broader Topix fell substantially less.

Kioxia, the NAND flash producer, was among the heaviest fallers, dropping sharply on the combination of the CXMT listing and general memory weakness. SoftBank Group also declined materially, which is worth noting because SoftBank sits at the intersection of several strands of this story: it owns Arm, it is the parent of SB Energy — the developer of the Ohio campus at the centre of the Nvidia–OpenAI backstop talks — and it holds a large position in OpenAI itself.

The currency channel offered no offset. A weaker yen normally supports Japanese exporters by inflating the domestic-currency value of overseas earnings, and Japanese equities have historically rallied on yen weakness. On this occasion the semiconductor losses overwhelmed it entirely. When a currency tailwind fails to cushion an equity market, it is usually a signal that the selling is being driven by foreign investors reducing sector exposure rather than by domestic investors repricing earnings.

Korean chipmakers fell alongside their Japanese counterparts. The pattern across Asia was consistent: the more a market’s index depended on semiconductors, the worse it did. Hong Kong, whose index is weighted toward financials, property and internet platforms rather than chip hardware, finished slightly higher.

Oil, Iran, and the unwinding of a war premium

The commodity story running underneath Tuesday’s equity rotation is arguably more consequential for the global economy than the chip story, and it has been developing since February.

The 2026 conflict began on February 28, when the United States and Israel conducted airstrikes on Iranian military targets. Iran responded by closing the Strait of Hormuz to foreign shipping — the chokepoint through which roughly a fifth of global seaborne oil normally transits. Traffic was severely disrupted, fuel shortages appeared in parts of Asia, and the effects rippled through global freight and insurance markets. On March 19 the United States began an aerial campaign specifically aimed at reopening the strait.

Brent crude, which had been trading around $70 a barrel before the conflict, moved well above $100. United States retail gasoline prices rose roughly 45% between late February and early May.

A two-week ceasefire mediated by Pakistan was agreed on April 8 and extended indefinitely later that month. On June 17 the two governments signed a fourteen-point memorandum of understanding after two months of stalled negotiation. Just over a week later, a drone strike against a vessel in the Strait of Hormuz, interpreted by Washington as a violation, prompted a further round of contained American strikes. Hostilities resumed for a period before the United States paused its strikes again on Friday, July 24. Tehran suspended its retaliatory operations in response and entered discussions with Oman over navigation through the strait.

That pause held into a second day over the weekend, and it is the proximate reason oil fell on Monday and again on Tuesday. President Trump characterised talks with Iran as going well and said a deal was possible, while warning that strikes would resume if negotiations failed.

Price readings on Tuesday varied by contract and by the hour, as they will in a market repricing a geopolitical premium. West Texas Intermediate for September delivery traded in the low $80s, down roughly 1% to 3% depending on the point of measurement during the European session; Brent traded in the mid-$80s. A supply-side development reinforced the move: crude loadings resumed at the Caspian Pipeline Consortium terminal on Russia’s Black Sea coast, a key export route for Kazakh barrels that had been disrupted by Ukrainian drone attacks.

The important point for investors is that this is a pause, not a settlement. The memorandum signed in June broke down within a fortnight. Nothing in the current arrangement is structurally different from the arrangement that failed in late June, and the President has stated explicitly that strikes would resume if talks collapse. Positioning that assumes a durable peace is positioning against a demonstrated pattern.

Gold’s retreat and what it says about the inflation trade

Gold has been one of the defining assets of 2026, and its behaviour on Tuesday deserves attention because it appears, superficially, to contradict itself.

Bullion traded below $4,100 an ounce on July 28, down more than 1% on the session, with readings in the $4,030 to $4,085 range depending on whether one looks at spot or the August futures contract and at which hour. That is a substantial decline from levels reached at the height of the conflict.

The standard explanation — receding safe-haven demand as the United States and Iran extended their pause — is correct as far as it goes. Gold accumulated a considerable war premium between February and June, and that premium is being released.

But there is a second mechanism operating in the opposite direction to the usual narrative. Gold is conventionally described as an inflation hedge, and the oil shock of the spring pushed inflation sharply higher. Yet gold fell during much of that period rather than rising. The reason is that the inflation caused by an energy shock raised expectations of monetary tightening, and higher expected policy rates increase the opportunity cost of holding a non-yielding asset. Gold competes with real yields, not with the consumer price index. When inflation rises for a reason that makes a central bank more hawkish, the rate channel typically dominates the hedging channel.

That dynamic is precisely what is in play ahead of Wednesday’s Fed decision. A committee that is debating whether to raise rates rather than cut them is an unhelpful environment for gold, regardless of where headline inflation prints. OCBC Bank has forecast gold declining through the end of 2026 on the combination of rising Treasury yields, a firmer dollar and weaker investor demand for precious metals — a view that is far from consensus but that follows logically from the rate framework.

Silver traded around $57.68 an ounce on the same session, also lower.

The Federal Reserve: Kevin Warsh’s second meeting, and the first that matters

The Federal Open Market Committee convened on Tuesday, July 28 and will announce its decision at 2:00 p.m. Eastern on Wednesday, July 29. There is no Summary of Economic Projections scheduled for this meeting, which removes one of the two main sources of information the market normally extracts from a Fed day and concentrates attention on the statement language and the press conference.

The target range for the federal funds rate stands at 3.50% to 3.75%, unchanged since before the June meeting.

This is Kevin Warsh’s second meeting as Chair. He was nominated by President Trump on January 30, 2026, and succeeded Jerome Powell, whose term as Chair concluded in May. Warsh served as a Fed governor from 2006 to 2011 and has spent the intervening years as one of the more prominent critics of the institution’s post-crisis balance-sheet expansion.

His first meeting, on June 17, produced no change in rates but a decisive shift in signalling. The minutes of that meeting record that participants generally observed inflation had increased further and remained well above the Committee’s 2% longer-run objective. The updated projections showed a median year-end 2026 federal funds rate of 3.8%, up from 3.4% in the March round, and nine of eighteen participants pencilled in at least one increase during the remainder of the year. For a committee that entered 2026 with markets expecting cuts, that was a reversal of some consequence.

Warsh has been unusually direct in public. In remarks on July 14 he described a policy “regime change” intended to eliminate what he characterised as an inflation tax on American households. During his confirmation process he called inflation “a choice.” He has also argued that the artificial-intelligence investment boom is disinflationary over time through productivity gains, and has advocated more rapid reduction of the Fed’s balance sheet than his predecessors pursued — a combination he has described as pairing productive dovishness on the rate path with aggressive quantitative tightening.

That framework has an interesting implication for the current market. If the Chair genuinely believes AI capital investment is a disinflationary force, then a market selloff in AI infrastructure is not straightforwardly good news for the inflation outlook, whatever it does for financial conditions.

What the market expects

Market-implied probabilities on Tuesday pointed to a hold as the overwhelmingly likely outcome, with prediction-market and futures-derived estimates of no change clustering around 70% to 80%, implying something between a one-in-five and a one-in-three chance of an increase. Those numbers should be read as a range rather than a point estimate: different instruments imply different probabilities, and the gap between them is itself a measure of genuine uncertainty.

An increase at this meeting would be a significant surprise but not an absurd one, which is a different situation from the one that prevailed at most meetings of the past two years. Warsh indicated in late July that the meeting would involve substantial debate among members about the appropriate path.

Why the rate path matters more to AI stocks than to Unilever

The interaction between Wednesday’s decision and the sector rotation described in this article is worth making explicit, because it is not merely coincidental timing.

The value of any equity is the present value of the cash it is expected to generate. When most of that cash is expected in the near term — as with a consumer goods company selling soap next quarter — the discount rate applied has a modest effect on the answer. When most of it is expected far in the future — as with an infrastructure buildout that consumes cash today and is projected to generate returns in the 2030s — the discount rate dominates. A one-percentage-point change in the assumed long-run rate moves the valuation of a long-duration asset far more than it moves the valuation of a short-duration one.

This is the arithmetic underneath the rotation. Investors selling AI infrastructure and buying consumer staples are, whether or not they describe it in these terms, shortening the duration of their equity portfolios in an environment where the expected policy rate has moved up rather than down. A Fed that raises rates would reinforce that trade mechanically. A Fed that signals patience would relieve it.

It also explains why a market can be simultaneously convinced that AI demand is real and unwilling to pay yesterday’s price for it. Both propositions can hold. The cash flows may arrive exactly as projected and still be worth less today than they were worth when the market expected policy rates to be falling by now.

The inflation picture the Fed is looking at

The June consumer price index, published on July 14, was the most encouraging inflation report of the year and also one of the easiest to misread.

Headline CPI fell to 3.5% year over year from 4.2% in May, the first decline in five months and below a consensus forecast near 3.8%. On the month, the index fell 0.4% against expectations of a 0.1% decline — the largest monthly drop since April 2020.

Almost all of that came from energy. Energy costs were still up 15.7% year over year, but that was down sharply from 23.5% in May. Gasoline was up 26.7% against 40.5% previously; fuel oil up 42.9% against 58.9%. Shelter inflation eased slightly to 3.3% from 3.4% and food to 3.0% from 3.1%.

Core CPI, which excludes food and energy, was flat on the month, putting the twelve-month core rate at 2.6%.

Three distinctions are worth drawing explicitly because they are routinely collapsed in coverage. First, a fall in the inflation rate is not a fall in prices. Headline CPI declining from 4.2% to 3.5% means prices are still rising, merely more slowly. The monthly figure of −0.4% does represent an actual decline in the index level, driven by energy, but that is a single month and largely a base effect from the retreat in crude. Second, the divergence between headline at 3.5% and core at 2.6% is the entire story: strip out the war-driven energy shock and underlying inflation is close to target. Third, the mechanism running in reverse is powerful. Just as rising energy prices pushed headline inflation from roughly target to above 4% between February and May, falling energy prices will pull it back down over the coming months as the year-earlier comparisons roll off.

This is why a Fed that raises rates in July would be doing something genuinely difficult to justify on the data alone. It would be tightening into decelerating headline inflation and an at-target core reading. The case for doing so rests on a different argument: that inflation expectations have drifted, that the labour market remains tight, and that a central bank which has spent five years above target cannot afford to be seen accommodating another shock. Whether one finds that persuasive depends on how much weight one places on credibility relative to data.

The dollar, the euro, and the rate differential nobody expected

Currency markets on Tuesday were quiet in a way that itself carried information. The dollar index sat at approximately 101.55, up a fraction of a percent and holding near a one-month high. The euro traded at about $1.1366, essentially unchanged.

The dollar’s firmness through a week in which oil fell sharply is the notable feature. A declining oil price is normally mildly negative for the dollar because it reduces United States inflation and therefore the expected policy rate. That transmission has been blocked because the market is simultaneously pricing a meaningful probability that the Fed raises rates — the higher-for-longer premium is offsetting the disinflationary impulse from energy.

On the other side of the pair, European Central Bank officials have maintained a hawkish tone, and euro-area composite survey data have been reasonably solid. That has kept the euro from weakening despite the interest-rate differential. The net result is a currency pair that has gone almost nowhere while the assumptions underneath it have shifted considerably.

For dollar-based investors holding European equities, the flat exchange rate means Tuesday’s gains in Unilever and Mercedes-Benz translate roughly one-for-one into dollar returns. For companies, a stable euro-dollar rate through a period of substantial commodity volatility is unusually helpful; much of the currency-translation noise that distorted 2022 and 2023 comparisons has faded.

Crypto’s quiet bear market

Digital assets barely registered in Tuesday’s coverage, which is itself a change from previous years. The market is in the ninth month of a drawdown that has been long, shallow and largely ignored.

Bitcoin traded around $63,100 on July 28, down roughly 2.9% over twenty-four hours. Ether was near $1,872, down about 3.6%; Solana around $73, down roughly 4.1%; XRP near $1.05, down about 4.4%. Total crypto market capitalisation was in the region of $2.2 trillion. Readers should note that these are point-in-time readings from a market that trades continuously, and that percentage moves quoted against a euro base, as some European data providers display them, will differ from dollar-denominated moves by the amount of the currency’s daily change.

The context is that Bitcoin peaked at $126,296 on October 6, 2025 and has spent most of 2026 well below it, falling under $70,000 in early June and briefly beneath $60,000 during the month. As of late June, it had traded below its 200-day moving average for 233 consecutive sessions, the fourth-longest such stretch on record — though the depth of the drawdown has been shallower than in previous cycles, which is why some analysts describe 2026 as the mildest bear market Bitcoin has experienced.

Three explanations recur in the analysis, and they are not mutually exclusive. The first is macro: crypto has been trading with roughly 84% correlation to the Dow, which suggests a broad risk repricing rather than a crypto-specific failure. Higher-for-longer rate expectations are straightforwardly negative for a non-yielding asset with no cash flows. The second is flows: United States spot Bitcoin exchange-traded funds have recorded a record streak of outflows, and leveraged positions have been liquidated in waves. The third, and the most interesting, is competition for the same capital. Since April, memory-chip exchange-traded funds have attracted roughly $12.7 billion while Bitcoin ETFs have seen more than $2 billion of outflows. The speculative dollar that once went into digital assets has been going into AI hardware instead.

If that third explanation is correct, it has an uncomfortable implication for both trades. A rotation out of AI hardware does not automatically rotate back into crypto — it rotated into Unilever this week — but the two assets have been drawing on a common pool of risk appetite, and that pool is currently shrinking. Traders positioning ahead of the FOMC decision have been treating the Fed as the near-term determinant of direction, which is consistent with a macro-driven rather than idiosyncratic market.

Nothing in this section should be read as a view on the merits of any digital asset. It is a description of price behaviour and observable flows.

The case that the AI trade is fine

Long-form market coverage has a tendency to describe whichever direction prices moved as vindicated. It is worth setting out the strongest version of both arguments, because the evidence genuinely supports either depending on which assumptions are given priority.

The constructive case has four planks.

First, the operating results are not deteriorating. Alphabet grew revenue 24% and expanded operating margin by two percentage points. Google Cloud grew 82%. That is not the profile of a company whose customers are pulling back. If AI demand were softening, cloud revenue growth would be the first place it appeared, and it is accelerating rather than decelerating.

Second, the physical constraint is real. The bottleneck in AI capacity is not customer appetite but electricity, grid interconnection, transformers, cooling and construction labour. Data centres now consume an estimated 70% of global memory output and memory remains in acute shortage with SK Hynix’s chief executive publicly warning the imbalance may persist past 2030. Shortages of that kind are not consistent with a demand illusion.

Third, the China lithography story is being over-extrapolated. Five machines in 2026 and twenty in 2027, at a 28-nanometre design point, built with imported Japanese components, does not displace ASML in any commercially relevant timeframe. EUV remains a genuine monopoly and the technical gap to it is measured in decades of accumulated process knowledge rather than in engineering budget.

Fourth, the negative free cash flow at Alphabet is a timing artefact of a company choosing to build ahead of demand. Capital expenditure is not a loss. If the assets are utilised, the depreciation is matched by revenue and the cash flow recovers. Amazon spent years running negative free cash flow while building AWS, and that decision produced one of the most profitable businesses in corporate history.

The case that the AI trade is in trouble

The skeptical case is equally coherent and rests on different evidence.

First, the financing structure has changed in a way that degrades revenue quality. Vendor financing — a supplier providing the capital with which a customer buys the supplier’s product — has a long and unflattering history in technology, most memorably in the telecommunications equipment sector around 2000. The specific concern is not that Nvidia’s revenue is fictitious but that some portion of it is being pulled forward by capital that Nvidia itself supplied, and that the credit risk of the entire ecosystem is being concentrated on one balance sheet. Reported widening in Nvidia’s own credit spreads suggests bond investors share that concern.

Second, the aggregate capital numbers have reached a scale that requires external financing rather than internal cash generation. Combined 2026 capital spending across Alphabet, Microsoft, Amazon and Meta at roughly $724 billion, rising toward $950 billion in 2027 on analyst estimates, exceeds what these companies can fund from operations while maintaining buybacks and dividends. That means debt issuance, and debt issuance means the bond market gets a vote. Bond markets have begun exercising it.

Third, the depreciation problem is real and arriving. Assets purchased in 2025 and 2026 begin depreciating immediately. If the useful-life assumptions applied to AI accelerators prove optimistic — and hardware generations in this field have been turning over in roughly two years rather than the five or six commonly assumed — then reported earnings across the hyperscalers are currently overstated. This is a genuine accounting judgement, not an accusation; the assumptions are disclosed and auditable. But small changes to them produce large changes in reported profit.

Fourth, competitive erosion is happening faster than the consensus assumed on multiple fronts simultaneously. China building its own lithography tools, CXMT raising $8.6 billion to expand DRAM capacity, custom silicon programmes at every major hyperscaler — each individually is manageable. Occurring together, they compress the window during which the incumbents’ margins are defensible.

Fifth, the market’s own behaviour is evidence. The Philadelphia Semiconductor Index reached record highs in late June and has since fallen more than 20% from that peak. A decline of that magnitude in the sector at the centre of the investment cycle, occurring while the underlying companies are reporting record results, is the market telling you it has changed its discount rate rather than its earnings forecast.

What would settle the argument

Several observable developments over the next two quarters would materially favour one interpretation over the other.

• Cloud revenue growth rates. If Azure and AWS growth hold or accelerate in the reports due this week and next quarter, the demand-illusion thesis weakens considerably. Deceleration would strengthen it.
• Free cash flow trajectory at Alphabet. A return to positive free cash flow in the third or fourth quarter would confirm the timing-artefact reading. A second and third negative quarter would not.
• Whether the Nvidia–OpenAI backstop is signed, and on what terms. A structure in which third-party lenders take the majority of the risk would be reassuring; one in which Nvidia’s guarantee is the primary credit support would not.
• Useful-life disclosures. Any hyperscaler shortening the depreciation schedule on AI hardware would be a significant signal, and would mechanically reduce reported earnings.
• Actual delivery of Chinese DUV tools. The reported target is five units in 2026. Whether they ship, to whom, and whether SMIC moves them into production in 2027 is verifiable within eighteen months.
• Memory contract pricing into 2027. If DRAM and NAND contract prices continue rising through CXMT’s capacity ramp, the supply-glut thesis fails.

Until then, both narratives will be argued from the same set of facts.

The export-control architecture, and why it produced this outcome

Understanding why a report about a single lithography startup in Shanghai can remove tens of billions of dollars of market value from Dutch, Japanese and American companies requires some familiarity with the policy structure that has governed semiconductor equipment trade since 2019.

The controls have three layers. The first restricts the sale of the most advanced manufacturing equipment to Chinese customers — initially EUV systems, later extended to the most capable immersion DUV models and to a widening list of deposition, etch and metrology tools. The second restricts the sale of advanced logic and memory chips themselves, most visibly the highest-performance AI accelerators. The third, and most consequential in practice, restricts the provision of servicing, spare parts and engineering support to Chinese fabs operating equipment they already own.

The Netherlands, Japan and the United States coordinated these measures, which is why ASML, Tokyo Electron, Applied Materials, Lam Research and KLA all carry similar exposure profiles. The stated objective was to slow Chinese progress toward advanced-node manufacturing capability with military applications.

The measurable effects have been substantial and, from the perspective of the controls’ designers, mixed. Chinese purchases of permitted equipment surged as customers stockpiled ahead of anticipated restrictions, which is why ASML’s China revenue share reached roughly 36% in the fourth quarter of 2025 and about 33% for the full year, before falling to approximately 19% in the first quarter of 2026 as that pull-forward unwound. Chinese output at mature nodes expanded rapidly. And Chinese state investment in domestic equipment development accelerated sharply, funded through vehicles including the National Integrated Circuit Industry Investment Fund.

The Yuliangsheng scanner, if the reporting is accurate, is a direct product of that investment. Whether the controls succeeded therefore depends entirely on the counterfactual and the time horizon chosen. They demonstrably delayed Chinese access to advanced lithography. They also converted a commercial dependency into a national priority with effectively unlimited funding, and eliminated the commercial incentive that had previously made domestic development economically unattractive when ASML tools were freely purchasable.

This is not a novel dynamic. Export restrictions on jet engines, on satellite components, and on high-performance computing have each produced a similar pattern: a period of genuine constraint followed by indigenous substitution at somewhat lower capability and considerably higher cost. What is unusual in semiconductors is the extreme concentration of the choke point. There is precisely one supplier of EUV systems on earth, and its capability rests on a supply chain — Zeiss optics, Trumpf lasers, Cymer light sources — that is itself concentrated to an almost unreasonable degree.

Investors evaluating ASML are therefore evaluating a geopolitical position as much as a business. That is a materially different exercise from evaluating a capital-goods company, and it is why the shares respond so sharply to news that has minimal near-term revenue implication.

Who else is exposed, and how

The chip complex is often traded as a single instrument, but the companies inside it have quite different sensitivities to the specific developments of the past week. Distinguishing between them is useful.

Lithography and equipment. ASML is the most directly exposed to the Chinese DUV story, both in near-term mature-node revenue and in terminal-value perception. Japanese peers Tokyo Electron and Advantest are exposed to the same theme through different product categories — coaters and track systems in the first case, test equipment in the second — and were among the heaviest fallers in Tokyo. Nikon and Canon, which produce dry DUV lithography systems, occupy an unusual position: they are competitors to ASML in older technology and would be the most direct commercial casualties of a credible Chinese entrant at that node.

Memory. Samsung Electronics, SK Hynix, Micron Technology, SanDisk and Kioxia are exposed to CXMT’s capacity expansion. The exposure is not symmetric: CXMT competes in DRAM, so Micron, Samsung and SK Hynix face the more direct threat, while SanDisk and Kioxia, which are NAND-focused, are exposed principally through sentiment and through the possibility that CXMT’s proceeds fund NAND expansion later. All five fell on Monday.

Logic and accelerators. Nvidia and Advanced Micro Devices are largely insulated from the Chinese lithography development — nothing in a 28-nanometre-class DUV scanner threatens their product positions — but are fully exposed to the financing debate and to any moderation in hyperscaler capital budgets. Their weakness on Monday was driven by the circular-financing story rather than the China story, which is why they fell alongside but for different reasons.

Foundry. Taiwan Semiconductor Manufacturing Company occupies a strange position. It is simultaneously the largest customer for ASML’s EUV systems, the manufacturer of Nvidia’s accelerators, and the company with the most to lose from any credible Chinese advanced-node capability. It is also, for the same reasons, the most exposed to escalation in the Taiwan Strait, a risk that does not appear in any of this week’s headlines but sits underneath all of them.

Beneficiaries. SMIC, Hua Hong and CXMT are the direct commercial beneficiaries of domestic tool availability, and Chinese semiconductor equities have generally outperformed their Western counterparts during this episode. Whether they can convert equipment access into competitive yields and costs is a separate and unresolved question.

Historical comparisons, and their limits

Two historical analogies have been circulating in commentary this week. Both are informative and both are imperfect.

The first is the telecommunications equipment bubble that peaked in 2000. Lucent Technologies, Nortel Networks and others extended substantial vendor financing to competitive local exchange carriers, which used the money to buy equipment from those same vendors. Revenue grew impressively until the carriers could not raise follow-on capital, at which point the receivables proved uncollectible and the revenue proved to have been an advance against losses. Lucent’s vendor financing exposure reached approximately $8 billion; the subsequent write-downs contributed to one of the largest destructions of shareholder value in American corporate history.

The parallel to Nvidia’s arrangements is obvious and is being drawn explicitly by critics. The differences are also substantial. The telecommunications carriers of 1999 were largely start-ups with no revenue building networks for demand that did not yet exist. OpenAI, Microsoft, Meta and Amazon have enormous existing revenue and, in the hyperscalers’ case, some of the strongest balance sheets in the world. The fibre laid in 1999 sat dark for a decade; the accelerators being installed in 2026 are running at capacity. A closer parallel might be to say that the risk is not that the demand is imaginary but that the price being paid for it is too high.

The second analogy is to the succession of Chinese technology breakthroughs that have periodically shocked Western markets — SMIC’s 7-nanometre Kirin processor in 2023, the DeepSeek model releases in early 2025, and now the Yuliangsheng scanner. The pattern in each case has been a violent one-to-three-day repricing followed by partial recovery as the technical limitations became better understood. Investors who bought the dip on the first two occasions did well.

The limitation of that analogy is survivorship reasoning. That a pattern has repeated twice is weak evidence that it will repeat a third time, particularly when the direction of travel — Chinese capability improving, Western export controls tightening in response, Chinese investment accelerating in response to that — has been consistently one-way. The question is not whether the market overreacts to individual headlines. It plainly does. The question is whether the cumulative trend the headlines describe is real, and on that the evidence is less comforting.

How analysts and investors responded

Sell-side reaction to Monday’s ASML decline was, by most accounts, more sanguine than the price action implied. Several analysts and a substantial share of retail investors were reported to have dismissed the selloff as an overreaction, on the grounds that the technology gap remains wide and the volumes trivial. That is a defensible position on the facts as reported.

Reaction to the Nvidia financing story was less forgiving. Michael Burry, whose criticism of AI-sector accounting has been persistent, responded to the OpenAI backstop reports with a public remark on the circularity of the arrangement. CNBC characterised the potential backstop as another strike against the AI trade. The credit-market response, reported widening in Nvidia’s spreads, is more meaningful than equity commentary, because bond investors are paid to think about downside scenarios and have no upside participation to distract them.

On Big Tech capital spending, the shift in analyst framing has been pronounced. Where 2024 and 2025 notes treated rising capital expenditure guidance as evidence of demand, the July 2026 notes have treated it as a cost requiring justification. Wall Street is now, as one preview put it, demanding receipts on a $700 billion spending programme. That is a change in the grading rubric rather than a change in the underlying numbers, and it explains why Alphabet could report an excellent quarter and lose 7% of its value.

Readers should treat all of the above as attributed opinion. Analyst positioning has been a poor predictor of returns throughout this cycle in both directions.

Risks and uncertainties

Several material risks sit around the current situation, and it is worth stating them without dramatisation.

• The Iran pause is fragile. The June memorandum collapsed within two weeks. A resumption of hostilities would reverse the oil decline, re-accelerate headline inflation, and remove the disinflationary impulse the Fed is currently observing. Saipem’s €70 million of additional regional operating costs is a reminder that the effects reach well beyond the crude price.
• A Fed surprise is genuinely possible. With implied probabilities of an increase in the one-in-five to one-in-three range, the market is not fully positioned for a hike. An increase would tighten financial conditions into an already fragile technology tape.
• Depreciation assumptions across the hyperscalers are a live accounting judgement. If useful lives on AI hardware prove shorter than currently assumed, reported earnings across the group are overstated. No evidence of impropriety exists; the risk is that reasonable assumptions turn out to be wrong.
• Concentration risk in the indices is extreme. A handful of AI-linked names constitute an unusually large share of the S&P 500 and of global equity benchmarks. Investors who believe they are diversified by holding an index fund have more semiconductor exposure than they may realise.
• Nvidia’s contingent commitments are not fully disclosed. The scale of guarantees, backstops and purchase commitments across the AI ecosystem is not visible from published financial statements in any consolidated form. That opacity is itself a risk factor.
• Chinese execution is unproven. The bear case on ASML assumes Yuliangsheng delivers. Domestic supplier delays have already slowed production this year and key components remain imported. Execution risk cuts both ways.
• Memory is a cyclical industry at peak margins. SK Hynix operating margins reported as high as 72% are not a steady state. Any industry earning those returns attracts capacity, and CXMT has just raised $8.6 billion to add some.
• Taiwan. None of this week’s news touched the Taiwan Strait, but the concentration of advanced manufacturing there remains the largest single unpriced risk in the sector.

What happens next: the confirmed calendar

The following are scheduled events with fixed dates, distinct from forecasts.

• Wednesday, July 29, 2:00 p.m. ET: FOMC policy statement, followed by Chair Warsh’s press conference. No Summary of Economic Projections at this meeting.
• Wednesday, July 29, after the U.S. close: Microsoft and Meta Platforms report quarterly results.
• Thursday, July 30: Apple and Amazon report quarterly results.
• Wednesday, August 5: SanDisk reports quarterly results, the first read on memory pricing since the CXMT listing.
• Mid-August: July consumer price index release from the Bureau of Labor Statistics, the first full month reflecting the post-ceasefire energy decline.
• Later in 2026: first reported deliveries of Chinese immersion DUV systems to SMIC, Hua Hong and CXMT, if the reported schedule holds.

Beyond the calendar, several things are expectations rather than commitments and should be labelled as such. Analysts expect combined hyperscaler capital spending to rise toward roughly $950 billion in 2027. UBS expects DRAM prices to rise 32% and then 18% over the next two quarters. SK Hynix’s chief executive expects the memory shortage to persist beyond 2030. None of these are guaranteed, and all three would be revised quickly if hyperscaler capital budgets were cut.

The binding constraint is electricity, not silicon

One feature of the current cycle is consistently underweighted in market commentary, and it explains why several of this week’s headline numbers are as large as they are.

The scarce input in artificial-intelligence infrastructure is no longer chips. It is power, and the physical apparatus that delivers it.

Consider the scale implied by the deals under discussion. The Ohio campus that OpenAI is reported to be seeking to lease is described as a 10-gigawatt facility. For context, that is roughly the continuous output of ten large nuclear reactors, and comparable to the average electricity demand of a mid-sized European country. The Nvidia–SK Group arrangement in South Korea covers more than two gigawatts. These are not data centres in the sense that term meant a decade ago; they are industrial energy installations with computing equipment inside them.

The constraints that follow are physical and slow-moving. Grid interconnection queues in the United States now run to multiple years in most regions. Large power transformers have lead times measured in years rather than months. High-voltage cable, switchgear, industrial chillers and the specialised construction labour required to install them are all in shortage. None of these can be resolved by a capital allocation decision made this quarter.

This has three consequences that bear on the investment debate.

First, it substantiates the bull case in a specific and verifiable way. If AI demand were speculative, capacity would be idle. Instead, the reported bottleneck across every hyperscaler is the inability to energise sites fast enough. Memory data points the same way: with data centres consuming an estimated 70% of global memory output and both DRAM and NAND in shortage, the constraint is manifestly on the supply side. It is difficult to construct an account in which demand is illusory and every physical input is simultaneously rationed.

Second, it explains the shape of the financing. If the bottleneck were chip fabrication, Nvidia would have no reason to guarantee lease obligations on buildings. It is guaranteeing them because the buildings and their power contracts are what limit how many accelerators can be deployed, and because the developers of those sites — SB Energy in the Ohio case — require credit support to finance construction at this scale. Seen through that lens, the backstop is less a disguised sale than an attempt to buy forward capacity in the genuinely scarce resource. The criticism does not disappear, but it is a different criticism from the one usually made.

Third, it introduces a risk that is rarely discussed in market coverage: political. Electricity demand of this magnitude has begun to show up in residential utility bills in several American states, and data-centre siting has become contested at the local level in a number of jurisdictions. Regulatory intervention in how large loads are priced and connected is a plausible development over the next two to three years, and it would fall directly on the economics of the buildout. That is a policy risk of a kind semiconductor investors have not historically had to model.

Europe’s position: earnings without the technology exposure

The relative resilience of European indices this week is not an accident of sentiment; it is a consequence of index composition, and it has been a recurring feature of 2026.

The STOXX 600 and the major national indices carry far less weight in semiconductors and AI-linked hardware than the S&P 500 or the Nasdaq Composite. ASML is the significant exception, which is why the Euro STOXX 50, where its weight is largest, lagged the broader STOXX 600 on Tuesday while the FTSE 100, dominated by energy, financials, pharmaceuticals and consumer staples, posted a solid gain. When the AI trade sells off, European indices lose one large constituent while their banks, insurers, luxury houses and consumer goods companies carry on.

That characteristic has been a persistent drag during periods of AI enthusiasm and a persistent cushion during periods of doubt. Investors who spent 2024 and 2025 lamenting Europe’s absence from the artificial-intelligence boom have discovered in July 2026 that the absence has a second side.

The earnings evidence this week supports a more constructive reading than simple index arithmetic. Unilever’s volume-led acceleration, Mercedes-Benz beating consensus adjusted EBIT by roughly 39% across every division, and Philips raising both margin and free-cash-flow guidance are not the results of companies in a struggling economy. Euro-area composite survey data have been reasonably solid, and European Central Bank officials have maintained a hawkish tone that suggests they see no urgency to support activity.

Two qualifications belong alongside that. The first is China. Mercedes-Benz cut its unit sales outlook explicitly on Chinese weakness, and the same dynamic runs through the luxury, automotive and industrial sectors that constitute a large share of European market capitalisation. European corporate earnings are more exposed to Chinese consumption than European gross domestic product is, which is a distinction investors frequently miss.

The second is that some of this quarter’s European margin strength is not organic. Philips’s headline margin included roughly 4.2 percentage points from a United States tariff refund. Tariff-related items — refunds, provisions, reversals and pass-through pricing — have distorted comparability across a range of European exporters through 2025 and 2026, and they will continue to do so. Reading a European earnings season without adjusting for them produces a materially more flattering picture than the underlying trend supports.

What the market reaction does and does not prove

A caution is warranted about the interpretive habit that dominates coverage of days like this one.

Price movements are evidence about what investors currently believe. They are not evidence about whether those beliefs are correct. Alphabet’s shares fell more than 7% after reporting 24% revenue growth and 82% cloud growth; that tells you the market repriced its expectations, not that Alphabet’s strategy is wrong. Unilever rose 6% on a strong quarter; that tells you the market was positioned for less, not that consumer staples are now a superior asset class.

The specific error to avoid this week is treating the rotation as a verdict. Money moving from semiconductors into consumer staples is a portfolio decision made under uncertainty by participants with varying time horizons, leverage constraints and mandates. Some of it is fundamental. Some of it is risk management ahead of a Fed meeting and four megacap earnings reports in forty-eight hours. Some of it is systematic strategies responding mechanically to volatility. Disentangling those contributions in real time is not possible, and analysis that assigns a single cause to a complex move is generally overconfident.

What can be said with reasonable confidence is narrower and more useful. The dispersion between sectors on July 27 and 28 was unusually wide relative to the movement in the broad indices. That pattern — flat index, violent sector rotation — is characteristic of a market reallocating rather than one de-risking. Investors are not leaving equities. They are changing which equities they own.

Whether that reallocation proves prescient will not be knowable for several quarters. The relevant evidence — cloud growth rates, free cash flow trajectories, memory contract prices, and whether five Chinese lithography systems actually ship — arrives on its own schedule.

Frequently asked questions

Why are chip stocks falling in July 2026?

Three overlapping reasons. Reports that a Chinese state-backed group has begun mass-producing immersion DUV lithography machines have raised questions about ASML’s long-term position. The Shanghai listing of Chinese DRAM producer CXMT, which surged around 470% and raised roughly $8.6 billion to expand capacity, has raised concerns about future memory supply. And reports that Nvidia is arranging more than $750 billion of AI deals, including a possible $250 billion backstop for an OpenAI data-centre lease, have revived the argument that AI demand is partly financed by the chip vendor itself. The Philadelphia Semiconductor Index has fallen more than 20% from its late-June record.

Has China actually broken ASML’s monopoly?

No. The reported Chinese machines are immersion deep-ultraviolet systems, which are a generation behind the extreme-ultraviolet systems ASML uses for the most advanced chips. No Chinese entity is close to producing EUV, and ASML remains the sole global supplier. The reported production target is roughly five DUV units in 2026 and about twenty in 2027, and some critical components are still imported from Japan. The market reaction reflected concern about the long-term trend rather than near-term revenue loss.

What is the difference between DUV and EUV lithography?

Deep-ultraviolet lithography uses 193-nanometre light; immersion DUV adds a water layer between the lens and wafer to improve resolution. Extreme-ultraviolet uses 13.5-nanometre light generated inside a vacuum chamber and enables far finer features in a single exposure. DUV can approach advanced nodes using multi-patterning — repeated exposures — but at higher cost, longer cycle time and lower yield. EUV is what makes leading-edge chips economically viable at scale.

What did the Federal Reserve decide at the July 2026 meeting?

The meeting began Tuesday, July 28 and concludes Wednesday, July 29, with the decision published at 2:00 p.m. Eastern. At the time of writing no decision has been announced. The target range stands at 3.50%–3.75%. Market-implied probabilities pointed to a hold as the most likely outcome, with roughly a one-in-five to one-in-three chance of an increase depending on the instrument used. There is no Summary of Economic Projections scheduled for this meeting.

Who is the current Federal Reserve Chair?

Kevin Warsh. He was nominated by President Trump on January 30, 2026 and succeeded Jerome Powell, whose term as Chair concluded in May 2026. Warsh previously served as a Fed governor from 2006 to 2011. He has adopted a notably hawkish posture, describing inflation as “a choice” and promising a policy “regime change,” and has advocated faster reduction of the Fed’s balance sheet.

What is the current US inflation rate?

Headline consumer price inflation was 3.5% year over year in June 2026, published July 14, down from 4.2% in May. The index fell 0.4% on the month, the largest monthly decline since April 2020, driven almost entirely by energy. Core inflation, excluding food and energy, was flat on the month and 2.6% year over year. The July figures are due in mid-August.

Why did Philips shares fall despite beating expectations?

Two reasons. Roughly 4.2 percentage points of the reported 16.4% adjusted EBITA margin came from a one-off United States tariff refund rather than underlying performance; excluding it, the margin was close to what analysts had modelled. More importantly, comparable order intake fell 1%, with mid-single-digit declines in Connected Care. For a medical-equipment maker, orders are the forward indicator and revenue is the backward one. Philips also raised full-year margin guidance to 13.5%–14.0% and free cash flow guidance to €1.5–1.7 billion.

How strong were Unilever’s second-quarter results?

Underlying sales growth was 5.8%, of which 5.5 percentage points came from volume — the company’s strongest quarterly volume performance in more than a decade. First-half underlying sales growth was 4.8%, split 4.2% volume and 0.6% price. The company now expects full-year growth within its 4%–6% multi-year range rather than at the bottom of it. Shares rose about 6% in London, the largest one-day move in two years.

What is “circular financing” in the AI industry?

It describes arrangements in which a supplier provides capital to a customer that the customer then uses, directly or indirectly, to buy the supplier’s products. Nvidia has made equity investments in AI companies, and is reported to be discussing a backstop of as much as $250 billion for an OpenAI data-centre lease plus separate financing for GPU purchases. Critics argue this inflates apparent demand and concentrates ecosystem credit risk on one balance sheet. Nvidia’s position is that these are investments in removing physical bottlenecks — power, land, construction — rather than disguised sales. No accounting impropriety has been alleged by regulators.

Why did Alphabet’s free cash flow turn negative?

Capital expenditure on data centres and AI infrastructure roughly doubled to $44.9 billion in the second quarter, exceeding the cash generated by operations and producing free cash flow of negative $5.85 billion — the first negative quarter since Alphabet’s 2004 flotation. Reported profits remained strong; capital spending does not hit the income statement immediately but is expensed through depreciation over subsequent years. Management raised 2026 capital expenditure guidance to $195–205 billion from $180–190 billion.

Why is Bitcoin falling in 2026?

Bitcoin traded near $63,100 on July 28, well below its October 2025 peak of $126,296. Analysts point to three drivers: expectations that the Fed may raise rather than cut rates, which is negative for a non-yielding asset; a record streak of outflows from United States spot Bitcoin ETFs alongside leveraged liquidations; and competition for the same speculative capital from AI-related equities and funds. Bitcoin has been trading with roughly 84% correlation to the Dow, suggesting a macro-driven move.

What should investors watch over the next week?

The FOMC statement and press conference on Wednesday, July 29; Microsoft and Meta results the same afternoon; Apple and Amazon on Thursday, July 30. The specific metrics that matter are Azure and AWS growth rates, any change to capital expenditure guidance, and any revision to depreciation assumptions on AI hardware. SanDisk reports on August 5, providing the first read on memory pricing since the CXMT listing.

Final assessment

The most useful way to read July 28 is not as a bad day for technology stocks but as the point at which two separate arguments, each building for months, arrived at the same market simultaneously.

The first argument is about competition. For three years the working assumption underlying semiconductor valuations has been that export controls would preserve a durable Western technological lead, and that Chinese manufacturers would remain confined to mature nodes with imported equipment. That assumption has been eroding steadily and, in the space of forty-eight hours, took two visible hits: a domestic immersion DUV scanner entering production, and a Chinese DRAM producer raising $8.6 billion in the largest mainland listing in sixteen years. Neither event changes 2026 revenue for any Western company by a material amount. Both change the shape of the distribution of outcomes after 2028, and that is what long-duration equities are priced on.

The second argument is about capital discipline. Alphabet’s negative free cash flow quarter was the clearest possible statement that AI infrastructure spending has reached a scale where it can consume the cash generation of one of the most profitable businesses ever built. When investors saw a company grow revenue 24%, expand margins by two percentage points, and still burn $5.85 billion, they stopped treating capital expenditure as a proxy for demand and started treating it as a cost. Nvidia’s reported willingness to guarantee $250 billion of a customer’s lease obligations arrived into that changed mood and confirmed the worst reading of it.

The strongest evidence against panic is operational. Cloud revenue is accelerating, not decelerating. Memory is in genuine physical shortage with prices forecast to keep rising. Data centres are power-constrained rather than demand-constrained. Companies do not run into electricity shortages serving imaginary customers.

The strongest evidence for concern is structural rather than operational. The financing has become circular in ways that were not true two years ago; the aggregate capital requirement now exceeds internal cash generation and requires bond markets to cooperate; and the depreciation assumptions underpinning reported hyperscaler earnings rest on useful-life estimates that hardware refresh cycles have been challenging. None of these is a scandal. All of them are reasons for a lower multiple.

What genuinely changed this week is the burden of proof. Through 2024 and 2025, a company announcing higher AI spending was presumed to be capturing opportunity, and the burden fell on skeptics to explain why not. Since July 22 the presumption has reversed. Microsoft, Meta, Apple and Amazon will report into that reversed presumption over the next two days, and the reaction to their results will say more about where this cycle stands than any single headline about lithography or backstops.

Meanwhile, the money that left semiconductors did not leave the market. It went into a soap company that sold more units and a carmaker that beat estimates while shrinking. That is not a flight to safety. It is a preference for cash flow that exists over cash flow that is promised — which, after a three-year period in which the opposite preference was extraordinarily profitable, is at least an interesting change of mind.

This article is provided for general informational purposes and does not constitute financial, investment, tax, or legal advice.

Sources

• Associated Press: How major US stock indexes fared Monday, July 27, 2026
• The Motley Fool: Stock Market Today, July 27, 2026
• Bloomberg: ASML shares drop after report of China producing DUV chipmaking tools
• Tom’s Hardware: China begins mass production of homegrown immersion DUV lithography machines
• TrendForce: China reportedly starts mass-producing immersion DUV tools
• Bloomberg: Japanese, Korean stocks tumble as chip selloff intensifies
• CNBC: CXMT’s 466% market debut makes it the most valuable China-listed company
• Fortune: CXMT shares surge 466% in the biggest mainland Chinese IPO since 2010
• 24/7 Wall St.: Memory stocks fall as China’s CXMT IPO rattles the sector
• Bloomberg: Nvidia’s $750 billion in deals reignite circular AI financing fears
• CNBC: Nvidia and OpenAI in talks for up to a $250 billion AI backstop
• Axios: Nvidia reignites circular AI concerns as it weighs an OpenAI financing guarantee
• Forbes: Nvidia and OpenAI discussing a $500 billion data centre
• Alphabet: Second Quarter 2026 Results announcement, July 22, 2026
• SEC EDGAR: Alphabet Form 8-K exhibit, second quarter 2026 results
• CNBC: Alphabet earnings takeaways — revenue beats, shares fall on capex
• Fortune: Big Tech earnings slam into a market in revolt over AI spending
• Forbes: Big earnings week tests Wall Street’s AI spending fears
• Unilever: 2026 first-half results announcement
• Unilever: Q2 2026 results full announcement (PDF)
• RTÉ: Unilever tops quarterly sales growth estimates as volumes rise
• Mercedes-Benz Group: Second Quarter 2026 results
• Euronews: Mercedes-Benz cuts sales forecast as profit rises despite China weakness
• Philips: 2026 Second-Quarter Results
• Investing.com: Philips shares tumble as weaker orders overshadow a headline beat
• Saipem: quarterly financial reports
• MarketScreener: Saipem adjusted first-half profit down 6.4%
• Reuters via Business Recorder: European shares inch up as earnings counter banks and tech slide
• Federal Reserve: FOMC minutes, June 16–17, 2026
• Federal Reserve: Chair Warsh’s press conference transcript, June 17, 2026 (PDF)
• CNBC: Warsh pledges Fed policy regime change on inflation
• PBS News: Fed Chair Warsh emphasises independence, signals focus on inflation
• CNBC: Consumer price index inflation report, June 2026
• CBS News: Inflation eased more than expected in June as gas prices fell
• CNBC: Oil prices as the U.S.–Iran fighting pause holds, July 28, 2026
• UK House of Commons Library: US–Iran ceasefire and nuclear talks in 2026
• ABC News: How the US–Iran ceasefire and memorandum broke down — a timeline
• Yahoo Finance: Gold remains below $4,100 ahead of the Fed meeting
• Reuters via Yahoo Finance: Dollar hits one-month high on lingering chances of a Fed hike
• Fortune: Three reasons Bitcoin is stuck in a bear market
• Bitcoin.com News: Crypto traders push BTC near $60,000 as the 2026 bear market persists

Affiliate disclosure: Businessfinance.news may earn compensation from qualifying actions completed through selected links on this website, at no additional cost to the reader. Affiliate relationships do not influence our editorial reporting, analysis, or conclusions.