Research Finder
Find by Keyword
Were July Semi Market Declines Demand Warning, or Execution Timing?
July ran opposite to the technology data. We think the disagreement is about execution timing across a supply chain with no buffer left.
Key Highlights:
- The Philadelphia Semiconductor Index fell 21% in July, its worst month since October 2008, with intraday swings of at least 2% in all 22 sessions and roughly $2.2 trillion in market value erased.
- Samsung told analysts on 30 July that the memory shortage worsens in 2027 and persists into 2028, and that it has signed supply agreements of five years or longer with the top five data center operators covering 60% to 70% of its capacity, with upfront payments and floor pricing.
- Microsoft reported commercial remaining performance obligations of $678 billion on 29 July, up 84%, with all sequential growth coming from customers outside the frontier model companies, and management indicated it expects to remain capacity constrained into next year.
- CXMT holds roughly 8% of DRAM against Samsung at 38%, SK Hynix at 29% and Micron at 22%, and without EUV access it appears to need about 30% more wafers for equivalent output, with HBM production targeted from the end of 2026.
- Moonshot AI's Kimi K3, released 16 July and priced by markets as an efficiency shock, can measure as comparable to or more expensive per completed task than GPT-5.6 Sol at medium reasoning, according to Artificial Analysis, because per-token pricing understates what token-heavy reasoning models consume.
The News:
In July, semiconductor equities recorded their sharpest monthly drawdown since the 2008 financial crisis. The Philadelphia Semiconductor Index fell 21% and roughly $2.2 trillion in market value bled off the sector. The decline ran opposite to the technology data reported during the same weeks, as TSMC raised both spending and revenue outlooks, Intel posted its strongest year-over-year growth in more than a decade, and Samsung extended its memory shortage guidance into 2028. Two open-weight model releases from Chinese labs, a Chinese memory listing, and a report of possible domestic Chinese lithography production supplied the narrative pressure. We read the resulting volatility as a repricing of execution timing rather than a repricing of end demand.
Analyst Take:
The semiconductor and AI tell was TSMC. The company fell 7% in a single session despite reporting a 77% jump in quarterly operating profit. When a print that good produces a move that bad, the market is not arguing about the quarter. It is arguing about everything scheduled after it.
The reasonable objection deserves a full examination. A sector that has run this hard for this long accumulates crowded positioning, and crowded positioning unwinds violently for reasons that have nothing to do with fundamentals. That reading fits the Korean leveraged product liquidations and the forced selling in flash names. We think it explains the shape of the move and the speed of the recovery. It does not explain why the same investors punished capital expenditure increases they rewarded two quarters earlier, or why the drawdown concentrated in names with the longest execution runways. Positioning explains the amplitude. Something else set the direction.
What It All Means
Nothing major actually happened in July, which provides the analytically interesting part. No big misses, regulatory or litigation moves or tech breakthroughs. The swans were not even really all that grey. The July sequence consisted of three events, none of which changed a shipped product or a booked order.
On 16 July, Moonshot AI released Kimi K3, described as the largest open-weight model published to that point and claimed to perform competitively against frontier Western systems. DeepSeek's V4 was already available, and coverage tied the pair to the worst week for United States chip equities in more than a year. The market appeared to price this as a demand event, on the reasoning that cheaper capable models require less silicon to serve.
On 27 July, CXMT listed in Shanghai, rose 466% on debut and raised $8.6 billion. In the same window, The Information reported that a Chinese firm has begun producing immersion DUV lithography systems for delivery to SMIC and CXMT. Neither development altered installed capacity or qualified output during the quarter.
Underneath, the technology data moved the other way. Samsung disclosed long-dated capacity commitments to hyperscale buyers, covering 60% to 70% of its capacity with upfront payments and floor pricing. Intel narrowed its foundry operating loss materially while data center revenue grew sharply. Applied Materials, Lam Research and the wider equipment complex showed no sign of order deterioration.
Having sat on the vendor side of supply-constrained enterprise hardware businesses in the US, Europe and Asia, we recognize the pattern. When allocation rather than demand sets the quarter, the operative risk stops being whether customers want the product and becomes whether every upstream commitment lands on its promised date.
Market Analysis
The industry is running without shock absorbers. Memory is contracted years forward. Advanced packaging is spoken for. Leading-edge capacity is allocated. Microsoft's contracted backlog of $678 billion is the clearest statement available that the demand side has been pre-sold, and management continues to describe itself as capacity constrained.
That condition changes what news does to price. In a chain with inventory, a one-quarter slip in HBM4 qualification or a delayed grid interconnect gets absorbed. In a chain with the buffers drained, it propagates. Investors are therefore forced to underwrite a schedule across HBM4, 2nm ramps, rack-scale integration and power delivery, and any headline touching a date carries disproportionate weight. That, we think, is what the volatility is measuring.
The demand side of the ledger looks more durable than July's price action suggested. Inference workloads are projected to compound at roughly 35% through 2030 and per McKinsey eventually represent 30% to 40% of data center demand. The majority of this growth will be driven by reasoning and agentic patterns that consume substantially more compute per completed task than the single-pass workloads they replace.
Which returns us to the open-weight question. The efficiency argument is genuinely strong: cheaper capable models widen access and reduce the cost of any given inference. The measurement so far cuts against the conclusion drawn from it. Token-heavy reasoning models can cost more per finished task even when they cost less per token, and every efficiency gain to date has expanded the set of economically viable tasks faster than it has reduced aggregate consumption. Cheaper tokens are not fewer tokens.
Looking Ahead
The key trend we'll be monitoring is whether execution proof points arrive on schedule, because that is the variable the market has chosen to price. This week supplies an unusually broad cross-section. AMD reports Tuesday against Street expectations near $11.3 billion, where MI350 shipment cadence, MI450 / Helios rack-scale timing, and gross margin against memory input costs matter more than the headline. Sandisk gives the NAND pricing read Wednesday, with Western Digital (WD), now hard drives only, indicating whether storage capacity is keeping pace with the compute buildout. The equipment complex follows Thursday and Friday, with ACM Research offering a direct look at Chinese fab spending. Applied Materials reports 13 August. NVIDIA is scheduled for 26 August. Should those prints show order books intact and schedules holding, the July drawdown will likely be remembered as a timing dispute. Should two or more slip, the repricing was early rather than wrong.
Stephen Sopko | Analyst-in-Residence – Semiconductors & Deep Tech
Stephen Sopko is an Analyst-in-Residence specializing in semiconductors and the deep technologies powering today’s innovation ecosystem. With decades of executive experience spanning Fortune 100, government, and startups, he provides actionable insights by connecting market trends and cutting-edge technologies to business outcomes.
Stephen’s expertise in analyzing the entire buyer’s journey, from technology acquisition to implementation, was refined during his tenure as co-founder and COO of Palisade Compliance, where he helped Fortune 500 clients optimize technology investments. His ability to identify opportunities at the intersection of semiconductors, emerging technologies, and enterprise needs makes him a sought-after advisor to stakeholders navigating complex decisions.