404K SEMI-AI Evening Brief, July 13, 2026 — AI Capex Estimates Rise Again, Memory Enters a Contract-Driven Re-Rating, and 800V and Optical Interconnect Take Over
目录
Pre-Market Highlights
Full AI/Semiconductor Supply Chain
AI Models, Applications, and Capital Expenditure
CSP and Cloud Capital Expenditure
AI Cloud and Data-Center Operators
GPUs, CPUs, ASICs, and Memory
Foundry, Advanced Packaging, and Equipment
Optical Communications, High-Speed Interconnects, and Power
Internet/Platforms
Software/SaaS
Consumer Electronics / Smart Vehicles
Investment Bank Target Price Changes Over the Past 12 Hours
The AI buildout continues to redirect cash flow from cloud providers toward chips, memory, equipment, and data-center infrastructure. TSMC’s revenue, Meta’s data-center investment, and SanDisk’s long-term agreements provide concrete validation. The sharp correction in memory stocks is also a reminder that earnings upgrades, valuation volatility, and leverage unwinds can occur simultaneously.
404K SEMI-AI | 2026-07-13
Pre-Market Highlights
The strongest pre-market signals continue to come from AI infrastructure. Meta has expanded its Hyperion project in Louisiana to 5GW, with disclosed investment exceeding $50 billion. TSMC’s June revenue reached NT$442.68 billion, up 67.9% year over year. Capital continues to flow into compute, wafers, packaging, and power infrastructure, while the near-term debate has shifted toward returns on construction and delivery bottlenecks.
Memory has entered a phase in which rising prices, more long-term contracts, and trading-driven deleveraging coexist. DDR4 is appreciating faster than DDR5 as mature-node capacity contracts, while SanDisk has increased future revenue visibility through five long-term agreements. Below-consensus earnings forecasts for SK hynix and an ADR supply shock have further amplified share-price volatility. Prices, product mix, and capital expenditure must all be monitored going forward.
Incremental demand is spreading into system-level components across the supply chain. CPO mass production remains some distance away, allowing copper interconnects, long-haul optical networks, 800V rack power, and high-end passive components to capture demand first. On the consumer side, devices are absorbing higher memory prices, while foldables, autonomous driving, and robotics still require validation through mass production, orders, and yields.
Full AI/Semiconductor Supply Chain
AI Models, Applications, and Capital Expenditure
OpenAI
1) GPT‑5.6 Sol is expected to increase available usage by approximately 10% through inference optimization. The 372k context window has temporarily been rolled back to 272k because of high usage-based charges, while the five-hour limit remains suspended.
2) Efficiency improvements reduce the cost per task but may also increase call volumes. Key validation points include subscription usage, progress in restoring the larger context window, and total token consumption; per-inference pricing alone is insufficient.
“GPT‑5.6 Sol is expected to increase available usage by approximately 10% through inference optimization, while the 372k context window has temporarily been rolled back to 272k because it generated excess usage charges.”
Anthropic
Enterprise customers remain willing to pay a premium for frontier capabilities, but adoption exposes issues involving workflows, error correction, and evaluation standards. The investment implications center on private evaluations, memory, traces, and orchestration control: the deeper models penetrate business processes, the more customers need to retain accumulated learning within their own systems. Key indicators include paid enterprise adoption, renewals, and demand for private deployment.
“Enterprises should control their own learning loops, including private evaluations, memory, traces, fine-tuning, and orchestration choices.”
Frontier-Model Economics
The weighted price of frontier models rose from approximately $1.07 in January to $1.62 in late June, while other models declined from $0.22 to $0.18. This divergence indicates that complex tasks still command a capability premium, while lower-cost models expand the range of economically executable tasks. The key questions are whether token volume can grow faster than unit prices decline and whether model-provider margins can support continued training investment.
“Cheaper, better models do not reduce total spending; they expand the range of things people are willing to try.”
CSP and Cloud Capital Expenditure
Meta
1) The planned scale of the Hyperion project has increased to 5GW, with disclosed investment rising from approximately $10 billion initially and $27 billion last October to more than $50 billion. Local businesses have already secured over $1.6 billion in contracts, and Meta plans to invest more than $1 billion in road, water-supply, and drainage improvements.
2) The Alberta project is approximately 1GW, with investment of roughly C$13 billion. The current intensity of construction does not support the view that compute demand has peaked; risks lie in power availability, depreciation, and project returns.
“The Hyperion data-center supercluster will reach 5GW, with total project costs exceeding $50 billion.”
Microsoft
Consensus estimates place Microsoft’s free cash flow at approximately $55 billion in 2027, rising to $129 billion in 2029 and $165 billion in 2030. The company is viewed as having no need to issue debt to fund capital expenditure. Capex weighs on cash flow in the near term; over the longer term, the question is whether Azure and AI revenue can cover depreciation. Balance-sheet capacity remains a relative advantage.
AI Cloud and Data-Center Operators
CoreWeave
The market interpreted potential compute resales by Meta as new competition, sending CoreWeave down as much as 10.8%. Meta’s subsequent 5GW expansion, third-party compute purchases, and plans for additional data centers continue to indicate a structural shortage. CoreWeave’s strengths are its existing revenue base and backlog, while risks include customer concentration, financing costs, and rental-rate pressure after new supply comes online.
Nebius
Nebius fell 12.4%, and as much as 17% intraday, amid the narrative that Meta would sell surplus compute capacity. However, the company’s comparative investment case still rests on rapid growth and a relatively strong balance sheet. The approaching earnings release is merely an event catalyst; the real validation points are revenue growth, data-center commissioning, secured power, capital expenditure, and cash burn.
IREN
IREN has reportedly secured the most power among three neocloud operators, totaling 5GW. Power resources establish an upper bound for data-center expansion, but secured capacity is not the same as commissioned capacity. Investors should monitor grid-connection timelines, construction capex, customer contracts, and utilization, while avoiding the assumption that power reserves translate directly into revenue. Project-financing maturities should also be reviewed.
GPUs, CPUs, ASICs, and Memory
NVIDIA
On-demand rental pricing for B200 GPUs from non-hyperscale cloud providers stands at $5.81 per GPU-hour, versus $5.08 for the composite Blackwell index and $2.67 for H100. Immediately available B200 capacity has fallen to zero. Frontier GPUs continue to command scarcity premiums, while Rubin is driving further upgrades in HBM, packaging, and power. Risks include repricing of rental rates, utilization, and equipment life as new supply comes online.
Broadcom
Broadcom is involved in Meta’s custom Iris chip, which reportedly reached production readiness within six months and is scheduled for volume production in September. CSP-developed chips can reduce some third-party GPU costs while reallocating value toward design, IP, foundries, and high-speed interconnects. Key indicators are Iris volume production, the scale of Meta’s deployment, and the realization of Broadcom’s AI-related revenue.
FuriosaAI
The Renegade inference NPU consumes approximately 200W. Production is planned to increase from 20,000 units this year to 40,000–50,000 next year, while FuriosaAI is developing the third-generation Stoke high-speed chip interconnect with Broadcom. Low power consumption provides differentiation in the inference market, but volume production, customer expansion, and the software ecosystem will determine whether its efficiency advantage translates into orders.
SK hynix
Q2 operating profit is forecast at KRW60.4 trillion, up 61% quarter over quarter and 556% year over year, but approximately 8% below the KRW65 trillion consensus. Korea Investment & Securities cut its 2026–2027 operating-profit forecasts by 9%–11% while maintaining a Buy rating and KRW3.8 million target price. Long-term HBM agreements improve visibility, while near-term pressure comes from sales mix, ASP growth, and the HBM4 ramp.
“The average fixed transaction price of DDR4 8Gb 1Gx8 rose from $11.50 in January to $21 in June, nearly doubling within six months.”
SanDisk
1) Evercore raised its target price from approximately $1,916 to $3,100, maintained its Outperform rating, and set a $4,000 bull-case target.
2) Five long-term agreements represent at least $62 billion of committed revenue, $42 billion of remaining performance obligations, and more than $11 billion in guarantees and prepayments. These contracts improve earnings visibility, while NAND pricing continues to determine earnings sensitivity.
DRAM Industry
The average fixed transaction price of DDR4 8Gb 1Gx8 rose from $11.50 in January to $21 in June, while prices for 8Gb DDR4 could rise by more than another 50% in Q3. Incremental mature-node capacity remains limited, and HBM is consuming additional wafer capacity, creating a structural shortage in legacy specifications. Key indicators include quarterly contract prices, commodity DRAM allocation, and new capacity additions.
Foundry, Advanced Packaging, and Equipment
TSMC
1) June revenue reached NT$442.68 billion, up 6.2% month over month and 67.9% year over year. First-half revenue totaled NT$2.404 trillion, up 35.6% year over year.
2) Bank of America expects Q3 revenue to increase 11%–15% quarter over quarter, with a gross margin of 67%–68%, and forecasts capital expenditure of $58 billion, $78 billion, and $83 billion for 2026–2028. Revenue validates demand, while risks center on 2nm depreciation, overseas expansion, and customer concentration.
“TSMC’s monthly CoWoS capacity will increase from 120,000 wafers in 4Q26 to 180,000 in 4Q27, while SoIC capacity will rise from 20,000 wafers in 4Q26 to 50,000 in 4Q28.”
Samsung Foundry
Tesla’s AI5 has already taped out and is planned for production at Samsung Electronics’ Taylor fab using the 2nm 2P or 2P+ process, with 2P+ expected to enter volume production in 2027–2028. AI5 will also be manufactured by TSMC. Samsung has secured validation from a physical-AI customer, but its share will still depend on yields, the volume-production schedule, and allocation between the two suppliers.
ASML
Q2 sales are expected to reach €9 billion, 3% above consensus, while EPS is expected to exceed consensus by 5%. EUV shipment forecasts for 2026–2028 have been raised to 66, 95, and 109 systems. Capital expenditure by TSMC and at advanced nodes supports demand, while physical delivery, customer-fab readiness, and equipment acceptance collectively determine the pace of revenue recognition.
TES
Order backlog reached a record KRW145.5 billion at the end of Q1. Q2 revenue is expected to reach KRW115.1 billion, up 18% quarter over quarter and 40% year over year, with operating profit of KRW28.8 billion, up 30% quarter over quarter and 41% year over year. Faster equipment move-ins at Samsung Electronics and SK hynix are supportive, while the pace of backlog conversion into revenue provides direct validation of the equipment cycle.
Wonik IPS
Order backlog stood at approximately KRW400 billion at the end of Q1. Q2 revenue is expected to reach KRW240.4 billion, up 46% quarter over quarter, while operating profit is forecast at KRW21.6 billion, up 102% quarter over quarter. Domestic revenue is being recognized first, with overseas revenue deferred to Q3. The improving cycle should therefore still be assessed alongside year-over-year profit growth, order delivery, and customer production schedules.
King Yuan Electronics
Investment in the company’s first US testing facility could reach $1.4 billion, while 2026 capex has been raised to NT$94.9 billion, nearly double the original plan. The expansion is intended to capture back-end demand from TSMC’s US wafer production. Risks include the construction cycle, utilization, and customer qualification. Key indicators are equipment move-ins, trial production, order conversion, the depreciation ramp, and cash recovery.
Advanced-Packaging Glass Substrates
Glass substrates will require renewed qualification of TGV, wet-processing, coating, electroplating, inspection, automated handling, and cutting equipment. A 10μm deviation on a large-format panel could result in scrapping the entire panel. The thematic trade must ultimately be validated through demonstrations, qualification, approved-vendor status, orders, repeat purchases, reliability testing, capacity, and full-panel yield.
“There are many demonstration systems but few approved suppliers; delivery creates an opportunity, but only qualification creates an order.”
Optical Communications, High-Speed Interconnects, and Power
Ciena
Q2 2026 revenue reached $1.57 billion, up 40% year over year, while full-year guidance was raised to $6.3 billion. Q1 cloud revenue grew 76%. Long-haul data-center interconnect has already begun contributing ahead of CPO. Key questions are how the $7.8 billion backlog converts into revenue, margins, and cash flow, and whether component supply can keep pace.
Amphenol
Q1 2026 revenue reached $7.62 billion, up 58% year over year, with organic growth of 81% in the AI data-center business. Large-scale CPO production may still not arrive until 2028–2029, leaving backplane connectors, copper cables, and high-speed electrical interconnects responsible for in-rack transmission. The risk is that changes in rack architecture redistribute content value per system.
“Every quarter that CPO is delayed gives copper interconnect another quarter of uncontested and irreplaceable demand.”
Texas Instruments and Analog Devices
UBS believes adoption of 800V rack-power architectures is accelerating. Texas Instruments is using GaN and 300mm manufacturing to target sidecar power shelves and high-voltage DC-DC conversion, while Analog Devices’ acquisition of Empower adds integrated voltage regulators and silicon capacitors. Racks above 600kW shift the bottleneck toward final conversion near the package. Market share and revenue disclosed in earnings reports are the key validation points.
Murata
Prices for low- and mid-capacitance MLCCs, including 1μF, 2.2μF, and 4.7μF products, have declined modestly, while high- and ultra-high-capacitance products remain tight or continue to appreciate. Consumer-electronics inventory adjustments are weighing on mature specifications, while AI servers and electric vehicles continue to drive high-end demand. Key indicators include product mix, lead times, the pace of high-end capacity expansion, blended pricing, utilization, and inventory.
Internet/Platforms
Shopify
Jefferies upgraded Shopify from Hold to Buy and raised its price target from $140 to $160. The analyst positions Shopify as the infrastructure layer for agentic commerce, with agents entering product discovery and purchasing workflows potentially providing a modest boost to GMV and strengthening merchant retention. The August 2026 partner commission reform will test channel stability and long-term unit economics.
“Becoming the infrastructure layer for agentic commerce and providing merchants with an ‘agent-enablement’ toolkit.”
AppLovin
BofA believes the valuation of approximately 17x calendar-year 2027 EBITDA is supported by the gaming business, which has the potential to deliver more than 20% year-over-year revenue growth. However, new e-commerce pixel additions fell from approximately 950 in May to around 750 in June, with no material acceleration in adoption following the open launch on June 22. The next growth curve depends on customer adoption, not merely the ad-tech narrative.
“At 17x CY27 EBITDA, valuation appears reasonable, with the gaming business alone capable of delivering more than 20% year-over-year growth.”
Amazon
Combined free cash flow at Amazon, Google, Meta, Microsoft, and Oracle is expected to swing negative from a peak of approximately $260 billion in 2024, driven by roughly $1.8 trillion in AI infrastructure investment during 2026–2027. Whether platform businesses can continue funding this buildout depends on operating cash flow, cloud revenue, and debt costs; capital-expenditure beneficiaries receive more cash in the near term.
Google
1) Google is both participating in hyperscale AI investment and purchasing compute capacity from third parties. External procurement indicates that its internal resources remain insufficient, while platform cash flow bears the upfront cost of data centers, chips, and power. Key indicators include cloud growth, capital intensity, depreciation as a share of costs, and third-party procurement pricing; external purchases or resales should not be interpreted in isolation as evidence of excess supply.
2) Google Search: Search activity and clicks have long helped Google assess answer quality, while generative AI extends this feedback loop into more specialized judgments and workflows. The platform can use feedback to improve its products, while enterprises need to retain control over private evaluations and memory. From an investment perspective, AI search usage, advertising monetization, query costs, click quality, and user retention must improve concurrently.
X
Real-time news, finance, and public discourse continue to concentrate on high-frequency information-feed platforms, but the available content only indicates posting activity and provides no data on time spent or active users. Platform value must be validated through user growth, advertising monetization, creator retention, content governance, paid conversion, and server costs; a single trending event cannot substitute for operating metrics.
“For real-time conversation, news, and finance, 𝕏 seems to be where the discussion happens.”
YouTube, TikTok, and Instagram
A creator maintained daily posting across all three platforms, yet views fell sharply and simultaneously after a sustained increase. This shows that consistent content supply does not guarantee traffic realization; algorithm changes, content quality, and cross-platform competition all affect distribution. For the platforms, the key indicators are creator retention, recommendation stability, and improvements in monetization tools.
“The trend had been steadily rising, then suddenly collapsed... If it is happening across every platform, the problem may be with my content itself.”
YouTube
Creators may still receive copyright warnings or takedowns even when using music they believe to be compliant, with the issue most pronounced on YouTube. Copyright enforcement protects content assets but also raises creators’ operating costs. Key indicators include false-positive rates, appeal efficiency, licensing tools, speed of revenue reinstatement, processing times, and creator churn.
Airbnb
More than 52,000 short-term rental listings have been added across US World Cup host cities, up 12% year over year. Hotel room rates have risen approximately 20%, yet occupancy has declined, while projected net World Cup revenue for New York hotels has been cut from $300 million to approximately $160 million. Expanding platform supply can absorb event-driven demand, but pricing, occupancy, and regulation will determine actual monetization.
Creator Subscriptions
A controversial investment-content product is priced at $439 per year or $49 per month. A 5% paid-conversion rate would imply approximately 15,000 paying subscribers and roughly $6.6 million in annual subscription revenue; a 10% conversion rate would imply approximately 30,000 paying subscribers and around $13.2 million in annual subscription revenue. These estimates exclude taxes and operating expenses. Key validation points are the actual paid-conversion rate, renewal rate, and customer-acquisition cost.
Agentic Commerce Entry Points
As agents increasingly enter search, discovery, comparison, and purchasing workflows, platform value will concentrate around product catalogs, payments, merchant tools, and fulfillment interfaces. Shopify’s rating upgrade provides an investment signal, but the real operating validation remains incremental GMV, merchant retention, partner-driven transactions, and margins following the commission reform.
“Capping commission revenue streams should improve the long-term unit economics of partner-driven transactions.”
Meta Advertising Platform
The generative-AI data issue extends the learning mechanism of advertising platforms: browsing and conversions help platforms optimize ad delivery, while customers’ corrections and business judgments carry greater commercial value. Meta’s platform advantage derives from its scale of data, while enterprise concerns will increase as models penetrate deeper into business operations. Key indicators include advertising conversion, adoption of AI tools, and data-isolation options.
TikTok
After one creator maintained daily posting across platforms, TikTok views shifted from sustained growth to a rapid decline in parallel with other platforms. A single platform’s algorithm cannot fully explain the synchronized change; content quality and production methods are also variables. The platform must reduce supply volatility through recommendation stability, creator retention, audience matching, and monetization tools.
Instagram
Instagram views weakened in parallel with YouTube and TikTok, showing that creators cannot secure stable distribution merely by replicating content across platforms. For the platform, abundant short-form video supply must still translate into effective viewing and commercial returns. Key indicators include recommendation efficiency, creator earnings, content differentiation, and advertising demand.
Cross-Platform Creator Distribution
Publishing daily content across three platforms diversifies single-platform risk but cannot eliminate common shocks from algorithms, copyright enforcement, and content quality. Platform competition will center on creator tools, copyright handling, revenue sharing, audience data, and cross-platform operating efficiency. Creators’ ability to post consistently also depends on whether these costs remain manageable.
Software/SaaS
Palantir
In a comparison of software growth rates, Palantir is expected to deliver 63.5% revenue growth, significantly ahead of AppLovin at 43.1% and Oracle at 33.3%. High growth supports scarcity value but also raises the bar for justifying the valuation. Key indicators include commercial-customer growth, contract expansion, margins, and cash flow—not merely revenue-growth rankings.
