404K SEMI-AI Evening Report 2026-07-16 — TSMC Capacity Expansion, Lower AI Inference Costs, Accelerating Optical Interconnects
目录
Pre-Market Highlights
Full AI/Semiconductor Value Chain
AI Models, Applications, and Capital Expenditure
AI Cloud and Data Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundry and Semiconductor 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
TSMC and ASML have extended demand visibility beyond 2027. The bottleneck in AI hardware is also shifting from a simple scramble for GPUs toward memory, packaging, fiber, power, and software efficiency. Internet and software companies are beginning to use traffic, token budgets, and cloud-platform revenue to validate whether AI is generating real business.
404K SEMI-AI | 2026-07-16
Pre-Market Highlights
The strongest signal tonight remains in semiconductor manufacturing. TSMC posted record second-quarter revenue and profit and raised its 2026 capital-expenditure budget to US$60-64 billion, while ASML upgraded its full-year outlook for the third time. Orders are no longer confined to GPUs, with demand spreading more visibly to EUV, advanced packaging, HBM, optical interconnects, and power equipment.
Inference costs continue to decline, but cheaper compute does not automatically translate into higher returns. The DeepSeek V4 deployment shows that a mature software stack can deliver orders-of-magnitude improvements on the same hardware within weeks. Enterprises, meanwhile, are shifting token budgets from fixed per-seat fees to pooled consumption models. The key metrics ahead are utilization, output per user, and cloud-platform gross margins—not just model leaderboards.
Divergence across consumer and platform businesses is increasingly reflective of underlying economics. AI-driven traffic skews more heavily toward brands’ direct websites, while Amazon and DoorDash continue to defend their gateways through fulfillment networks. Foldable devices, robotics, and autonomous driving all have catalysts, but capacity, yields, fleet utilization, and mass-production timelines remain the more rigorous validation metrics.
Full AI/Semiconductor Value Chain
AI Models, Applications, and Capital Expenditure
OpenAI
1) After Codex and ChatGPT Work reached 9 million daily active users, weekly usage allowances were reset again, indicating that demand growth is beginning to test service reliability.
2) Following the removal of the five-hour limit for Codex Plus and Pro, the product issue has shifted from availability to the allocation of weekly allowances. If supply cannot keep pace, both user experience and inference costs will come under pressure.
"We actually reached 9 million active users earlier today, but then got distracted by the several million things the team had to handle to keep the system running properly and reliably."
Anthropic
1) The company is reportedly in talks to increase its bank credit facility by several billion dollars ahead of a potential listing, as funding requirements rise alongside compute infrastructure investment.
2) Its annualized revenue increased from US$9 billion to US$30 billion in the first quarter, with enterprise APIs accounting for approximately 75%-85%. Growth is rapid, but credit facilities, long-term compute contracts, and forward plans exceeding 100GW are also driving cash requirements materially higher.
DeepSeek: On its first day running on MI355X, V4 delivered only approximately 1-2 tokens per second per user; after 26 days of optimization, performance improved by more than 100x. Disaggregated serving on GB200 delivered gains of up to 5x versus B200, while B300 achieved a further approximately 3x improvement within one week. The model itself matters, but software adaptation speed and throughput per watt are what ultimately determine revenue.
AI Cloud and Data Center Operators
CoreWeave
1) The company has disclosed a US$99.4 billion backlog, confirming that demand visibility remains extended.
2) CoreWeave is considering using put options to hedge against declines in memory prices. Evercore views this as evidence that long-term supply agreements are becoming more binding, rather than a sign that DRAM, NAND, and HBM demand is about to lose momentum. The real risk is the simultaneous lock-in of high-priced contracts and customer concentration.
"DRAM/NAND supply shortages will intensify toward the end of CY26 and may persist through most of CY27."
Nebius: Its agreement with Meta is valued at US$27 billion. The AI cloud model can generate US$7-13 million in annual revenue per megawatt, above the US$1.2-2.3 million from pure colocation leasing, but capital expenditure also rises to US$40-60 million per megawatt. Order value must be assessed alongside financing costs and power-delivery execution.
IREN: Its 200MW Microsoft-related project represents approximately US$1.94 billion in annualized revenue. It also has a 60MW Nvidia project involving 18,400 GPUs and a five-year contract value of approximately US$3.4 billion. The two contracts convert power capacity into auditable revenue, but construction timelines, GPU delivery, and customer concentration will continue to determine the pace of cash recovery.
GPU/CPU/ASIC
Nvidia
1) A Japanese partnership project will deploy 13,750 Vera CPUs and 27,500 Rubin GPUs, with targeted power capacity of approximately 140MW. Sovereign AI projects and financial institutions building their own data centers are emerging as incremental demand beyond hyperscalers.
2) The Kyber controversy relates to Rubin Ultra, while the company has confirmed that Vera Rubin is already in mass production. The two timelines should not be conflated.
"That is not accurate. Vera Rubin has entered mass production and will continue to be produced at scale."
AMD
1) During the DeepSeek V4 deployment, MI355X performance improved by more than 100x from approximately 1-2 tokens per second per user on the first day, demonstrating that software optimization can still materially alter hardware economics.
2) The MEXT architecture increases memory’s share of hyperscale data center spending from approximately 8% in 2023-2024 to approximately 30% in 2026. The memory wall is increasingly determining whether GPU clusters can reach full utilization.
Intel
1) TSMC’s 2nm outsourced share of Nova Lake production has reportedly been cut from the originally planned 60%-70% to below 20%, with more than 80% potentially shifting back to Intel 18A.
2) The 18A process uses RibbonFET and PowerVia. Any return of orders must be validated through mass-production yields, rather than pilot-line progress alone.
Qualcomm: The company is targeting more than US$40 billion in non-handset chip revenue by FY2029, including over US$15 billion from data centers. In FY2027, at least two hyperscale customers are expected to contribute no less than US$1 billion each. The revenue trajectory outside handsets is steeper, with the performance and customer ramp of the first data center chips serving as the key validation points.
Groq: Groq 3 racks are scheduled to ship in the second half of 2026. Each rack contains 256 LPUs, 32 liquid-cooled trays, and 52-layer PCBs. Forecast demand for 2026-2027 is 12,000-13,000 racks. Copper-clad laminate content is worth approximately NT$3.84 million per rack, as low-latency inference shifts value toward PCBs and liquid cooling.
"Each rack requires 384 CCL sheets priced at NT$10,000 each, bringing the CCL content value per system to NT$3.84 million, or approximately US$120,000."
HBM/DRAM/NAND/SSD/HDD
Micron: Management guided to revenue of US$33.5 billion, plus or minus US$750 million; gross margin of approximately 81%; and EPS of US$19.15, plus or minus US$0.40. HBM consumes approximately three times as many wafers per bit as DDR5. Earnings leverage is substantial, but risk is also concentrated in the duration of high-priced contracts and the pace of new supply additions.
Samsung Electronics
1) Goldman Sachs expects the average selling price of HBM to rise to approximately US$2.90 in 2027, up approximately 90% year over year.
2) Equipment orders for the Pyeongtaek P5 facility will exceed KRW10 trillion. The project is planned to include six cleanrooms, with capacity more than 50% higher than existing fabs. The first line could begin operating as early as 2027, but the capacity ramp remains dependent on equipment lead times.
"Most equipment suppliers for the first-phase P5 production line have already been selected, while supplier selection for the second and subsequent phases is also progressing."
SanDisk: New long-term supply agreements are beginning to adopt fixed and floating pricing structures with both price floors and ceilings. Customers are consequently exploring financial hedges, indicating that price risk is moving from the spot market into contracts. Near-term supply remains tight; the next indicators are NAND contract pricing, enterprise SSD deliveries, and whether customers seek to renegotiate terms. The earliest disconfirming signals would be premature contract renegotiations, deteriorating inventory turnover, and customers reducing committed volumes.
NAND and SSDs: KV caches for AI inference can reach tens of terabytes, with individual transfers of 10-100GB, requiring approximately 100-microsecond read latency and large-block sequential access. Samsung Electronics and Kioxia together hold approximately 43% market share, while YMTC’s share has increased from 8% to 13%. Beyond pricing, firmware and endurance will determine enterprise market share.
Foundry and Semiconductor Equipment
TSMC
1) Second-quarter revenue reached US$40.2 billion, with a 67.7% gross margin and 60.3% operating margin. The company raised its 2026 capital-expenditure budget to US$60-64 billion.
2) Demand for advanced nodes and advanced packaging continues to drive capacity additions. With capital intensity rising, the next question is whether higher wafer pricing and utilization can absorb the resulting depreciation.
3) TSMC advanced packaging: Industry-wide CoWoS monthly capacity is expected to reach approximately 200,000 wafers by the end of 2026, including approximately 120,000 wafers at TSMC. By 2028, estimated advanced-wafer revenue is projected to rise from US$90-100 billion in 2025 to US$160-180 billion. Demand is very strong, but packaging yields and repeat customer orders require continued verification.
"Given continued strong structural demand from customers, including the emerging agentic AI market, we have decided to raise our FY2026 capital budget to US$60-64 billion and will continue investing aggressively to support customer growth."
ASML: Quarterly sales reached €9.326 billion and net income €2.918 billion. Full-year revenue guidance was raised to €43-45 billion, with a gross margin of 54%-56%. Annual Low-NA EUV capacity is approximately 65 systems and is planned to increase by approximately 30% in 2027. Order conversion remains dependent on the pace of customer fab construction and equipment acceptance.
Applied Materials: Demand for selective etch and deposition equipment is increasing with the complexity of 3D structures. Every 100,000 wafers of monthly capacity represents an approximately US$1 billion opportunity for etch and deposition equipment, and the related business is growing by more than 30%. Transistor stacking and backside power delivery will add process steps, but equipment lead times exceeding one year could also delay revenue recognition. Customer acceptance, chamber deliveries, and advanced-node fab starts must be assessed together.
Optical Communications, High-Speed Interconnects, and Power
Coherent
1) Third-quarter revenue reached US$1.806 billion, including US$1.362 billion from data center and communications, with an adjusted gross margin of 39.6%.
2) Six-inch InP capacity is scheduled to double by year-end and more than double again by 2027. Its serviceable addressable market exceeds US$50 billion by 2030, but platform expansion must be validated through revenue from 1.6T, CPO, and optical networking.
"Demand already exists. The pace of revenue recognition depends on how quickly supply is unlocked."
Lumentum: For short-reach scale-up connectivity, VCSEL costs at scale could decline to US$100 per Tbps, versus approximately US$50 per Tbps for copper interconnects. Global compound-semiconductor capacity can reportedly support approximately twice the required scale-up optics demand. The technology competition will not be determined by speed alone; fiber cost, thermal stability, and yields will also determine market share.
Corning: First-quarter optical communications revenue reached US$1.85 billion, up 36% year over year, while segment net income increased 93%. A GPU rack requires approximately 36 times as much fiber as a CPU rack. Preform production lines require 18-24 months to build, and the top five suppliers control nearly 60% of capacity, making upstream glass the true constraint behind the shortage.
"If something were not scarce, no one would lock in supply for multiple years and help fund new factory construction."
Credo: The acquisition of DustPhotonics is valued at approximately US$1.3 billion. Active copper cables consume approximately 2W per end over a three-meter connection, compared with up to approximately 20W per end for active electrical cables over the same distance. In-rack connectivity is shifting from passive to active copper, increasing content value, but high-speed SerDes bit-error rates and customer qualification will determine the volume ramp.
Marvell Technology: First-quarter revenue reached US$2.418 billion, up 28% year over year, with data center revenue of US$1.833 billion, or 76% of the total. The company expects scale-up optics revenue to exceed US$300 million in FY2028 and custom silicon revenue to exceed US$10 billion in FY2029. Order visibility is strong, while product transitions and customer concentration remain the main variables.
800V DC Power: A 600kW rack requires approximately 12,500 amps at 48V, versus approximately 750 amps at 800V, reducing facility power consumption by approximately 5%. Relevant AI data center capacity is estimated to reach 39GW by 2030. Power-system content is worth approximately US$3.6-4.8 million per megawatt, with safety standards and scaled deployment serving as the key validation points.
Internet/Platforms
Shopify: Traffic growth from AI platforms to brand-owned sites is approximately 4–8x that from traditional marketplaces, versus a roughly 2x gap in traditional search; product discovery is shifting from large, centralized platforms toward direct-to-consumer brand sites. The company’s real moat lies in payments, catalog synchronization, and omnichannel interfaces. The risk is that AI platforms ultimately control checkout and customer relationships.
“brand.com is outperforming what I call traditional marketplaces by 4–8x”
Amazon
1) Same-day delivery now covers more than 2,000 U.S. cities, with the business growing approximately 40x.
2) AI revenue rose from approximately 2% of AWS revenue in Q1 2024 to approximately 10% in Q1 2026; Bedrock’s annualized revenue is approximately $5.5 billion. Physical fulfillment and the cloud platform are defending Amazon’s consumer and enterprise gateways, respectively.
DoorDash: The platform now covers approximately two-thirds of U.S. restaurants, its grocery business is growing by approximately 30%–40%, and it operates more than 100 DashMart locations. AI may reshape the search gateway, but it cannot immediately replicate local inventory, delivery density, or merchant relationships. Investment validation hinges on order frequency, fulfillment costs, and grocery gross margins.
Etsy: An estimated more than 100% of EBIT comes from on-platform advertising, with an overall take rate of approximately 25%. If AI search directs traffic straight to brand sites, platforms with heavier advertising burdens and weaker fulfillment capabilities will be more exposed to pricing pressure. Key metrics are ad load, active buyers, and merchant retention. If traffic declines and ad pricing cannot offset the shortfall, earnings will come under pressure before GMV does.
“Stronger intent and higher conversion rates have already been highly effective in lifting CPC to a sufficiently high level”
Google
1) Google and Shopify are jointly advancing the Universal Commerce Protocol, allowing merchants to retain their existing payment arrangements. There is currently no evidence that Google will directly capture payment revenue.
2) The Steel River project plans 2.5 GW of solar capacity and 2.9 GWh of energy storage, underscoring how competition for the AI gateway is now tied to power procurement.
Meta: The AI R&D; organization has approximately 3,000 engineers, roughly 70% of whom are new graduates. Hyperion is planned at 1.5 GW, the Iowa project at 1 GW, and Prometheus at more than 3 GW, with a network-bandwidth target of approximately 22 Pbps. Model progress must be validated through advertising conversion and returns on capital expenditure, not cluster scale alone.
“It has been more than a year since the disastrous Llama 4 launch prompted Zuck to rebuild the entire AI organization.”
Waymo: Daily trips per vehicle are approaching the low 20s. The model indicates EBIT per mile could turn positive at just above 30 to the mid-30s. The company imports approximately 300 Ojai vehicles per month. Fleet expansion can dilute remote-assistance and maintenance costs, while utilization, municipal permits, and relationships with partner platforms remain risks. If additional vehicles do not raise order density in parallel, depreciation and operating costs will instead weigh on returns.
Uber: Internally, the company has set its generative-AI budget at $1,500 per employee per month, indicating that the platform is turning AI into an operating tool. Externally, it faces the risk that autonomous-driving providers bypass aggregators. Key metrics include AI-driven savings in customer service, development, and dispatch costs, as well as the stability of the Waymo partnership. Internal cost reductions and external gateway risks must be assessed separately; tool adoption cannot substitute for profit improvement.
Microsoft: Azure scored 53 out of 54 in the enterprise AI stack assessment, reflecting its integration of models, data, identity, and developer tools. However, the cloud platform’s actual profitability depends on GPU utilization and incremental software revenue. If customers merely migrate inference costs without adding workloads, returns on capital will come under pressure. The most important metric is whether usage growth can outpace increases in depreciation and electricity costs.
Apple Services Ecosystem: Personal agents capable of invoking apps directly could weaken traditional gateways, but the history of automatic replenishment shows that consumer behavior changes slowly. Amazon’s Subscribe & Save accounted for only approximately 1%–3% of GMV in 2023. Apple’s advantages remain its devices, payments, and permission layer; the risk is that weaker model capabilities shift the gateway elsewhere.
Brand-Owned Sites: Consumers purchase approximately 250–300 types of products annually, making demand more complex than the “automatic replenishment” narrative suggests. AI is more likely to become a high-intent search channel first than to replace platforms entirely. Merchants able to connect catalogs, inventory, and payments to models in real time will benefit first, with repeat-purchase and conversion rates providing more robust evidence. Gateway shifts must ultimately translate into transactions, returns, and customer-acquisition costs—not traffic alone.
“The infrastructure behind the e-commerce technology stack is highly complex”
Grocery Platforms: A typical grocery store carries approximately 40,000–50,000 products, far more than the curated inventory of on-demand delivery services. Instant delivery can capture high-frequency demand, but long-tail categories still require marketplace aggregation. Key metrics should include items per order, out-of-stock rates, and delivery subsidies rather than city coverage alone. The broader the assortment, the more likely inventory accuracy and picking costs become profit bottlenecks.
Advertising Platforms: AI search may reduce clicks while increasing intent and conversion per visit, allowing ad pricing to potentially offset traffic declines. Amazon has time to embed sponsored prompts into the AI shopping experience. If ad revenue per search declines faster, platform margins will come under pressure before GMV does. Click volumes, conversion rates, CPC, and ad load should be tracked together on the same operating dashboard.
Payment Protocols: The Universal Commerce Protocol allows merchants to bring their own payment arrangements, weakening the single-path assumption that AI platforms will inevitably take over payments. Revenue attribution will depend on who controls product catalogs, fraud prevention, and after-sales responsibility; fee rates, chargeback rates, and conversion rates merit closer attention than protocol names. The party bearing refund, risk-control, and customer-service costs is the party that truly controls the transaction economics.
Fulfillment Networks: Local warehouses, driver density, and inventory synchronization remain physical barriers for internet platforms. AI can redistribute traffic, but it cannot quickly replicate same-day delivery across more than 2,000 cities or a network of over 100 local warehouses. Valuation dispersion will increasingly favor platforms capable of turning digital gateways into on-time delivery. Order density, delivery times, and contribution profit per order remain the most direct validation metrics.
“Counterevidence to consumer automatic replenishment includes the fact that delivery accounted for only 30% of U.S. milk consumption in the early 1960s and subsequently fell to zero”
Software/SaaS
Anthropic
1) Annualized revenue rose from $9 billion to $30 billion in Q1, with APIs accounting for approximately 75%–85% and net revenue retention at approximately 500%. Growth is being driven by increased usage among large customers, but exceptionally high retention also implies revenue concentration and a parallel increase in compute costs. Key validation points are gross margin and the sustainability of contribution per customer.
2) Claude Code: Enterprises are willing to pay approximately $150–$250 per developer per month, well above traditional developer-tool seat pricing. This pricing is justified only if the product reduces debugging and delivery time. If automation errors create rework, permission incidents, or code leakage, security-review costs will offset efficiency gains. Actual returns must be measured through delivery cycles, defect rates, and manual-review time.
“Net dollar retention—the increase in spending by customers that had already been corporate clients at least one year earlier—was 500%”
