404K SEMI-AI Evening Tech Brief 2026-07-01 — Memory Price Increases, Packaging Bottlenecks, AI Cloud Expansion
目录
Pre-Market Key Points
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
CSP/Cloud Capex
AI Cloud/Data Center Operators
GPU/CPU/ASIC
HBM/DRAM/NAND/SSD/HDD
Foundry and Advanced Packaging
Semiconductor Equipment/Testing
MLCC/Passive Components and Power Supplies
Optical Communications/High-Speed Interconnect
Robotics/Autonomous Driving
Internet/Platforms
Software/SaaS
Consumer Electronics/Smart Vehicles
Investment-Bank Target Price Changes in the Past 12 Hours
The AI trade continues to spread from GPUs into memory, advanced packaging, passive components, optical interconnects, and data center power. Tonight’s strongest marginal changes are: South Korea’s semiconductor exports hit a new high, with DRAM/HBM/SSD volumes rising together; TSMC raised its CoWoS capacity plan, but near-term allocations were cut; Yageo and Samsung Electro-Mechanics are transmitting AI server demand into MLCC pricing and long-term agreements.
Pre-Market Key Points
The main thread in AI infrastructure tonight looks more like “bottleneck migration”: demand has not disappeared, but the market is starting to separate constraints from GPUs themselves into memory, packaging, testing, power, and passive components. South Korea’s June exports exceeded USD 100bn for the first time, with semiconductor exports reaching USD 44.8bn; this data set pulls the memory price-upcycle narrative back from company-specific stories to industry momentum.
The second thread is the dual signal from advanced packaging. Nomura says TSMC’s 2027 CoWoS capacity target was raised from 1,100kpcs in 2026 to 2,000kpcs; meanwhile, iM Securities cut its 2026 global AI accelerator CoWoS allocation from 1,380K to 1,096K. Capacity is expanding, but delivery cadence remains constrained, which will affect share expectations for NVIDIA, Broadcom, AMD, and Google TPU.
Software and platforms are more divergent. Salesforce and ServiceNow were upgraded by Guggenheim with higher price targets, not because AI monetization is immediate, but because valuations already reflected excessive pessimism. Microsoft’s layoff signal shows hyperscalers are still reallocating labor costs toward AI, cloud, and high-priority projects; platform companies need margin discipline to absorb capex pressure.
Full AI/Semiconductor Value Chain
AI Models/Applications and Capex
OpenAI/frontier-model compute
1) Inference demand continues to amplify compute consumption. Industry estimates suggest token demand could reach 24x by 2030, while HBM supply only doubles and HBM demand reaches 5x.
2) OpenAI-related inference optimization is described as reducing the operating cost of existing models by more than half, with Q1 gross margin at 39% and a year-end target of 52%. The money ultimately flows to GPUs, HBM, power, and cloud leasing. If this is wrong, the key question is whether unit inference cost continues to decline.
"agentic workloads can consume 100x-1,000x more tokens"
Claude/Fable
Access to Claude Fable 5 and Mythos 5 has been restored. After the Commerce Department lifted export controls, the company redeployed the models with new safety classifiers; some normal coding and debugging tasks will fall back to Opus 4.8. The investment implication is direct: frontier-model commercialization depends not only on capability, but also on regulatory clearance, safety false-positive rates, and developer experience.
"Claude Fable 5 will be available again globally tomorrow"
Amazon
1) Amazon’s head of Agentic AI said Q1 2026 Bedrock request volume had already exceeded the total of all prior years. This proves enterprise model calls are entering nonlinear growth, but the material does not disclose revenue per request, inference cost, or gross margin. For now it only shows AWS’s AI application entry point is heating up, not direct profit elasticity.
2) AWS forward-deployed engineering organization: AWS is investing USD 1bn to build a Forward-Deployed Engineering organization, embedding engineers with customers and delivering production-grade code in around 45 days. Cloud competition is moving from “selling model APIs” to “helping customers implement workflows,” which will improve stickiness but also increase labor and delivery costs.
CSP/Cloud Capex
Google
Google’s next-generation Humufish TPU is reportedly using Intel EMIB-T rather than TSMC CoWoS. The key point is not switching suppliers, but bypassing CoWoS interposer reticle limits, lowering cost, and securing a second source. If EMIB-T yields cannot ramp on schedule, Google will still have to return to the tight CoWoS queue.
"Google’s next TPU, codenamed Humufish, is set to use Intel’s EMIB-T instead of TSMC CoWoS"
Meta
Meta’s USD 27bn, five-year agreement with Nebius remains the core anchor for new-cloud long-term contracts. For the market, the question is not whether Meta wants compute, but whether new-cloud operators can secure power, GPUs, and financing, then deliver usable capacity on customer timelines.
Amazon
Amazon is absorbing enterprise AI requests through Bedrock while also being cited in the material as continuing to add cloud and AI infrastructure investment. Tonight, the more important signal is on the demand side: request volume exceeded the historical cumulative total, indicating that inference entry points are expanding. The next validation points are AWS revenue growth and capex returns, not call volume alone.
Microsoft
Microsoft is reportedly cutting about 2.5% of its workforce, or roughly 5,500 positions. This does not mean cloud demand is weakening; it looks more like reducing lower-priority labor costs during an AI and cloud capex cycle. It helps margins, but the material does not disclose departmental distribution, so the direct impact on Azure or Office cannot be judged.
AI Cloud/Data Center Operators
Nebius
Nebius’s story remains that “power and operating capability are scarcer than bare GPUs.” Company threads point to seven sites, 3.5GW of contracted power, and USD 2bn of strategic support from NVIDIA; on-demand GPU pricing was raised 29%, and it continued signing customers even after raising preemptible capacity pricing. The risk is that long-term customer contracts, financing, and delivery cadence must all be executed in sync.
CoreWeave
CoreWeave is discussed within the OpenAI, Claude, and new-cloud compute demand chain. High-intensity inference will increase GPU leasing demand, but the market will increasingly focus on margins and returns, not just revenue growth; if model customers gain bargaining power, returns on new-cloud assets will be compressed.
IREN
The debate around IREN is execution transparency. Short interest is about 64.4mn shares, roughly 19% of public float, with short value of about USD 3.9bn and days to cover of around 1.3 days; the share price fell about 30% in June. This shows the market is not only looking at power resources, but also asking whether the project roadmap, customer contracts, and capital allocation are clear.
Bloom Energy
Bloom Energy and Brookfield expanded their AI infrastructure power partnership from USD 5bn to USD 25bn. What data centers are buying is time: PJM averages 40 months from application to operation, ERCOT’s queue is about 226GW, while Bloom’s on-site power delivery is around 90 days. The gap is that the announcement did not disclose MW, sites, customers, or project economics.
"partnership from 5B to 25B"
GPU/CPU/ASIC
NVIDIA
For NVIDIA, the near-term focus should be advanced packaging rather than demand itself. iM Securities cut its 2026 NVIDIA AI GPU output estimate from 11.14mn units to 9.24mn, and Rubin from 3mn to 1.5mn; however, total AI accelerator volume in 2026 is still expected to grow 51% YoY. Demand is strong, but delivery is constrained by CoWoS size, yield, and HBM stacking.
"Rubin GPU...from 3 million units down to 1.5 million units"
Intel
Intel’s incremental signals tonight come from EMIB-T and EUV photomasks. Google TPU’s use of EMIB-T shows Intel has won a high-signal advanced-packaging use case; meanwhile, the company is expanding manufacturing and utility facilities in Santa Clara by about 107,000 square feet to strengthen internal photomask production. Execution risk remains yield and scale-up.
Broadcom
Jefferies maintains Broadcom at Buy with a USD 550 price target, arguing C28 visibility is improving and ASIC customers are expanding from Google to OpenAI, Meta, Claude, and other model customers. The report estimates OpenAI’s Jalapeno XPU is 50% lower cost and 40% lower power than Blackwell, with shipments before year-end and volume production in 2H27. The key for Broadcom is whether customer diversification can offset dependence on a single TPU customer.
AMD
UBS says AMD’s 2027 CoWoS demand may grow 232%, driven by Venice CPU volume and MI450/MI455 adoption; AMD’s share of total CoWoS demand is expected to rise from 6% in 2025 to 17% in 2027. This is a server CPU and AI accelerator resonance story, not a single-product story.
"CoWoS demand to surge 232% in 2027"
Qualcomm
Qualcomm’s Dragonfly AI accelerator uses a high-bandwidth compute approach, stacking LPDDR on logic chips through through-silicon vias and bypassing traditional HBM, claiming a 6x bandwidth-per-watt improvement for inference. Initial shipments are planned for FY2027, with targets of USD 5bn in data center revenue in 2027 and USD 15bn by 2029.
Inference-chip startups
Etched completed A0 tape-out on TSMC N4P, with customer contracts exceeding USD 1bn and cumulative financing of USD 800mn; Cerebras WSE-3 reached 2,522 tokens/s in Llama 4 Maverick inference, above NVIDIA DGX B200’s 1,038 tokens/s. The competitive focus is shifting from “faster single chip” to whether deployable racks can be sold.
HBM/DRAM/NAND/SSD/HDD
Micron
Micron is the core of tonight’s memory thread. Company interviews point to AI demand still accelerating, USD 22bn in long-term customer commitments, roughly USD 100bn in remaining contracted revenue obligations, gross margin of 84.6%, next-quarter guidance around 86%, Q4 revenue guidance of USD 49bn-51bn, and EPS of USD 30-32. The validation point is whether LTAs can truly reduce cyclical downside.
"Memory is no longer just a commodity. It’s becoming one of the biggest bottlenecks in AI."
Samsung Electronics
Kyobo Securities raised its Samsung Electronics price target to KRW 500,000, expecting 2Q26 revenue of KRW 176.2tn and operating profit of KRW 80.3tn. The fundamental anchors are DRAM and HBM pricing, and the technology anchor is HBM4E reliability-test yield above 70%, but mature yields are typically around 80%, so yield ramp remains a risk.
SK Hynix
Kyobo Securities raised its SK Hynix price target to KRW 4,000,000, expecting 2Q26 revenue of KRW 83.5tn and operating profit of KRW 63.5tn. Another financing thread says the company plans to issue about 17.79mn shares in ADS form, equal to 2.5% of shares outstanding, with proceeds directed to the Yongin fab, Cheongju advanced packaging, and EUV equipment.
SanDisk
Bernstein raised its SanDisk price target from USD 1,700 to USD 3,000 and lifted its FY27/FY28 base-case EPS to USD 243/272, with a bull case of USD 350/400. The core assumption is that NAND LTAs are extending to 3-5 years and adding price guardrails and financial guarantees.
"While these LTA’s do not completely remove risk of future downcycles"
Memory industry pricing
TrendForce expects 3Q26 conventional DRAM ASP to rise 13%-18% QoQ, PC DRAM up 15%-20%, server DRAM up 13%-18%, and NAND up 10%-15%. South Korea’s June DRAM including modules exports were USD 21.846bn, +385% YoY and +17% MoM; MCP/HBM exports were USD 12.681bn, +171% YoY.
Foundry and Advanced Packaging
TSMC
TSMC CoWoS is today’s most important bottleneck indicator. Nomura says the 2027 CoWoS capacity target was raised from 1,100kpcs in 2026 to 2,000kpcs; meanwhile, iM Securities cut 2026 global AI accelerator CoWoS allocations. Capacity expansion validates demand, while the near-term cut signals delivery risk.
"CoWoS production capacity to 2,000kpcs"
OSAT/packaging and testing
Packaging and testing orders are already locked through 2027. ASE, Amkor, and JCET together account for more than 60% of global OSAT sales, with utilization around 90%; ASE’s 2026 packaging and testing unit prices are rising 5%-20%, and memory-related OSAT price increases can reach 30%. The back end is no longer a low-bargaining-power contractor.
Ajinomoto ABF
Citi maintains Ajinomoto at Buy with a JPY 6,400 price target. Management said functional materials shipments are strong and outperforming the FY3/27 sales growth plan of +11%; ABF continues to be driven by larger and more multilayered package substrates. The risk is alternative materials, but the report judges existing product pricing is unlikely to decline when new products launch.
Semiconductor Equipment/Testing
ASML, Lam Research, Applied Materials
Susquehanna raised price targets for the three equipment companies: ASML from EUR 1,475 to EUR 2,350, Lam Research from USD 385 to USD 475, and Applied Materials from USD 575 to USD 900, all rated Positive. The core assumption is increasing backlog, with wafer fab equipment spending potentially reaching as high as USD 300bn.
"wafer fab equipment spending is expected to reach as high as USD 300bn"
Advantest/Teradyne/FormFactor
As AI chip complexity increases, testing is no longer just a pre-shipment process, but a bottleneck for yield and reliability. The material places Advantest, Teradyne, and FormFactor in the same chain as KGD, silicon photonics burn-in, and probe-card pricing; the next question is whether orders continue passing from GPUs/HBM into testers and probe cards.
Aehr
Aehr’s silicon photonics test orders have reached USD 37.2mn, with single-quarter growth above 500% and book-to-bill above 3.5x. Under Goldman Sachs’ framework, the AI optical networking market could rise from about USD 15bn in 2026 to about USD 154bn in 2028, with CPO exceeding USD 90bn.
MLCC/Passive Components and Power Supplies
Yageo
Yageo raised prices across its full capacitor lineup starting July 1, covering about 50% of revenue, with an official increase of around 50%; MLCC capacity utilization is above 80%, and high-end MLCC is above 90%. This is direct evidence of AI server demand spilling over into basic components. The risk is downstream inventory adjustment after prices rise too quickly.
