Daily Update - July 3rd, 2026
CXMT filed a $4.3 billion IPO and secured a $3 billion DRAM deal with Tencent, positioning it to expand its memory chip production. Elsewhere, SK Hynix committed $64 billion to memory chip expansion. Separately, SoftBank plans a 10-gigawatt AI cloud platform in the U.S., while Kioxia emerges as the premium provider of NAND.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast playe
CXMT Files $4.3B IPO, Secures $3B Tencent DRAM Deal
ChangXin Memory Technologies (CXMT), China’s largest DRAM maker, filed for a 29.5 billion yuan ($4.3 billion) listing on Shanghai’s Star Market and disclosed a $3 billion DRAM supply agreement with Tencent. The Hefei-based chipmaker is pursuing the offering during a global memory pricing upcycle driven by AI computing demand, with US export controls and competition in high-bandwidth memory flagged as key risks in its prospectus. (StorageNewsletter, SCMP)
Vik: What a great time to IPO! At the peak of the memory market. CXMT Q1 revenue was 700% YoY. They have inked deals with Tencent and ByteDance to provide memory. If the USG approves, who knows they might supply to Apple too!
SoftBank Plans 10-GW AI Cloud Platform in U.S.
SoftBank Corp. CEO Junichi Miyakawa announced plans to offer AI compute services in the United States at a scale of 10 gigawatts, positioning the company as an AI cloud provider to meet surging domestic demand. The initiative would place SoftBank among the largest AI infrastructure operators in the U.S., though specific timelines, capital commitments, and data center partners were not disclosed in the announcement. (Bloomberg)
Vik: More competition to neoclouds like CoreWeave and Nebius but I still don’t buy the overcapacity argument even as Meta is now selling compute too. We haven’t scratched the surface of what AI can do, and lots of compute still needed.
Nokia Wins Sole-Supplier Deal for Orange Belgium Optical Network Overhaul
Nokia has been selected by Orange Belgium as the sole supplier for a multi-year transformation of the operator’s transport infrastructure. The project will unify fixed and mobile networks into a single converged optical transport network across Belgium, with the goal of building a more resilient and scalable platform to handle growing bandwidth demand from consumer and business services. (Nokia)
Vik: Orange Belgium is upgrading their networks to support the next generation of communications infra. Nokia is solely entrusted to install optical networks, with AI-powered hardware, to spruce up performance.
SK Hynix Commits $64B to Memory Chip Expansion
SK Hynix announced a $64 billion investment plan to expand memory chip manufacturing capacity, framed as part of a broader push to meet AI-driven demand. The South Korean company plans to build new production facilities, with spending phased over multiple years across domestic and international sites. (Reuters)
Vik: This is part of South Korea’s broader push to double memory production capacity in 5 years. Michael Burry (of The Big Short fame) is concerned with Government level investments going into memory calling it “the beginning of the end.”
Oriole Networks, AMD Deploy First Commercial Photonic AI Network
UK startup Oriole Networks has deployed what it calls the world’s first large-scale commercial photonic AI network, built in partnership with AMD as part of the UK Advanced Research & Invention Agency’s (ARIA) Scaling Inference Lab. The system moves AI infrastructure optimization from processors to the optical interconnect fabric, connecting AMD accelerators via photonic links rather than conventional electrical interconnects. (EE News Europe)
Vik: Oriole has a great approach to optical networking, and its interesting that they are partnering with AMD. With their PRISM Ultra platform, they can hook up 1,000 nodes with zero hops! Clos networks be damned, this is radical! Look up their website, seriously.
Infineon Closes ams OSRAM Sensor Portfolio Deal
Infineon Technologies has completed its acquisition of ams OSRAM’s non-optical analog and mixed-signal sensor portfolio, a deal first announced in February 2026. The transaction, which cleared all required regulatory approvals, adds sensor technologies, engineering talent, and development sites to Infineon’s Edge Systems division, expanding its capabilities in automotive, industrial, and medical sensing. (EE News Europe)
Kioxia, Sandisk Begin Sampling 10th-Gen BiCS10 3D NAND
Kioxia and Sandisk have begun sample shipments of their 10th-generation BiCS10 3D NAND flash memory, targeting AI data center workloads. Sandisk’s BiCS10 1Tb TLC device delivers up to 4.8 Gb/s interface speed and a 59% bit density improvement over BiCS8, along with improved power efficiency. (Sandisk, Reuters, Bloomberg.com, TechPowerUp)
Vik: Some nice quotes from the Reuters article:
“Kioxia is two to four years ahead of rivals with NAND performance and power consumption due to strengths such as its wafer bonding technology, said Kazuyoshi Saito, an analyst at IwaiCosmo Securities.”
"They haven't been able to respond to the current NAND boom at all. That is why demand is now concentrated on Kioxia alone,"
Anthropic in talks with Samsung to produce custom AI chips
Anthropic is in discussions with Samsung to manufacture a custom AI chip, according to reports from TechCrunch and The Korea Herald. Financial terms, chip specifications, and a production timeline have not been disclosed.
--- (TechCrunch, The Korea Herald)
Vik: So Anthropic considering using Samsung’s 2nm process for their chip? I thought the major issue was yield. Or is TSMC's capacity so constrained that Anthropic will hedge elsewhere? There was some talk last month that Samsung could make Google’s chips too.
Global AI Sales Justify Data Center Costs in Q1, But Margins Remain Thin
Global AI sales excluding China hit $25B in Q1, narrowly exceeding the estimated $21B in data center and chip depreciation costs, but margins remain structurally thin as the industry confronts the unit economics of inference. This financial pressure is accelerating enterprise demand for cheaper open-source models, custom silicon, and aggressive inference optimization, evidenced by OpenAI’s reported success in halving inference costs. The shift from simple chatbots to multi-step agents has increased token consumption while simultaneously triggering cost discipline among buyers who are now routing work to cheaper models rather than paying premium rates. This dynamic marks a critical inflection point where revenue growth is justifying the physical buildout but failing to capture sufficient margin to sustain current capital expenditure levels without significant efficiency gains. The competitive landscape will increasingly favor companies that can deliver efficiency gains or own the software routing layer, rather than those relying on model capability alone to command premium pricing. (theinformation.com)
Vik: The era of mogging frontier tokens for the f*ck of it is done. The future belongs to those who can deploy AI “intelligently” while routing the right requests to the right models.
Hyperscaler CapEx Management and Investor Sentiment
Goldman Sachs projects AI infrastructure companies will account for 57% of S&P 500 EPS growth in Q2, signaling a structural pivot in market profitability toward the physical layer of AI. Concurrently, secondary signals regarding declining rental rates for Blackwell and Hopper platforms suggest hyperscalers are beginning to price in capacity normalization or increased merchant supply. Meta’s CapEx defense, arguing internal overbuild can be monetized externally, highlights the fragility of the current investment thesis as unit economics shift from capability to profitability. The market’s punitive reaction to Meta’s spending underscores growing investor sensitivity to the return on deployed capital. This marks a transition from pure capability expansion to a rigorous scrutiny of operating margins and capacity utilization. (theinformation.com)
Meta Monetization via AI Agents and WhatsApp
Meta is developing a cloud infrastructure business to sell AI compute and model hosting services, directly challenging Amazon, Microsoft, and Google, while simultaneously pushing internal teams to adopt its own development tools over external coding assistants. This marks Meta’s first major push into merchant AI infrastructure, shifting from a capital-heavy consumer model to a profit-generating compute provider. The move capitalizes on Meta’s excess GPU capacity and custom silicon efficiency to create a new high-margin revenue stream alongside its core advertising business. Watch for hyperscaler response and whether Meta can replicate AWS’s scale in the specialized AI cloud segment within 12-18 months. (bloomberg.com)
Vik: “Profit generating provider” is a good cash cow for Meta after all that coin they have spent in building out compute. Some will even argue its high time all those investments translate into revenue. Elon first ran this playbook. Seems like Zuck is on the train. Real question is — will others like Microsoft, Google, AWS also profit off their compute spend?
By the Numbers
Closing moves, Jul 2:
Up: RIVN +8.4%, AAPL +4.8%, PLTR +2.8%
Down: AEHR -17.1%, SITM -14.5%, SNDK -14.1%, AMKR -12.9%, KLAC -11.5%, IREN -10.4%
Highs/lows: RIVN at a 90-day high; AEHR, SITM at 30-day lows
Why:
AEHR -17.1% — Aehr Test’s deletion from the Russell index forced mandatory index-fund selling, compounded by concerns over stretched AI-orders valuation.
AAOI -15.2% — Sector-wide optics valuation selloff dragged AAOI down alongside Coherent (-9%) and Lumentum (-8%) as investors questioned stretched multiples.
VICR -19.2% — CEO and Chairman Patrizio Vinciarelli sold $7.5M (20,000 shares) via Rule 10b5-1, spooking investors despite strong recent guidance.
UCTT -17.8% — Ultra Clean issued soft forward guidance that disappointed investors and drove the selloff.
VECO -18.5%, ACLS -19.0%, ACMR -15.5%, MXL -17.1% — no clear public catalyst yet