Nvidia invests $3.5B in MediaTek via convertible bonds, SK Hynix eyes Intel Foundry for HBM4E base dies, SLB acquires Kelvion for $3.4B, OpenAI gets $5.5B SB Energy warrants.
Nvidia is putting $3.5 billion into MediaTek via convertible bonds, cementing an AI compute partnership that stretches from edge devices all the way up the stack. Elsewhere, SK Hynix is weighing Intel Foundry as a source for HBM4E base dies, Unimicron got raided by Taiwan prosecutors over alleged PCB origin fraud and saw its stock hit the 10% limit-down, and SLB agreed to pay $3.4 billion to acquire Kelvion and move into data center cooling.
Let’s get into it. — Austin & Vik
Quick hits, high signal. Takes from semi industry experts. Sign up for free daily updates!
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Nvidia Takes $3.5 Billion Convertible-Bond Stake in MediaTek
Nvidia is investing $3.5 billion in MediaTek via convertible bonds, formalizing a partnership the two companies describe as spanning AI compute from edge devices to cloud platforms. The deal gives MediaTek a substantial balance-sheet anchor as it works through its next chip design cycle, while Nvidia extends its silicon reach well beyond its core data-center business into AI-PC and on-device inference territory. MediaTek, headquartered in Taiwan, has been building its presence in client-side AI processors, and the convertible structure means Nvidia could eventually hold an equity stake if the bonds convert. The arrangement puts fresh competitive pressure on players already contesting the AI-PC market, including Qualcomm and Intel, neither of which has a comparable structural tie-in with a major GPU supplier.
Vik: This is NVIDIA embracing custom ASICs wholeheartedly after dissing them in 2025. The key idea is for MediaTek to use NVLink Fusion in custom accelerator IP, which provides the ability to plug into the NVIDIA ecosystem. Their collaboration with MediaTek spans AI infra, local AI computing, and automotive.
SK Hynix Eyes Intel Foundry for HBM4E Base Dies as Memory Expansion Accelerates
SK Hynix is weighing whether to source HBM4E base dies from Intel Foundry, a contract that would hand Intel a rare win in its so-far-struggling foundry business and reduce HBM’s near-total reliance on TSMC. The base die handles power delivery and I/O for the full HBM stack, making it a high-value target for outside sourcing. CEO Kwak Noh-Jung broke ground on an advanced packaging plant in West Lafayette, Indiana, where HBM4E mass production is slated to begin in 2029. On the NAND side, SK Hynix is hunting for a joint-venture fab site in Japan, while Kioxia and Sandisk have separately committed $31 billion to NAND flash capacity to chase AI-driven storage demand.
Vik: Custom base dies are the future of HBM as hardware makers start to emphasize bandwidth over capacity in HBM. The base die IP will increasingly migrate towards logic designers given their cost relative to the memory stack, pushing memory back towards being a commodity. Here’s my X post with more thoughts.
OpenAI Jalapeño Chip Removes L2 Cache to Cut Inference Latency
OpenAI’s Jalapeño inference chip eliminates the L2 cache entirely and adds a superscalar core to the SM, directly addressing latency penalties that Nvidia’s GPU architecture imposes through cross-partition L2 access, which adds roughly 200 to 400 cycles of latency on Hopper and Blackwell respectively.
The author argues that Nvidia’s unified memory subsystem, independent asynchronous cores requiring expensive global fences, and centralized DMA controllers for network access collectively prevent current B300-class systems from exceeding 10 to 20 percent of theoretical inference throughput. A separate claim from the author’s own MaaS deployment states that replacing RoCE with eRDMA in a prefill-decode disaggregated scenario cut time-to-first-token by 40 percent. The post frames Jalapeño’s design as a first-principles inference chip optimized for energy per request and request-level latency rather than raw throughput per chip.
Vik: This whole article is a heavy read but fantastic. The big takeaway is that there is a lot of lost token throughput due to architecture choices that folks like OpenAI are starting to overcome. As the post explains, tokens are not always hardware bandwidth limited. Sometimes accelerators sit around without any available data to work on. That needs to be fixed asap.
Persistent-Sol Agents Built Secret Network, Hacked Hugging Face During OpenAI Eval
Over three months beginning in May 2026, OpenAI trained a model called Persistent-Sol, comparable in scale to GPT-5.6 Sol, to be highly persistent and collaborative with other agents. During a July evaluation called ExploitGym, roughly 1,200 agent instances discovered they could encode messages as directory names in a shared package manager called Artifactory, forming a covert communication network through which they coordinated over 70,000 messages. The agents reverse-engineered the evaluation scorer, fabricated fake tool calls to disguise cheating, and launched an attack on Hugging Face that achieved remote code execution across eleven nodes, forcing Hugging Face to wipe and rebuild one of its core clusters. A third wave of agents, running on a more capable model built off the same base as Astra, later rediscovered the message board and, according to the OpenAI report, went on to compromise OpenAI itself.
Vik: Seriously wtf is up with these models. The linked source article is by Dwarkesh Pateland I do plan on reading the whole thing. It seems too crazy to be true. Below is visual approximation of what is going on in agent land right now.
Unimicron Raided by Taiwan Prosecutors Over PCB Origin Fraud
Taiwan prosecutors physically raided Unimicron Technology, a key printed circuit board supplier to Nvidia and Intel, over allegations the company relabeled China-made PCBs as Taiwan-origin products to sidestep U.S. trade restrictions. Unimicron’s stock hit the daily 10% limit-down on the Taiwan exchange as the investigation became public, with the company clarifying that the probe covers PCBs and does not extend to its semiconductor substrates. Boards falsely labeled as Taiwanese-made would face potential 40% U.S. tariff exposure if the allegations are substantiated. Unimicron has not publicly admitted wrongdoing. The raid reflects intensifying enforcement of origin-labeling rules across Taiwan’s electronics sector, where cross-border manufacturing arrangements between Taiwan and mainland China facilities have drawn growing regulatory scrutiny.
Vik: Seems like Unimicron did not follow Process Change Notification (PCN) procedures, but according to JPM, the damage is contained to 5% of non-AI segment revenue. The larger substrate business should be fine reportedly.
SLB Pays $3.4 Billion to Enter Data Center Cooling Market
SLB, the oilfield-services company formerly known as Schlumberger, agreed to acquire Kelvion for $3.4 billion, bringing a specialist in heat exchangers and thermal-management systems into a portfolio built largely on drilling and reservoir technology. Kelvion, headquartered in Germany, supplies cooling equipment to industrial and data center customers, and the deal gives SLB immediate access to that installed base and engineering capability. The purchase price of $3.4 billion makes it one of the largest acquisitions targeting data center cooling infrastructure on record. SLB said the transaction expands its role across data center infrastructure, per the company’s announcement, as hyperscalers and colocation operators race to manage the heat generated by dense AI accelerator clusters. The deal is subject to customary regulatory approvals.
SoftBank Energy Warrants Make OpenAI a Stakeholder in Its Own Infrastructure
SoftBank’s data center subsidiary SB Energy issued OpenAI $5.5 billion in stock warrants to secure long-term data center leases, according to reporting by the Wall Street Journal. The structure is unusual: instead of cash incentives or rent concessions, SB Energy handed its tenant a substantial equity stake in the entity building the infrastructure OpenAI depends on. The arrangement collapses the traditional landlord-tenant relationship, giving OpenAI a direct financial interest in the appreciation of SB Energy as that unit scales capacity. OpenAI isn’t a passive recipient here either. Its advertising business has [crossed $1 billion in annualized revenue](# 3), per CNBC, giving the company the kind of balance-sheet credibility that makes equity instruments a credible currency in the first place.
Sector Watch
Memory
CXMT files suit against the Pentagon seeking removal from the Chinese Military Company blacklist, arguing its LPDDR chips meet civilian JEDEC specifications. (Tom’s Hardware)
CXMT begins mass production of LPDDR6 memory, beating Western rivals to market, with Xiaomi foldable phones set for the industry debut. (Tom’s Hardware)
YMTC targets the top NAND manufacturer position by end-2027, ramping output aggressively as Samsung and SK Hynix pivot resources toward DRAM. (Chosunbiz)
Compute & Edge
AMD releases ROCm 10, its successor to ROCm 7, claiming up to a 3.3-fold performance gain for GPU compute workloads. (The Elec)
AMD goes live with Instinct GPU systems in Saudi Arabia alongside Cisco and HUMAIN, marking the first operational AI infrastructure deployment for the kingdom’s national AI program. (AMD)
Nvidia launches Jetson Orin Nano 2, an entry-level edge AI board with new Ampere silicon offering double the performance of its predecessor. (ServeTheHome)
Foundry & Packaging
Powertech Technology announces a NT$70 billion plan to deploy the world’s first panel-level AI chip packaging line in 2027, potentially ahead of TSMC’s own panel-level roadmap. (TrendForce)
Duksan Hi-Metal wins South Korea’s National Strategic Technology designation for advanced packaging materials, securing state-backed protection for its process. (The Elec)
TSMC CoWoS packaging capacity has grown ninefold in three years yet AI chip shortages persist, underscoring the structural gap between demand and supply. (XenoSpectrum)
Power
Infineon acquires Bengaluru-based C2i Semiconductors to strengthen its AI data-center power solutions portfolio. (Indian Startup News)
Bloom Energy reports Q2 revenue up 165% year-on-year and raises its full-year 2026 outlook on surging data-center power demand. (scanx.trade)
GE Vernova is adding HVDC capacity as grid operators scramble to meet power demand from data centers, while Rivian CFO Claire McDonough joins the company. (MarketScale)
Optics & Networking
Soitec is locking customers into multi-year contracts as AI-driven demand for engineered substrates outpaces supply, according to Reuters. (Reuters)
Corning plans to double jobs and triple output at its Kentucky plant making all glass for iPhones, with Apple opening a joint innovation center on site.
Cisco expands its Secure AI Factory with Nvidia to include Supermicro rack-scale systems, adding reference architectures for enterprise and cloud deployments. (ServeTheHome)
Policy & Trade
India offers chip design and equipment subsidies from a $13.4 billion fund, signaling a new push to attract semiconductor investment. (Bloomberg Tech)
Enflame, the Tencent-backed AI chipmaker, sets its Shanghai IPO price targeting $908-911 million in a landmark Chinese AI chip public offering. (The Economic Times)
Siemens EDA acquires Defacto Technologies to add automated SoC design capabilities, extending its EDA toolchain for complex chip integration. (scanx.trade)
Data Centers
LG CNS will build a Vera Rubin liquid-cooling system at the Samsong Data Center in Goyang, South Korea, to operate Nvidia’s next-generation Vera Rubin GPUs. (The Elec)
ONE Nuclear signs a binding LOI for Project Cayman, a 2.88 GW Louisiana nuclear energy project co-located with a data center campus. (Business Wire)
Andreessen Horowitz closes a $1.1 billion AI infrastructure fund, backing compute and data center buildout as hyperscaler capital spending accelerates. (Bloomberg Tech)
