Hello world, it’s Friday, August 28th.
Marvell beat estimates and raised its annual forecast, but shares dropped anyway as investors grew anxious about the timing of Google revenue. Anthropic signed a six-year, $45 billion compute deal with Nscale, SK Hynix broke ground on its $4 billion Indiana HBM fab targeting 2029 production, and Nvidia has paused revenue-sharing deals with cloud partners as memory costs pressure margins.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Tell us how to make Semi Doped better! Give us your feedback! ❤️
Marvell Earnings Beat Fails to Calm Google Revenue Timing Fears
Marvell Technology posted a strong fiscal second quarter and raised its annual revenue forecast, yet shares fell as investors fixated on a single uncomfortable question: when does the reported Google ASIC commitment actually show up in reported revenue. The anxiety isn’t idle. The Next Platform reports the widely cited $120 billion figure reflects a multi-year cumulative estimate, not a near-term booking, which leaves a sizable gap between the headline number and anything Marvell can recognize soon. Q2 revenue came in at $1.87 billion, beating consensus, with data center sales carrying the bulk of the upside. CEO Matt Murphy pointed to accelerating custom silicon demand as the engine behind raised guidance, but analyst questions on the earnings call kept returning to Google deal pacing. The beat was real. The timeline remains genuinely murky.
Anthropic Commits $45B to Nscale Compute, Drops MatX Buyout
Anthropic has signed a six-year, $45 billion compute agreement with British infrastructure firm Nscale, anchored at the Monarch Compute Campus in Mason County, West Virginia, securing an reported 460MW of Nvidia Vera Rubin capacity. The deal gives Anthropic guaranteed GPU access without owning the silicon. That calculus apparently shaped its separate decision on MatX: talks toward a $7 billion outright acquisition of the chip startup collapsed, and Anthropic is now pursuing a partnership instead, according to sources cited by Reuters. The company had weighed buying MatX partly to reduce Nvidia dependence, but backed away. Anthropic is expected to publicly file its IPO prospectus after Labor Day.
SK Hynix Breaks Ground on $4 Billion Indiana HBM Fab, Production Set for 2029
SK Hynix held a groundbreaking ceremony on August 27 at Purdue University’s Holloway Gymnasium in West Lafayette, Indiana, marking the start of construction on its first U.S. AI memory production hub, a facility spanning roughly 133 acres that will house an HBM production line alongside advanced packaging R&D. CEO Kwak Noh-Jung told Reuters the Indiana fab won’t begin AI chip output until 2029, and that a memory shortage is expected to persist through 2030. The project carries a $4 billion price tag, according to Bloomberg. Kwak also said the company is exploring deeper cooperation with Japanese customers and suppliers, including potential closer ties with Kioxia in NAND, extending SK Hynix’s strategic reach across both major memory segments.
Austin: What a win for Purdue. Amazing. I assumed it was just a little R&D fab, can’t believe it’ll have a production line. Reason enough to enroll in Purdue for undergrad or grad school right? Get an internship over at the fab, maybe participate in some R&D, etc?
Nvidia Pauses Cloud Revenue Deals as Memory Costs Squeeze Margins
Nvidia has paused revenue-sharing arrangements with AI cloud partners, according to a Wall Street Journal report cited by Reuters, a move that reshapes the financial terms underpinning its hyperscaler relationships. The timing is awkward: the company lowered its gross-margin outlook in its fiscal 2027 second-quarter earnings, citing surging memory-chip prices that are pushing up costs. Nvidia also warned that revenue growth will trail demand growth, blaming supply-chain bottlenecks spanning foundry capacity, data-center power, and available sites. Both pressures arrive together, and together they put the cost structure of Nvidia’s cloud business under a degree of scrutiny it hasn’t faced in its recent run of outsized quarters.
OpenAI Jalapeño Benchmarks Land as Georgia Power Deal Clears Review
First independent benchmarks for OpenAI’s Jalapeño ASIC show 1.5 to 1.9 times better efficiency per watt than Nvidia’s current inference hardware, according to EE Times. Sam Altman’s summary was characteristically spare: “We made a chip and it is fast.” The chip is a clean-sheet general-purpose AI accelerator, not a derivative design, which makes the efficiency figures meaningful rather than the usual apples-to-oranges marketing claim. On the infrastructure side, the Georgia Public Service Commission approved Georgia Power’s power contract for OpenAI’s $20 billion data center, with commissioners adding consumer-protection provisions requiring OpenAI to absorb cost overruns rather than passing them to residential ratepayers. Site plan review is the next procedural step before construction can advance.
Austin: Interesting macro mashup here — AI can help model labs make awesome chips, but will state and local governments allow permit them to be deployed and powered on?
TSMC Widens Foundry Lead as AI Chip Demand Lifts Market 29%
The global foundry market expanded 29% in Q2 on surging demand for AI silicon, lifting total revenue well above year-earlier levels and handing TSMC a share lead it continues to stretch. Samsung, by contrast, sits at roughly 7% market share, a gap wide enough to raise real questions about its advanced-node roadmap. TSMC’s dominance isn’t an accident of timing: years of disciplined capacity investment in leading-edge nodes and CoWoS advanced packaging have made it the default supplier for every major AI chip designer. That position gives TSMC considerable leverage heading into 2027 capacity allocation talks, where demand for CoWoS slots already outpaces available supply.
Sector Watch
Memory
SK Hynix CEO says the company is actively weighing closer operational ties with Kioxia, signaling potential NAND collaboration beyond their existing relationship. (Bloomberg.com)
CXMT reports first-half revenue up 873% year-on-year with net profit topping 77 billion yuan, its first earnings release as a public company. (WSJ)
Samsung Electro-Mechanics diversifies its FC-BGA substrate customer base as AI server demand sustains the semiconductor substrate boom, per Maeil Business. (매일경제)
Compute
Cerebras details the Nexus system architecture at Hot Chips 2026, tripling rack-scale performance, and previews the CS-6 wafer with integrated stacked DRAM. (Tom’s Hardware)
MediaTek confirms Dimensity 9600 Series on 2nm process will power the Vivo X500 Pro series, marking a high-profile design win for its next-generation mobile SoC. (91mobiles.com)
AMD adds native UCIe 1.1 connectivity to selected Versal RF Series adaptive SoCs, enabling direct chiplet interconnects for mixed-signal AI and communications designs. (EE News Europe)
Foundry & Packaging
Siemens EDA tools receive certification for Intel’s 14A and 18A-P process technologies, expanding EDA support for Intel Foundry’s advanced nodes.
Diodes Incorporated completes acquisition of ElevATE Semiconductor, expanding its footprint in AI-driven semiconductor test equipment. (semivision)
Unimicron, a major ABF substrate supplier, is under investigation in a Taiwan origin-labeling probe, creating a potential supply-chain compliance risk. (digitimes)
PCB & Components
Chinese MLCC makers report rising sales and profits as AI server buildout drives inventories to record lows, with SCMP citing broad gains across domestic capacitor producers. (SCMP)
Samsung Electro-Mechanics shares jump over 4% as investors tie the FC-BGA and MLCC supplier directly to Nvidia’s AI infrastructure demand surge. (theconnectmoney.com)
Vishay Intertechnology expands its power transformer portfolio with the IFBT flyback series, targeting wide-input industrial and AI infrastructure power conversion designs. (Vishay)
Data Centers
Lambda closes a $926 million senior secured term loan B facility, its second major debt financing this year, backing committed GPU deployments for an investment-grade customer. (Lambda)
IREN reports FY26 results with $4 billion in contracted ARR for 2026 capacity and a new multi-year contract with a leading frontier AI lab. (IREN)
Volta AI secures a JPMorgan-led $5 billion debt facility to fund its AI infrastructure buildout, one of the largest single debt raises for an independent AI compute provider. (Bloomberg)
Policy & Trade
Vietnam formally urges Qualcomm and Samsung to deepen AI and chip investment in the country as Hanoi accelerates its semiconductor industry upgrade strategy. (Reuters)
China’s NDRC declares the domestic semiconductor supply chain significantly safer as self-sufficiency efforts gain traction across materials, equipment and design. (SCMP)
Nvidia forms an employee-funded US political action committee as the AI chip leader moves to build direct influence in Washington policy circles. (Reuters)
Optics & Networking
Meta publishes details of MetaRoCE, a new RDMA-over-Ethernet protocol reshaping AI cluster networking by extending custom silicon strategy from compute into fabric. (Intelligent Living)
LG Innotek wins a supply contract to provide automotive camera modules for Amazon Zoox robotaxi deployments, with volume production planned. (The Elec)
SoftBank is seeking an additional $10 billion loan to help refinance debt tied to its investment in OpenAI, extending leverage exposure to the AI buildout. (Bloomberg Tech)
Tell us how to make Semi Doped better! Give us your feedback! ❤️