Firmus, the Australian company that pivoted out of Bitcoin mining last year, raised $2B in equity at a $10.5B valuation — nearly double its April round — from Blackstone, Coatue, Jane Street, and existing backer Nvidia, days after Blackstone and Coatue arranged a separate $10B debt facility for the same AI-factory buildout across the Asia-Pacific. Nvidia's equity locks in compute demand while Blackstone and Coatue's debt funds the land, power, and concrete underneath it — two capital layers with different return requirements financing the same buildout, the model Coatue's own Next Frontier venture is now replicating for US data-center land. (Bloomberg)
Private Companies
Lanciumfundraise
Nvidia agreed to invest up to $3B in Lancium, the Texas power developer whose Abilene campus is the first live site of the Stargate buildout — a $2B initial tranche for a ~20% stake, with $1B more tied to grid-connection milestones, valuing Lancium's land and power portfolio near $10B. It extends the equity-for-demand-security logic Nvidia already runs with neoclouds down to the physical power layer underneath them, and sets Lancium up for a targeted 2027 IPO. (Yahoo Finance)
Source Foundryfundraise
Leopold Aschenbrenner's Situational Awareness fund put another $400M into Source Foundry this week, taking its stake in the two-year-old startup to roughly $500M and valuing the company near $5B. Founded by Stanford researchers Abdulmalik Obaid and Joe Burg with backing from Sequoia, Source Foundry is building lithography machines aimed at breaking ASML's monopoly on the extreme ultraviolet tools every leading chip depends on — the same bottleneck TSMC, Samsung, and Intel all still buy from one supplier. (WSJ)
Public Markets
SK Hynix's board approved a $38B plan for two new memory fabs, and the stock fell anyway — investors are weighing AI-era shortage against 2027 oversupply risk, the same capex debate now shadowing Micron's build-out plans. (Yahoo Finance)
Marvell extended its post-FMS-2026 rally on new CXL memory-pooling and photonic shared-memory hardware — a bet that disaggregated memory, not just more GPUs, drives the next leg of data-center capex — clawing back share from Broadcom's optical DSP lead. (TradingKey)
Emerging
Memory supply chain: Apple is testing DRAM and NAND chips from China's CXMT in iPhones and MacBooks sold in China and is seeking White House approval to use them, as the AI-driven memory shortage that has already squeezed SanDisk, SK Hynix, and Samsung pushes device makers toward a supplier a bipartisan group of US senators explicitly asked Apple to avoid. It's the shortage reaching past the data center into consumer hardware, and the clearest sign yet that a device maker's memory sourcing has become a live US-China political test, not just a cost problem. (Semafor)
Quantum error correction: A new paper proposes OmniQEC, an AI system that discovers quantum error-correcting codes automatically, using large language models to coordinate code generation, syndrome-extraction design, and decoder evaluation as one iterative search loop instead of hand-designed code families. It reports codes competitive with or better than the best human-designed constructions across several noise models tested, evidence that AI-driven design search is becoming a legitimate tool for closing the gap between today's qubit counts and a fault-tolerant machine. (arXiv)