OpenAI is negotiating a new funding round that would value the company at as much as $1.5T, roughly double the $730B mark it set in February, after investors first approached with an offer closer to $1.2T. The gap between what investors are offering and what OpenAI believes it is worth is opening at the same moment the company is pushing its IPO to 2027 over what CEO Sam Altman calls unresolved AI-safety risk, which only sharpens the question of how much of a $1.5T valuation is really compute capacity paid for in advance. (Bloomberg)
Private Companies
Crusoecustomer
Crusoe signed a multi-year partnership under which Perplexity will run its full model lifecycle exclusively on Crusoe's infrastructure: training on dedicated NVIDIA GB300 NVL72 clusters and serving production traffic through Crusoe's Managed Inference service. The deal backs Crusoe's vertically integrated pitch — one AI-native cloud company handling both the compute-hungry training phase and the latency-sensitive inference phase for a fast-scaling model developer, rather than splitting the workload across specialized providers. (GlobeNewswire)
Public Markets
Shares rallied after reports that SK hynix is exploring a lease or joint venture to make memory chips at Intel's stalled Ohio campus, as Washington pushes chipmakers to add U.S. capacity amid a persistent HBM and DRAM supply crunch. (TechCrunch)
Meta$680.74▲ 1.1%Mkt Cap: $1.71T
Shares extended their rally as Bloomberg detailed testing results on Meta's third-generation Arke inference chip, built with Broadcom and made by TSMC, landing within 2-3% of simulated targets ahead of a first-half-2027 rollout aimed at cutting Nvidia dependence. (Bloomberg)
Emerging
Quantum's own foundry problem: Anderon, an IBM-backed pure-play quantum wafer foundry in Albany, New York, finalized a $1B CHIPS Act R&D award with the Commerce Department, matched by another $1B from IBM itself, with first wafers already running through the line. The award treats quantum-hardware manufacturing as a supply-chain bottleneck distinct from AI-chip fabrication, betting an independent foundry can serve the whole quantum industry the way TSMC serves classical chip design. (IBM Newsroom)
Memory in flash, not DRAM: A new paper, LLM Inference in a Flash!, runs large language model inference directly on flash-based compute-in-memory hardware, pairing integer-only quantization with a dictionary-compressed KV cache to cut dynamic cache traffic 15x on Llama-3.1-8B and Qwen-2.5-7B with limited accuracy loss. It joins this year's growing memory-wall literature — alongside Kepler's ferroelectric composite and TeRAM's 3D SRAM — searching for a cheaper substitute for the DRAM and HBM capacity AI inference keeps running out of. (arXiv)