Reflection AI released Beam, a 501B-parameter mixture-of-experts open-weight model with 23B active parameters, pretrained on 23.8T tokens, and says it needs 3x to 4x less inference compute than leading Western open models, with weights due this month. Behind it sit two summer compute deals worth more than $7B for Nvidia GB300 capacity through 2029: SpaceX at $150M a month from July 1 at Colossus 2, about $6.3B in total and cancellable by either side on 90 days' notice after the first three months, and more than $1B from Nebius. Reflection has no data centers of its own, so its pace on the larger model it is already training depends on lease terms with SpaceX and a neocloud rather than on owned capacity. (TechCrunch)
Private Companies
Firmusfundraise
Firmus priced its ASX IPO at A$11.00 a share to raise A$7.1B ($4.94B), valuing the equity at about $30.6B, with an over-allotment option that lifts the raise to as much as $5.5B. Roughly $30B of debt puts enterprise value near $60B, and about half the shares go to existing holders including Nvidia, Coatue, Blackstone and Jane Street. Book building closes this week and trading opens October 23, on the second-largest IPO in Australian history. (Reuters)
Vincifundraise
Vinci raised a $250M Series B at a $1.5B valuation led by Advent, Temasek and Xora Innovation, ten months after its $46M seed and Series A. Its physics foundation model simulates heat and package warpage for chip design, and the company says it runs up to 1,000x faster than conventional solvers. Vinci plans to grow pilots from two customers to twenty, putting thermal sign-off for dense AI packages in competition with Cadence and Synopsys. (Vinci)
DeepSeekfundraise
DeepSeek is close to raising at least 80B yuan, about $12B, in a round led by Tencent and battery maker CATL, Bloomberg reported. The raise could reach 100B yuan, up from an initial 50B yuan target after demand rose following the company's latest model launch. An earlier 2026 round raised about 50B yuan at a $52B to $59B valuation, and the company has not announced this one. A cloud operator and an energy-storage manufacturer now anchor funding for China's largest open-model lab. (Bloomberg)
Public Markets
Applied Materials and Intel will co-develop transistor, interconnect and Foveros 3D-stacking packaging technology for AI chips at Applied's new EPIC Center and Intel's Hillsboro campus, with Intel a founding partner. No financial terms were disclosed. (Applied Materials)
Schneider agreed to buy PTC for $205 a share in cash, a $22.6B equity value, financed with up to €17B of debt and €6B of new shares. Product-design software moves inside a leading data-center power and cooling supplier. (Euronews)
Emerging
Flash as a memory tier: A UC Berkeley paper, Characterizing High Bandwidth Flash for LLM Serving, models an HBM, high-bandwidth flash and host-memory hierarchy with buffered, cache-aware scheduling. It cuts completion time 36% to 88% against HBM-only systems and energy use by up to 59%, and the scheduler stretches flash write lifetime from 1.21 to 14.82 years in the test configuration. The scheduler's endurance gain is what makes flash a plausible capacity tier for inference, which ties memory vendors' high-bandwidth-flash roadmaps to serving-software design. (arXiv)