Before we dive in… listen to today’s episode of The Compute 100 podcast!
Barun and Rajiv are multi-time founders in networking silicon. Between them they helped build Cavium, Innovium, and more before co-founding Auradine, where they incubated Upscale before spinning it out. Upscale builds the switch silicon, systems and software that connect AI chips to each other using open standards.
We dig into why AI clusters need a network built around memory reads and writes between chips, not the packet switching that cloud networks were designed for, and why Barun and Rajiv chose open standards over a proprietary fabric like NVLink. They explain how a networking startup works with hyperscalers on their custom chips, why it sells complete systems to new clouds and enterprises, and how the AI buildout could make networking a trillion dollar market. Give it a listen below!
SpaceX's third-quarter results put its AI business at the center of the company. The Super Intelligence segment, which holds Grok, the Cursor coding tool and a neocloud, brought in $7.56B of $14.6B in revenue, and AI took $16.4B of the quarter's $19.6B in capex. Nameplate compute reached 1.7 GW at September 30, up from 1.4 GW in June, and on the earnings call Musk said SpaceX is bringing roughly 200,000 Nvidia GB300s online a month through year-end, with next year primarily Vera Rubin. CFO Bret Johnsen guided fourth-quarter capex to $35B to $40B. Musk placed the constraint on deployment in power, cooling and infrastructure and said SpaceX is casting turbine components itself because turbine makers face 12 to 18 month delays on metals. (SpaceX)
Private Companies
Flapping Airplanesfundraise
Flapping Airplanes, a year-old AI research lab, raised $1B at a $6B valuation, Bloomberg reported. Index Ventures and Lux Capital led, with Sequoia Capital, GV and Kleiner Perkins participating. The valuation is four times the $1.5B set by its $180M seed round in January, and the company has no public product. It works on training models with less data and less compute, so the round funds research into reducing the compute each model needs. (Bloomberg)
Fireworksfundraise
Fireworks is in early talks to raise at a $30B to $40B valuation, about double the $17.5B set by its $1.5B Series D in July, Business Insider reported. The inference platform said in July that annualized revenue had passed $1B and that it serves more than 40 trillion tokens a day, so the reported range is 30 to 40 times that run rate. The terms could change and the company declined to comment. It is also managing an incident this month in which internal credentials were used without authorization to reach some customer accounts. (Business Insider)
Public Markets
Nvidia introduced DSX POD, a reference design for complete Vera Rubin AI factories that partners build and validate inside modular containers before shipping them to a prepared site. Nvidia says this cuts deployment at that site from six months to as little as 14 days, with units due worldwide in early 2027 from partners including Dell, HPE, Lenovo, Schneider Electric, Vertiv, Crusoe and Foxconn's Ingrasys. The release names no European site or customer and gives no capacity, pricing or baseline for the six-month figure. Integration of compute, cabling, rack-level power and liquid cooling moves from the construction site to the partner's factory floor, so the schedule depends on power and cooling at the site being ready first. (NVIDIA)
Emerging
Carbon-aware photonics: A paper from Bharat Bhatta and Ishan Thakkar, Green Light: Carbon-Aware Design of Photonic Transformer Accelerators at the Edge, counts manufacturing carbon alongside operating energy for microring-resonator transformer accelerators. Moving from analog to stochastic signaling raises die area 4.3x to 4.6x and embodied carbon 4.4x to 4.8x while cutting operational energy about 91%, so embodied carbon equals or exceeds operational carbon per inference. The fastest of the six organizations is not the most carbon-efficient, and the abstract reports no fabricated device. Edge inference designs will need to be judged on lifecycle carbon as well as energy per token. (arXiv)
Optical chiplet manufacturing: NTT, NTT Innovative Devices, Japan Semiconductor, Furukawa Electric and Shinko Electric Industries were selected under Japan's NEDO program to develop mass-production methods for photonics-electronics convergence devices. The target is more than a tenfold gain in manufacturing productivity over R&D-stage processes, plus a pilot production line; Japan Semiconductor covers silicon photonics, Furukawa the compound-semiconductor optical devices and Shinko the packaging. No budget was disclosed. The program funds the step between optical chiplet demonstrations and volume supply for AI data centers. (NTT)