FRI 16 OCT 2026 · 2 MIN READ
A thin Friday: TSMC says AI demand is strong, and a Korean neocloud lays out a 200,000-GPU plan.
TSMC guided Q4 revenue to $52.3-54.3B. Vessl AI targets 200,000 GPUs by 2030. NVIDIA released two open Nemotron models. Nothing new in power or buyer signals inside the window.
General AI Infrastructure
TSMC guides Q4 revenue to $52.3-54.3B
TakeawayTSMC reported third-quarter 2026 results and guided higher for the fourth.
- Q3 net revenue: US$46.72 billion; gross margin 66.4%.
- Q4 guidance: revenue US$52.3-54.3 billion, gross margin 65.0-67.0%.
- Q4 operating margin guided to 56.5-58.5%.
AngleThe guide implies roughly 12-16% sequential growth (computed from $46.72B against the $52.3-54.3B range). Leading-edge foundry output still sets the pace for accelerator supply, and the guided margin dip is the number to watch.
Neocloud & GPU Cloud Moves
Vessl AI plans 200,000 GPUs by 2030
TakeawayKorean GPU cloud Vessl AI laid out an expansion plan at a Seoul briefing, per Seoul Economic Daily.
- Targets 50,000 GPUs next year and 200,000 by 2030; about 7,000 GPUs today via partners.
- Eight AI data centers by 2028, 50MW combined; Vera Rubin cluster in the first half of 2027.
- Seeking 400 billion won in GPU financing; 2026 revenue of 250 billion won expected.
AngleThese are company targets, not signed financings. The company says 30,000 GPUs of demand is confirmed, 70-80% of it from overseas, and that GPUs sell out within one to two weeks of being secured.
Model Releases & Compute-Demand Drivers
NVIDIA releases Nemotron 3.5 Super VL and Bolt
TakeawayNVIDIA expanded its open Nemotron family with a 120-billion-parameter multimodal model and local-first Bolt models.
- Super VL was trained on about 26 trillion tokens; NVIDIA claims up to 5x speed over the next best model in its class.
- Bolt 74B-A9B uses 2-bit quantization to fit in 24GB for RTX and DGX Spark.
- A 4GB Bolt E4B model is due October 23.
AngleThe speed and ranking claims are NVIDIA's own. The product direction is the signal: capable agentic models sized for local hardware shift some inference demand off rented GPU hours.
How this is made. Compiled each weekday morning by the Rack Rate desk's research agents from primary sources, with every item linked to the document it came from, and reviewed before it publishes. The Angle on each story is the desk's own read; it does not recommend suppliers, and Rack Rate's own data is labelled as such. Corrections: daily@rackrate.ai.