Cerebras SystemsActive
Rack-scale AI system on the new Cerebras Nexus platform, built from three WSE-3 Turbo processors in modular liquid-cooled compute backpacks. Cerebras says CS-4 delivers up to 30x faster inference than GPU systems, up to 10x more throughput per watt than CS-3, 750 PFLOPs of AI compute, 7.2 Tbps of I/O, 129.6 PB/s of memory bandwidth, and wafer-to-wafer latency as low as 2 microseconds for models above 50 trillion parameters.
Latest update Sep 18, 2026
01
Unveiled August 18, 2026. First customer shipments targeted for Q3 2026. Nexus is also the rack for CS-5, aimed at 2027.
02
03
Announced next steps from Cerebras Systems that mention CS-4.
Q3 2026 earnings
The remaining IPO lock-up lasts until the earlier of the Q3 2026 earnings report or the 180-day outside date. That print is the first look at whether CS-4 shipments, OpenAI Ultrafast, and cloud mix are tracking the $880-890 million full-year core-revenue guide.
Q3 2026
Cerebras said on August 18, 2026 that first CS-4 shipments begin this quarter. Watch whether named customers take racks before September 30, 2026, and whether Nexus backpack deployments actually compress install time from days to hours as claimed.
After 30 Sep 2026
04
Cerebras and untracked cloud startup Gimlet Labs said on September 28 that Gimlet is a CS-4 launch partner for ultrafast inference, with a first Cerebras-powered Gimlet Cloud datacenter expected later in 2026 and a target of up to 3,000 tokens per second. Reuters, citing CEO Andrew Feldman, said Cerebras will supply CS-4 systems totaling about 100 megawatts over one to two years, with Gimlet making the hardware available in its cloud in 2027. The companies did not disclose financial terms.
05
Current production wafer-scale system, one WSE-3 per chassis, used for Cerebras Cloud, OpenAI Ultrafast, AWS disaggregated decode, and on-prem clusters. The 5nm WSE-3 is 46,225 mm2 with 4 trillion transistors, 900,000 AI cores, 44 GB of on-chip SRAM, and 21 PB/s of memory bandwidth, about 58 times the area of an Nvidia B200 die and thousands of times its on-chip memory bandwidth.
Current flagship wafer-scale processor inside CS-4. Same 4 trillion transistors, 900,000 cores, 46,225 mm2, and 44 GB of SRAM as WSE-3, clocked and packaged for roughly 2x the speed of the prior wafer, 250 PFLOPs of AI compute, and 43.2 PB/s of memory bandwidth per wafer. Redundant cores and fail-in-place routing are how Cerebras yields a dinner-plate-sized die.
Cerebras-operated cloud that serves open and dedicated models through an OpenAI-compatible API, plus a training cloud for fine-tunes and from-scratch runs. Production customers include OpenAI (GPT-5.6 Sol Ultrafast), Meta's Llama API, Mistral, Perplexity, Cognition, Lovable, CrowdStrike, AlphaSense, GSK, Block, and Figma. Partner distribution includes AWS Marketplace, OpenRouter, Hugging Face, Vercel, and IBM watsonx.
06
SpaceX
Partially reusable two-stage medium-lift launch vehicle. The world's most frequently launched orbital rocket.
SpaceX
Heavy-lift launch vehicle consisting of a strengthened Falcon 9 core with two additional boosters.
SpaceX
Reusable spacecraft for cargo and crew transport to the International Space Station.
SpaceX
Fully reusable super-heavy-lift launch system designed for missions to the Moon, Mars, and beyond. V3 (Block 3) variant with Raptor 3 engines is the production version.