TL;DR
On October 27, 2026, Cerebras and Oracle announced wafer-scale inference coming to OCI in 2027: published tests show CS-4 delivering 4,400+ tokens/sec/user on GPT-OSS-120B, up to 30x faster than GPU solutions, and independent benchmarks put CS-3 at 2,700+ tokens/sec/user on gpt-oss-120B versus 900 tokens/sec on NVIDIA's B200.
Comparison Overview
On October 27, 2026, Cerebras and Oracle announced that Cerebras high-speed inference will launch on OCI in calendar 2027, targeting AI labs, enterprises and developers. Cerebras's Wafer-Scale Engine (WSE) integrates compute and memory on a single wafer, reducing data movement and inter-die communication, and targets long inference sequences and multi-step agentic workflows. This article uses Cerebras's published CS-4 and CS-3 measurements plus Artificial Analysis independent benchmarks to quantify the tokens-per-user speed gap between wafer-scale chips and mainstream GPUs (NVIDIA B200).
Comparison Table
| Dimension | Cerebras WSE (CS-4 / CS-3) | NVIDIA GPU (B200 etc.) |
|---|---|---|
| Chip architecture | Wafer-Scale Engine: compute and memory integrated on a single wafer, cutting data movement and inter-die communication; weights stay in on-chip SRAM (WSE-3/3T) | Mainstream GPU (B200), the Artificial Analysis benchmark reference; architecture parameters not listed within this research scope |
| Benchmark model | GPT-OSS-120B (CS-4); gpt-oss-120B (CS-3); Llama 4 Maverick 400B (CS-3) | Head-to-head measurements on the same models |
| Tokens/sec per user | CS-4: 4,400+ tokens/sec/user on GPT-OSS-120B with identical prompts; CS-3: 2,700+ tokens/sec/user on gpt-oss-120B, 2,522 tokens/sec on Llama 4 Maverick (400B) | B200: 900 tokens/sec on gpt-oss-120B (580 at 10-way concurrency); 1,038 tokens/sec on Llama 4 Maverick (400B) |
| GPU compared against | Serves as the baseline reference | NVIDIA B200 (Artificial Analysis independent benchmark) |
| Speed advantage | CS-4 up to 30x faster than GPU solutions (Cerebras release); CS-3 about 2.4x ahead on Llama 4 Maverick 400B | Baseline |
| OCI availability | Cerebras high-speed inference slated to launch on OCI in calendar 2027; OCI already has Dedicated AI Clusters inference foundation | OCI GPU inference details not covered in this research |
| Typical workloads | Long inference sequences and multi-step agentic workflows; agentic workflows run at effective batch size 1, where SRAM-resident weights (WSE-3/3T) win on decode latency | General-purpose GPU inference |
Evidence and Positions
Cerebras's official line: CS-4 delivers 4,400+ tokens/sec/user on GPT-OSS-120B with identical prompts, up to 30x faster than GPU-based solutions (Cerebras release, August 19, 2026); Cerebras high-speed inference is scheduled to launch on OCI in calendar 2027, on top of OCI's existing Dedicated AI Clusters; the WSE integrates compute and memory on a single wafer for long inference sequences and multi-step agentic workflows (Cerebras release, October 27, 2026).
Independent benchmarks: per Artificial Analysis (as cited in Cerebras's blog), CS-3 reaches 2,700+ tokens/sec/user on gpt-oss-120B while NVIDIA's B200 manages only 900 tokens/sec (580 at 10-way concurrency); per 36Kr's citation of Artificial Analysis, CS-3 hits 2,522 tokens/sec on Llama 4 Maverick (400B) versus B200's 1,038 — about a 2.4x lead.
Paper and disclosures: per an arXiv paper (October 2026), agentic workflows run at an effective batch size of 1, and architectures with SRAM-resident weights (WSE-3/3T) win on decode latency; Cerebras's existing major customers include OpenAI, G42, MBZUAI and AWS (Cerebras 10-Q risk disclosure).
Interpreting the Disagreement
First, vendor self-test vs independent benchmarks: CS-4's 4,400+ tokens/sec and the 30x ceiling come from Cerebras's own release (August 19, 2026), a vendor self-test figure; the CS-3 numbers (2,700+ tokens/sec/user, 2,522 tokens/sec) and the B200 comparisons come from Artificial Analysis independent benchmarks (cited via Cerebras's blog and 36Kr). Measurement conditions, concurrency settings and model versions are not fully aligned; the 30x ceiling is a peak under a specific prompt-and-model combination, not a universal gap across scenarios.
Second, the B200 figure is concurrency-sensitive: on gpt-oss-120B, B200 goes from 900 tokens/sec single-stream to 580 tokens/sec at 10-way concurrency, so comparisons must state the concurrency basis.
Third, OCI delivery has not happened: wafer-scale inference reaches OCI only in calendar 2027; all current numbers come from published measurements on Cerebras's own clusters. Actual speed, capacity and pricing after OCI launch are not disclosed.
Data Sources
- Cerebras press release - Cerebras and Oracle team up to bring Cerebras AI inference to OCI
- Cerebras CS-4 release - Up to 30 times faster than GPU-based solutions
- Cerebras blog - Cerebras CS-3 vs NVIDIA DGX B200 Blackwell
- 36Kr - Report citing Artificial Analysis inference benchmarks
- arXiv paper - Decode latency analysis for agentic workflows (October 2026)