Run Four Serving Benchmarks Before Treating Cerebras CS-4 as a GPU Alternative

The trade-off is raw speed against verified economics. Cerebras' launch benchmark for GPT-OSS-120B reported more than 4,400 tokens per second per user, a result characterized as up to 30 times faster than GPU solutions. Turn that claim into a purchase decision by dividing the per-user rate by your target concurrency, then comparing the resulting cost per million tokens with your current GPU fleet. That division is the test.

The launch number is a condition, not a production SLA

Cerebras announced the CS-4 rack-scale accelerator on Aug. 18, 2026, and said first shipments would begin that quarter. The system is built around three Wafer Scale Engine 3 Turbo processors. The launch specifications list 750 PFLOPs of AI compute, 129.6 petabytes per second of memory bandwidth, and 7.2 terabits per second of I/O. The launch footnote warns that throughput is not fixed; it moves with model architecture, context length, numeric precision, and serving configuration.

Run the four-number serving stress test

Score any accelerator, including CS-4, against these four production numbers before comparing it with your current stack. Keep the four numbers in one table so the cost line is visible.

• p50/p95 time-to-first-token (TTFT). Measure first-token latency at your real context length, not a short prompt.
• Sustained tokens per second per user and aggregate throughput. Run the same prompt set until the rate stops improving, record the stable value, and multiply it by concurrent users.
• Cost per million tokens at target concurrency. Price the same workload on the new hardware and your current fleet, including idle capacity and power.
• Model compatibility. Confirm quantization, context length, licensing, and serving configuration before trusting any speed ratio.

TTFT shows the user-visible delay. Sustained rate exposes hardware limits. Cost surfaces idle capacity and power. Compatibility gates the validity of any speed ratio.

Use the same prompt set, output length, and concurrency on both stacks. Record the stable rate, not the first sample, and note the concurrency at which it stabilizes. If the vendor supplies a demo, ask for the raw log, prompt distribution, and serving configuration.

Context length is where the benchmark decays. Short prompts can look clean, while long prompts change KV cache pressure and scheduler behavior. Test the longest context you will serve, not the shortest one that looks good.

Quantization changes the arithmetic. Lower-precision models can run faster, but your product may need higher precision for quality. If the vendor's number uses a different precision, the comparison is invalid.

Licensing is a common failure mode. Your model may run on the hardware but not under your production license, or only under terms that block your use case. Put the license terms in the same spreadsheet as the cost per million tokens.

Power and topology change the cost line

Cerebras said CS-4 would deliver up to 10 times more throughput per watt than the previous CS-3 system. That changes the power line, but not the invoice. Bursty workloads can show a good per-watt number while cost per million tokens stays high because utilization stays low.

Cerebras CEO Andrew Feldman said a disaggregated inference setup could provide inference 10 times faster than GPUs and 5 times more throughput than Cerebras alone. The topology changes the request path, so the test must name the exact serving configuration. Comparing a single rack to a GPU cluster is misleading unless the request path is the same.

Wafer-to-wafer latency as low as two microseconds is meant to support clusters for models with more than 50 trillion parameters. Smaller models may not show that advantage in your test. Larger models may change the serving configuration enough to invalidate a short benchmark.

Feldman said OpenAI's GPT-5.6 Sol Ultrafast mode reaches up to 14 times the speed of OpenAI's standard offering for select customers. Speed modes can be customer-specific, model-specific, and workload-specific. Name the exact model, mode, and customer tier in the test.