There is no evidence-backed universal winner. Cerebras publishes high per-user generation speeds for selected models and API prices for named models; NVIDIA publishes Blackwell cost-per-token results tied to particular software and benchmark configurations. Those measures are not directly interchangeable. The right choice depends on your model, latency target, usage pattern, and whether you want managed API access or to run an inference stack yourself.
What the published figures actually compare
The available numbers answer different questions. Cerebras’s head-to-head table reports output speed for selected models, while its pricing page lists developer API rates and approximate speeds for particular models. NVIDIA’s published Blackwell figures are benchmark costs for named hardware and software setups. A per-user speed, an API charge, and a benchmark infrastructure cost describe different aspects of serving—not a matched comparison of what every buyer will pay or experience.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
ASUS Dual GeForce RTX 5060 Ti 16GB GDDR7 OC Edition Gaming Graphics Card | $786.37 | Buy on Amazon |
| 2 |
|
ASUS TUF Gaming GeForce RTX™ 5080 16GB GDDR7 OC Edition Graphics Card | $1,831.31 | Buy on Amazon |
For that reason, treat the figures below as dated, configuration-specific reference points. They do not establish that all Cerebras systems outperform all NVIDIA GPUs, or that one provider is cheaper for your production workload.
How fast are Cerebras and GPU systems on the reported models?
The following output-speed figures appear in Cerebras’s Form S-1/A, filed May 4, 2026. The filing attributes the comparison to Cerebras internal measurements and an Artificial Analysis benchmark published April 14, 2026. The table identifies the figures as tokens per second; they belong to that comparison’s models and test conditions, not to every NVIDIA deployment or an independently matched test of all current systems.
#1 Best Overall
- AI Performance: 767 AI TOPS
- OC mode: 2632 MHz (OC mode)/ 2602 MHz (Default mode)
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Axial-tech fan design features a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- A 2.5-slot design maximizes compatibility and cooling efficiency for superior performance in small chassis
| Model | GPU output speed (tokens/s) | Cerebras output speed (tokens/s) |
|---|---|---|
| Qwen-3 235B | 262 | 873 |
| MiniMax M2.5 | 223 | 1,039 |
| GLM 4.7 | 245 | 1,164 |
| OpenAI GPT-OSS-120B | 795 | 1,735 |
| Llama-3.3 70B | 164 | 2,457 |
In this specific set, Cerebras reports higher output-speed figures for each listed model. That is useful evidence about those tests, but it is not enough to predict the speed an application will deliver: model version, serving configuration, prompt and response lengths, concurrency, and latency requirements all affect the result. Read the dated Cerebras S-1/A comparison.
A newer CS-4 claim is a separate comparison
In an August 18, 2026 announcement, Cerebras said CS-4 delivers more than 4,400 tokens per second per user on GPT-OSS-120B using identical prompts, and claimed up to 30 times the speed of GPU solutions. Those are vendor claims about CS-4; they should not be combined with the S-1/A table as though the model, GPU baseline, and test conditions were the same.
The announcement also quotes Dylan Patel, founder and CEO of SemiAnalysis: “CS-4 makes dramatic improvements in system deployability, reliability, and networking, which enables scaling performance to larger models for large scale token factories.” This is a quote hosted in the vendor’s announcement, not an independent benchmark finding. See Cerebras’s CS-4 announcement.
What do the published cost figures mean?
Cerebras’s developer pricing is an API rate charged per million input or output tokens for named models. NVIDIA’s figures are reported infrastructure benchmark costs per million tokens for particular Blackwell inference configurations. They are different cost bases: do not compare them as if both were a provider’s all-in customer price.
| Published figure | What it covers | Important qualification |
|---|---|---|
| GPT OSS 120B: approximately 3,000 tokens/s; $0.35 per million input tokens and $0.75 per million output tokens | Cerebras developer API pricing and listed approximate model speed | Pricing page accessed October 7, 2026; the page says performance varies by model and configuration. |
| Qwen 3.8 27B: approximately 1,850 tokens/s; $0.99 per million input tokens and $1.49 per million output tokens | Cerebras developer API pricing and listed approximate model speed | Pricing page accessed October 7, 2026; the page says performance varies by model and configuration. |
| NVIDIA B200 with GPT-OSS-120B: $0.02 per million tokens; $0.11 per million at launch | NVIDIA’s report of SemiAnalysis InferenceX benchmark cost using TensorRT-LLM | NVIDIA describes a fivefold improvement through software optimization. These are benchmark infrastructure figures, not API rates or a buyer’s complete total cost of ownership. |
| NVIDIA GB300 NVL72: $0.123 per million tokens at 116 tokens/s per user | NVIDIA’s report of SemiAnalysis InferenceX benchmark results using NVIDIA Dynamo and TensorRT-LLM | The reported cost is tied to the named stack and the stated per-user speed; it is not a general GB300 price or guaranteed production cost. |
The Cerebras rates and model speeds are listed on its inference pricing page. NVIDIA’s Blackwell and GB300 cost figures are presented on its performance benchmarking page, which attributes them to SemiAnalysis InferenceX. Software is part of those benchmark configurations: serving software can materially change performance and economics without changing the accelerator hardware.
Why an API rate and a GPU benchmark cost are not like-for-like
An API rate tells you what the listed service charges for metered input and output tokens. An infrastructure benchmark cost per token is a benchmark result for a particular deployed configuration; it is not necessarily the price offered to a customer or the full cost of operating that deployment.
Rank #2
- Powered by the NVIDIA Blackwell architecture and DLSS 4. System Requirements: Minimum 850W PSU with 16-pin 12V-2x6 (12VHPWR) connector required. Verify before purchasing.
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability. Compatibility: 348mm (13.7") length, 3.6 slots, 4.3 lbs. Confirm case clearance and slot spacing. GPU bracket included.
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.6-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
For an infrastructure estimate, a buyer may need to account for matters the benchmark figure does not establish for their case, such as utilization, capacity, operations, and commercial terms. For an API estimate, use the actual input and output volumes because the listed rates differ by token type. Compare costs only after the service boundary is clear: what is included, who operates the system, and which usage assumptions apply.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to make a fair comparison for your workload
Ask providers to measure the same serving job, rather than relying on a headline tokens-per-second number or a benchmark cost in isolation. Cerebras itself cautions that speed comparisons vary by workload, configuration, date, and model.
- Match the model and precision. Confirm the exact model version and precision on each platform. A result for one model is not a proxy for another.
- Match the request shape. Use the same input prompt length and generated output length; token mix affects both latency and metered API charges.
- Set concurrency and a latency target. State how many requests run at once and the service-level target the application must meet.
- Measure interactive speed, not just aggregate output. Record time-to-first-token and per-user decode speed, as well as aggregate throughput at the required latency target.
- Use a consistent cost basis. Separate metered API charges from amortized infrastructure cost, and document what each estimate includes.
- Check deployment realities. Verify capacity, availability, geography, applicable terms, and who is responsible for operating the serving software and hardware.
This makes it possible to identify the actual trade-off: for example, a system may provide higher per-user output speed in a test, while another configuration may have a lower reported benchmark cost. Neither observation alone answers whether your service will meet its latency target at an acceptable all-in cost.
Managed API or infrastructure you operate?
Cerebras access options
Cerebras distinguishes developer access for exploration from enterprise production capacity, for which it lists quote-based pricing. Its pricing page also lists AWS Marketplace, OpenRouter, Hugging Face, and Vercel as access partners. These are distribution paths; the page says features, models, capacity, and performance depend on availability and applicable terms. Confirm the specific service, capacity, and commercial arrangement you would use rather than assuming every model or feature is available through every path.
NVIDIA benchmark context
The cited NVIDIA results concern Blackwell GPU inference with TensorRT-LLM; the GB300 NVL72 result additionally names NVIDIA Dynamo. The benchmark figures therefore describe hardware together with its software stack, not a hardware-only property. A buyer evaluating managed GPU capacity or a self-operated deployment should obtain a cost and performance estimate for the actual provider, configuration, and operating assumptions.
Which platform should you choose?
Choose based on the evidence that matches your deployment, not on a single headline. Cerebras’s dated head-to-head figures provide model-specific evidence of higher reported output speed in that selected comparison; its API page also gives concrete input and output rates for named models. NVIDIA’s cited Blackwell results provide specific benchmark cost figures with named model and software configurations. None of these sources supplies a single matched, independently audited comparison of current all-in production costs across both providers.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchQuick Recap
- If you need managed inference, compare the actual API or service charges, model availability, capacity, and latency at your expected traffic.
- If you plan to operate infrastructure, compare the exact hardware and software configuration at the required concurrency and service level, including utilization and operating costs.
- If interactive response speed matters, require time-to-first-token and per-user output measurements in addition to aggregate throughput.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
