NVIDIA Blackwell Sets $0.02 Per Million Token Floor for 120B LLM Inference

NVIDIA Publishes Concrete Cost Floor for Enterprise LLM Inference

NVIDIA's latest Blackwell B200 GPU benchmarks establish a verifiable cost floor for large-scale LLM inference: $0.02 per million tokens for a 120-billion-parameter model at production throughput. The company's total cost of ownership analysis shows inference providers using Blackwell can cut costs by up to 10× compared to Hopper-based deployments for identical workloads. For enterprise buyers, this number changes the math on build-versus-buy decisions and creates a pricing wedge against API vendors that haven't passed through hardware efficiency gains.

The benchmark targets GPT-OSS-120B, a reference model comparable to production-grade enterprise LLMs. NVIDIA reports throughput of approximately 60,000 tokens per second per GPU at this cost point. Early adopters named in the analysis — Baseten, DeepInfra, Fireworks AI, and Together AI — are already deploying Blackwell-backed inference at scale. The data gives enterprise buyers a concrete number to use in RFPs and contract negotiations when evaluating LLM API pricing.

Budget Implications for Enterprise LLM Deployments

At $0.02 per million tokens, a 50-billion-token monthly workload costs approximately $1,000 in raw inference at the hardware layer, excluding platform margin and networking overhead. Compare this to current SaaS LLM pricing: OpenAI GPT-4 Turbo charges $10 per million input tokens, Anthropic Claude 3.5 Sonnet charges $3, and Cohere Command R+ charges $3. The gap between hardware cost floor and retail API pricing now ranges from 50× to 500×, depending on the model.

This spread puts pressure on two groups. First, API vendors that don't control their own infrastructure or negotiate Blackwell access must explain why their per-token costs remain 10-50× above the published hardware baseline. Second, enterprises running large-scale inference on older GPU fleets (Hopper, Ampere) face a choice: migrate to Blackwell-backed platforms or accept a permanent cost disadvantage against competitors who do.

For procurement teams, the NVIDIA data supports specific contract requirements: disclose the underlying accelerator type, commit to passing through cost savings as hardware improves, or cap per-token pricing relative to independent benchmarks. Buyers can now treat $0.02 per million tokens as a realistic planning assumption for 120B-class models, not an aspirational target.

Inception Mercury 2.5 Tests Low-Price, High-Context Enterprise Model Market

Inception launched Mercury 2.5 on September 8, 2026, with a promotional input rate of $0.04 per million tokens — an 80% launch discount from a nominal $0.20 list price. The model delivers 1,107 tokens per second in production and supports a 260,000-token context window. Inception positions Mercury 2.5 as an enterprise-grade alternative to mid-price general-purpose LLMs, distributed via its own API, Baseten, and OpenRouter. The company is offering 100 million free tokens to new developers and supports dedicated capacity, autoscaling, and configurable data retention for enterprise deployments.

Mercury 2.5's pricing sits between NVIDIA's hardware cost floor and the retail pricing of established enterprise LLMs. At $0.04 per million tokens during the launch window, it undercuts Anthropic Claude and Cohere Command R+ by 75× while maintaining a margin above Blackwell's $0.02 baseline. The 260,000-token context targets use cases currently served by long-context models like Gemini 1.5 Pro and Claude 3.5 Sonnet.

The launch tests whether enterprises will adopt a newer, less-established model at a significant price discount versus paying incumbents for brand trust and proven reliability. Inception's distribution via Baseten — a Blackwell early adopter — means the model can theoretically run at or near NVIDIA's cost floor, leaving room for both platform and model provider margins. For buyers evaluating Mercury 2.5, the key questions are model quality relative to incumbents, data governance for proprietary workloads, and whether the launch pricing is sustainable or a customer-acquisition tactic.

Competitive Pressure on Inference Platforms and API Vendors

The combination of NVIDIA's published cost floor and Inception's aggressive pricing creates a squeeze. Inference platforms that secure Blackwell capacity gain a structural cost advantage over those relying on Hopper or non-NVIDIA accelerators. API vendors selling general-purpose LLMs at $1-10 per million tokens must either justify the premium with superior model quality or risk losing high-volume enterprise accounts to platforms that pass through Blackwell economics.

AMD MI300/MI400 and Intel/Google TPU deployments face the challenge of matching Blackwell's publicly cited performance and cost metrics without equivalent marketing visibility. For enterprise buyers, this means hardware-agnostic procurement strategies become more important: require vendors to commit contractually to tracking NVIDIA's TCO curve or demonstrate equivalent economics on alternative accelerators.

What to Watch

Track whether major API vendors (OpenAI, Anthropic, Cohere) reprice their models in response to Blackwell economics or defend current pricing with model differentiation arguments. Monitor whether Blackwell-backed inference platforms (Baseten, DeepInfra, Fireworks AI, Together AI) gain enterprise market share from incumbents that don't control their own infrastructure. Watch how AMD and Intel respond with competing benchmarks for their accelerators — without equivalent data, they cede the cost narrative to NVIDIA.

For procurement teams, use $0.02 per million tokens as a baseline when evaluating 120B-class LLM inference proposals. Require vendors to disclose accelerator type and commit to passing through cost improvements as hardware evolves. If a vendor's per-token pricing sits 20× or more above the hardware floor, demand a detailed explanation of where the margin goes — and whether it's defensible against competitors who operate closer to cost.