Enterprises do not simply need more AI capacity. They need a reliable way to give teams and applications access to the right model, control who can use it and understand what each successful task costs.
GPU-hours and tokens are important units, but neither is a business outcome. A cheap GPU can sit idle. A cheap model can create retries, exceptions and additional human review. The real economics of enterprise AI emerge only when infrastructure, model performance, access controls and workload quality are measured together.
GPU-hours and tokens are inputs, not outcomes
Demand for model output is expanding quickly. One national market illustrates the scale: China’s National Data Administration reported daily token usage rising from 100 billion in early 2024 to 100 trillion at the end of 2025 and more than 140 trillion in March 2026. These figures are not a proxy for global demand, but they show how quickly inference can move into products and workflows.
Trillions of tokens per day. Source: China National Data Administration, reported March 2026. National figures, not global totals.
Growth does not make every unit interchangeable. A GPU-hour varies by accelerator, network, region and commercial term. A token varies by model, tokenizer, context, caching and whether it is input or output. Selected published H100 prices from Crusoe, Lambda, Nebius and CoreWeave ranged from roughly US$3.85 to US$6.16 per GPU-hour in September 2026, while some major-cloud on-demand configurations were materially higher.
The useful question is therefore not “What is the cheapest token?” It is “What does it cost to complete this task at the quality, latency and risk level the business requires?”
Scroll horizontally to compare all columns →
| Buying lens | Useful measure | What it does not tell you |
|---|---|---|
| Infrastructure | Cost per available or committed GPU-hour. | How much useful model work the capacity will complete. |
| Model access | Cost per million input and output tokens for a named model. | Whether responses meet the required quality and latency. |
| Business workflow | Total cost per successfully completed task, including retries and human review. | Whether the wider process change will be adopted and sustained. |
Cooma decision framework. Lower infrastructure or token prices do not necessarily produce a lower cost per successful task.
Model choice is becoming an operational decision
Frontier models can provide fast access to advanced capabilities without requiring a buyer to operate the underlying infrastructure. Open-weight models create another path: greater control over deployment, the ability to optimise for a specific workload and less dependence on a single provider’s interface or commercial model.
Neither route is universally better. A coding assistant, document workflow, customer-facing agent and restricted-data application can justify different models and hosting arrangements. The right choice can also change as model quality, prices and workload volumes change.
Commercial terms matter as much as model selection. Pay-as-you-go access, subscriptions, reserved infrastructure and long-term capacity commitments transfer different risks. CoreWeave, for example, reported that committed contracts accounted for more than 98% of its 2025 revenue and that customers generally purchased capacity on a take-or-pay basis. Read the filing.
An organisation can avoid owning GPUs while still accepting substantial demand, supplier and concentration risk. Portability and clear usage attribution should therefore be designed into the operating model, not added after a provider decision has been made.
Open-weight models change the economics
With open-weight models, software efficiency becomes part of the infrastructure strategy. Model selection, quantisation, batching, caching and serving configuration can change how much useful output the same underlying compute produces.
For selected open-weight models, Cooma applies proprietary serving optimisation to improve how efficiently compute is converted into useful model output. The objective is not to claim that one model is always cheapest. It is to find the best fit for a defined workload and make its economics measurable.
This matters because headline token prices can hide operational cost. A cheaper model that produces more exceptions may increase the cost of processing an invoice. A faster model may create little value if a downstream approval queue remains the bottleneck. A valid comparison uses the same workload, quality threshold and service level.
Enterprises need a control layer
As teams adopt more models, direct provider accounts create operational sprawl. Credentials become difficult to rotate, usage sits in separate consoles and policy changes must be repeated across tools. Model choice becomes coupled to whichever integration was completed first.
Cooma TokenFactory provides one governed access layer for selected frontier and open-weight models. Applications and users connect with Cooma-issued keys while provider credentials remain separate.
- One familiar API. Connect applications, coding tools and agents through supported Responses and Messages APIs.
- Access by project. Choose which models each key can use, then expire, suspend or revoke access centrally.
- Usage boundaries. Set daily and monthly token limits appropriate to each person, application or agent.
- Attributable consumption. Review token usage and recorded charges by key instead of reconciling disconnected provider accounts.
The purpose is practical: make model choice easier to change while keeping access, usage and cost understandable as adoption grows. Explore Cooma TokenFactory.
What enterprise buyers should do next
- Start with the workload. Separate experiments, predictable production demand, latency-sensitive services and restricted-data workflows.
- Benchmark complete tasks. Compare models using representative inputs, an agreed quality threshold and the total cost of retries and human review.
- Keep model choice portable. Use stable interfaces and preserve the evaluations, prompts and operational records needed to change providers or models.
- Put limits close to access. Attach model permissions and usage boundaries to the key used by each person, application or agent.
- Review the economics regularly. Reassess model quality, token prices, serving efficiency and capacity commitments as the technology changes.
The opportunity is not simply to secure more compute or buy more tokens. It is to build an AI capability that can choose the right model, control its use and connect cost to useful work. That is how AI infrastructure becomes business advantage.
Cooma TokenFactory
Put AI to work. Keep control.
Talk to Cooma about the models, access controls and usage limits your business needs.