When OpenAI’s chief financial officer speaks about return on investment, the industry listens. Sarah Friar has introduced a scorecard designed to grade the AI age: no longer just raw power or model parameters, but four concrete levers — useful work, cost per successful task, dependability, and return on compute.

The move is more than an accounting exercise. It signals the end of the phase where companies could afford to experiment without asking what each generated response actually cost. At its core is the concept of “useful work”: how many of the tasks an LLM performs achieve the intended goal without requiring manual corrections? In a factory, an office, or a contact center, the gap between a completed task and one that produces unusable output directly affects productivity. Friar shifts the yardstick from raw token throughput to tangible outcomes.

The second pillar, cost per successful task, is where infrastructure takes center stage. In the cloud, the math is often opaque because fees bundle storage, networking, and add-on services. On on-premise hardware, by contrast, the cost per inference is measurable with near-surgical precision: add up server amortization, energy consumption, cooling, and software licenses, and divide by the number of tasks that actually succeeded. This transparency allows teams to optimize GPU allocation, choose the quantization level that balances latency and quality, and decide whether a 70-billion-parameter model truly delivers a proportional advantage over a smaller one.

Dependability, the scorecard’s third point, hits a raw nerve in local deployments. A self-hosted model can suffer performance degradation if request loads spike without adequate orchestration, or if available VRAM cannot sustain extended context windows. Measuring real uptime, response consistency, and the ability to serve peak demand without degradation is critical for assessing total cost of ownership and, above all, for preventing AI from becoming a bottleneck instead of an accelerator.

The fourth piece, return on compute, closes the loop: every dollar spent on hardware must produce measurable output. In an era where GPUs are the most contested resource, the difference between a well-sized on-prem cluster and an underutilized one translates into margins that can determine the competitiveness of a product or an internal service.

Beyond the numbers: who wins and who loses

Friar’s scorecard is not just a tool for CFOs. It puts pressure on the entire AI supply chain. Hardware vendors will need to demonstrate not just teraflops, but efficiency in real-world scenarios: throughput per watt, production stability, ease of integration into continuous inference pipelines. Companies that invested in models without defining success metrics risk discovering that their projects have a negative ROI. Conversely, those that have already embraced a DevOps culture for ML — with monitoring, logging, and canary releases — can read that scorecard as validation and further refine their stacks.

For those evaluating on-premise deployments, the challenge is twofold: you must build an environment where these four dimensions are observable and optimizable. Without precise data on actual GPU utilization, cost per successful task remains an abstraction. Without dependability metrics, “useful work” crumbles when the system fails. This is where the convergence of observability tools, orchestration, and automation becomes the real differentiator.

Ultimately, the scorecard adds nothing that a good engineer doesn’t already know: you measure what matters. But having it come from the CFO of the company that symbolizes generative AI carries the weight of institutional endorsement. The next wave of investment will no longer settle for impressive demos. It will demand numbers — and those who can produce them starting from their own hardware will gain a negotiating and operational advantage that is hard to bridge.