Find the Bottleneck Before Buying More AI

Ask the technology and operations owners to trace one important workload from request to usable result. Record elapsed time, waiting for data, computing time, correction work and total cost. Use an ordinary busy period as well as a quiet one. Keep failed requests in the record.

Then identify the dominant constraint. If the system is waiting for permission to retrieve documents, more processors may do little. If demand exceeds available computing capacity while data arrives promptly, additional capacity may be justified. These are diagnostic examples, not claims about how most businesses operate.

Finance should set the comparison: the current system, an improvement to its data or serving setup, and added capacity. Hold the task and quality requirement constant. A cheaper answer that creates more repair work is not necessarily a cheaper service.

At NetApp INSIGHT, the NVIDIA programme includes sessions on shared model memory and access to enterprise data. CoreWeave’s agenda covers capacity choices, storage and operational monitoring. Ask which part of your measured delay each offer would remove, what integration it requires and which charges remain outside the quoted price. ¹²

Bring the same workload to competing discussions. Request the hardware, model, workload mix, concurrency and quality conditions behind any improvement claim. A faster internal operation is not the same as a faster customer journey. Ask whether the result includes data movement, failed attempts and human correction.

The AI Conference’s San Francisco programme, with workshops on 29 September and conference sessions on 30 September–1 October, provides a parallel place to question evaluation and infrastructure assumptions. Use it to seek counter-evidence, not another collection of product demonstrations. ³

Two revised research preprints give a reason to test alternatives to buying more capacity. TIDE studies adapting a model-serving method to changing workloads, including selected non-English datasets. NearOracleKVSelection studies reducing the model memory retained during generation. Both report results under specified experimental conditions; neither establishes savings for your organisation. ⁵⁶

A new speech-recognition preprint supplies a caution. In experiments on Whisper models, some compression methods worsened error differences between groups, while distillation sometimes narrowed them. Its estimate of correction time uses an assumed time per error, not observed employee productivity. Lower computing requirements therefore need a separate quality and repair-cost check. ⁷

The practical direction is promising: better use of existing systems may release capacity. What remains unproven is the net benefit for your mix of tasks, languages, traffic and service commitments. Do not convert a research speedup into a budget saving without that bridge.

A US vendor event is not evidence of worldwide readiness. Connectivity, available cloud regions, power reliability, data permissions and language coverage can change the preferred investment. A central team should compare those constraints explicitly for each operating market rather than impose one infrastructure answer.

Zurich’s AI Policy Summit on 29–30 September and AI+X Summit on 1 October offer a different institutional perspective within the Zurich AI Festival. Conference discussion is not enacted law or regulatory clearance. Use policy specialists to identify the applicable obligations; use operating evidence to assess performance. ⁴

For teams in Africa, the Middle East, Latin America, Asia-Pacific, Europe and the United Kingdom, the missing evidence is often organisation-specific. The studies cited here do not support a regional ranking of cost or capability.

Close the week with a short decision record: the measured constraint, the alternatives considered, the remaining uncertainty and the next commitment. Buy where a comparable test supports the case. Fund a bounded trial where the opportunity is credible but unproven. Defer where the proposed purchase does not address the constraint.

The Monday and Friday dates are Institute planning suggestions, not external deadlines. The objective is a better allocation of capital, not a purchase by the end of conference week.