PP&A built an interview guide that followed the AI stack from end to end: how teams integrate AI with their general IT stack, whether they build models from scratch or adapt foundation models, how architecture and performance needs differ across training, tuning and inference, how they choose vendors, which design features they require, from cloud or on-premises hosting to GPUs, other accelerators and liquid cooling, and how they purchase AI infrastructure.

Between December 19 and 27, 2023, PP&A led five interviews with senior AI practitioners in the United States and Europe: the head of AI for drug discovery at a global pharmaceutical company, the head of autonomous operations and emerging technologies at a large industrial manufacturer, a data science manager in the AI center of excellence of a bank, the vice president of AI products at a building technology company, and the engineering leader who runs the AI platform at an enterprise software company. The team synthesized the transcripts into a structured interview summary that set the five perspectives side by side across eight themes, and delivered it to the client on December 28, 2023.

No single hosting pattern held. The enterprise software company ran all of its AI in the public cloud and had no plans to move on-premises. The other four kept or were growing their own footprint. The bank ran a hybrid model, with on-premises systems for compliance, regulatory and high-security workloads and the cloud for exploration and scale. The building technology company trained about 80 percent of its models in the cloud but expected to shift toward owned hardware as it scaled, because on-premises capacity costs less at scale and can be amortized. The industrial manufacturer expected customer demand for edge solutions to push it the same way.

AI-dedicated hardware remained a minority of infrastructure and spend, and views on its growth diverged. The pharmaceutical leader put it at about 20 percent of the corporate estate, and the bank at 30 to 40 percent with little change expected. The industrial manufacturer spent about 30 percent on specialized hardware and expected that share to flip to 70 percent within a few years. Generative AI raised the bar: models grew from 0.5-2 gigabytes to 65-100 gigabytes, training runs on top-end GPUs took weeks per iteration, and at the bank inference at scale needed more compute and storage than training.

Buyers put compatibility with existing IT, solution quality, ease of integration, technical support and cost transparency ahead of absolute price. The bank and the industrial manufacturer preferred to buy directly from vendors rather than through channel partners or managed service providers. The bank signed multi-year contracts after a selection process of about a year, and the building technology company planned on three- to five-year purchase cycles. The pharmaceutical leader warned that vendor support for AI hardware had shrunk from five years to two, forcing faster depreciation. The summary gave the client a practitioner-level map of these trade-offs to test its AI infrastructure product plans against.

• --