Making AI an asset, not an expense

Cheri Williams, in HPE-produced content carried by MIT Technology Review, argues that as AI moves from pilots to production portfolios—assistants, retrieval systems, agentic apps—consumption pricing can become a hard-to-forecast monthly line. Deloitte’s 2026 State of AI in the Enterprise is cited: worker AI access rose 5% in 2025, and the share of companies with at least 40% of AI projects in production is expected to double within six months.

The piece frames a workload-by-workload decision: when demand is steady and capacity stays productive, owning infrastructure can beat buying requests one at a time. Leaders should find their crossover utilization, then run an operating model that onboard workloads, governs use, and fills idle capacity—or ownership never pays.

We’re filing this as sponsored strategy, not newsroom investigation—and the core question is still fair. Token tabs feel fine in pilot season and nasty when agents loop all day.

Useful AI at scale needs honest unit economics. Retrieval-heavy and agentic workloads burn context differently than chat toys; generic benchmarks lie. The crossover-point framing is how CFOs should actually talk—provided they count energy, ops talent, and model freshness, not only sticker GPU prices.

I’m watching for bait-and-switch: “own your capacity” that quietly means buy one vendor’s stack. Ownership without an adoption machine is just a depreciation schedule. The three questions in the piece—steady demand, crossover math, productive utilization—are the right gate.

If enterprises get this wrong, they either drown in variable spend or strand capital on idle clusters. If they get it right, AI stops being a surprise OpEx line and becomes a planned productive asset. That is a win for useful deployment—even when the messenger is a hardware seller.

MIT Technology Review, September 29, 2026. Explicitly labeled as produced by HPE, not written by MIT Technology Review’s editorial staff.