On Monday, VCs open the week shifting AI bets toward power, efficiency, and retention

Venture firms began the week by steering new checks toward AI efficiency and energy constraints, according to investor priorities set in Monday partner meetings. The focus is moving away from foundation model training and into power, data pipelines, and application retention, as data center capacity and cost discipline define near term outcomes.

The timing matters because power, not chips alone, is now the gating factor for AI scale. Hyperscalers have locked long term power deals and capacity, leaving startups to win with efficient inference, predictable margins, and faster payback to enterprise buyers.

Why the timing matters

Grid bottlenecks and siting delays continue to slow new data center builds, pushing investors toward software that improves GPU utilization, model routing, and caching, along with companies that help developers monitor cost per query and quality. Capital is also flowing to startups that unlock or manage power, including demand response aggregation, transmission software, and tools to finance or optimize on site generation tied to AI workloads.

Rates remain higher than the last cycle’s zero cost era, which compresses risk budgets for long duration model bets. Open source models and commodity access to inference have narrowed defensibility at the base model layer, so investors want application level lock in through workflows, proprietary data, and retention, not model ownership.

What investors want now

Term sheets are clustering around clearer unit economics. Partners are asking for gross margins above 60 percent, net dollar retention at or above 120 percent by month 18, and payback within 12 months for midmarket accounts, with transparent evaluation frameworks rather than benchmark anecdotes.

Specialist climate and deep tech firms have dry powder for the grid and power side, while generalist funds are adding AI efficiency specialists to diligence. Startups pitching training centric stories are being pushed to show a path to profitable inference, customer hosted options, and resilience to model churn.

For the broader landscape, this tilt pulls early stage dollars toward the seams between energy, compute, and enterprise software. The winners are likely to be companies that treat power availability and cost as product features, and that can prove sticky workflows even as underlying models and hardware continue to change.