Across industries, the narrative that rising compute expenses are the main barrier to AI adoption is being rewritten. Instead of wrestling with budgets, top engineers at companies such as Wonder and Recursion are prioritizing how fast they can ship models, how flexible the infrastructure is, and whether they can sustain high‑throughput workloads. These priorities stem from real‑world constraints: latency spikes, sudden capacity limits, and the need to iterate on models at breakneck speed.
Wonder, a cloud‑native food‑delivery giant, reports that AI adds only a few cents per order—2 to 3 cents today, climbing to 5–8 cents as demand grows. Yet the company’s real pain point is not the margin impact but the rapid exhaustion of cloud capacity. What began as an assumption of unlimited resources forced Wonder to shift to a second region within months, a move that surprised the team. Even more striking is the ambition to build individualized micro‑models for each user—a strategy that would slash personalization but is currently too expensive to scale. The company’s budgeting process is described as “art rather than science,” reflecting the unpredictable economics of token‑based AI services.
Recursion’s approach mirrors a hybrid strategy, blending on‑prem GPU clusters with cloud inference to meet diverse compute needs. Early on, the lack of robust cloud offerings pushed the biotech firm to build its own infrastructure, starting with gaming GPUs and moving to high‑performance A100s and H100s. Today, large foundation‑model training runs on a 200‑petabyte image repository on-prem, while lighter inference jobs are pre‑emptively scheduled in the cloud. From a cost perspective, on‑prem workloads can be ten times cheaper in the short term and halve the five‑year total cost of ownership. Recursion’s CTO warns that companies hesitant to commit to long‑term compute investments end up paying on demand and stifle innovation.
These stories underscore a broader industry shift: when AI is operating at scale, economics are no longer the decisive factor. Instead, the conversation has moved to how quickly and reliably models can be deployed, how flexible the infrastructure is, and whether capacity can keep pace with demand. Companies that embrace hybrid or multi‑regional strategies and invest in scalable, fast‑to‑market solutions are positioning themselves to outpace competitors who lock themselves into costly, static architectures.
Want the full story?
Read on VentureBeat →