Modal chief executive Erik Bernhardsson used a Latent Space interview to explain how compute for AI workloads is bought, sold and scheduled. The discussion centres on elastic GPUs, the idea that demand for graphics processors comes in blocks of a given number of chips for a given length of time, which is why capacity is hard to match to real usage. The conversation also covers the rapid growth of sandboxes, isolated environments where AI agents run code and tasks. Agent workloads tend to be bursty and unpredictable, which Bernhardsson argues puts new pressure on how cloud providers price and allocate capacity. The episode frames these shifts as the early economics of an emerging "agent cloud" rather than a settled market. The show notes are brief, so specific figures and vendor comparisons should be checked against the full article. For Australian businesses and professionals, this changes little on its own. It matters if your team runs AI models or agents on GPU cloud platforms, where per-second or elastic pricing could reduce idle spend but makes costs harder to forecast. If personal information is sent to offshore compute or sandbox services, the Privacy Act 1988 and its cross-border disclosure rules (APP 8) still apply, so check where data is processed before adopting any such service. A small, low-risk test of pricing and performance is worth doing before committing budget.