The state of play of data center development

Sharp starts from the workload, because the workload dictates what infrastructure a site needs. Training a frontier model is the part that can travel: you feed it one large fixed data set, so the site mostly needs power, land, labor and water. That did force broader regional deployment, and he thinks that wave is leveling out. Inference is what he sees driving region-specific growth from here, and it lands inside the existing cloud availability zones, the clusters near major city centers built to hold uptime and performance commitments for enterprise customers. Kann offers the standard lay version of this, training anywhere and inference near users because latency matters, and then the more interesting speculation on top of it: split inference into latency-sensitive and latency-insensitive halves, and send the insensitive half out to cheap, clean, possibly intermittent power in the middle of nowhere.

Sharp grants that this works, then explains why it is the shrinking half. Latency is not really the binding property, because as long as latency is consistent most workloads tolerate it. Throughput is the constraint, meaning the sheer volume of data that has to move. Simple text-to-text inference can serve all of North America from two or three markets, but he calls that an early-innings picture. More advanced reasoning models do not generate one token against one prompt and stop; a token may pass through several models to check for hallucination or to route through a mixture of experts, so AI infrastructure increasingly wants to be near other AI infrastructure. Real-time inference also wants to be near the data it learns from, which sits in the tier-one markets. Chip economics push the same way, since GPUs are expensive and training is spiky, so operators want to backfill an installation with inference to raise utilization and return on invested capital. Kann’s summary, which Sharp endorses, is that going where the power is retains real validity while the relative share of what can be built that way dries up.

The consequence is that the constraint binds where it hurts most. Northern Virginia is a multi-gigawatt market at roughly 0.5% vacancy, and Sharp is careful about what is actually short there: not necessarily generation, but distribution. Concept to delivery for a versatile data center runs about 24 months, some interconnections now run longer than that, and utilities are asking for four-year forward load projections. Transformer lead times are past 50 weeks and gas turbines are backlogged beyond 2029. His two answers are unglamorous. One is vendor-managed inventory, buying switchgear and long-lead equipment ahead of need. The more important one is master planning with the utility over five years or more, which he frames as a mutual credit test: a utility will not overbuild unless it believes the load will show up and stay, so the developer’s job is to bring creditworthy customers who intend to occupy the asset for ten years or more, and then actually take the power down when it said it would.

Two intuitions get overturned along the way. If inference arrives in five-megawatt chunks, why not build a scatter of 120-megawatt sites, each easier to interconnect than one two-gigawatt campus? Sharp says customers really do think in five-megawatt blocks but want them contiguous, plugged into a single hundred-megawatt hall for operations and resiliency; the customer indifferent to location is the outlier. And like the grid, you have to size for the peak. The second is the long-standing hope that some workloads could skip backup power and be sited anywhere. Sharp wanted to build that facility and never saw one come to market with an acceptable service-level agreement, and liquid cooling has made it worse rather than better, because the heat does not displace itself and a stopped pump ruins the chips. Bridge power lands the same way. Running a data center off site generators until the interconnection arrives is an outlier; the common version is a negotiated deal with the utility, bringing your own generation or batteries under something like an interruptible tariff in exchange for faster power. His reason is strategic rather than technical. Becoming a generator is a short-term gap nobody would attempt absent the constraint, and he would rather invest in the utility doing the thing it is already good at.