What Mattered
Only two sources this week, so treat these as two separate signals, not a consensus.
SemiAnalysis forecasts roughly +50GW of global datacenter capacity in 2027. That is about equal to the entire cumulative global market through 2024. The evidence goes beyond a model output. More than 24GW across about 40 campuses is already in advanced construction for 2027 delivery. Slightly under $1T in binding, non-cancellable 15 to 20 year leases has been signed in 2026 to date, and most of the associated debt is already raised. Leases come before debt, and debt comes before construction, so most of 2027 is already locked in. Prefabricated designs from Google, Meta, AWS and DayOne cut build times below 10 months and MEP labor by up to 63%. SemiAnalysis treats power as broadly solved, because on-site generation skips 5 to 10 year grid queues. Their call is that the binding constraint moves to memory and logic in 2027 and 2028.
Nate B. Jones makes a narrower argument with no evidence behind it. As AI collapses the cost of producing artifacts, output volume stops signaling value. The thing worth optimizing is the outcome for a named person, and shipping is a midpoint, not the finish.
Why It Matters
The two pieces meet at one point. If SemiAnalysis is even half right, inference gets materially cheaper and new models arrive faster through 2028. Cheap tokens make artifacts nearly free to produce, and that is exactly when Jones's warning applies: teams with agents that measure PRs and features will produce unused software at scale.
For infrastructure planning, the bottleneck shifting from power to silicon matters more than the headline GW figure. Capacity contracted in 2026 should push 2027 and 2028 pricing in buyers' favor. Even so, scarce HBM and leading-edge logic may limit GPU allocation however much floor space exists. Interest rates that rise to fund the capex also make a client's own on-prem build more expensive compared with renting.
Discount both sources. SemiAnalysis sells the model these numbers come from, its site data is proprietary, and it raises the oversupply question without answering it. Jones offers a heuristic from his own observation, and he does not address the scope and accountability costs of working past role boundaries.
Next Move
1. Test client architectures against 2027 inference prices well below today's. Avoid multiyear commitments priced at current rates. Favor short terms and portability across providers. 2. Plan around accelerator and memory supply as the 2027 to 2028 risk, not datacenter space. Ask vendors about allocation, not capacity. 3. Rebuild AI program scorecards around adoption, task completion and gains in end user capability. Stop using PRs merged and features deployed as primary metrics. 4. Watch for confirmation in hyperscaler capex guidance, lease announcements and HBM pricing. Two sources give a hypothesis, not a trend.