Microsoft is planning to more than triple its data center capacity over the next six years. According to a Bloomberg report published September 10 and covered the next day by Data Center Dynamics, the company operates roughly 12GW of data center capacity today and is targeting 38GW by 2032.
That figure comes from people Bloomberg described as familiar with the plans, and Microsoft declined to comment. The scale is visible in the build already finished. DCD reporter Georgia Butler noted that the company brought 88 data centers online in fiscal 2026, 31 of them in the fourth quarter alone, and that capital spending climbed from $55.7 billion in 2024 to $145.3 billion in 2026.
The buildout is no longer gated by demand or by money. It's gated by electricity, and the electricity isn't showing up on schedule. Texas has paused new large-load interconnections pending an audit of a 474-gigawatt queue, more than five times the roughly 90GW the grid pulled at its summer peak. That arithmetic is why the most interesting work in cloud right now is happening at the edge, where the goal is not more capacity but less waste.
The capacity math stopped adding up
The spending forecasts keep getting bigger. Goldman Sachs expects the major hyperscalers to spend about $581 billion in the U.S. this year on AI infrastructure, including data center construction. In the same analysis, the bank projects U.S. data center power demand to more than double from 2025 levels to 66GW in 2027.
Only about half of the data center capacity scheduled for the next one to two years is expected to arrive on time, Goldman Sachs said, citing delays and cancellations. The Electric Power Research Institute put data centers at 9 to 17 percent of U.S. electricity demand in 2030, and as much as 20 percent by 2035.
The gap between what's promised and what's wired up is now measured in billions of dollars. The energy advisory firm Relae, formerly Carbon Direct, reported in June that more than $170 billion in planned AI data center capacity has been blocked, withdrawn, or stalled by community opposition since January 2024.
Pablo Vegas, president and CEO of the Electric Reliability Council of Texas, said in an April statement that ERCOT believes its own forecast "to be higher than expected future load growth," pointing to unresolved questions about how large load gets identified and verified for long-range planning. That's a grid operator telling you the queue is padded with projects that will never draw a kilowatt.
What the edge router is doing instead
Cloudflare released its Auto Router in public beta on September 30, free while in beta, and the mechanism tells you a lot about where the industry's pressure point has landed. Developers set their model to cloudflare/auto and a classifier running on Workers AI across Cloudflare's edge network sorts the request before any expensive model sees it.
The published benchmark is narrow but specific: 97 tasks, three samples per model per task, 291 trials on a workspace benchmark covering email, calendars, files, travel, and finance. cloudflare/auto finished 86.6 percent successful at a total cost of $2.10, or $0.0084 per success. Anthropic's Claude Opus 5.5 hit 96.6 percent at $0.0210 per success. OpenAI's GPT-6 Sol came in at 84.2 percent and $0.0108. The router delivered 80 percent the cost of Sol and 35 percent the cost of Opus.
Cloudflare credits Michelle Chen, Alex Reneau, and Kevin Jain on the writeup. Their framing is that the savings come from not paying frontier rates for work that isn't frontier work, and that the router's job is scoring "utility = expected quality - adaptive cost penalty" across a pool of eligible models.
The engineering detail worth reading twice is cache accounting. Within a single turn the cache is hot, so switching models rarely pays. Across turns, the router applies a switching penalty that grows with the number of tokens already in context, because throwing away another model's cache forces a full context rewrite. Cloudflare's own conclusion is blunt: cheaper tokens don't always mean a cheaper answer.
The routing layer becomes the meter
Here's the part that should make anyone building on a model API uncomfortable. AI Gateway sits on every request, and a router on that path is choosing which company gets paid for the inference. Layer on the budgets, spend limits, and identity-aware analytics Cloudflare already ships, and the edge network is no longer just a delivery layer. It's the procurement decision.
That's a different business than content delivery, and it carries a quality ceiling. A 10-point gap against Opus is real, and it will show up somewhere in a workflow that the benchmark doesn't simulate. The savings also have a shelf life. If frontier model prices fall, the router's margin shrinks with them, and the whole case rests on the assumption that cheap tokens stay cheap enough to matter.
Where the new capacity is actually landing
The subsea picture moved in the same window. Google announced three new cable systems on August 12, per DCD's Dan Swinhoe: Alisios linking the Dominican Republic, Panama, and Chile; Canoa linking the Dominican Republic to Bermuda; and OlaLuz running the Dominican Republic straight to Florida. A new branch of Firmina will land in the Dominican Republic as well.
Bikash Koley, Google's VP of global infrastructure, wrote on LinkedIn that the cables are meant to create "geographically diverse rings" across the Pacific Coast, the Caribbean, and the Atlantic. The Dominican Republic has three data centers and six cables across five landing points, according to Data Center Map, which makes the country a small hub built almost entirely out of connectivity rather than power.
Utility Dive reporter Brian Martucci documented the regulatory side on September 23, tallying the large-load tariffs, minimum contract terms, and upfront impact-study fees that states now use to weed out speculative load. Utility Dive reports that some states went past tariffs and paused projects outright.
That's the tell. When generation is the scarce input, the asset you buy is a route and a router, not a campus. Google has invested in more than 30 subsea cables over 15 years, more than any other hyperscaler, and Cloudflare now counts automated traffic as larger than human traffic on the networks it watches.
What to watch next
Cloudflare has said it will expand the models behind cloudflare/auto, add zero-data-retention filtering, account for provider capacity when picking a model, and release a cloudflare/auto-best variant that drops the cost tradeoff and picks on quality alone. That last one is the interesting line: the company is selling the same classifier twice, once to save money and once to buy accuracy.
The other date to hold is the ERCOT audit, which will tell you how much of that 474GW queue is real. Microsoft said it needs a quarterly average of just over 1GW to reach 38GW in 2032, and it's been hitting roughly that. The number that decides whether this holds is not on Microsoft's roadmap. It's on a utility's interconnection worksheet.
More cloud and edge coverage is at the desk's cloud and edge archive.