DEV Community: Jack H

AI Is Redrawing the US Data Center Map

Jack H — Sun, 18 Oct 2026 04:18:20 +0000

Power access is shaping where new campuses rise, when they come online, and the networks connecting them.

October 18, 2026

In June 2024, construction began on the Stargate campus in Abilene, Texas. Within a year, its first two buildings had been energized. By September 2025, Crusoe said the site was running early training and inference workloads on Oracle Cloud Infrastructure. The planned campus would eventually contain eight buildings. Crusoe’s account captures both the speed of the AI infrastructure boom and its defining complication: large campuses become useful in stages.

That progression is changing the geography of computing. Developers need sites where land, electricity and equipment can come together on a workable schedule. As established markets fill up, new locations become attractive, and connecting distant facilities becomes an engineering problem of its own.

The spending curve shows how quickly this has become a national construction story. In November 2022, private US data center construction was running at an annualized pace of $13.9 billion. By August 2026, that rate had reached $85.0 billion, roughly six times as much. These Census Bureau figures cover data centers broadly and measure construction work, including fixed electrical and mechanical systems. They exclude racks and servers, whose costs sit on top of the building investment.

Figure 1. Monthly construction spending expressed as a seasonally adjusted annual rate, in current dollars. The August figure is a spending pace, rather than money already spent during 2026. Source: US Census Bureau.

Construction spending was already increasing before ChatGPT’s launch. Because the figures are in current dollars, the curve combines changes in construction prices with changes in the amount being built. It establishes the growing scale of spending without measuring computing capacity.

The next question is where that work can happen. In its August report, JLL placed 77% of North American data center capacity under construction in frontier markets, highlighting West Texas, Ohio, Louisiana, Indiana and the Carolinas. Its market analysis describes a buildout extending well beyond the industry’s established hubs.

Texas illustrates the shift. JLL reported 26 gigawatts of existing and under-construction capacity across the state, compared with 13 GW in Virginia. Northern Virginia remained the largest individual market, with nearly 7 GW of inventory. The state totals show the direction of investment while including projects that still have to become operating facilities.

Figure 2. State totals combine existing capacity and capacity under construction, including hyperscalers’ own facilities. They do not measure electricity currently being consumed. Source: JLL, Midyear 2026.

A large pipeline creates a planning problem for utilities. They must decide how much infrastructure to build before they know which projects will arrive, when their servers will be installed, or how quickly customer workloads will fill them. A speculative request can therefore influence planning long before it becomes a substantial electricity load.

AEP Ohio’s experience shows how that pipeline changes as commitments become binding. Its new tariff required developers to advance through paid studies and financial commitments. About one-fifth of the initial requested capacity reached signed contracts. AEP attributed the narrowing to the filtering of speculative or uncertain projects. The utility still had 17.9 GW of total contracted data center projects, including earlier contracts, scheduled to come online progressively through 2035. AEP’s February update distinguishes those future obligations from operating demand.

Figure 3. Successive stages of one AEP Ohio intake cohort, as reported in February 2026. The final bar excludes contracts signed before the tariff. Source: AEP Ohio.

Texas faces its own version of the credibility problem. ERCOT’s September board update documented a pause in energization approvals for relevant new data center and cryptocurrency facilities of at least 75 MW while verification and community-impact reviews proceeded. Projects received provisional classifications; those classifications did not constitute final permission to connect. The operator expected to publish its community-impact report by December 10.

For developers, that makes the location of available power consequential to computing design. Microsoft offers a concrete example. Its Fairwater facilities in Wisconsin and the Atlanta area are linked through a dedicated network intended to let sites cooperate on large AI workloads. Microsoft explicitly identifies land and power availability as reasons to spread work across locations. Its description of the architecture connects the choice of geography to the need for new networking infrastructure.

Distance introduces costs of its own. Large training jobs require frequent communication among processors; slow exchanges can leave expensive hardware waiting. Microsoft describes dedicated fiber and changes to network protocols designed to reduce bottlenecks between sites. Finding land and power therefore creates another task: making the resulting facilities operate efficiently together.

There are also ways to change the terms on which a facility receives electricity. In a report published October 14, the International Energy Agency examines connection agreements that allow consumption to be curtailed, phased in or adjusted to grid conditions. Such arrangements can help projects connect sooner by changing the reliability or timing of their access.

The tradeoff is practical. Interrupting expensive AI hardware can be costly, and a facility operator may have limited control over customers’ workloads. The IEA’s analysis identifies coordination among computing customers, facility operators and electricity networks as a condition for greater flexibility. Onsite generation also brings equipment lead times and engineering requirements of its own.

The buildout is consequently creating an architecture that extends beyond each campus. Transmission studies, substations, fiber routes and electricity agreements increasingly influence where AI work can run. Abilene’s first buildings became useful when several of those pieces came together. Repeating that achievement across a much larger pipeline will determine how quickly the new map translates into computing capacity.

Data: Figures use the research snapshot checked October 18, 2026. Each chart retains its source’s geography, period and definition. Charts are original graphics based on the cited figures.

OpenAI Took Aim at Claude—and Shot Itself in the Foot

Jack H — Tue, 13 Oct 2026 19:32:34 +0000

Astra put OpenAI back in the fight with Anthropic. Then it changed the deal.

AI-generated editorial illustration of Claude and OpenAI competing.

On September 10, OpenAI closed the door on people trying to buy its $200 Pro subscription. A week earlier, it had launched Astra. Now demand was putting enough strain on its systems that product leader Tibo Sottiaux announced a pause in new subscriptions. Existing subscribers could keep using the service while OpenAI added capacity. For a company trying to gain ground against Claude among coding developers, that was an encouraging problem. People wanted what OpenAI was selling.

By early October, some of those customers had a different request: keep selling them the deal they already had. In a reply to Sottiaux’s recovery campaign, one user asked to retain the old plan instead of receiving more resets. Somewhere between Astra’s arrival and DevDay, the conversation had shifted from the model’s capabilities to the terms of access.

Claude’s position made this an offensive OpenAI had good reason to pursue. In coding, the rival had built a substantial lead: Menlo Ventures’ 2025 enterprise research estimated Anthropic at 54% of coding model spending against OpenAI’s 21%. An August ZDNET survey found roughly three quarters of 138 respondents using Claude Code, compared with just over a third using Codex. For developers already getting work done with Claude, an OpenAI launch had to offer a compelling reason to reconsider. Winning them over meant changing a working habit, and Astra gave OpenAI an opening to do that.

Generous access helped turn that opening into a credible comeback, and the timing favored OpenAI. Anthropic’s four-month promotion had raised Claude Code’s weekly limits by 50% from May 13 through September 13. From September 14, a permanent 25% increase over the original baseline replaced it, leaving users with about 17% less weekly allowance than during the promotion. Astra launched on September 3, and eligible subscribers received additional resets during the launch week. The early offers compensated for launch delays, but they also gave users more room to discover the model’s value as Claude’s summer generosity was winding down. In a Claude-community discussion, some described switching to Astra or making it their main coding tool while retaining Claude; some cited the resets. A capable model could attract attention, but sufficient allowance let customers build it into their work. The amount of access became part of the appeal—and part of the expectation OpenAI would later disturb.

The temptation was to turn that renewed interest into a much bigger sale. OpenAI’s business was already accelerating before Astra: its overall revenue run rate grew more than 20% in July, while enterprise growth reached 32%. By September’s end, CNBC reported approximately $50 billion in annualized revenue, with Q3 run-rate growth of 77% overall and 107% for enterprise. Anthropic’s reported run rate had exceeded $65 billion by July. CFO Sarah Friar’s September account of OpenAI’s business described customers spending more through subscriptions and usage-based offerings as AI took on more work. That ambition explains the attraction of persistent agents: OpenAI wanted a larger role in customers’ working lives, with more opportunities to earn revenue.

DevDay was where that broader ambition began to collide with the bargain that had helped bring developers back. At the September 29 event, OpenAI introduced Dots alongside Sol 6.1, Astra Ultrafast and a $500 subscription. Dots promised persistent Astra-powered agents with cloud computers and connected apps. Alongside the expanded bundle, OpenAI announced that Pro 200’s allowance would fall from 20 times Plus’s to 10 times, while the price stayed at $200 a month. Sottiaux had explained the economics before the keynote: the revised plan would provide “half the dollar in API spend,” although more efficient models could still accomplish more work. TokenUsage’s subscription-value comparison shows Astra’s token allowance on the $200 plan halved. For coding subscribers, Dots therefore had to help justify losing part of the benefit they already valued. Adding an agent would only make that trade attractive if the agent proved useful.

Slow performance made the trade look worse and brought the disagreement to a head. Early Dots users described simple tasks taking minutes, while Sol’s rollout left coding users waiting too. Sottiaux acknowledged that Sol’s first two days had been slow, attributing the problem to a large load spike. Customers frustrated by those waits were being asked to accept less included coding access while testing a new agent’s usefulness. Developer discussions included cancellations and plans to return to Claude, and one user reported moving several subscriptions and an API budget to Anthropic. These accounts show individual reactions, but the complaint is easy to understand. OpenAI had made the subscription harder to defend precisely when it needed returning users to make it a habit.

The recovery campaign reads like an acknowledgment that the coding experience had to come first again. Sottiaux apologized and announced another reset, acknowledged demands for simplicity, and promised 28 days of improvements or full resets. On October 5, he announced approximately 50% faster default subscription speeds for Astra and Sol. On October 8, OpenAI rolled out Sol 6.1 Ultrafast, extending the premium speed offering already introduced for Astra at DevDay. The public effort was now focused on improving the tools subscribers used for daily work. That addressed the performance complaints while preserving the announced allowance reduction. Faster models could make the smaller allowance more useful, but OpenAI still had to persuade customers that the revised deal deserved their payment.

Dots’ continuing speed complaints exposed the unfinished part of that persuasion. Two days after the default coding-speed announcement, OpenAI recorded delayed Dots responses, failed turns and reduced proactivity. That incident was marked resolved, yet an October 8 user still described long silences and iteration slower than Codex or Work. October 9 updates addressed delegation, browsing and usability, showing OpenAI was still working on the agent. It does reveal the difficulty of the pitch: the agent meant to broaden the subscription’s value still needed repair while OpenAI worked to restore confidence in coding. Astra had given OpenAI a way to challenge Claude. Keeping that opening requires making the subscription dependable enough that developers can plan tomorrow’s work around it.

Disclosure: I run tokenusage.dev, which is cited in this article. This story was researched and drafted with AI assistance.