Over the past two years, when people looked at AI infrastructure, most eyes were fixed on chips. Companies like Amazon had to buy GPUs, build clouds, and run models; without enough cards, work couldn’t get started. As late as Q1 2026, TSMC was still saying that supply for AI and high-performance computing remained tight. Nvidia itself sells cards. What it now needs to worry about is the other end: after customers buy the cards and take them home, is there power and a data center to actually put those cards to work?
In the same week of August 2026, both of these companies put money and planning into power. Reuters, citing a report by The Information, stated that Nvidia plans to invest an initial $2 billion in Texas power and campus developer Lancium for about a 20% stake; once power is connected and agreed milestones are met, it will add up to another $1 billion. In the same week, Amazon confirmed it will build a data center campus in Pecos County, planning to generate power on-site next to the campus first, and later connect to the public grid according to schedule.
Lancium finds land, builds campuses, and connects to the grid in West Texas. If Nvidia’s investment goes through, it will take a stake in the entity bringing power online. Once power is connected, the developer becomes more valuable, and Nvidia’s equity becomes more valuable as well. The second tranche of funds will only be paid once energization progress is made.
In Abilene, some are already using these campuses. Oracle announced that AI workloads are already live. Crusoe stated that it is initially designed for over 200 megawatts, scaling up to 1.2 gigawatts in phases.
Amazon is taking a different path. It plans to generate power on-site next to the campus, bypassing the queue for the main Texas grid for now. Musk launching Colossus in Memphis using gas turbines was the same playbook, as written before. Pacifico framed the design as: initially neither purchasing electricity from the grid nor feeding power back. Amazon stated it still intends to connect back to the grid in the future. Permit filings list 35 gas turbines with a nominal output of 5,000 megawatts, with calculated emission limits exceeding 33 million short tons based on full-load annual hours.
One sells cards; the other buys cards to build clouds. What both are competing for at the same time is electricity delivered to a specific data center on a specific plot of land by a specific date.
Chips arrive earlier in the construction chain. A data center’s shell can be built within a year or two, but connecting power to that shell often takes a long time. Laying high-voltage lines and constructing dedicated substations routinely takes years; according to an industry survey from Q2 2025 reported by Power Magazine, average lead times for large power transformers have extended to about 128 weeks, while generator step-up transformers require around 144 weeks.
Microsoft’s characterization of this mismatch is remarkably direct. In the Q1 FY2026 earnings report, CFO Amy Hood pointed out that the bottleneck facing the company over past years has been the space or power required to house chips. CEO Satya Nadella added that chips might already be sitting in inventory, but without fully commissioned data center facilities ready for power-on.
Data from the main Texas grid illustrates the gap between application volume and reality. In its June 2026 Operational Overview, ERCOT disclosed that large load interconnection requests totaled approximately 474 gigawatts, about 90% of which were designated as data centers; however, approved capacity for power delivery stood at only 8.926 gigawatts. The actual peak load observed by the grid was 3.966 gigawatts, calculated by summing individual monthly peak loads across projects.
The bottleneck here does not stem from a sudden inability of the U.S. to generate electricity. Chips keep arriving, and buildings can still be constructed; the bottleneck is whether power can enter the room on the agreed date. Only after someone locks in that date can the cards start working.
If Nvidia completes the transaction as reported, it is taking a stake in a developer that is already building campuses and securing grid connections in Texas. Founded in 2017, Lancium stated from the start that it aimed to relocate power demand to where wind generation is most concentrated. As reported by the Houston Chronicle, its Fort Stockton campus was once registered as a load capable of ramping power consumption up or down within seconds. The company later stated to the state government that approximately 1.2 gigawatts in Abilene were under contract and construction, with another 3 gigawatts having secured interconnection approval.
Amazon, on the other hand, chose to build on-site power generation to bypass queue delays on the main grid. E&E News reported on this emerging trend of constructing natural gas power plants right next to campuses. This approach effectively swaps one set of engineering hurdles for another: gas turbines require their own procurement queues, gas lateral lines require permits, air emissions must undergo state environmental review, and fuel prices together with carbon emission costs will persist across the power plant’s entire lifecycle. Amazon stated that self-generation will not raise electricity rates for Texas residents, while confirming its long-term plan to connect back to the grid.
Both companies selecting West Texas for their projects also reflects the rationale of placing load directly near power sources. West Texas combines surplus wind node capacity, natural gas, and inexpensive land. Lancium leveraged the inherent advantage of renewable energy surplus, while Amazon relied on natural gas supply, land, and produced water infrastructure. Virginia remains the largest existing market in the U.S., while Utility Dive ranks Texas second. Companies head to Texas because it is easier to convert public transmission grid bottlenecks into internally controllable engineering variables. By contrast, building new backbone transmission lines in Northern Virginia is estimated by PJM to take more than five years.
Regionally, Texas remains one of the fastest-expanding markets at the margin, enabled by developers’ ability to relocate, build in phases, and self-generate power. However, this pace of expansion is facing policy adjustments. In June 2026, the Governor of Texas directed data centers to cover their own infrastructure costs to mitigate impacts on residential electricity rates; in August, new grid interconnection approvals were paused following audit requirements. The PJM and Northern Virginia markets display greater rigidity: capacity auctions tightly mirror power system reliability requirements, projected incremental demand from data centers far exceeds newly cleared generation capacity, and community as well as transmission permitting reviews are far more stringent.
Compared to regional breakdown, the maturity tier of a given megawatt plays a far more decisive role in asset valuation. Planned campuses, contracted capacity, shell structures under civil construction, and fully energization-approved, operational facilities carry vastly different levels of cash flow certainty. The market is already revaluating powered assets that are close to steady-state operations and leased to major hyperscalers. The three Northern Virginia assets transacted between Digital Realty and Blackstone totaled 288 megawatts, valued at $7.8 billion gross. Upstream equipment manufacturers are also seeing customers place orders years in advance to lock in transformers and switchgear. Power-on dates are already being priced.
Going forward, when seeing how many chips a company has ordered or how many gigawatt-scale campus plans it has announced, consider four more specific execution questions first: Has power delivery to the site been officially approved? Are lead times for critical equipment locked in in advance? Have environmental and community approval processes been completed? If energization is delayed by a year, who is contractually responsible for paying the penalty costs?
Chips are still shipping. The next visible battle is over who gets those cards plugged into data centers that are already energized and ready to power up.