AI 数据中心的电，开始用卡车运

According to a Bloomberg report on October 8, Oracle is using trucks to transport compressed natural gas (CNG) to its own AI data centers. A facility outside Salt Lake City has been operating this way for over a year, waiting for pipelines to be completed and connected; the campus being built in Shackelford County, Texas, for OpenAI is also supported by the same method during its initial phase; New Mexico's 'Project Jupiter' is considering replicating this approach to power up some facilities before the pipeline becomes operational. It should be noted that the report cites informed sources; Oracle did not directly respond to Bloomberg's request for comment, but it subsequently posted on social media thanking suppliers, effectively confirming the existence of this arrangement.

This practice is known in the industry as virtual pipelines: gas is drawn from existing pipelines, compressed into specialized trailers, transported to the site, depressurized, and fed into the campus's power generation equipment. It delivers fuel, not electricity; the site still relies on generators or fuel cells to convert the gas into electricity. The supplier for the Utah project is Setarus under Superior Plus, while the supplier for the Texas project is Volta Grid.

The costs are twofold. In terms of price, according to East Daley Analytics, including labor, specialized equipment, and truck fuel, the cost of CNG delivered to the site is approximately four times the price at major pipeline hubs. Physically, it's even more challenging: SemiAnalysis energy analyst estimates that supporting just 100 MW of power consumption would consume the gas in one large trailer in about 40 minutes. With Project Jupiter's cap at 2.45 GW, linear extrapolation at full load equates to nearly 900 trips per day. SemiAnalysis also provided a scale assessment: verified CNG truck deliveries can roughly only reach the 200 MW level.

What's bottlenecking Project Jupiter is the pipeline. This line is the Green Chile project by Energy Transmission Company, 17 miles long, with a design capacity of 400 million cubic feet per day. The New Mexico State Land Office twice denied right-of-way across state trust lands in March and July, citing emissions and water consumption; the route was forced to shift to federal land, pushing the start date from August 15 to February 1, 2027.

The power supply plan itself has already been changed once. On April 27, 2026, Oracle and BorderPlex Digital Assets jointly announced: Project Jupiter abandons the original gas turbine plus diesel backup units, switching to a Bloom Energy solid oxide fuel cell microgrid of up to 2.45 GW. The company stated that nitrogen oxide emissions would decrease by about 92%, accompanied by closed-loop non-evaporative cooling. However, the fuel cells still burn natural gas; if the pipeline doesn't flow for a day, this solution cannot get fuel for a day; the air quality permit corresponding to the fuel cells has also not yet been approved.

On the contract level, Oracle has already distanced itself first. On September 24, it issued a force majeure notice to the developer, STACK Infrastructure under Blue Owl, claiming that pipeline delays constitute an event beyond its control. According to Bloomberg, if both parties confirm the establishment of power-related force majeure, Oracle could delay rent payments for three years after payment begins, but must still pay full rent for the entire lease term. The company's external stance is that 'Project Jupiter is still progressing as planned, fully committed to New Mexico'; Blue Owl and Bloom Energy also separately stated that their financial and supply commitments remain unchanged. The approximately $18 billion construction loan is provided by about 20 banks. According to Bloomberg reports, the current trading price in the secondary market is below 90% of par value, meaning banks have already priced in 'later opening'.

Oracle's financials are amplifying this time gap. For the quarter ending August 31, revenue was $19.3 billion, cloud infrastructure revenue was $7.4 billion (+121% YoY), adjusted EPS was $1.63, beating market expectations of $1.39; remaining performance obligations, i.e., contracted but unconfirmed revenue, reached $664 billion. Capex for the same quarter was $28.5 billion, compared to $8.5 billion last year; free cash flow was -$5.4 billion. The numbers look good, but they correspond to revenue that can only be realized after powering up the server rooms first.

On October 8, Oracle's stock fell about 5.5% to around $135.7. Attribution requires caution, as at least three things overlapped that day: Bloomberg's truck logistics report, market concerns over the gap in OpenAI's annualized revenue figures (media estimates around $5 billion vs. previously disclosed $7 billion), and reports that the company was negotiating with Apollo and Goldman Sachs to use independent entities to finance the purchase of AI chips. Attributing the drop solely to this logistics news does not hold up.

There are four counter-evidences. First, truck logistics are a transitional measure, not the final form: Oracle's own data center page describes the Texas campus as Volta Grid-provided power infrastructure and the New Mexico campus as gigawatt-scale fuel cells; reading this as a long-term power supply route is a misinterpretation. Second, nearly 900 daily trips are arithmetic extrapolations based on full-load linearity, not the company's plan; 100 MW accounts for only 4% of Project Jupiter's cap, lighting up only the first building. Third, the force majeure notice currently only buys time, not rent exemption, and all three parties—Oracle, Blue Owl, and Bloom Energy—publicly claim progress as planned. The tension between public statements and leaked information indicates the project hasn't stalled. Fourth, publicly verifiable data center contract sizes for suppliers are 60 MW and 135 MW; these are lower bounds at the announcement level, not capability caps, but indeed indicate that commercially viable scales are currently in the hundreds of megawatts.

The real change lies in the standard for the phrase 'powered on'. Previously, looking at a campus's power supply capacity focused on installed capacity and quotas; now it depends on whether fuel can be delivered hourly without interruption. Capacity is written on paper, logistics happen every day. The same constraint is visible elsewhere: regarding gas turbines, GE Vernova's backlog plus reserved capacity totals about 116 GW, with 2030 capacity sold out and over half of 2031 sold (according to Chinese financial media); Oracle's own 1.3 GW campus in Port Washington, Wisconsin, is believed by third-party analysis cited by Morgan Stanley to potentially still face power shortages by 2028; on October 2, it signed with We Energies to subscribe to 10% to 20% of NextEra Point Beach nuclear plant generation, approx. 125 to 250 MW, still pending approval from the Wisconsin Public Service Commission. Looking ahead, three things suffice: whether Project Jupiter's pipeline will become operational on Feb 1, 2027; when the fuel cell air quality permits will be approved; and whether Oracle's Investor Day at the end of this month can provide the figure for 'powered-on capacity'.