Data Centers

The On-Ramp Is Gas. The Destination Is Still Nuclear.

TeraContext.AI Team 19 min read

Editorial cartoon titled "THE ON-RAMP": at a roadside station signed "FOSSIL FUEL & SNACKS," a weary attendant fills a giant server cabinet labelled "AI FACTORY – 600 kW RACK" and "NATURAL GAS ONLY" from a pump marked "NATURAL GAS: MARKET PRICE (FOR SERVERS ONLY)," while cars wait in line. On the hills behind, a nuclear plant with cooling towers sits beside a sign reading "SMR NUCLEAR – THE DESTINATION (EST. 2030s)."

One rack, 600 kilowatts — and why that rewrites the nuclear engineer’s job.

One Nvidia rack is now specified at about 600 kilowatts.

That is the figure Jensen Huang put on each Rubin Ultra “Kyber” rack in his GTC 2025 keynote — roughly the continuous load of 500 American homes, packed into a single liquid-cooled enclosure. Even in 2025, Uptime Institute found the typical data-center rack averaged under 9 kW. Today’s Vera Rubin NVL72 cabinets already sit at about 190–230 kW. Kyber was originally slated for the second half of 2027 on 800-volt DC distribution. In July 2026, SemiAnalysis reported a slip into 2028 over midplane manufacturing; Nvidia said its roadmap is intact. Either way, the facility problem is the same: racks are becoming small power plants.

Multiply that density across a gigawatt-scale campus and you have the engineering problem of this decade. As the third quarter of 2026 closed, two clocks were running at once. AI load is climbing now. Advanced nuclear is commercializing on a slower, more deliberate schedule. Gas turbines are what sit in between.

That collision is rewriting the career map for power professionals. The rest of this post follows the sequence the market is actually using: the demand math, the gas bridge, the nuclear permits that landed this quarter, the first-of-a-kind contract that did not, the fuel bottleneck behind advanced reactors, and where a nuclear engineer fits.

The grid cannot connect this load on the old timetable

JLL’s 2026 Global Data Center Outlook expects nearly 100 GW of new data-center capacity worldwide between 2026 and 2030, taking the global base to about 200 GW. BloombergNEF’s July 2026 U.S. outlook is more specific, and more startling: 118 GW of installed U.S. data-center capacity by 2030 — a 52 percent upward revision in seven months.

BNEF also estimates that the most new data-center demand the U.S. grid has ever connected in a single year is about 10 GW. JLL puts the average wait for a grid connection in primary data-center markets at more than four years.

Read those last two lines together. Demand is growing by tens of gigawatts. The grid’s best year connected ten. For the past decade, grid planners forecast peak-demand growth of only about 0.5 percent a year. It cannot close that gap on interconnection queues alone.

So the buyers are changing shape. Hyperscalers are becoming energy developers — “shadow utilities” that fund, own, or anchor generation instead of waiting in line:

  • Amazon bought Talen Energy’s Cumulus data-center campus next to the Susquehanna nuclear plant in 2024, and is backing Energy Northwest’s Cascade project in Washington: a first phase of four X-energy Xe-100 units totaling 320 MWe, with an option to expand to 12 units and 960 MWe.
  • Meta signed with TerraPower in January 2026 to fund early development of two Natrium units and secure energy rights on up to six more — as many as eight units and 2.8 GW of baseload by the mid-2030s if the later plants are built. That is an option set, not eight shovels in the ground.
  • Google is taking clean-energy attributes from Kairos Power’s Hermes 2 plant in Oak Ridge through a power-purchase agreement between Kairos and TVA, and has separately pioneered “clean transition” tariff structures that let a large buyer fund new firm power through its utility.

The halls themselves are learning to flex. Nvidia’s Vera Rubin DSX AI Factory reference design includes DSX Flex, software that lets an AI factory adjust draw to grid conditions and coordinate with on-site generation. Nvidia and Emerald AI argue flexible AI loads could unlock up to 100 GW of capacity on the existing U.S. grid. The first commercial deployment is to be the 96 MW Aurora AI factory at Nvidia’s research center in Manassas, Virginia — billed as the world’s first power-flexible AI factory and developed with Digital Realty, EPRI, and PJM. It was slated to open in the first half of 2026 and has not yet opened.

Utilities already feel the direction, if not yet the full volume. For the nine months through June 30, 2026, TVA reported about $10 billion in operating revenue and $965 million in net income — net income up $220 million year over year. Electricity sales rose only about 1 percent. TVA attributed the sales gain mainly to data processing, web hosting, and related services; revenue rose mainly on higher fuel-cost recovery rates. The signal is the customer mix, not a demand spike on the monthly bill.

Gas is winning the next five years

SMRs dominate headlines. On-site gas turbines are winning the interconnection race.

BloombergNEF counts 124 GW of announced on-site gas capacity for U.S. data centers and says on-site gas “must play a significant role” if its forecast is to be met. Unless the grid connects new data centers faster than its best year to date, about 10 GW, roughly 48 GW of on-site gas must be built by 2035. The near-term workhorse is the aeroderivative turbine — a jet engine adapted for ground power. GE Vernova’s LM2500XPRESS delivers about 35 MW per unit, arrives roughly 95 percent factory-assembled, installs in weeks, and starts in five minutes. Crusoe has ordered 29 of them across two orders (December 2024 and June 2025), nearly 1 GW.

Speed is the reason. A new SMR is still targeted at five or more years from construction permit to power, and closer to a decade from project start. A modular aero unit with a manufacturing slot can be producing in months. Developers call that “Phase 0”: energize the first wave of servers on turbines, then hand the load to the grid or to nuclear in the 2030s.

One caveat: speed depends on the slot. GE Vernova’s gas-equipment backlog and slot reservations grew from 100 to 116 GW in the second quarter of 2026 alone. On its earnings call, the company said it is already taking reservations for 2031 deliveries.

The most interesting projects do not choose. They sequence.

Blue Energy’s Port of Victoria, Texas plan is the template made literal: two GE Vernova 7HA.02 turbines delivering about 1 GW to a nearby Crusoe data-center campus by 2030, then up to five BWRX-300s adding about 1.5 GW from around 2032. On September 16, Blue Energy filed Part 1 of its NRC construction-permit application for the first nuclear unit — just over four months after it announced the BWRX-300 choice. Gas and nuclear are not rivals in that design. They are sequential phases on the same site, the same interconnection, and the same customer.

Nuclear is moving. Light-water SMRs moved first.

Capital commitments are one thing. Permits and electrons are another. The third quarter produced a run of light-water SMR milestones — reactors that burn the same low-enriched uranium today’s fleet already uses:

Date (2026) Milestone Scale Status
Sept. 29 NRC issues a construction permit for TVA’s Clinch River BWRX-300 — the first U.S. SMR construction permit 300 MWe (first unit) No construction start date yet; operating license still required
Sept. 16 Blue Energy files Part 1 of an NRC construction-permit application for its Victoria, Texas BWRX-300s Up to ~2,500 MW gas-plus-nuclear FID expected 2027; Part 2 of the application expected in 2027
Sept. 3 Studsvik, GE Vernova Hitachi, and Samsung C&T agree to advance a four-unit Swedish project ~1,200 MWe Pre-FID; first unit mid-2030s

Clinch River is the U.S. headline. The NRC finished a 14-month review, four months ahead of its own schedule. The project team pairs TVA and GE Vernova Hitachi with Bechtel and Sargent & Lundy. A construction permit is not a construction start, and TVA has said it is still working through cost-sharing with its partners. It is still the first time a U.S. utility is permitted to build this class of reactor.

The same week, a different nuclear story landed in Reston.

The Natrium handoff is what FOAK risk looks like in a contract

In September, Bechtel told Virginia regulators that it and TerraPower could not agree on the next phase of the roughly $4 billion Natrium project in Kemmerer, Wyoming, and would move forward separately.

The news surfaced through a September 15 WARN Act filing covering about 200 positions at Bechtel’s Reston, Virginia headquarters, with reductions expected to begin on or about November 16. Bechtel has said it does not expect any actual layoffs and is reassigning staff to other nuclear projects. It is also staying on site to finish the Sodium Test & Fill Facility and support an orderly transition. TerraPower says it will maintain its schedule — construction complete in 2030, commercial operation in 2031 — and it will rebid the EPC contract.

What Bechtel delivered under the earlier phase is not in dispute. Selected as TerraPower’s engineering and construction partner in 2020, it supported the project through:

  • the NRC construction permit issued in March 2026 — the first the NRC has ever issued for a commercial non-light-water power reactor, and the first such U.S. permit in more than 50 years
  • completion of the Sodium Test & Fill Facility’s exterior enclosure
  • placement of major procurements
  • the start of foundation work for Kemmerer Unit 1

Neither company has published the contract, but Bechtel says the work to date “was performed under an earlier phase of the contract, prior to full EPC scope,” and Lincoln County Board of Commissioners Chairman Kent Connelly told WyoFile that Bechtel wasn’t expressly under contract for the next “nuclear island” phases. The parting came at that phase boundary. Neither company has disclosed the commercial sticking points, or said which side said no.

First-of-a-kind nuclear is where cost estimates get stress-tested. Whichever side said no, a clean handoff at a phase gate beats an unworkable contract signed early.

That knowledge is scarce, which is why the same firm’s other book of work matters as context rather than consolation. Georgia Power and Southern Nuclear brought Bechtel in in 2017 to complete Vogtle Units 3 and 4 — the Westinghouse AP1000 units that were the first new U.S. large reactors in more than three decades. On September 23, a week after the WARN filing, the Bechtel–Westinghouse consortium agreed on the main commercial terms of the EPC contract for Poland’s first nuclear plant, three AP1000 units; both sides aim to finish the full contract by year-end. And Bechtel remains on TVA’s integrated project team for Clinch River, where the NRC permit arrived on September 29 — contracted to plan the unit and positioned to build it if TVA proceeds.

Bechtel’s website still describes it as the only EPC with current projects across the full spectrum of reactor technologies, citing Poland, Clinch River, and Natrium. Unless it wins another advanced-reactor job, that last leg ends when the Sodium Test & Fill Facility is finished. The Kemmerer handoff shows how the firm is choosing where to take FOAK risk.

Fuel is why advanced reactors are harder to contract than light-water plants

Any honest discussion of sodium-cooled or high-temperature designs has to include fuel. Natrium runs on high-assay low-enriched uranium (HALEU) — uranium enriched to between 5 and 20 percent U-235 — and needs roughly 15 to 20 metric tons of it for a first core. Before 2022, Russia was what TerraPower called the only commercial source of HALEU fuel. When that door closed, TerraPower announced a delay of at least two years, and the West started building a fuel chain almost from scratch.

That chain is taking shape, slowly:

  • Enrichment. TerraPower has supply agreements with ASP Isotopes intended to support Natrium’s first core and up to 150 metric tons over 2028–2037, tied to a planned enrichment plant at Pelindaba, South Africa. In the U.S., Centrus’s Piketon, Ohio demonstration cascade — which Centrus calls the only HALEU production it knows of outside Russia — made about 900 kg of HALEU (as UF6) a year, more than 1,900 kg in all, until DOE-funded production ended in mid-June 2026; Centrus intends to keep running it commercially. Under a $900 million DOE task order awarded in January 2026 and finalized in July, Centrus is to add commercial-scale capacity starting around 2029 and deliver its first metric ton by July 2032.
  • Conversion and fabrication. Framatome has completed a uranium metallization line at its Richland, Washington plant and produced metal pucks from depleted uranium — a step toward HALEU metal for Natrium fuel. Making HALEU there will need a further license amendment. The Natrium Fuel Facility is under development in Wilmington, North Carolina.

Nothing public ties fuel supply to the Bechtel–TerraPower decision. Fuel uncertainty is still one reason an owner–contractor risk split on an advanced reactor is harder to write than on a light-water plant that uses today’s LEU.

The talent map is already moving. The BLS table is not.

Read the federal statistics alone and the profession looks flat. The Bureau of Labor Statistics projects 0 percent employment growth for nuclear engineers from 2025 to 2035, on a base of about 16,600 jobs, with roughly 800 openings a year from retirements and transfers. The 2025 median wage is $133,970.

Look at BLS’s own reasoning: it attributes the flat outlook to an increased focus on renewable energy and limited construction of new nuclear plants. This quarter produced the year’s second U.S. power-reactor construction permit (the first for an SMR), a 2.5 GW gas-plus-nuclear plan entering NRC review, and a four-unit Swedish SMR program. Occupational projections are built on history. They lag inflection points.

The demand is shifting to four places:

  1. Hyperscaler energy teams. Tech companies need engineers who can evaluate reactor designs, audit counterparties, manage nuclear supply chains, and integrate behind-the-meter generation with data-center architectures such as Nvidia’s 800-volt DC designs.
  2. The SMR and microreactor vanguard. Venture-backed developers — X-energy, Oklo, Kairos Power, Valar Atomics and others — are moving from design into hardware: fuel plants, factory-built components, site construction, and — for Valar — a test reactor that reached criticality this summer. That creates a vacuum for people with execution and NQA-1 quality-assurance experience.
  3. EPC integration. Bechtel, Fluor, Kiewit, and Sargent & Lundy are building nuclear practices for the coming wave of light-water SMR projects. The rarest profile is the one that has taken a reactor from permit to power.
  4. The hybrid power-systems architect. With gas as Phase 0, the most valuable engineer can design a 100 MW turbine microgrid today — and lay out the substation, interconnection, and site so an SMR can take over the load in the 2030s.

Turbines now, reactors next is not a detour

The smartest move for a nuclear engineer in 2026 may be a few years on gas turbines. That sounds like leaving the field. It is not.

A nuclear plant is a heat source bolted to a conventional power plant. In a BWRX-300, reactor steam drives the turbine directly. Natrium stores reactor heat in molten salt and raises steam for a separate energy island; TerraPower plans a commercially available steam turbine there, and NRC staff agreed that running it is not a reactor control only a licensed operator may manipulate — though, absent an exemption, a licensed operator must still know of and consent to energy-island manipulations that may affect the reactor. Everything downstream of the reactor — turbine-generator, condenser, feedwater, switchyard, grid synchronization — is work a combined-cycle gas plant demands too.

Skill In a nuclear plant In a gas-fired data-center plant
Thermodynamic cycles Rankine steam cycle Brayton cycle, plus Rankine again in combined-cycle bottoming
Turbomachinery Main steam turbine, coolant pumps Gas turbine, compressor, HRSG steam turbine
Heat transfer and fluid systems Steam supply, condensers, decay-heat removal Heat-recovery steam generators, cooling, fuel-gas systems
Electrical integration Generator, switchyard, offsite power Same, plus islanded microgrids serving AI loads
Quality and safety culture NQA-1, Appendix B, probabilistic risk assessment Hazard analysis and high-reliability operations, where nuclear rigor stands out
Project delivery Modular, first-of-a-kind construction Modular, factory-built packages on compressed schedules

The calendar makes the case. Blue Energy’s Texas turbines are slated for about 1 GW by 2030, with reactors following around 2032. TerraPower targets 2031 at Kemmerer. Clinch River is an early-2030s plant if TVA pulls the construction trigger. Sweden’s first BWRX-300 is mid-2030s. That leaves a five-to-seven-year window in which most new power-plant engineering work is on the gas side.

An engineer who spends that window commissioning turbine fleets on data-center campuses arrives at the SMR wave with something scarce: hands-on experience integrating generation with AI loads — often at the sites where the reactors will go. On gas-plus-nuclear projects like Blue Energy’s, the bridge is literal: the same team, the same site, a different heat source.

The gaps are real. Gas work adds combustion, emissions controls, air permitting, and the maintenance cycles of aeroderivative engines. Reactor physics and NRC licensing sit idle unless the engineer keeps them current — through SMR licensing support, ANS involvement, or professional licensure.

For employers the logic runs the other way. A nuclear-trained engineer on a gas project brings a safety and quality culture few turbine teams have, and is already on staff when the site is ready for its reactor.

What to watch in Q4

  • Who wins the Natrium EPC rebid — and how the contract allocates FOAK risk. (In August, TerraPower named Hyundai E&C as EPC partner for up to eight follow-on Natrium units.)
  • Whether Poland’s Bechtel–Westinghouse AP1000 EPC is signed by year-end, now that main terms are agreed
  • Clinch River’s construction decision, and how TVA settles cost-sharing
  • Blue Energy’s Part 2 application and 2027 FID, and whether Energy Northwest files the Cascade construction-permit application in the first half of 2027, as X-energy now guides (slipped from an end-2026 target)
  • HALEU milestones at Centrus and ASP Isotopes

What the quarter actually proved

A contractor handoff on a first-of-a-kind reactor is not a failure of the renaissance. It is what an industry looks like when it moves from theory to practice with real contracts and real accountability. The last U.S. nuclear revival, in the 2000s, stalled when shale gas turned cheap, alongside the recession and Fukushima. This one is using cheap gas as the on-ramp.

AI demand for firm power is not infinite. For the rest of this decade it will still outrun what the grid can connect. For nuclear engineers, the work no longer lives only inside traditional utilities. It lives in the projects that pair turbines today with reactors tomorrow, in the contracts that allocate FOAK risk honestly, and in the fuel chain being rebuilt from scratch.

The engineers who can navigate commercial risk, fuel logistics, and fast microgrid deployment will not just find jobs. They will build the plants the racks are waiting on.

Works Cited