The MetroVolt Data-Center Burner: Direct-DC Campus Power, Q_E Closure, and the Plug Requirement in a Low-Neutron D–³He Tandem Mirror

The load that a hyperscale artificial-intelligence campus imposes on the grid is native direct current, growing faster than transmission can be built, and increasingly sited where interconnection queues are measured in years. We present MetroVolt, a data-center power plant built on the Kronos deuterium–helium-3 (D–^3He) tandem-mirror burner, whose direct energy converter (DEC) delivers high-voltage direct current natively—the same physical form the compute hall consumes—so that the plant couples to the load through two-to-three power-conversion stages rather than the four-to-six of a thermal generator feeding an alternating-current (AC) grid and an uninterruptible supply. At the frozen design point the burner reaches a plant energy multiplication =1.318 at fusion power 4298.5 MW with a neutron fraction of only 5.44 % of the total power—the low-activation advantage of the fuel—and a length ladder delivers net electric output of +104/+850/+2832 at 55/440/1400 m device length, with a Monte-Carlo uncertainty quantification returning P(>1)=1.0 over the design ranges. The binding physics gate is the end-plug potential: confining the central-cell ions requires a plug-to-central-cell density ratio n_p/n_c=16 built from an ion-confining potential _i=248.75 kV against a carrier temperature T_e=89.7 keV, a requirement-class target more than two orders of magnitude beyond the demonstrated tandem-mirror record and retired only by a dedicated plug-potential experiment; direct energy conversion is a bounded upside lever, taken at a deliberately conservative module efficiency η_DEC=0.49, not the closure mechanism. A second, independent modelling choice—the synchrotron radiation transport—is shown to be equally load-bearing: the same machine reads =0.39 optically thin, =1.32 under the reflective-cavity Albajar–Fidone–Johner model adopted as canon, and =4.56 under Trubnikov, so that a first-wall reflectivity R_wall0.83 is a hard specification, not a convenience. Reduced to the data-center product, the direct-DC path raises the source-to-chip efficiency from 75.9 % (conventional AC) to 88.1 % (DEC-direct medium-voltage DC); a plasma-independent electronic buffer follows synchronized GPU training swings at 80 MW s^-1—about 105× any plasma throttle—while holding the DC bus to ±0.40 %; waste-heat-driven absorption cooling takes the facility from PUE 1.35 to 1.17; and the neutron-lean 90 %-aneutronic burn licenses against a mobile tritium source term of 3.6 g, roughly 275× below a deuterium–tritium plant of the same power. We separate throughout what is robust—the neutron-source performance and the power-electronics topology—from what is conditional on retiring the plug gate. This is a comprehensive, review-length treatment: twenty-six figures, full governing-equation derivations, a GPU-accelerated computational methodology, a global sensitivity and uncertainty study, and a benchmarking section against the published tandem-mirror, direct-conversion, synchrotron, and data-center-distribution literature. Key results (frozen anchors): T_e = 89.7 keV; P_fus = 4298.5 MW; P_n = 233.95 MW; T_i = 90 keV; mathrmfus-P_n = 4064.6 MW; P_ax = 2719.7 MW. Methods & codes: Kronos toolkit, OpenMC. Live verification: 16 gate validator(s) with live-recompute cards (11 reproduced, 5 revised). See the Verification section and data/verification.csv. Related: Paper page · De-risking register · 3D model · Learn more about Kronos Part of the 2026 Kronos publication series; independently re-run and stamped in the Kronos de-risking register (DOI 10.5281/zenodo.22645689). All numerical values are frozen design-point anchors; see the register. Public research artifact. No proprietary, financial, or supply-chain information is included.