Ep. 015 - DG Matrix on 800V DC vs Legacy AC in Data Centers
Summary
800 V DC is being pulled into AI data centers by rack density. Haroon Inam argues that raising voltage is cheaper than raising current: moving from legacy 240 V AC distribution toward 800 V can deliver almost triple the power over the same copper while reducing conductor weight, cost, and I²R losses. His shorthand is categorical: “raising voltage to get more power is far cheaper than raising current to get more power.”
The choice of 800 V reflects semiconductor and supply-chain economics as much as electrical physics. Inam’s “educated guess” is that EV investment and 1,200 V devices used in motor drives and wide-bandgap applications made 800 V a practical, safely derated landing point. The transition has therefore already started through engineering, manufacturing, and supply-chain preparation, even before the NVIDIA architecture that drives demand arrives.
DG Matrix’s SST thesis depends on collapsing systems, not replacing a durable transformer with expensive electronics. Inam calls his earlier AC-to-AC solid-state-transformer effort “one of the dumbest things we could have done”; the value emerged only after integrating rectification and then adding isolated, bidirectional AC and DC ports. The resulting multi-port design could replace a STATCOM, UPS, rectifier, energy-management system, and behind-the-meter aggregation layer, but required at least 700,000 engineering hours.
The adoption curve is credible in direction but explicitly unbankable in timing. SemiAnalysis’s chart rises from nearly zero in 2026 to almost 80% of incremental capacity and more than 30 GW by 2030, but Inam endorses only its shape: sidecar use should decline over time as native facility-wide DC takes root. A delay in 600 kW or 1 MW rack roadmaps could shift adoption right, while Inam stresses that it “could go up,” last longer, or fall faster.
Mixed hardware is the central commercial risk—and power fungibility is the proposed hedge. CPUs are “so back,” ASIC requirements may diverge, and Amazon or Microsoft must support broad cloud fleets even while building for AI users such as OpenAI and Anthropic. Multi-port SSTs can allocate full rated capacity between AC and DC, but the hosts correctly press that conversion still requires different protection, connectors, and whips even when the copper remains usable.
The longer-run architecture is a software-defined power fabric fed increasingly by behind-the-meter generation. Inam expects grid bottlenecks and multiyear upgrades to favor deployable 10–20 MW “cellular power” blocks, arguing that “speed to power or the speed to compute will win out.” He imagines dynamic power routing, GPU scheduling, superconducting links carrying 5–6 MW, and possibly token-generation auctions—while candidly saying, “I’m smart enough to know that I’m not that smart.”
DG Matrix attaches concrete milestones—and concrete execution risks—to that vision. It has 400 kW units that can be paralleled today, is looking at a 6 MW SST for 2027, and plans a containerized 10 MW system for 2028 with 35 kV input and programmable 800 V or 1,500 V output. Silicon carbide is currently preferred over gallium nitride at megawatt-class power. Its current silicon-carbide sources are in the United States, potentially Japan, and Europe rather than China, while some non-CPU, non-software electromechanical parts use a China-plus-one strategy. Inam warns that software-defined power must become “cybersecurity proof.”
Deep dive
1. Rack density makes 800 V a copper-economics necessity
Inam’s premise is that GPU power and the synchronicity required among them are outrunning legacy AC cables. Moving from 240 V AC single phase across three phases toward 800 V, “root mean square to root mean square,” can deliver almost triple the power through the same copper—provided the distribution system can handle the voltage.
Nicolas Bontigui tests every possible driver—busbar weight, copper inflation, efficiency, and I²R losses—and Inam accepts the bundle. Current is the binding constraint; once creepage and clearance are addressed, “raising voltage to get more power is far cheaper than raising current to get more power.”
Why stop at 800 rather than 950 or 1,200 V? Inam offers an explicitly hedged “educated guess”: EVs built an 800 V ecosystem, while motor drives and wide-bandgap semiconductors were built around 1,200 V devices. That combination makes 800 V a practical operating level with semiconductor headroom, power density, and an established component base.
Adoption has “started” before the NVIDIA architecture that drives demand arrives. Manufacturing methods must mature and EV-oriented supply chains must migrate toward data centers; as Inam puts it, customers cannot wait for that moment and then say, “I’m going to invent 800-volt architecture now.”
2. Multi-port integration is the SST’s economic reason to exist
Inam’s formative failure came in 2011–2013, when his team spent “millions and millions of dollars” building an AC-to-AC SST. Replacing iron, copper, and insulation that survive 40–50 years and repeated thermal cycles with costlier, more delicate electronics was, in retrospect, “one of the dumbest things we could have done”—a science experiment without a value proposition.
The breakthrough was to absorb the downstream conversion. Integrating AC-to-DC rectification makes the overall system cheaper and potentially more reliable through integration and less margin stacking because it comes from one company; adding more isolated, bidirectional AC and DC ports then allows one system to replace a STATCOM, UPS, rectifier, energy-management system, and behind-the-meter energy aggregator.
That “Holy Grail” was not easy integration theater: controls, cooling, electromagnetic interference, and density produced repeated “brick walls,” requiring at least 700,000 engineering hours before deployments could begin broadly. Inam’s end-state metric is “power in, tokens out”—collapse copper, iron, and “junk,” eliminate stranded capacity, and drive the lowest token cost per kilowatt-hour.
3. Unipolar 800 V trades a conductor for harder fault management
Inam reaches back to the Dreamliner for the best analogy. At 30,000–40,000 feet, insulation degrades differently above roughly 300 V, so the aircraft used plus 270 V and minus 270 V around a common conductor: 540 V of delivery without exposing each side to that full potential. He wonders—without claiming certainty—whether arc-flash concerns similarly inspired plus/minus 400 V.
The key question is whether loads remain balanced across the positive and negative rails. If not, bipolar distribution behaves like two circuits; imbalance can create a runaway condition in some converters, collapsing one voltage relative to the other, while faults, grounding, and non-isolated conversion become substantially more complicated.
Unipolar 800 V may therefore win on one fewer copper conductor and simpler return grounding, provided arc flash is solved. Inam sees the likely answer in rapid detection and the ability to “quench the source” before it continues feeding a fault; he hopes that capability can support one harmonized architecture.
DG Matrix is intentionally topology-agnostic because every port is galvanically isolated. Its outputs can float at minus 800 V or plus 800 V, or ground the midpoint for plus/minus 400 V, accommodating NVIDIA reference grounding schemes. When pressed on other cost differences, Inam gives an honest non-answer: there may be many, but “generally” the third copper conductor is the significant one.
4. Hardware uncertainty controls the pace of DC adoption
Jordan Nanos frames four phases: white-space retrofit, native DC compute, facility-wide DC, and an SST-centered end state. SemiAnalysis’s chart takes 800 V DC from virtually nothing in 2026 to almost 80% of annual incremental capacity and more than 30 GW by 2030; Inam agrees that sidecar use should decline as native DC takes root, but says, “We don’t have that crystal ball.”
Sidecars remain useful in AC-dominated brownfields and as insurance if DC migration does not happen quickly enough. Conversely, if 600 kW and 1 MW rack systems slip, the curve follows the hardware to the right. Inam refuses false precision: adoption “could go up,” persist longer, or decline faster, even if its general shape remains plausible.
SemiAnalysis’s chiller analogy exposes why theoretically superior designs stall: buyers were told liquid-cooled systems could run chillers at 45°C, yet the share doing so remains extremely low because their future workload mix was uncertain. With CPUs “so back,” plus GPUs, TPUs, ASICs, storage, and other loads, Amazon and Microsoft need fungible fleets even when OpenAI or Anthropic consumes the AI capacity; direct self-builders may eventually choose more specialized facilities.
The hosts’ pushback is that multi-port hardware cannot erase decisions already embedded in switchgear and distribution. Inam’s narrower answer: hybrid builds can put full rated load on either AC or DC, provided the combined load stays within the machine rating; a later switch from AC to DC preserves the copper but changes protection, connectors, whips, and the relevant output port. Hyperscalers may remain conservative, while neoclouds and developers may accept more risk for faster payback.
5. Fifteen-year facilities need power fabrics, not fixed rack assumptions
Nanos’s rack roadmap runs from roughly 12 kW for traditional air-cooled CPU racks, to 30–40 kW around 2020, to 130–140 kW for current GB200 and GB300 systems. The next step is 360 kW per rack next year or potentially at the end of this year, with 600 kW per rack from Vera Rubin by the end of 2027; 1 MW is discussed for the 2028–2030 period. He says the chart needs a log scale; one facility he visited devoted “well over 80%” of its area to power and cooling.
Inam compares the future data center to an 80-foot robot controlled by “a little kitty cat”: the power architecture stays massive while the GPU “brain” keeps shrinking. He imagines a bicycle-wheel geometry fed from multiple points, superconducting strands moving 5–6 MW, and software routing power around failures so no capacity is stranded, coordinated with GPU job scheduling and perhaps auctions for token generation.
Optical interfaces or optical computing might reverse or flatten power-density growth, Inam allows, but he distinguishes brainstorming from execution and declines to predict when. Nanos counters with Jevons’ paradox: efficiency gains may simply induce more consumption. Inam’s durable conclusion is flexibility, not a wattage forecast—“I’m smart enough to know that I’m not that smart.”
6. Distributed generation expands the opportunity—and the attack surface
Central grids were designed for one-way power flow, and placing 100 MW at one node can choke surrounding transmission even when generation exists elsewhere. Upgrades take multiple years and hundreds of millions of dollars; behind-the-meter “cellular power” can instead arrive in 10–20 MW blocks and scale toward a distributed gigawatt. Inam’s call is that “speed to power or the speed to compute will win out,” though he revises “disruption” to the more cautious “augmentation” of utilities.
The product roadmap is specific: 400 kW units can be paralleled today; DG Matrix is looking at a 6 MW SST in 2027; and a large-container 10 MW unit, taking 35 kV input and producing programmable 800 V or 1,500 V output, is slated for 2028. It belongs outside white space; bringing medium voltage closer to compute saves copper but faces safety and national zoning barriers, with Inam speculating China might move first.
On devices, DG Matrix is agnostic in principle but uses silicon carbide because it is more mature from hundreds of kilowatts through megawatts; gallium nitride currently fits hundreds of watts through kilowatts. The company has experimented with both for over 10 years and sources silicon carbide from the United States, potentially Japan, and Europe rather than China. It uses China-plus-one alternatives such as Mexico or Vietnam for some electromechanical parts.
Inam’s closing warning is that adaptive power becomes critical-infrastructure software. DG Matrix draws on prior transmission-grid security work, but he argues the skill set must expand across the industry, including background checks on every entity touching the electronics or developing the software. Nanos supplies the memorable downside case: “We don’t want Stuxnet in any of these new, big data centers.”