Synthos Think Pieces · the power layer up close · July 19, 2026
The 800-volt data center: what Jordi Visser is actually talking about
When Jordi Visser mentions "Nvidia's 800V DC power side," he's pointing at something specific: NVIDIA is rewiring the AI factory. The plan replaces today's chain — 415/480-volt AC distribution stepped down to a 54-volt copper busbar inside each rack — with 800-volt direct current from the building's edge straight to the rack, timed to the Kyber rack generation (Rubin Ultra) NVIDIA targets for 2027. The reason is physics: a megawatt-class rack at 54 volts means currents so large they'd need up to 200 kg of copper busbar per rack. The reason it's investable is the supply chain: 800 volts is the EV industry's voltage class, so the silicon-carbide and gallium-nitride power chips built for cars get a second, bigger customer — and NVIDIA has published a named partner list three tiers deep. Our conviction signal backs the theme: datacenter power is the only layer of the AI stack still Very Bullish after this cycle's cooldown. The honest caveats: every headline number is an unaudited NVIDIA marketing claim, the hyperscalers spent 2025 writing a rival ±400-volt spec, and a July report says Kyber may slip to 2028 (NVIDIA denies it).
Synthos Research · synthosresearch.com · Think Piece · drawing on 584 power-and-energy-stack claims in our knowledge base; conviction labels from the 958-claim window ending 2026-07-14 · technical claims verified against primary sources listed below · educational only, not investment advice↷ The bigger story inside the story: Visser's real claim is an information-arbitrage claim — this trade moved on a local Korean news report before any sell-side desk wrote it up.
- Datacenter power
- Very Bullishheld · −9 within band, still the hottest layer
- AI capex / build-out
- Bullishwas Very Bullish · −17 this cycle
- Semiconductors / GPUs
- Bullishheld · −5, essentially flat
- Memory / HBM
- Very Bullishwas Bullish · +33, turned more bullish
What Visser actually said
Two claims anchor this piece. On May 4, 2026, Visser connected the edge-computing wave — "autos, phones, computers, and humanoids… that needs different semiconductors for the power side" — to a second leg most people missed: "a report from Nvidia over their need for data center equipment from South Korea for their new 800 V DC power side — and they need power semis." Power semiconductor stocks, he noted, were up double digits that week. Then on June 17 he gave the same platform its optical read: "optical is very early; Marvell goes higher because Vera Rubin — which drives optical demand — hasn't started and its earnings surprise hasn't hit yet."
But listen to what he's actually claiming, because it's only half about voltage. His point was that this story was in no newspaper: "It is a story by a local person in Korea… maybe it's in Korean. You hit the translate button… I don't think sell-side research from the banks can keep up with this." The 800-volt trade, in Visser's telling, is a live demonstration of where investment edge now comes from — global, real-time, machine-translated primary sources — which happens to be the premise this research engine is built on. He is also not blind to the cycle: by July 5 he was warning that "the build won't be as easy or fast as it has been" and rotating attention toward the application layer. Hold both thoughts; the map below does.
What 800 VDC actually is — the plumbing, in one section
Today's data center feeds a rack like this: medium-voltage grid power → step-down to 415/480-volt three-phase AC → AC UPS → power distribution units → AC-to-DC power supplies inside every rack → a 54-volt DC copper busbar that the compute trays sip from. NVIDIA's own materials put the end-to-end efficiency of that chain below 90%. It was fine when a rack drew 20 kilowatts. A Rubin Ultra-class Kyber rack is a different animal — roughly 600 kilowatts, with 1 megawatt the stated design target. At 54 volts, a megawatt is ~18,500 amps: NVIDIA says feeding it would take up to 200 kg of copper busbar per rack — about 200 metric tons per gigawatt-scale facility just to move electrons the last meter.
The 800 VDC architecture deletes most of that chain: convert 13.8-kilovolt grid AC directly to 800-volt DC at the data center perimeter (industrial rectifiers today; solid-state transformers are being studied, not committed), run two DC conductors to each rack, and step down once — a high-ratio 64:1 converter taking 800 V to 12 V immediately next to the GPU. NVIDIA's claimed wins, all carrying an "up to" hedge and none independently audited: ~5% better end-to-end efficiency, 70% lower maintenance cost (fewer power-supply failures), 30% lower total cost of ownership, ~45% less copper, and the same wire gauge carrying 157% more power than 415 VAC. One flag from our verification pass: that last figure was "85% more" in NVIDIA's May 2025 blog and "157%" by October, against subtly different baselines — treat the direction as real and the decimals as marketing.
The named supply chain — three tiers, and one honest caveat
NVIDIA published its partner list twice — May 2025, expanded at OCP in October 2025. Read it as a map of where the sockets are:
The caveat: with 25+ companies named across three tiers, being on the list is table stakes, not a moat. Our verification pass found no source that quantifies per-rack power-semiconductor dollar content or design-win share — the single number that would tell you who actually gets paid. Until someone discloses 800-VDC-attributable revenue, position sizing on this theme is faith in a direction, not a number. That is exactly the kind of gap our Frontier Gap Board exists to flag.
Where the experts agree — and where they don't
The tracked panel converges on the direction: power is the binding constraint, and capital concentrates into bottlenecks. The disagreement is about duration — whether the power bottleneck (and its premium) persists, or gets solved like every bottleneck before it.
"Edge computing — autos, phones, computers, humanoids — needs different power semiconductors… plus Nvidia's 800V DC datacenter power side. That thesis is playing now."
Jordi Visser · May 4, 2026 · conviction 70
"Vera Rubin — a full rack-scale agentic system — is now in full production; its supply chain is twice as large as Grace Blackwell." And on the interconnect: "the copper wall is moving into the rack itself… intra-rack links go optical, driving an explosion in optical demand."
Jensen Huang · NVIDIA · June 1–2, 2026 · management lane — an interested party, weighted accordingly
"Power is now the clear bottleneck (turbines, transformers, data-center delays); because the system solves for intelligence per unit energy, capital concentrates into power to clear it."
Raoul Pal · Real Vision · June 11, 2026 — echoed by Pompliano (July 16): "grid interconnect is now a 5–7 year wait… value flows to whoever can hook up to power or route around grid limits"
"Power was the biggest constraint in 2024–25 but people have woken up; ~15–18 GW added this year, ~30 GW next year — almost a doubling… fab/TSMC capacity will 'fully beat' power as the binding constraint again in 2027."
Dylan Patel · SemiAnalysis · February 3, 2026 — if the bottleneck rotates back to fabs, the power layer's scarcity premium compresses
"China has built a real semiconductor industry… microcontrollers and power chips now matching and undercutting Texas Instruments / ST Micro, though still behind on leading edge."
Dylan Patel · February 5, 2026 — the margin risk under the power-semi trade: this is trailing-edge silicon, and China is aggressively insourcing it
"Semiconductor bottlenecks — naphtha, helium, petrochemicals for lithography — limit output; algorithmic and agentic efficiency means projected capex may never fully happen, risking a bubble."
Jordi Visser · June 23, 2026 — yes, the same Visser. The thesis author is himself cycle-aware; by July 5 he was rotating attention to the application layer.
The hyperscalers didn't wait for NVIDIA's spec: Google, Meta and Microsoft spent 2025 co-authoring the OCP Mt Diablo ±400 VDC standard — Google explicitly anchoring to the 400-volt EV supply chain — before an October update broadened it to "±400 or 800 VDC."
Verified from primary sources (Google Cloud blog, Apr 2025; OCP Diablo spec) — not a KB voice. A softer conflict than it looks, but proof NVIDIA doesn't dictate the facility side.
Where it stands, where it's going
Headline-driven and already fast-moving: Navitas jumped ~25% on the May 2025 announcement and another ~20% around Computex 2026 — with an immediate after-hours fade the second time, a sign the easy news is priced on contact. Watch the Kyber timeline dispute (2027 vs 2028) resolve; a confirmed slip hits the whole chain, hardest at the pure plays.
The build phase. Perimeter-conversion orders (Vertiv, Eaton, GE Vernova tier) and first 800-VDC-attributable revenue disclosures from power-semi names are the tell — this is when the theme either shows up in numbers or gets reclassified as marketing. Visser's Marvell/optical leg rides the same Rubin ramp on the data side.
If NVIDIA's architecture wins, the DC data center becomes the second act of the EV power-silicon supply chain and the power layer compounds through the 100-GW-scale build Jensen projects. If Patel is right, power gets solved, the bottleneck rotates back to fabs in 2027, and the scarcity premium moves with it. Both can be partially true — volume up, margins down is the trailing-edge default.
The affected map
| Name | Short | Medium | Long | Why |
|---|---|---|---|---|
| NVDANVIDIA · Buy — Tactical, FV $245 | ▲ tailwind | ▲▲ strong | ▲ tailwind | The architecture's author. 800 VDC isn't a product NVIDIA sells — it's the moat-deepener: every partner in three tiers now builds to NVIDIA's reference design, and performance-per-watt ("performance-per-watt is revenue" — Huang) is the metric the whole spec optimizes. Kyber timeline is the near-term swing. |
| VRTVertiv · Hold, FV ~$340 | · contested | ▲▲ strong | ▲▲ strong | Named in the infrastructure tier from day one; the perimeter-conversion and DC-busway socket is its home turf. Contested short-term because the AI premium is rich and jumpy — an ~11% single-day drop in June on no news was pure positioning. The medium-term order book is the thing to watch. |
| ETNEaton · Hold, FV ~$415 | ▲ tailwind | ▲▲ strong | ▲ tailwind | Same tier as Vertiv with a broader electrical-equipment base — less AI-pure, less AI-fragile. The 800-VDC perimeter plus the grid-side buildout (transformers, switchgear) both land in its lap. |
| GEVGE Vernova · Hold, FV ~$1,130 | ▲ tailwind | ▲▲ strong | ▲ tailwind | Added to NVIDIA's infrastructure tier in October — and the company Jensen names first when he says "energy is the bottom layer of AI." Gas turbines, grid equipment, and now a socket in the DC conversion story. The most direct "Jensen said so" beneficiary in our covered set. |
| MPWRMonolithic Power · Hold, FV ~$1,385 | ▲ tailwind | ▲▲ strong | · contested | Named silicon partner and the highest-quality US-listed power-stage franchise — already the vertical-power incumbent on NVIDIA boards. Long-term contested for the same reason it's excellent: premium multiple on content-per-GPU growth that competitors (and China's trailing-edge push) will attack. |
| TXNTexas Instruments · Hold, FV ~$270 (below price) | · neutral | ▲ tailwind | · contested | Named partner with the broadest analog/power catalog — but this is exactly the trailing-edge terrain where Patel says Chinese power chips are "matching and undercutting TI." The 800-VDC socket is real; the margin durability is the open question. Our FV sits below the current price. |
| ONonsemi · Hold, FV ~$95 | · neutral | ▲ tailwind | ▲ tailwind | The clearest EV-to-datacenter crossover: its SiC franchise was built for 800-volt cars, and this architecture points the same silicon at AI factories just as EV demand wobbles. A named partner where the new socket could matter to a depressed base — if SiC wins share in the rectifier/converter stages. |
| MRVLMarvell · Hold, FV ~$250 | · contested | ▲▲ strong | ▲ tailwind | Visser's named pick on this platform — the optical leg of the same Rubin story ("Vera Rubin, which drives optical demand, hasn't started"). Huang agrees on the physics ("the copper wall is moving into the rack"). Priced for a lot of it already; a Kyber slip delays the surprise Visser is waiting for. See also LITE and COHR on the same wave. |
| FLEXFlex · Hold, FV ~$117 (below price) | · neutral | ▲ tailwind | · contested | Named in the power-systems tier (sidecars, shelves, embedded power). Real socket, contract-manufacturer economics — the volume shows up before the margin does. Our FV is below the current price after a strong run. |
| NVTSNavitas · Avoid, FV ~$7.50 (below price) — dive added 2026-07-19 | · watch | · contested | · contested | The most-leveraged pure play: NVIDIA-selected (May 2025, 8-K filed), honored at NVIDIA's 2026 partner ceremony, demoing a single-stage 800V→6V GaN board that skips the 48V bus entirely. But one of 14+ silicon partners, revenue −45% in 2025 with negative gross margin in two of the last three quarters, +53% dilution since 2022, and insiders sold into the $28-32 May spike (Form 4s). Our full deep dive says Avoid — the theme is real, the instrument isn't. |
| STMSTMicroelectronics · Hold, FV ~$65 — dive added 2026-07-19 | · neutral | ▲ tailwind | · contested | Named silicon partner with a real SiC franchise and its own 800-VDC release (March 2026) — and the name Patel's China warning cites directly. It WAS the cheap way in at $21-30; at ~$62 (+193% off the low) the recovery is largely pre-paid. Our dive: Hold, with a $50-54 add zone. |
| Infineon · Delta · Schneiderforeign-listed partners · context | ▲ tailwind | ▲ tailwind | · contested | The non-US remainder of the map: Infineon is the SiC scale leader; Delta is the PSU-tier heavyweight; Schneider anchors the infrastructure tier with Eaton/Vertiv. Same thesis, same trailing-edge margin question, listed elsewhere. |
What we're watching (the falsifiers)
- Does Kyber ship in 2027? The SemiAnalysis-reported slip to 2028 (midplane yields) vs NVIDIA's "roadmap is intact." A confirmed slip delays the entire 800-VDC production ramp — and Visser's Marvell earnings-surprise leg with it.
- Does 800-VDC revenue actually get disclosed? The theme graduates from story to number the first time a power-semi or infrastructure partner breaks out attributable revenue or design-win content per rack. Until then, no source we verified can size who gets paid.
- Whose voltage do facilities standardize on? OCP's Diablo spec now reads "±400 or 800 VDC." Watch what Google, Meta and Microsoft actually pour concrete for — NVIDIA racks can win while NVIDIA's facility voltage doesn't.
- Patel's 2027 rotation. If ~30 GW of new power lands next year and the binding constraint swings back to fabs, the power layer's scarcity premium — the reason this is the last Very Bullish topic on our board — compresses. Our own signal going off Very Bullish on datacenter power would be the panel telling us first.
- China's trailing-edge push. Power chips are not leading-edge silicon. Evidence of Chinese power-semi content in Western DC infrastructure — or price cuts at TI/ST/onsemi attributed to it — tests the margin half of the thesis.
Sources
Expert claims are distilled from the shows and voices in the Synthos knowledge base and attributed to the named speaker + date above: Jordi Visser (May 4, Jun 17, 23, Jul 5, 2026), Jensen Huang (Jun 1–2, 10, 2026 — management lane), Dylan Patel / SemiAnalysis (Feb 3–5, 2026), Raoul Pal (Jun 11, 2026), Anthony Pompliano (Jul 16, 2026). Technical and ecosystem claims were verified against primary sources: NVIDIA's 800 VDC announcement (May 2025), NVIDIA's OCP ecosystem update (Oct 2025), gigawatt AI factories post, Navitas's announcement and GTC 2026 demo release, Google Cloud's Mt Diablo / ±400 VDC post, the OCP Diablo 400 spec, and trade press (DataCenterDynamics, Power Electronics News). Kyber-slip report: SemiAnalysis via CNBC, July 6, 2026 — disputed by NVIDIA. All efficiency/copper/TCO figures are NVIDIA or partner marketing claims, labeled as such. Conviction labels are from our topic-sentiment signal (window ending 2026-07-14); company verdicts and fair values from our deep dives dated 2026-07-03. For how voices are weighted, see our methodology and verified voices.
Go deeper: the NVIDIA deep dive · Vertiv · Eaton · GE Vernova · Monolithic Power · Texas Instruments · onsemi · Marvell · Flex · Navitas · STMicroelectronics · the companion piece on the AI build-out at large · the Frontier Gap Board.