Before You Read: What You’re Looking At
This Is the Actual Output
Sawbuck Intelligence produces 10-module Edge Reports on demand for any market you’re evaluating. The reports below — on Wolfspeed SiC in terrestrial AI data center infrastructure and its orbital compute extension — are real reports generated through our system. Not summaries. Not edited excerpts. The full output.
The format is consistent across every report: executive summary with a clear thesis, 10 implementation areas each with a logic-gated signal assessment, structured risk/reward tables with current state and 12-month catalysts, hard-stop callouts for constraints that money cannot solve, and a strategic conclusion mapping where margin concentrates versus where risk concentrates.
No metaphors. No filler. Capacity utilization trajectories, capex commitments, technical hard stops. If it’s not actionable, it’s not in the report.
The Wolfspeed SiC / AI data center case study was selected because it touches every dimension the Edge Report format was designed to handle: a constrained supply chain with sequential bottlenecks, multiple competing players at different layers, significant capital commitments creating revenue visibility, technical hard stops that are physics-based rather than capital-based, and a 12–24 month catalyst window where positioning decisions matter. It is also a market where the gap between headlines (“AI needs more power”) and actionable supply chain intelligence (“which specific device, at what voltage tier, supplied by whom”) is large enough to be commercially meaningful.
If you want this format applied to your market — competitive intelligence for a segment you’re evaluating, a TAM analysis for an investment thesis, or a full supply chain map for a hardware or semiconductor play — the order form is at the bottom of this page. $10. Under 10 minutes.
Scaling 800V+ Infrastructure in the AI Data Center Supply Chain
Executive Summary
AI workloads have permanently broken the power and cooling assumptions that defined cloud data centers from 2010 to 2023. The migration from 400V to 800V+ DC bus architectures is no longer a roadmap discussion — it is an engineering requirement being executed in real time, with hyperscaler capital commitments exceeding $600 billion for 2026 alone.
This transition creates a discrete set of winners and bottlenecks across the supply chain. Wide-bandgap semiconductors — specifically silicon carbide (SiC) — are the critical enabling technology across multiple layers simultaneously: rack power conversion, grid interconnection, solid-state transformers, and direct-to-chip thermal management. Wolfspeed’s restructured balance sheet, paired with its Gen 4 TOLT portfolio and first-commercially-available 10kV SiC MOSFET, positions it as the foundational case study for how this technology moves from prototype to gigawatt-scale deployment.
The bottlenecks are real and sequential. Wafer supply, advanced packaging throughput (CoWoS), grid interconnection queues, and local permitting are each capable of independently capping deployment velocity. Understanding which constraint bites first — and where in the supply chain the margin concentrates — is the analytical task this report addresses.
10 Module Signal Summary
Each module receives a signal assessment based on a logic-gated evaluation of capex trajectory, capacity utilization, key players, and technical hard stops. Full module detail is available in the complete terrestrial Edge Report.
Wolfspeed’s January 2026 breakthrough in 300mm SiC substrate technology enables step-function cost reduction but is not yet in production. Near-term supply runs through the Mohawk Valley fab (150mm) with the JFK expansion (200mm) still ramping. The U.S.-based manufacturing chain has become a differentiated procurement argument as hyperscalers seek supply chain resilience. Hard stop: negative gross margins persist; 300mm volume production is a 2027–2028 event.
The migration from 400V to 800V+ DC bus architecture is an engineering requirement, not a roadmap item. Wolfspeed’s TOLT portfolio (top-side cooled, Gen 4 MOSFET) is purpose-built for this migration. Nvidia’s 800VDC standard with Vertiv for Vera Rubin is the architectural forcing function — once the dominant chip platform mandates 800V, the entire downstream supply chain must conform.
Vertiv: $15B backlog, 252% YoY organic order growth, 2026 production sold out. At AI rack densities exceeding 120kW+, direct-to-chip liquid cooling is the only viable thermal management path. This is not a cycle — it is an architectural constraint. Risk: trailing P/E ~70x prices near-perfect execution; any hyperscaler capex pullback creates multiple compression.
TSMC CoWoS capacity is the current binding constraint for AI GPU supply. Applied Materials (AMAT) guided semiconductor equipment business to grow over 20% in calendar 2026. TSMC’s CoWoS is sold out through 2026. Every AI GPU requiring 2.5D integration runs through a Taiwan-concentrated bottleneck that 18–24 months of new capacity announcements will not resolve until 2027.
In high-EMI 800V+ environments, copper-based interconnects face fundamental limitations. Marvell’s 1.6 Tb/s PAM4 optical DSP (Ara T) is the first-to-market solution for the transition to optical scale-out networking. Nvidia’s $2B equity stake in Marvell (NVLink Fusion, March 2026) is the external validation. Custom ASIC revenue projected to scale from $1.5B to $9–$11B in AI ASIC revenue by 2026.
The “edge” narrative is being redefined by centralization. Cerebras’s $20B multi-year OpenAI contract for 750MW of compute is not edge — it is a centralized AI supercomputer. The IPO at $56.4B fully diluted valuation (68% pop on debut) validates premium pricing for specialized inference infrastructure. Risk: G42/OpenAI customer concentration with lock-up expiration H2 2026.
Wolfspeed’s 10kV SiC MOSFET — the industry’s first commercially available device at this voltage — enables solid-state transformers (SSTs) that connect directly to medium-voltage grids without intermediate step-down. Key claims: 99% conversion efficiency, 30% lower system cost vs. silicon IGBT, 300% higher power density. This is the longest-duration opportunity in this report — 2026 design wins appear as 2027–2028 revenue.
The semiconductor supply chain is bifurcating into two ecosystems. China’s domestic AI chip self-sufficiency ratio reached ~41% by 2025 (projected 86% by 2030). Wolfspeed’s CFIUS clearance (Renesas equity, January 2026) validates the national security positioning of U.S.-based SiC manufacturing. Taiwan CoWoS concentration remains unresolved until 2027.
Global data center power exceeding 1,000 TWh by end 2026 equals Japan’s entire annual consumption. Grid interconnection queues, transformer manufacturing backlogs (12–18 months), and 27 states advancing their own data center legislation are constraints that capital cannot overcome. A greenfield AI factory in a constrained market faces 18–36 months from site selection to full power delivery. This is the primary governor on AI deployment velocity 2026–2028.
Camtek and KLA occupy the inspection layer that scales with every technology complexity increase. SiC crystal defect characterization, 2.5D packaging inspection, and CoWoS yield management each require specialized tooling that did not exist at production scale five years ago. As SiC moves to 300mm, inspection requirement per wafer increases non-linearly — a structural volume multiplier independent of which chip architecture wins.
The Binding Sequence — Terrestrial
Technology does not limit this buildout. Capital does not limit this buildout. The binding constraints are, in order of near-term severity: (1) advanced packaging throughput (CoWoS), (2) grid interconnection queue and transformer supply, (3) permitting and regulatory timelines, and (4) SiC wafer production capacity at 200mm scale.
The companies that win are not necessarily those with the best technology — they are those whose technology solves the binding constraint at the moment it becomes the bottleneck. Vertiv’s cooling solutions are winning now because thermal management is the current acute constraint. Marvell’s optical interconnects are the next-cycle winner as bandwidth becomes the bottleneck. Wolfspeed’s 10kV SiC wins after that, as grid-scale SST deployments accelerate. Sequence matters more than absolute quality.
| Company | Position | Signal | Primary Risk |
|---|---|---|---|
| Vertiv (VRT) | Thermal management + 800V power; $15B backlog, 2026 sold out | ▲ HIGH CONVICTION | Trailing P/E ~70x; any hyperscaler capex pause = multiple compression |
| Marvell (MRVL) | Optical DSP + custom ASICs; supplying all 5 U.S. hyperscalers | ▲ HIGH CONVICTION | 3nm wafer allocation constraint; TSMC N3 bottleneck limits ASIC output |
| AMAT | Process equipment leader at every technology inflection | ▲ HIGH CONVICTION | WFE spending cycle timing; guided >20% semiconductor equipment growth |
| Wolfspeed (WOLF) | SiC substrate + 10kV MOSFET + TOLT; foundational but financial risk | ◆ WATCH | Negative gross margins; 300mm production is 2027–2028 event |
| Cerebras (CBRS) | Specialized inference infrastructure; $20B OpenAI contract | ◆ WATCH | Customer concentration; lock-up expiration H2 2026 |
The “Vacuum-Ready” Supply Chain: Scaling 800V–10kV SiC for Orbital AI Clusters
Executive Summary
The convergence of three forces — unlimited orbital solar energy, plummeting launch costs, and SiC’s unique material properties (radiation hardness, thermal conductivity, efficiency at extreme temperatures) — makes space-based AI compute an engineering problem rather than a physics problem. Google’s Project Suncatcher and Anthropic’s expressed interest in “multiple gigawatts of orbital AI compute capacity” with SpaceX signal that orbital data centers have crossed from speculation to capital allocation.
SiC is the enabling material at every critical junction: power conversion in vacuum where waste heat cannot convect away, thermal management through radiative cooling where top-side heat dissipation is the only viable path, radiation survival where wide bandgap provides inherent Single Event Burnout resistance, and mass optimization where every kilogram of power electronics displaces a kilogram of compute payload.
The economics remain unproven. SpaceX’s own pre-IPO filing acknowledges “significant technical complexity and unproven technologies.” But the capital allocation is real, and the supply chain decisions being made today will determine who captures margin if orbital compute scales.
Where the Physics Amplifies the Terrestrial Thesis
Orbital Module Signal Summary
5 of 10 orbital modules rated bullish (physics advantages confirmed), 4 neutral (depends on cost curve), 1 cautious (regulatory vacuum). Full module detail: The “Vacuum-Ready” Supply Chain →
| Module | Orbital Signal | Key Constraint |
|---|---|---|
| SiC Substrate Dominance | ▲ Structural Advantage | 10kV space-grade qualification not yet initiated; multi-year process |
| Power Conversion in Vacuum | ▲ Strongest Signal | Radiative cooling = the only thermal path; 99% efficiency is not optional |
| SWaP-C Optimization | ▲ Structural Tailwind | Launch cost must reach <$500/kg for economics to close |
| Radiation Survival | ◆ Watch Carefully | No 10kV SiC flight heritage; SEB rates at this voltage theoretical, not measured |
| Inter-Satellite Optical Links | ▲ Enabling Technology | 1.6 Tbps demonstrated in lab over 200m; orbital deployment adds pointing challenges |
| Sovereign Supply Chain | ◆ Bifurcating | ITAR compliance for orbital AI compute not yet codified |
| Solid-State Launch Switching | ▲ Structural Tailwind | Paschen minimum arcing eliminates mechanical switches during ascent; SiC is the only viable solution |
| Advanced Packaging Synergy | ◆ Early Stage | -150°C to +150°C thermal cycling survivability for CoWoS-style packages unproven |
| Orbital TCO | ◆ Speculative but Directional | Economics don’t close today; crossover requires $200/kg launch + 99% SiC efficiency realized |
| Regulatory & Kinetic Risk | ▼ Uncharted Territory | No regulatory framework for commercial orbital data centers; FCC/FAA/NOAA/DoD jurisdiction overlap |
What the Combined Analysis Tells You
The terrestrial and orbital reports are connected by a single underlying question: which supply chain nodes capture disproportionate margin as AI compute scales from megawatt to gigawatt to terawatt-scale deployments? The answer, across both environments, points to the same set of companies — but on different timelines and with different risk profiles.
High-Conviction Positions
- Vertiv (VRT) — thermal management and power infrastructure. $15B backlog. 2026 production sold out. Nvidia Vera Rubin lock-in.
- Marvell (MRVL) — optical interconnect + custom ASICs for all five U.S. hyperscalers. NVLink Fusion + Celestial AI CPO.
- Applied Materials (AMAT) — process equipment at every technology inflection. >20% equipment growth guided 2026.
- Wolfspeed (WOLF) — sole supplier of commercial 10kV SiC MOSFET. TOLT top-side cooling. U.S.-sovereign supply chain. CFIUS cleared.
- Grid infrastructure (GE Vernova, Siemens Energy) — power delivery to AI factories regardless of which chip wins.
Monitored Risks
- Wolfspeed financial stability — negative gross margins. 300mm production is a 2027 event. Any customer concentration = margin pressure before AI revenue ramp.
- CoWoS single-source risk — Taiwan concentration for AI GPU packaging unresolved until TSMC Arizona operational (2027).
- Permitting timelines — 18–36 months from hyperscaler commitment to power delivery. Not solvable with capital alone.
- Vertiv valuation — trailing P/E ~70x. Any hyperscaler capex pullback = multiple compression risk from a demanding entry.
- Orbital compute economics — don’t close today. 5–10x cost premium vs. terrestrial until launch costs hit $200/kg and SiC 99% efficiency is proven at scale.
- SiC qualification cycles — 12–24 months from design win to revenue. Demand signal today = production revenue 2027–2028.
The Investment Thesis, Compressed
AI hardware captures the headlines. Infrastructure captures the cash flows. The companies anchoring this report — Wolfspeed, Vertiv, Applied Materials, and Marvell — sit at the physical layer of AI that cannot be virtualized, cannot be upgraded with a software patch, and cannot be sourced from a different vendor without 18–24 months of qualification work.
The migration to 800V+ architectures is not a question of whether — it is a question of sequence, timing, and which supply chain nodes capture disproportionate margin at each inflection. SiC’s material advantages (efficiency, thermal conductivity, radiation hardness) are amplified, not diminished, as compute density increases and the deployment environment becomes more extreme. That is a structural tailwind that compounds with every GPU generation increase.
The terrestrial thesis is executable now. Vertiv’s backlog provides revenue visibility. Marvell’s hyperscaler relationships provide demand certainty. The orbital thesis is positioning now for a 2028–2030 inflection. The companies making supply chain decisions today — which SiC devices to qualify, which packaging to adapt for thermal cycling, which launch vehicle to design around — will determine who captures margin if orbital compute scales. The “if” is the key qualifier. But the capital allocation signals from Google, Anthropic, and SpaceX are real, even if the revenue is not.
Read the Full Reports
The summaries above are exactly that — summaries. Each full Edge Report contains the complete per-module analysis, risk/reward tables, hard-stop callouts, and sourced data. Both are available in full, free to read.
Your Market, Not Ours
This Wolfspeed report is what Sawbuck Intelligence produces on demand — for any market you’re evaluating. Describe your company and the competitive landscape you need mapped. Receive the same format: 10 modules, logic-gated signals, risk/reward matrix, hard-stop identification. In under 10 minutes. For $10.
AI infrastructure investors, power electronics analysts, semiconductor supply chain operators — this is the analytical layer that sits between headlines and positions.
Order Your Report — $10