9 Ventures
Research-driven coverage of the physical infrastructure supporting AI: power architecture, rack-level scale, and the suppliers positioned to benefit from rising datacenter power density.
Past bets that played out
Flagship analysis argues AI datacenter rack power density is rising from tens of kW to hundreds of kW and beyond, forcing a transition to 800V DC distribution at the rack/facility level. The work cites NVIDIA’s 800V DC requirements for next-generation “AI factories,” the Open Compute Project Mt. Diablo (Diablo 400) ±400VDC/800VDC spec, and NVIDIA Rubin Ultra NVL576 “Kyber” (mid-2027) and “Feynman” (2028) racks as industry milestones. The research highlights a potential ~10x rise in power-infrastructure cost per rack and durable margin/BOM share gains for semiconductor content in 800V systems.
Post argues $PENG delivered a record Q3 FY26 with broad-based beat, expanding margins, and raised guidance; management’s preliminary FY27 view is characterized as conservative (“sandbag”), implying upside to estimates. Mix shift toward AI-driven businesses (Memory + non-hyperscaler AI infrastructure) is highlighted, with backlog building into Q4.
Post argues that AI chip performance is now constrained by interconnect/packaging (data movement), not raw compute. Conventional bump/wire methods can’t deliver required bandwidth/power efficiency. Thermo-compression bonding (TCB) is the current high-volume solution enabling advanced AI packages. Implicit thesis: the market is underestimating how early we are in the adoption curve for next-gen approaches like hybrid bonding (title), suggesting a multi-year runway for advanced packaging capex and
Post argues that AI chip performance is now constrained by interconnect/packaging (data movement), not raw compute. Conventional bump/wire methods can’t deliver required bandwidth/power efficiency. Thermo-compression bonding (TCB) is the current high-volume solution enabling advanced AI packages. Implicit thesis: the market is underestimating how early we are in the adoption curve for next-gen approaches like hybrid bonding (title), suggesting a multi-year runway for advanced packaging capex and
What this channel is watching now
Primary focus: the evolution of AI datacenter power architectures and their downstream effects on power-electronics suppliers and semiconductor content. Top tickers mentioned in coverage: NVDA, MSFT, GOOGL, META.
Latest videos and market context
Recent YouTube posts from this source. Create an account to unlock live alerts and the full research trail.
$FCEL: The Counterparty Picture Just Got a Lot Cleaner
Post argues FCEL’s counterparty risk improved because “Fit Energy” (a CEPA counterparty/partner) appears to be connected to a credible (“legit”) data center player and can plausibly source ~380 MW of U.S. data center sites. Implies reduced execution/credit risk and improved viability of FCEL’s data-center-related pipeline.
$FCEL Signs 380 MW Deal With Fit Energy. Here’s What To Actually Make Of It.
Post discusses FuelCell Energy (FCEL) filing an 8-K (June 22, 2026) announcing a Capital Equipment Purchase Agreement (CEPA) with Fit Energy USA LP for up to 380 MW of carbonate fuel cell block systems (2.5 MW blocks), delivered in four phases, intended for baseload power for data centers. The author frames it as potentially tape-moving but emphasizes there is “nuance” and unspecified due-diligence items (no economics, timing, financing, or cancellation terms provided in the excerpt).
A Physical AI Darling
Post argues public markets are underpricing an unnamed semiconductor supplier positioned at the intersection of automotive ADAS sensing/processing, humanoid robotics hardware stacks, and quantum photonics. Core claim: humanoid robotics is shifting from a software bottleneck to a hardware/supply-chain bottleneck, and ADAS-proven silicon (vision processors, radar, LiDAR, sensor fusion) transfers to robots with low incremental engineering cost. Mentions a “hottest robotics IPO of the year” coming in weeks as a potential attention/catalyst, but provides no company/ticker identifiers.
$PENG Q3 FY26 Review: The Sandbag Gets Bigger
Post argues $PENG delivered a record Q3 FY26 with broad-based beat, expanding margins, and raised guidance; management’s preliminary FY27 view is characterized as conservative (“sandbag”), implying upside to estimates. Mix shift toward AI-driven businesses (Memory + non-hyperscaler AI infrastructure) is highlighted, with backlog building into Q4.
Proof-backed call history
Recent published items include deep-dive analysis on 800V DC at the rack level, a title-only note suggesting a carbon-capture angle on $FCEL, and a high-level piece on favorite names across layers of the Physical AI trade. Coverage is largely thematic and infrastructure-oriented; some posts are high-level or paywalled and do not always include single-stock actionable picks in public excerpts.
Post argues FCEL’s counterparty risk improved because “Fit Energy” (a CEPA counterparty/partner) appears to be connected to a credible (“legit”) data center player and can plausibly source ~380 MW of U.S. data center sites. Implies reduced execution/credit risk and improved viability of FCEL’s data-center-related pipeline.
Post discusses FuelCell Energy (FCEL) filing an 8-K (June 22, 2026) announcing a Capital Equipment Purchase Agreement (CEPA) with Fit Energy USA LP for up to 380 MW of carbonate fuel cell block systems (2.5 MW blocks), delivered in four phases, intended for baseload power for data centers. The author frames it as potentially tape-moving but emphasizes there is “nuance” and unspecified due-diligence items (no economics, timing, financing, or cancellation terms provided in the excerpt).
Post argues public markets are underpricing an unnamed semiconductor supplier positioned at the intersection of automotive ADAS sensing/processing, humanoid robotics hardware stacks, and quantum photonics. Core claim: humanoid robotics is shifting from a software bottleneck to a hardware/supply-chain bottleneck, and ADAS-proven silicon (vision processors, radar, LiDAR, sensor fusion) transfers to robots with low incremental engineering cost. Mentions a “hottest robotics IPO of the year” coming i
Post argues $PENG delivered a record Q3 FY26 with broad-based beat, expanding margins, and raised guidance; management’s preliminary FY27 view is characterized as conservative (“sandbag”), implying upside to estimates. Mix shift toward AI-driven businesses (Memory + non-hyperscaler AI infrastructure) is highlighted, with backlog building into Q4.
Post argues that Nvidia’s next-gen “AI factories” require a shift from legacy AC distribution (415/480VAC) toward 800VDC distribution for data centers due to extreme rack power density (claims ~370kW/rack for Vera Rubin vs ~120kW/rack on Hopper). This implies a multi-year capex cycle in high-voltage DC power distribution equipment and a potential mispricing of key suppliers, but the post does not name the alleged “backbone supplier.” Only Nvidia and Siemens are explicitly referenced as architect
Post argues that Nvidia’s next-gen “AI factories” require a shift from legacy AC distribution (415/480VAC) toward 800VDC distribution for data centers due to extreme rack power density (claims ~370kW/rack for Vera Rubin vs ~120kW/rack on Hopper). This implies a multi-year capex cycle in high-voltage DC power distribution equipment and a potential mispricing of key suppliers, but the post does not name the alleged “backbone supplier.” Only Nvidia and Siemens are explicitly referenced as architect
Post argues Silicon Motion (SIMO) delivered a major earnings beat and raised profitability trajectory: management now expects to exit 2026 at >30% operating margin versus author’s prior 2028 28% base case. Highlights: revenue beat vs guidance, GM >50%, OM above guide, EPS above consensus; author says thesis intact and ramp faster than modeled.
Post summarizes a management meeting with Broadwind ($BWEN) CEO/CFO. Key points: strong demand/backlog commentary, confidence ("not hedging"), focus on growth + margin expansion, and explicit target to return to historical best margins in Gearing and Industrial Solutions, framed as a long-term "Supercycle" thesis. No explicit valuation, numbers, guidance, or near-term catalyst is provided, so actionability is moderate.
Post argues AI datacenter rack power density is rising sharply (40kW to 600kW+), forcing a shift to 800V DC distribution at the rack/facility level. Cites NVIDIA as mandating 800V DC for next-gen “AI factories” and Open Compute Project’s Mt. Diablo (Diablo 400) spec for ±400VDC bipolar/800VDC. Mentions NVIDIA Rubin Ultra NVL576 “Kyber” rack arriving mid-2027 as first production 800VDC >600kW/rack, with “Feynman” in 2028 expected >1MW/rack. Claims power-infrastructure cost per rack could rise ~10
Post argues AI datacenter rack power density is rising sharply (40kW to 600kW+), forcing a shift to 800V DC distribution at the rack/facility level. Cites NVIDIA as mandating 800V DC for next-gen “AI factories” and Open Compute Project’s Mt. Diablo (Diablo 400) spec for ±400VDC bipolar/800VDC. Mentions NVIDIA Rubin Ultra NVL576 “Kyber” rack arriving mid-2027 as first production 800VDC >600kW/rack, with “Feynman” in 2028 expected >1MW/rack. Claims power-infrastructure cost per rack could rise ~10
Post argues AI datacenter rack power density is rising sharply (40kW to 600kW+), forcing a shift to 800V DC distribution at the rack/facility level. Cites NVIDIA as mandating 800V DC for next-gen “AI factories” and Open Compute Project’s Mt. Diablo (Diablo 400) spec for ±400VDC bipolar/800VDC. Mentions NVIDIA Rubin Ultra NVL576 “Kyber” rack arriving mid-2027 as first production 800VDC >600kW/rack, with “Feynman” in 2028 expected >1MW/rack. Claims power-infrastructure cost per rack could rise ~10
Post argues AI datacenter rack power density is rising sharply (40kW to 600kW+), forcing a shift to 800V DC distribution at the rack/facility level. Cites NVIDIA as mandating 800V DC for next-gen “AI factories” and Open Compute Project’s Mt. Diablo (Diablo 400) spec for ±400VDC bipolar/800VDC. Mentions NVIDIA Rubin Ultra NVL576 “Kyber” rack arriving mid-2027 as first production 800VDC >600kW/rack, with “Feynman” in 2028 expected >1MW/rack. Claims power-infrastructure cost per rack could rise ~10
About this channel
9 Ventures focuses on research into the physical and infrastructure layers that enable advanced AI deployments. Analyses emphasize technology specs, deployment timelines, and cost structures—especially where hardware economics create durable supplier advantages.
Most recognized assets
Unlock the full track record
Follow 9 Ventures for ongoing analysis of AI infrastructure, rack-level power trends, and supplier implications. Contact or subscribe to access full write-ups and paywalled stock-specific recommendations referenced in some posts.
6 more thesis calls are available after sign-up.