Report date: 2026-08-17 (Asia/Shanghai) Market-data freeze: 2026-08-14 close, the latest completed U.S. trading session Research horizon: 12–24 months, with evidence gates through 2027 Status: Idea triage, not a final security recommendation Source posture: Company filings, investor relations materials and technical disclosures for operating facts; Yahoo Finance via yfinance for the frozen market screen.
The investable insight is narrower—and more useful—than “buy whatever Jensen mentions.” NVIDIA's purchase commitments, capacity rights, equity investments and production milestones are unusually strong leading indicators of where AI infrastructure will encounter a physical bottleneck. The repeatable strategy is to follow that procurement footprint, then buy only where the expected revenue and margin conversion are not already overwhelmed by valuation.
That distinction matters now. NVIDIA's March 2026 agreements with Lumentum and Coherent are hard evidence, not keynote rhetoric. Yet the optical basket has already repriced violently. The better prospective 2027 hunt is the 800 VDC power architecture, where deployment timing is explicit but supplier economics remain unsettled. Physical AI is a credible long-duration theme, but today it is closer to an ecosystem roadmap than a supplier-level procurement signal.
| Priority | Candidate / expression | Current actionability | Why it survives the first screen | First rejection |
|---|---|---|---|---|
| 1 | ON Semiconductor (ON) | Advance to full diligence | Direct 800 VDC collaboration, depressed industrial-cycle base and the least demanding headline multiple among named 800 VDC beneficiaries | NVIDIA ecosystem participation is not an order; absolute AI data-center revenue is undisclosed, and the proposed Synaptics acquisition adds integration and dilution risk |
| 2 | Eaton (ETN) | Advance as lower-purity, higher-quality alternative | Data-center orders and backlog already support the power-bottleneck thesis; broad installed base reduces single-program risk | Less pure 800 VDC upside, and the stock already embeds a meaningful data-center premium |
| 3 | COHR over LITE | Advance as relative-value research | Both have direct NVIDIA purchase commitments, but COHR screens near 9.0x sales / 23.4x forward P/E versus LITE near 27.2x / 28.1x | Lumentum could possess superior 200G/CPO mix, yield or pricing; headline multiples do not normalize cycle earnings or purchase accounting |
| 4 | NVIDIA (NVDA) | Core exposure / benchmark | Owns the architecture and can shift spend across suppliers; screens at roughly 17.6x provider forward earnings despite the largest direct profit pool | Consensus earnings may be extrapolating exceptional margins and capex growth; geopolitical and customer-concentration risk remain material |
| 5 | Vertiv (VRT) | Watch for evidence or price reset | Explicit 800 VDC product roadmap and participation in early large AI-factory designs | One-year return of roughly 121% and about 32.3x forward P/E leave less room for schedule slips |
| — | LITE, AAOI, AXTI | Valuation-gated watchlist | Optical/CW/compound-semiconductor demand is real | One-year gains of roughly 699%, 572% and 3,882% mean the market is no longer dismissing the bottleneck |
| — | Navitas (NVTS) | Reject for now | Technically relevant 800 VDC demonstrations and NVIDIA collaboration | Roughly 103x trailing sales, shrinking revenue and large operating losses without material program revenue |
Bottom line: the cleanest new research idea is ON as an underappreciated 800 VDC option, balanced by ETN as the quality expression. The cleanest relative-value question is COHR versus LITE. The highest-quality implementation may still be the obvious one—owning NVIDIA itself rather than paying extreme multiples for second-order beneficiaries.
| Signal | Information content | Current examples | Investment use |
|---|---|---|---|
| 1. Purchase commitment + capacity right + equity capital | Highest | NVIDIA–Lumentum; NVIDIA–Coherent | Confirms a strategic bottleneck and customer urgency; still requires valuation and margin work |
| 2. Product in production / named deployment | High | Spectrum-X Ethernet Photonics switches in production | Moves demand from roadmap to revenue-conversion monitoring |
| 3. Qualified partner + dated architecture transition | Medium-high | 800 VDC ecosystem; Kyber / Rubin Ultra timing in 2027 | Best hunting ground before program economics are fully known |
| 4. Product roadmap or engineering collaboration | Medium | Supplier reference designs and demonstrations | Watch for design win, order, backlog and revenue disclosure |
| 5. Theme declaration / ecosystem blueprint | Low for stock selection | Physical AI training and simulation blueprint | Defines a research universe, not a buy list |
The key analytical error is to treat every rung as equivalent. NVIDIA can validate a technology direction without guaranteeing that every named partner will earn attractive returns on capital.
Your cited returns are directionally correct, although exact percentages change with the chosen start date. Using unadjusted closes from 2025-08-15 to 2026-08-14, the frozen screen produces the following:
| Ticker | 2026-08-14 close | One-year price return | Market cap | Price / sales | Forward P/E | Read-through |
|---|---|---|---|---|---|---|
| NVDA | $225.16 | +24.8% | $5.45T | 21.5x | 17.6x | Architecture owner and benchmark |
| LITE | $926.14 | +699.4% | $82.1B | 27.2x | Direct NVIDIA commitment; expectations now extreme | |
| AAOI | $150.28 | +571.9% | $12.7B | 21.3x | Optical demand exposure, but negative trailing EBITDA and free cash flow | |
| COHR | $325.83 | +248.9% | $63.7B | 9.0x | Direct NVIDIA commitment; broader portfolio and lower sales multiple | |
| AXTI | $81.64 | +3,882.4% | $5.35B | 42.7x | Compound-semiconductor optionality; small revenue base and negative free cash flow | |
| ON | $82.66 | +61.8% | $32.2B | 5.2x | 18.3x | 800 VDC option with cyclical recovery exposure |
| ETN | $451.51 | +28.6% | $175.4B | 5.8x | 28.2x | Established power-system beneficiary |
| VRT | $293.84 | +120.8% | $113.1B | 9.9x | 32.3x | Direct data-center power/cooling exposure |
| NVTS | $14.45 | +108.8% | $3.77B | 103.3x | Collaboration has outrun commercial evidence |
Methodology: one-year return equals (2026-08-14 close / 2025-08-15 close) − 1; it excludes dividends and uses unadjusted daily closes. Market cap and forward multiples are Yahoo Finance provider fields captured after the 2026-08-14 close. Forward P/E is a consensus-derived snapshot, not a valuation model, and can change with estimates or provider methodology.
The conclusion is not that the market always dismisses Jensen Huang. It is that the market often underestimates the speed and magnitude of a bottleneck before order evidence appears, then overcorrects once the bottleneck becomes consensus. LITE, AAOI and AXTI are now evidence of the second phase, not proof that the first phase still exists.
On 2026-03-02, NVIDIA agreed to invest $2 billion in Lumentum, alongside a multi-billion-dollar purchase commitment and capacity rights for advanced laser components. Lumentum said the agreement supports a new U.S. fabrication facility and faster capacity expansion. This is direct evidence that laser capacity—not merely general optical networking—was strategically scarce. Lumentum–NVIDIA announcement filed with the SEC
The same day, NVIDIA agreed to invest $2 billion in Coherent at $256.80 per share, with another multi-billion-dollar purchase commitment and capacity rights spanning advanced lasers and optical-networking products across multiple product families. Coherent–NVIDIA announcement filed with the SEC
Coherent subsequently disclosed that it was expanding indium-phosphide capacity to address an industry-wide shortage. Its investor materials describe plans to double internal InP output by year-end and more than double it again by 2027, with six-inch production supporting EMLs, continuous-wave lasers and photodiodes. Coherent Q3 FY2026 filing Coherent investor presentation
NVIDIA has also moved co-packaged optics from roadmap to production: Spectrum-X Ethernet Photonics switches using CPO and 200 Gb/s SerDes are in production, with CoreWeave, Lambda and Oracle Cloud Infrastructure identified as early adopters. NVIDIA Vera Rubin production announcement
The public evidence proves that the bottleneck exists. It does not prove the terminal economics of each supplier. The remaining edge lies in four questions:
COHR screens at roughly one-third of LITE's sales multiple while its provider forward P/E is also lower. Coherent also has a broader optical and materials portfolio and has laid out staged InP capacity milestones. This creates a credible long COHR / underweight LITE research hypothesis: similar direct NVIDIA validation, materially different expectations.
This is not yet a ready-made pair trade. A valid implementation requires segment-level gross-margin normalization, commitment terms, capex intensity, share count and borrow/catalyst analysis. The decisive disconfirming evidence would be Lumentum demonstrating structurally higher content per system, superior yield or materially faster conversion of contracted capacity into free cash flow.
AAOI and AXTI may benefit from the same spending wave, but their current valuations require exceptional execution from much smaller economic bases. AAOI's trailing revenue was about $596 million at the snapshot, with negative EBITDA and free cash flow. AXTI's trailing revenue was about $126 million, with negative free cash flow. A real industry bottleneck can coexist with a poor security-level payoff when the stock price discounts several years of flawless conversion.
NVIDIA's 800 VDC architecture is intended to move power distribution closer to the rack at much higher voltage, reduce conversion stages and copper, and support megawatt-scale racks. NVIDIA explicitly connects full-scale production to Kyber rack-scale systems and the 2027 Rubin Ultra generation. Its ecosystem includes semiconductor, power-component and system vendors rather than a single exclusive supplier. NVIDIA 800 VDC technical overview
The architecture is real; the equity winners are not settled. This is exactly the phase in which the “NVIDIA leading indicator” framework is most useful.
| Candidate | Exposure proof | Expectations risk | What would make it investable | What would kill it | Next workflow |
|---|---|---|---|---|---|
| ON | Direct NVIDIA collaboration on 800 VDC power solutions; Q1 2026 AI data-center revenue more than doubled YoY and rose more than 30% sequentially | AI dollars, socket content and design-win economics are not separately disclosed; proposed Synaptics transaction complicates the base | Named design wins or backlog, quantified AI revenue/content, gross-margin accretion and evidence the core cycle has turned | Program delay, share loss, weak industrial/auto recovery, or acquisition dilution overwhelms AI upside | Build a sum-of-parts and earnings bridge separating core recovery, 800 VDC optionality and Synaptics |
| ETN | NVIDIA 800 VDC partnership; Q1 presentation cited data-center orders +240% YoY and backlog +44%; $1.5B capacity program | Broad exposure means lower purity; roughly 28x forward earnings already reflects data-center quality | Sustained book-to-bill, backlog conversion and incremental margins without working-capital deterioration | Order cancellations, lead-time normalization, aggressive industry capacity or slower hyperscaler capex | Model Electrical Americas data-center sales, backlog duration, margins and free-cash-flow conversion |
| VRT | Product portfolio planned for H2 2026, aligned with 2027 Rubin Ultra; company says it is in early design phases for several large AI-factory projects | About 32x forward P/E after a 121% one-year gain; “design phase” is not backlog | Named project awards, orders and revenue timing with margin discipline | Project delays, commoditization, working-capital strain or thermal/power competition | Track order conversion and separate 800 VDC revenue from general AI data-center growth |
| VICR / MPWR | Relevant high-density conversion and power-management capabilities; ecosystem positioning | Roughly 42x and 40x forward earnings respectively in the market screen | Platform-specific content disclosure and production awards | Alternative topology or internal/customer designs reduce addressable content | Technical architecture and teardown work before security underwriting |
| NVTS | NVIDIA collaboration and demonstrated 800 VDC boards | Roughly 103x sales; Q1 revenue $8.6M versus $14.0M a year earlier and GAAP operating loss $27.8M | Material production order, financing runway and credible gross-margin path | Continuing revenue contraction, dilution or programs remain demonstrations | No further work until commercial evidence clears a minimum threshold |
Sources: onsemi–NVIDIA 800 VDC collaboration, onsemi Q1 2026 results, onsemi–Synaptics transaction, Eaton 2025 annual report, Eaton Q1 2026 presentation, Vertiv 800 VDC update, Navitas Q1 2026 results, Navitas–NVIDIA collaboration.
ON combines three things the optical winners no longer offer simultaneously: a dated architecture transition, direct NVIDIA collaboration and a headline valuation that does not already require monopoly-like economics. Its Q1 disclosure that AI data-center revenue more than doubled is useful exposure proof, while the depressed industrial/automotive cycle creates operating leverage if the core business recovers.
But the thesis is not “ON is cheap.” It is a three-part underwriting problem:
Until those bridges are quantified, ON is a priority for diligence—not a buy recommendation.
Eaton offers less torque to a single architecture but better evidence that AI power demand is already reaching orders, backlog and cash flow. The company generated about $4.5 billion of operating cash flow and $3.6 billion of free cash flow in 2025 and is investing to expand capacity. This matters because a bottleneck winner that cannot finance inventory, factories and customer lead times may create revenue without creating per-share value. Eaton 2025 annual report
The trade-off is expectations. ETN is already recognized as a premium electrification compounder. The research question is whether data-center order growth can extend backlog duration and incremental margins enough to outrun the multiple.
NVIDIA describes physical AI as the next frontier and has published an ecosystem blueprint covering data generation, synthetic-data tooling, training, simulation and evaluation, with early users named. That validates a broad compute and software opportunity. It does not yet identify a scarce public supplier with contracted economics comparable to the Lumentum and Coherent agreements. NVIDIA physical-AI blueprint
The investable sequence should therefore be:
For now, most humanoid equities fail between steps one and three. Simulation usage can grow even if humanoid manufacturers never achieve attractive unit economics; that makes NVIDIA's platform exposure stronger than the average robot-OEM thesis.
The NVIDIA connection itself is public. It is not the edge. Potential mispricing sits in the gap between five variables:
| Variable | What consensus can observe | What still requires work |
|---|---|---|
| Architecture | CPO is in production; 800 VDC is targeted for 2027 | Ramp cadence, alternative topologies and qualification risk |
| Allocation | NVIDIA named partners and funded capacity | Supplier share, exclusivity and content per rack |
| Revenue | Purchase commitments, collaboration announcements and general AI growth | Take-or-pay terms, timing, cancellations and segment mix |
| Margins / cash flow | Company targets and consolidated results | Yield, depreciation, pricing, working capital and return on new capacity |
| Valuation | Current price and consensus estimates | Normalized earnings after bottleneck scarcity fades |
This produces three distinct investment structures:
NVDA remains the control asset. If a derivative supplier requires more optimistic assumptions than NVIDIA while offering less control over architecture, customer allocation and pricing, the second-order name fails an opportunity-cost test.
| Window | Evidence to demand | Bullish confirmation | Warning / invalidation |
|---|---|---|---|
| H2 2026 | Supplier design wins, order commentary, capex, inventory and backlog | Named 800 VDC production awards; optical commitments convert to shipments without margin dilution | Demonstrations remain non-revenue; inventory and receivables rise faster than sales |
| H1 2027 | Kyber / Rubin Ultra deployment readiness and customer schedules | Multiple customers retain 2027 deployment plans; power suppliers quantify content or orders | Architecture or customer deployment slips beyond 2027 |
| H2 2027 | Revenue, gross-margin and free-cash-flow conversion | New capacity earns attractive incremental returns and order coverage remains strong | Capacity utilization falls, pricing normalizes sharply or capex consumes cash flow |
| 12–24 months | Hyperscaler capex and system economics | AI-factory returns support continued cluster expansion | Capex growth weakens while industry capacity arrives simultaneously |
The broader leading-indicator thesis should be reconsidered if any of the following occurs:
Looking back from LITE, AAOI, COHR and AXTI selects the winners after the fact. A valid backtest must include every supplier NVIDIA mentioned or partnered with, including those whose revenue did not materialize. Otherwise the signal's hit rate will be overstated.
NVIDIA can fund redundant capacity, qualify multiple vendors and negotiate lower prices. That is excellent supply-chain policy and may be bad for long-run supplier margins. A large purchase commitment can lower volume risk while simultaneously capping scarcity pricing.
The 800 VDC architecture lists many partners. Being named proves technical relevance; it does not prove share. The more interchangeable the subsystem, the less valuable the logo slide.
AXTI's nearly 39-fold one-year move shows the danger. Even perfect industry forecasting can produce a poor investment if the stock discounts the terminal scenario before earnings arrive.
A consensus forward multiple is not a margin-of-safety calculation in a capacity-constrained industry. Normalized valuation must include post-shortage pricing, utilization and depreciation.
This section is a conditional research framework, not individualized advice.
Second-order suppliers should complement a distinct bottleneck exposure, not simply duplicate AI capex beta at a higher valuation. The best additions are those where NVIDIA creates a dated demand transition but current consensus has not yet capitalized several years of flawless execution.
| Topic | Primary source | Evidence used |
|---|---|---|
| Lumentum capacity | SEC exhibit, 2026-03-02 | $2B NVIDIA investment, multi-billion purchase commitment, capacity rights, U.S. fab |
| Coherent capacity | SEC exhibit, 2026-03-02 | $2B investment at $256.80, multi-billion purchase commitment and capacity rights |
| Coherent operations | Q3 FY2026 10-Q | InP expansion and industry shortage disclosure |
| Coherent capacity path | Investor presentation | Six-inch platform and staged InP expansion |
| CPO deployment | NVIDIA production announcement | Spectrum-X Photonics production status and early adopters |
| 800 VDC architecture | NVIDIA technical blog | 2027 Kyber / Rubin Ultra timing and ecosystem |
| ON operating evidence | Q1 2026 results | Revenue, margin and AI data-center growth statements |
| ON strategic complexity | Synaptics transaction | Approximately $7B all-stock transaction and physical-AI strategy |
| Eaton orders and capacity | Q1 2026 presentation | Data-center orders/backlog and 800 VDC positioning |
| Eaton cash generation | 2025 annual report | Operating cash flow, free cash flow and capacity investment |
| Vertiv timing | 800 VDC update | H2 2026 product plan and early AI-factory design work |
| Navitas evidence | Q1 2026 results | Revenue contraction, operating loss and 800 VDC demonstrations |
| Physical AI | NVIDIA blueprint | Ecosystem workflow and early users |
| Market screen | Yahoo Finance via local yfinance, captured 2026-08-17 |
2026-08-14 prices, one-year returns, market caps and provider forward multiples |
Your instinct is substantially right: NVIDIA is a powerful industrial leading indicator because it reveals bottlenecks through money, capacity rights and dated system transitions. The investable process, however, has two gates:
On that basis, CPO/CW lasers are proven but heavily repriced; 800 VDC is the most promising 2027 research frontier; and physical AI remains a theme awaiting procurement evidence. ON, ETN and COHR-versus-LITE deserve the next round of work. NVTS fails the current evidence-and-valuation screen. NVIDIA itself remains the hurdle rate against which every supplier idea should be judged.
Disclaimer: This report is for research and educational use only and does not constitute investment, legal or tax advice or an offer to buy or sell securities. Market data are best-effort third-party snapshots and should be refreshed before any decision. Forward estimates are uncertain; named collaborations do not guarantee orders, margins or shareholder returns.