NDI's performance is driven by a structured research and portfolio construction process designed to identify asymmetric opportunities and translate conviction into position sizing. The approach combines four core elements.

The performance attribution and case studies that follow illustrate how this framework was applied in practice across 2024 and 2025 — identifying expectation gaps, forming differentiated views, and sizing positions where the divergence between Market, Street, and our internal thesis was greatest.

NDI-FutureTech delivered chained annual returns of approximately 39.4% in 2024 and 38.9% in 2025, compounding an initial $100 investment to approximately $194 over the two-year period.

To isolate the sources of return, we decompose performance into stock-selection alpha and systematic factor contributions using a multi-factor regression model estimated on daily returns. The model uses four thematic factors representing distinct segments of the AI and digital economy cycle: SOXX (semiconductors and AI infrastructure), WCLD (SaaS and enterprise software), ARKK (used as a proxy for consumer internet/platform and consumer innovation exposure), and a proprietary Hyperscaler basket (cloud and mega-cap AI platforms, constructed as a market-cap-weighted index of AMZN, GOOGL, MSFT, and ORCL). Note that ARKK serves as a proxy only — while it provides directionally useful exposure to consumer internet and innovation-platform dynamics, it is not a precise match for the portfolio's specific consumer and platform holdings.

The broad market index (NDX) was excluded from both years' regressions following Variance Inflation Factor analysis. In 2024, NDX produced a VIF of approximately 17; in 2025, approximately 19 — both materially exceeding the threshold of 10, indicating that NDX movements were largely explained by the combination of the four thematic factors. Its inclusion created redundancy and unstable coefficient estimates. Removal improved model stability and ensures attributed performance reflects exposure to distinct AI-cycle themes rather than broad equity beta. Full regression output, diagnostic statistics, and VIF tables are provided in Appendix A.

Beyond return attribution, the portfolio's risk profile is evaluated through a VIX-based regime classification framework that segments the market environment into four volatility states — Low Volatility, Normal, Stress, and Crisis — each with defined VIX thresholds and confirmation-based transition rules. This framework enables regime-conditional analysis of maximum drawdown severity, Value at Risk (VaR), and tail-risk exposure (Conditional VaR), providing a structured view of how portfolio risk scales across market conditions. Investors interested in the detailed regime-level risk analysis, including drawdown recovery timelines and VaR estimates at the 95% confidence level, should refer to Appendix B.

** Statistically significant at the 1% level. The 2025 alpha estimate is economically substantial but not statistically significant at conventional levels (p = 0.205), reflecting higher daily return volatility and reduced signal-to-noise ratio. See Appendix for full regression diagnostics.
^† ARKK ETF is used as a proxy for consumer internet/platform and consumer innovation exposure; it is not a precise match for the portfolio's specific holdings in this segment.

Across both years, alpha remained the primary performance driver, contributing approximately 72% of returns in 2024 and 44% in 2025. The shift reflects the natural evolution of the AI investment cycle: in 2024, performance was concentrated in idiosyncratic, under-covered names where expectations lagged fundamentals; by 2025, the investable universe had broadened and systematic AI-cycle tailwinds contributed a larger share of returns. The four-factor model explains 87% of daily return variation in 2024 and 89% in 2025, indicating the portfolio maintains significant — and increasing — alignment with its thematic factor exposures, while generating meaningful excess return above what those exposures alone would predict.

Hyperscaler exposure was the most consistent factor tailwind, contributing approximately +7.4% in each year. The consumer internet and innovation proxy (ARKK) became the largest factor contributor in 2025 (+10.4%), reflecting the shift toward consumer internet and innovation-platform beneficiaries as AI adoption broadened. The SOXX coefficient flipped from marginally negative in 2024 to positive in 2025, consistent with the portfolio's increasing direct semiconductor exposure. WCLD turned from a positive contributor in 2024 to negative in 2025 as enterprise SaaS underperformed on margin compression concerns.

The portfolio also broadened significantly between the two periods, growing from 38 positions in 1Q24 to 63 in 2Q25, consistent with the shift from a concentrated alpha-driven profile to wider thematic participation.

Through 17 February 2026, NDI-FutureTech has returned −5.05% year-to-date. With approximately 33 trading days in the sample, the period is insufficient for formal regression-based performance attribution — coefficient estimates from such a limited dataset would carry unacceptably wide confidence intervals and produce unreliable factor decomposition. This section therefore provides headline return, volatility, and risk-adjusted metrics against the portfolio's benchmark universe. Full 2026 attribution will follow in subsequent reporting once a statistically meaningful sample has accumulated.

Data through 17 February 2026 (~33 trading days). Sharpe ratio = (YTD return − 4.15% risk-free rate) / Scaled to YTD standard deviations. Standard deviations annualised from daily returns; annualisation from a small number of observations can amplify estimates, particularly for indices that experienced extreme single-day moves during the period. The comparative ranking of volatilities is more informative than the absolute magnitudes.

The opening weeks of 2026 have presented a bifurcated and volatile environment across technology markets. Five of the six benchmarks — and the portfolio — have posted negative returns, with enterprise SaaS suffering a particularly acute dislocation (WCLD, −45.1%). The Hyperscaler basket (−11.7%) and ARKK (−8.6%) have also experienced meaningful drawdowns. Semiconductors (SOXX, +17.6%) stand as the sole positive performer — but, as the volatility chart above makes visually clear, that positive headline return has come at an extraordinary cost in daily price risk.

The most significant feature of the period is the dispersion in volatility profiles. NDI's Scaled to YTD standard deviations of 12.0% is a fraction of both SOXX (227.7%) and ARKK (226.9%). While capturing more of the semiconductor rally would have improved the headline return, doing so would have required absorbing a level of daily price volatility that is incompatible with disciplined portfolio management and most investors' risk budgets. On a risk-adjusted basis, SOXX's Sharpe ratio of 0.06 represents essentially zero excess return per unit of risk — confirming that the positive headline conceals an exceptionally rough and unpredictable path. ARKK, an actively managed innovation ETF, has delivered both a negative return and extreme volatility, underscoring the risk of concentrated innovation exposure without structural diversification and risk controls.

NDI's Sharpe ratio of −0.77, while negative, is the strongest among all strategies that posted negative returns — meaningfully better than NDX (−1.14), WCLD (−1.28), and the Hyperscaler basket (−2.12). The portfolio's ALL PATH construction framework — diversifying across distinct segments of the AI value chain rather than concentrating in a single theme — appears to be containing volatility effectively despite the portfolio's high-conviction technology positioning. The result is a modest drawdown with well-contained daily dispersion, preserving capital and optionality for recovery as conditions normalise.

NDI returned +10.1% in 1Q24, broadly in line with the strong risk-on environment (NDX +10.3%, SOXX +21.5%, Hyperscalers +14.4%). Two forces shaped the quarter: continued acceleration in AI infrastructure spending as hyperscalers expanded capacity aggressively, and a broadening of the AI narrative toward companies leveraging AI to improve operating efficiency and unit economics — particularly in consumer and platform businesses. NDI entered the year positioned across both layers: infrastructure beneficiaries and selective consumer and platform names where AI-enabled efficiency improvements were underappreciated. Performance was stock-specific rather than factor-driven, with the majority of gains coming from a small number of high-conviction positions.

Contributions are approximate: starting weight × quarterly simple return. 38 positions held during 1Q24. Contributions do not account for intra-quarter rebalancing.

The concentration of returns is striking: three positions — SMCI, HIMS, and CVNA — contributed +9.9% of the portfolio's +10.1% quarterly return. The remaining 35 positions approximately netted out, with meaningful gains from SE, NU, and DASH offset by losses in CHWY, UPST, and TSLA. This pattern is characteristic of a high-conviction portfolio in an early-cycle environment where a small number of correctly identified expectation gaps drive the majority of returns.

We selected SMCI as an asymmetric way to express our conviction in the AI infrastructure buildout. By late 2023, NVDA had become the consensus AI winner, and the market had already pulled forward aggressive revenue expectations two to three years ahead. The second derivative of NVDA's growth was likely approaching a peak, making the risk-reward increasingly balanced despite strong underlying fundamentals.

SMCI, as the leading server assembler and integration partner for NVDA's AI GPU clusters, offered a more attractive setup. Investors in NVDA were implicitly underwriting sustained high AI capex for years, yet expectations for SMCI remained anchored to near-term earnings. This created genuine asymmetry: if AI spending proved stronger and more durable than expected, SMCI's forward earnings and valuation multiple had significant room to re-rate, while its more modest starting valuation suggested limited incremental downside relative to NVDA.

On fundamentals, SMCI's engineering-first DNA distinguished it from legacy competitors like Dell and HPE that had increasingly shifted toward sales-and-marketing-driven models. Its modular server architecture allows rapid integration of next-generation GPUs, its focus on hyperscale customers aligns it with the largest AI spenders, and its deep NVDA partnership gave it a first-mover advantage on new platform launches. We believed AI infrastructure spending was still in its early innings, and SMCI offered a compelling way to capture that runway with a favourable risk-reward profile.

In early 2024, demand for the new generation of GLP-1 weight-loss drugs — treatments that work by mimicking a natural hormone to suppress appetite and improve blood sugar control — far exceeded supply. Obesity rates remain high in the US, and manufacturing capacity at Eli Lilly and Novo Nordisk could not keep pace with surging adoption. When a branded drug faces a shortage, the FDA allows licensed pharmacies to produce compounded versions — medically equivalent formulations made from the same active ingredient — to ensure patient access.

HIMS moved quickly to capitalise on this supply gap, offering lower-cost compounded GLP-1 treatments through its fully integrated telehealth platform: online consultation, prescription, and home delivery, removing the friction of traditional healthcare access.

Our thesis, formed in late 2023, was that the GLP-1 shortage would persist for multiple quarters, that demand was structurally larger than initial expectations, and that HIMS' digital distribution model positioned it as a scalable demand aggregator — creating a potential step-change in subscriber growth and revenue. Investor scepticism centred on the temporary nature of compounding once drug shortages normalise, regulatory uncertainty around compounded GLP-1s, and concerns that telehealth demand might prove cyclical. Our view was that even a temporary supply window could significantly expand HIMS' customer base, improve lifetime value, and accelerate brand awareness — creating durable growth beyond the shortage period. As GLP-1 demand flowed through the platform and subscriber growth accelerated, the market began to recognise this inflection, driving strong performance through the quarter.

The position was built around a full-cycle reset: a business whose differentiated model proved its value during the pandemic, then nearly collapsed under the weight of its scaling costs, and finally emerged with a cleaner balance sheet and a clearer path to profitability. Carvana's model combines the entire car-buying journey online — acquiring, inspecting, reconditioning, financing, and delivering vehicles directly to customers — in a fragmented industry historically defined by poor consumer experience and dealership friction. During COVID, this integrated approach created real value, but the stock fell approximately 97% in 2022 as the used-car boom reversed, rising rates made its leverage unsustainable, and the $2.2bn ADESA acquisition compounded liquidity concerns. The industry shakeout was severe: Vroom exited e-commerce and Shift filed for bankruptcy in October 2023, leaving Carvana as one of the few scaled pure-play online operators still standing.

By late 2023, the situation had changed materially. A comprehensive debt restructuring reduced near-term solvency risk, and management pivoted from growth-at-all-costs to unit profitability and cash generation. The stock rebounded roughly 1,000% in 2023 driven by short covering and improving fundamentals, but we believed the market was still underappreciating the forward setup.

Our thesis rested on three pillars. First, operating leverage from prior investment: Carvana had already built reconditioning and logistics capacity well ahead of current volumes and could grow into that footprint with limited incremental capex, improving free cash flow conversion. Second, industry consolidation strengthening competitive position: the collapse of online-first peers reduced competitive intensity, while Carvana's proprietary infrastructure created operational scale advantages that asset-light marketplaces cannot match. Third, a large structural market opportunity: despite its scale, Carvana remained a small-share player in a huge, fragmented market, with an integrated model bridging online convenience and physical execution — a difficult combination to replicate — supporting continued share gains as online adoption rises.

With solvency risk reduced, excess capacity in place, competition weakened, and the core model validated under stress, we viewed CVNA entering 2024 as a high-momentum turnaround with a clear path to continued multiple expansion as profitability visibility increased.

NDI returned +33.3% in 2Q25, significantly outperforming the broader market (NDX +17.6%, Hyperscalers +24.2%, SOXX +27.1%). Three forces shaped the quarter: tariff-driven volatility in the early weeks that resolved into a broader market focus on domestic reshoring and industrial capacity; the AI narrative shifting decisively toward the application layer as more capable coding agents and workflow automation tools drove practical enterprise adoption; and growing recognition of nuclear and strategic energy as structural themes, driven by AI-related power demand and US energy security policy. NDI entered the quarter positioned across strategic energy and nuclear beneficiaries, selective consumer and platform names with AI-driven operating leverage, and targeted AI application exposure. Performance was broad-based but conviction-driven, with meaningful contributions from a concentrated group of positions.

63 positions held during 2Q25. Case study positions highlighted in green. Contributions approximate: starting weight × quarterly simple return.

Returns were far more broadly distributed than in 1Q24: the top 10 positions contributed approximately +15.5% of the +33.3% quarterly return, with the remaining 53 positions contributing a further ~18%. This broader participation is consistent with the shift from a concentrated, alpha-driven profile toward wider thematic engagement as the AI cycle matured.

The nuclear theme was a high-conviction structural position entering 2025, driven by the intersection of AI-driven power demand, US energy security policy, and a multi-decade underinvestment cycle in domestic nuclear capacity. The rapid expansion of AI data centres has shifted hyperscaler competition from GPUs to power availability, exposing structural constraints in US grid capacity. While natural gas provides a near-term bridge, nuclear is increasingly recognised as the only scalable, reliable, carbon-free baseload option for long-term requirements. Policy momentum strengthened through early 2025, with the administration making domestic energy production and nuclear development strategic priorities, including efforts to reduce reliance on foreign fuel supply chains.

Despite this constructive backdrop, both names experienced significant drawdowns from December 2024 through March 2025 — LEU declined approximately 32% and OKLO approximately 8% — driven by risk-off conditions, policy uncertainty during the early transition period, and profit-taking in a crowded thematic trade. Importantly, no fundamental deterioration occurred. We viewed the drawdown as a positioning and sentiment reset rather than a thesis change, and maintained exposure based on improving medium-term fundamentals and policy visibility. This timing proved important as sentiment reversed sharply in 2Q25 alongside renewed policy clarity.

Oklo's small modular reactor (SMR) design focuses on simplified, factory-built systems that can be deployed faster and at lower cost than traditional large-scale plants. Its fast-reactor architecture allows use of recycled nuclear material and high fuel efficiency, positioning it as a leveraged beneficiary of next-generation nuclear deployment. Centrus targets the fuel supply constraint directly: the US currently has limited domestic uranium enrichment capacity and has historically relied on foreign suppliers, most notably Russia's state-owned Rosatom. Geopolitical tensions have made this dependence strategically untenable. Centrus is the only US-owned company licensed to produce high-assay low-enriched uranium (HALEU) — the specialised fuel required for most advanced reactor designs, including SMRs — giving it a critical position in a supply chain that must be rebuilt domestically.

As policy clarity improved and capital rotated back into strategic energy infrastructure during 2Q25, both positions rebounded sharply, validating the thesis and reinforcing nuclear's role as a core long-duration AI infrastructure theme.

Entering 2025, investor sentiment toward OSCR was weak due to concerns that the incoming Republican administration would roll back key provisions of the Affordable Care Act (ACA) — the US federal law that created subsidised health insurance marketplaces for individuals without employer-sponsored coverage. The specific fear was that reduced premium subsidies would materially shrink the addressable market for ACA-focused insurers.

Our analysis suggested this risk was overstated. Approximately two-thirds of ACA enrollees reside in Republican-led states, making meaningful subsidy reductions politically difficult without a viable replacement framework. With uninsured rates near historic lows and no clear alternative policy proposed, we viewed a sharp reversal as unlikely and largely priced in.

At the same time, OSCR was trading at approximately 3–4x free cash flow despite 20%+ expected revenue growth, rapid profitability improvement, and a scalable technology-first operating model. The core thesis rested on three elements. First, Oscar's integrated digital platform — telehealth, care navigation, data integration, and personalised patient engagement — improves outcomes while lowering administrative costs in a highly fragmented industry, with AI-driven automation offering further structural margin improvement over time. Second, Oscar is diversifying beyond the ACA marketplace into higher-quality growth vectors, including ICHRA (a framework allowing employers to fund individual insurance plans, supporting the shift away from traditional group coverage) and +Oscar, a SaaS platform providing technology infrastructure to other insurers — a capital-light, high-margin revenue stream. Third, the risk-reward asymmetry was compelling: with subsidy risks widely debated, management guiding conservatively, and expectations already low, any policy stability or continuation of current support had the potential to drive significant multiple expansion.

As policy fears eased and operational performance continued to improve through 2Q25, the stock re-rated accordingly.

Palantir was one of the highest-conviction positions held through 2025, despite valuation levels that many investors viewed as prohibitive. By early 2025, the stock was trading at approximately 60x EV/Sales, and the debate had shifted from business quality to whether a company of this quality could justify such extreme multiples.

Our view was that the market was underestimating both the immediacy and the scale of Palantir's role in the enterprise AI cycle. While AI adoption is straightforward for consumers through web interfaces and APIs, enterprise deployment is fundamentally different. Large organisations face structural barriers: fragmented and siloed legacy data, complex system dependencies and workflows, and strict requirements around security, governance, auditability, and control. For most enterprises, the challenge is not access to AI models but making those models operational inside real production environments.

Over more than 20 years, Palantir has built deep capabilities in data integration across complex, heterogeneous systems, ontology construction that maps business operations into machine-readable workflows, and privacy-first architecture with full access controls and auditability. With the launch of AIP (Artificial Intelligence Platform), these capabilities became the critical layer between foundation models and enterprise production. In our view, Palantir was — and remains — the only platform capable of securely operationalising AI at scale across large, complex organisations. This positioning made the company not just a software vendor, but essential infrastructure for enterprise AI adoption.

Entering late 2024, the NDI team formed the view that 2025 would become the "Year of AI Applications," driven by the emergence of more capable coding agents and workflow automation tools marking a step-change in practical enterprise adoption. We anticipated a growth reacceleration driven by rapid enterprise interest following AIP boot camps, shorter commercial sales cycles, expansion within existing customers as use cases moved from pilots to production, and a much larger effective TAM — as any organisation seeking to deploy AI securely would require a platform layer. This led us to conclude that the growth embedded in consensus expectations, despite high headline multiples, was too conservative.

Rather than reducing exposure on valuation concerns, we held the position through 2Q25 on the basis that Palantir's role as the secure AI operating system for the enterprise was becoming clearer, the shift toward real-world AI deployment was structural rather than cyclical, and growth was accelerating into the valuation rather than slowing into it. As 2025 progressed, commercial momentum accelerated, pipeline conversion improved, and operating leverage began to expand margins — confirming the growth inflection we had anticipated and resulting in meaningful gains for the portfolio.

The following case studies examine two positions that illustrate different dimensions of the portfolio's investment approach. Broadcom represents a long-duration thesis held since inception — a steady compounder where structural conviction in AI networking and custom silicon has compounded steadily across the full two-year period. Lumentum represents a high-conviction addition made at the start of 2025, where a differentiated view on the optical interconnect transition — catalysed by architectural breakthroughs from DeepSeek and Huawei — drove outsized returns within a single year. Together, AVGO and LITE demonstrate how NDI pairs durable, compounding positions with concentrated, catalyst-driven bets to build returns across different time horizons.

When the market was piling into NVIDIA, TSMC, and other obvious AI beneficiaries, we went deeper — asking what would be desperately needed as AI infrastructure scales beyond individual chips. The answer, fundamentally, is networking. The computation demand for AI is so immense that even the largest monolithic die cannot satisfy the latest training and inference workloads alone. Beyond building ever more powerful AI chips, the critical differentiator for AI end users becomes whether they can effectively link tens of thousands, then hundreds of thousands, and eventually millions of chips into one coherent cluster — with low cost, high performance, high reliability, low latency, and low maintenance overhead. We carefully chose AVGO over MRVL, and that selection has massively outperformed its weaker peer in the same rising tide that lifted all boats.

We have held two long-term theses on Broadcom since 2023. First, that AI ASICs will ultimately capture 30–60% market share in a world that initially ran on 100% merchant NVIDIA GPUs. Second, that the networking stack — not the individual AI chip — will ultimately determine the performance per dollar of an AI cluster.

When AI spending was still measured in single-digit billions, major cloud providers had no pressure to optimise cost. Their singular priority in the early phase was speed: NVIDIA's highly mature GPGPU and CUDA ecosystem was the definitively perfect fit for customers who were not cost-sensitive but time-sensitive. However, we expected inference to become the majority of AI compute, and inference is fundamentally easier to implement on custom ASICs. If a hyperscaler can run inference two to four times cheaper on a custom ASIC, the economic logic becomes overwhelming — it is simply a matter of time for the AI software stack to mature and for spending to reach a scale where optimisation becomes imperative. Eventually, every major cloud provider will find it compelling to shift inference workloads to custom ASICs to save cost, improve margin, and compete not just on intelligence but on cost per token in production environments rather than demonstrations.

For training, NVIDIA's CUDA moat is deeper. Training code evolves rapidly, and mapping workloads written for CUDA onto custom ASICs remains extremely difficult. That said, we also believe NVIDIA's GPGPU architecture will gradually converge toward something resembling an AI ASIC. NVIDIA cannot afford to waste die area on general-purpose CUDA cores indefinitely; competitive pressure will force it to allocate ever more silicon to tensor and AI-accelerating units — eventually making the GPGPU look increasingly like a purpose-built accelerator such as a TPU.

Broadcom has the strongest, most capable, and most mature custom ASIC team in the industry, having worked on Google's TPU for over a decade. Beyond silicon design, we identified Broadcom as having best-in-class SerDes IP, networking orchestration software, switches, NICs, and optical connectivity — spanning all major networking components where only NVIDIA can offer partial rivalry, and in areas like optical lasers, even NVIDIA must partner with Broadcom. In this domain, there is no room for tier-two players. Going cheap on fabless design means exponentially more cost and integration issues at the final AI cluster level. We expect Amazon and Microsoft will learn this lesson through choosing Marvell simply because it was cheaper and perceived as "good enough."

This thesis has been massively validated by the success of Gemini 3 and the launch of Google's TPUv7, which outperformed NVIDIA's Blackwell on performance, power, area, and cost metrics. Coming into 2026, we believe the next several quarters represent a critical tracking window — not only for TPUv7 shipment volumes approaching 30% or more of NVIDIA shipment levels, but also for Broadcom's ability to land and scale customers across the AI value chain, including Apple, SoftBank, OpenAI, ByteDance, and xAI — with the potential to pull Amazon, Microsoft, and Tesla into Broadcom's orbit as well.

As a portfolio holding, Broadcom has been a steady compounder, delivering market-beating returns in most quarters without going parabolic. Since inception, the position has returned approximately 200%, contributing 1.4 percentage points of the portfolio's roughly 90% cumulative return over the 2024–2025 period. This consistent, durable performance profile reflects both the quality of the underlying business and the strength of a structural thesis that the market has only gradually come to appreciate.

Lumentum was added to the portfolio at the start of 1Q25 and delivered a 339% return for the full year, contributing approximately 8 percentage points of the portfolio's 39% return in 2025 — making it the single largest contributor to annual performance. The total return on the position now exceeds 600%, and its contribution has only grown further into 2026.

The thesis was built on a firm conviction that optical interconnect is the endgame for AI networking, and that copper is merely an intermediate solution. While the Street faithfully followed Jensen Huang's copper-centric roadmap, we believed that roadmap could be wrong — and that optical represents a more viable path to building larger, higher-performance AI clusters linking more chips together with less overhead. Fundamentally, the future is light. Optical interconnect offers the best performance and cost trajectory for delivering high speed, high reliability, and long reach at scale. The challenge is engineering optimisation — making optical transceivers cheaper, simpler, and more performant. Copper cable, by contrast, has hit a physical wall that no amount of engineering can overcome. Active copper cables can stretch performance enough to make copper workable at 800G, but at 1.6T and beyond, the laws of physics drive copper's cost up exponentially. In short: copper is mature and ready but has no room to improve, while optical is immature but has vast room to improve and scale.

We chose Lumentum over Coherent because of Lumentum's EML focus. Lumentum is one of only three suppliers — alongside Broadcom and Sumitomo — capable of mass-producing 200G Electro-Absorption Modulated Lasers. Lumentum's 200G EML is the key component in Innolight's 1.6T transceiver, which was the only 1.6T module available in 1Q25, ramped into mass production through the year, and is positioned for majority share in 2026 as Innolight is the preferred reference design partner for both Google and NVIDIA. Coherent's VCSEL focus and lack of advanced EML capability means it remains a marginal player riding the tailwind of the optical boom rather than driving it. This deliberate selection of LITE over COHR has been rewarded: in the same rising tide that lifted all optical names, Lumentum has massively outperformed its weaker peer.

A pivotal catalyst emerged in early 2025 with DeepSeek's open-sourced inference architecture and Huawei's all-optical CM384 pod. DeepSeek had already demonstrated that it could emulate the performance of leading Western AI labs using a fraction of the training cost, achieving this on Huawei's Ascend clusters rather than NVIDIA hardware. But the deeper insight was architectural: by rearchitecting workloads and leveraging Huawei's optics-based scale-out fabric, clusters of hundreds of chips could operate efficiently across nodes without relying on NVIDIA's copper-bound, expensive, and increasingly constrained NVLink topologies. Huawei's all-optical superpods — linking hundreds and soon thousands of Ascend chips entirely through light — revealed that photonics is not just a performance enabler but a strategic weapon in the infrastructure race, allowing competitors with weaker individual chips to challenge NVIDIA's dominance through superior interconnect architecture. We identified this as a structural inflection point: the moment the industry recognised that data movement, not compute, is the bottleneck — and that the only medium capable of breaking that bottleneck is light.

The implications cascaded rapidly through the industry. Google announced plans to make its TPU clusters fully optical within two years, with the TPUv7 deployment paired with optical circuit switching demonstrating that tight software-hardware integration can deliver vastly superior performance per dollar compared to traditional electrical switching. NVIDIA, meanwhile, acknowledged the growing pressure through its Enfabrica acquisition — a networking silicon company designed to improve bandwidth efficiency inside scale-up systems. We interpreted this as Jensen Huang's attempt to reinforce NVIDIA's scale-up dominance without fully conceding that optics had already won the physics argument. The transition from electrical to optical signalling inside the data centre accelerated years ahead of expectations, pulling forward a generational opportunity for suppliers of the optical engines that make it possible.

The 2025 1.6T ramp, Coherent's failure to deliver high-volume 200G laser production, and Innolight's strong financial results — underpinned by Lumentum's 200G supply — demonstrated Lumentum's technological advantage in the market. Lumentum also stands to benefit from the secular shift toward optical circuit switching, though we note that competition from MEMS-based OCS providers presents a more contested landscape compared to Lumentum's dominant position in 200G EML. As these architectural shifts compound — from the DeepSeek-driven reappraisal of scale-out, to Google's all-optical TPU roadmap, to the broader industry recognition that optical is inevitable — Lumentum's position as the indispensable enabler of AI-scale optical connectivity has only strengthened, and its contribution to the portfolio reflects the depth of conviction we placed on this thesis from the outset.

Performance attribution is conducted using ordinary least squares (OLS) regression of daily portfolio returns against four thematic factor indices. The dependent variable is the daily log return of the NDI-FutureTech portfolio; the independent variables are the daily log returns of SOXX, WCLD, ARKK (used as a proxy for consumer internet/platform and consumer innovation exposure), and a proprietary Hyperscaler basket. The Hyperscaler basket is constructed as a market-capitalisation-weighted index of AMZN, GOOGL, MSFT, and ORCL, rebalanced daily using log returns applied to prior-day market-cap weights.

Stock-selection alpha is estimated as the annualised regression intercept, compounded over 252 trading days: annualised alpha = (1 + daily intercept)²⁵² − 1. Factor contributions are estimated as the product of the estimated beta coefficient and the corresponding factor's annual simple return over the matching performance period.

The broad market index (NDX) was initially considered but removed from both regressions following Variance Inflation Factor analysis. NDX produced VIF values of approximately 17 (2024) and 19 (2025), far exceeding the commonly applied threshold of 10. This severe multicollinearity arises because AMZN, GOOGL, MSFT, and ORCL collectively represent a substantial share of NDX's total market capitalisation, causing NDX movements to be largely explained by the Hyperscaler basket and, to a lesser extent, SOXX. Inclusion of NDX created redundancy and unstable coefficient estimates without materially improving explanatory power. Removal improved model parsimony and ensures that attributed performance reflects exposure to distinct AI-cycle themes rather than broad equity beta.

Significance codes: *** p < 0.001, ** p < 0.01, * p < 0.05. t-critical (two-tailed, 5%) ≈ 1.97 for both samples.

^† ARKK ETF is used as a proxy only for consumer internet/platform and consumer innovation exposure; it is not a precise match for the portfolio's specific holdings in this segment.

The evolution in betas from 2024 to 2025 is consistent with the portfolio's broadening from a concentrated, alpha-driven profile toward wider thematic participation — notably the shift from marginally negative to positive SOXX exposure and the increase in Hyperscaler beta reflecting the fund's growing large-cap AI platform allocation.

VIF analysis was conducted on the full five-factor model (including NDX) to test for multicollinearity before variable selection. A VIF exceeding 10 indicates severe multicollinearity warranting corrective action.

The NDX VIF values of ~17 (2024) and ~19 (2025) confirm that NDX is almost entirely explained by the four thematic factors. This is economically intuitive: AMZN, GOOGL, MSFT, and ORCL — the Hyperscaler basket constituents — collectively account for a substantial share of NDX market capitalisation, while SOXX, WCLD, and ARKK capture most remaining technology sector variation. After excluding NDX, all remaining factor VIFs fall below the threshold of 10, confirming the four-factor specification is well-conditioned.

The "Moderate" VIF classifications for SOXX and Hyperscalers reflect the structural overlap between semiconductor and mega-cap AI platform returns. This correlation is economically meaningful — Hyperscaler capex directly drives semiconductor demand — and the regression coefficients remain interpretable as distinct thematic exposures despite this moderate cross-correlation.

In 2024, the sum of alpha (28.2%) and total factor contribution (15.2%) equals 43.4%, exceeding the portfolio's 39.4% return by approximately 4.0 percentage points. In 2025, the sum (36.7%) falls short of the portfolio's 38.9% return by approximately 2.2 percentage points. These differences arise from several sources: factor correlation and overlap mean individual factor contributions are not strictly additive; the annualisation of the intercept via compounding introduces rounding relative to the actual daily return path; and nonlinear portfolio effects — including position sizing changes, intra-period rebalancing, and asymmetric payoff profiles — are not captured by a linear regression with constant coefficients.

Factor contributions should therefore be treated as indicative directional estimates of the portfolio's systematic exposure rather than strictly additive components of total return. The decomposition is most informative as a framework for understanding how alpha and factor exposure have evolved between the two years, rather than as a precise accounting identity.

Risk metrics are computed conditionally on a four-state volatility regime model derived from the CBOE Volatility Index (VIX). Each regime corresponds to a distinct range of implied equity volatility, with transitions governed by a 7-day confirmation rule to prevent excessive regime switching caused by intraday noise or single-day spikes. The regime definitions and the number of trading days observed in each state across the full sample (January 2024 – January 2026) are summarised below.

The asymmetric entry and exit thresholds — for example, entering Stress at VIX 20.5 but requiring VIX to fall below 19 to return to Normal — incorporate hysteresis to prevent whipsawing at regime boundaries. Over the full sample, the market spent the majority of time in Normal conditions (63%), with Low Volatility accounting for the entirety of the calm 1H 2024 period. The Stress and Crisis regimes were concentrated in the March–May 2025 tariff-driven selloff, totalling approximately 68 trading days combined.

Maximum drawdown is measured as the largest peak-to-trough decline in the portfolio's cumulative normalized price series, with the peak computed as a running all-time high from inception. Drawdowns are reported for each regime by filtering the drawdown series to dates classified under that regime. Because the drawdown is measured from the inception rolling peak (not reset per regime), deeper regimes naturally exhibit larger drawdowns as prior losses accumulate.

NDI's maximum drawdown of −34.5% during the Crisis regime reflects the tariff-driven sell-off of March–April 2025, with the trough occurring around 8 April 2025 before the sharp reversal triggered by the tariff pause announcement. Notably, the portfolio recovered to a new all-time high by approximately late July 2025 — a recovery period of roughly 80 trading days from trough — driven by the concentration in high-conviction positions (LITE, nuclear names, and PLTR) that benefited disproportionately from the post-crisis rotation into AI application and strategic infrastructure themes. NDI's deeper drawdown relative to NDX (−22.1%) and SPX (−18.4%) is consistent with its higher-beta, concentrated positioning, while the ARKK proxy experienced an even deeper −41.3% trough, reflecting its exposure to speculative growth names with weaker balance sheets and less structural differentiation.

The progression from Low Vol (−7.5%) through Normal (−13.4%), Stress (−24.9%), and Crisis (−34.5%) confirms that the regime framework captures meaningful escalation in downside risk, and that NDI's volatility profile is structurally higher than broad indices across all market conditions — a characteristic consistent with its concentrated, high-conviction approach.

Value at Risk is estimated using the historical simulation method at the 95% confidence level, computed separately for each regime from the empirical distribution of daily returns. VaR represents the loss threshold that is exceeded on only 5% of trading days — in other words, the boundary of the worst 1-in-20 daily outcomes. Conditional VaR (CVaR, also known as Expected Shortfall) measures the average loss on days that exceed the VaR threshold, capturing the severity of tail events rather than just their frequency.

1-Day VaR (95% Confidence)

Interpretation example: historically, about 1 in 20 Low Volatility days saw NDI losses of 2.6% or worse.

1-Day Conditional VaR (Expected Shortfall, 95% Confidence)

Interpretation example: in Normal regimes, when NDI's losses fell into the worst 5% of days, the average loss was approximately 4.1%.

21-Day Rolling VaR (95% Confidence, Since Inception Across All Regimes)

Interpretation: across all market conditions since inception, only about 5% of rolling 21-day periods saw NDI experience losses of approximately 11.6% or worse.

The VaR and CVaR metrics confirm the expected scaling of risk across regimes, with NDI exhibiting approximately 1.7–1.8× the daily tail risk of NDX in Normal and Stress environments, narrowing to approximately 1.3× in Crisis as correlations compress and all assets fall together. The gap between VaR and CVaR — which widens in Stress and Crisis regimes — indicates that tail events become more severe, not just more frequent, as volatility rises. In the Crisis regime, NDI's CVaR of −7.7% is notably 1.3 percentage points worse than its VaR of −6.4%, confirming that the worst days in extreme environments cluster toward exceptionally large losses rather than hovering near the VaR threshold.

The 21-day VaR provides a longer-horizon perspective: a −11.6% monthly loss boundary at the 95th percentile is consistent with the portfolio's concentrated, high-beta profile while remaining within the range expected for a strategy that targets 30–40% annualised returns. This loss magnitude was realised during the March–April 2025 tariff episode — the most severe drawdown in the sample — and was fully recovered within approximately four months.