Positioning Through the Transition

Opening

I. The map — what’s changing that an allocator can position against

I.1 Substrate rent capture and concentration

I.2 The rate-regime question — secular-stagnation r* anchor

I.3 Rent-capture leakage into consumption growth

I.4 Yield-curve control / financial repression as a regime risk

I.5 The reinstatement fork’s portfolio implication

II. Short-term (1–3y) — the substrate rent window

II.1 The substrate stack by layer

The chapter V map names the visible peaks — Nvidia, TSMC, ASML, the hyperscalers, the frontier labs. Allocator positioning requires the layer-by-layer cut underneath. Each layer has its own concentration shape, its own bottleneck, and its own access path for an outside investor. Nine layers are load-bearing.

Compute design. Nvidia held 80 to 86 per cent of the AI-accelerator market by revenue through 2025 (chapter V). Its data-centre revenue is increasingly anchored on a small number of buyers — four undisclosed customers were 61 per cent of Q3 FY2026 revenue. AMD’s MI300/MI350 has secured second place in the merchant market; Broadcom and Marvell own the custom-ASIC channel for the hyperscalers’ in-house chips (TPU, Trainium, MTIA). Outside Nvidia, the merchant-AI-silicon position is essentially a basket of AMD plus the ASIC enablers.

Leading-edge logic fabrication. TSMC holds roughly 90 per cent of leading-edge logic capacity at 3nm and below, with Samsung Foundry and Intel Foundry as the credible-but-trailing alternatives. The substrate-level constraint at the leading edge is not silicon — it is the next layer.

Advanced packaging. TSMC’s chip-on-wafer-on-substrate (CoWoS) is the binding bottleneck for the entire AI accelerator stack as of mid-2026. CoWoS capacity ran roughly 35,000 wafers per month at the end of 2024, 75,000 at the end of 2025, and is on track for ~130,000 by the end of 2026; demand is forecast at one million wafers in 2026 against that supply.¹ Nvidia has reportedly booked more than 50 per cent of TSMC’s 2026 CoWoS capacity; top customers collectively over 85 per cent. TSMC has begun outsourcing 240,000–270,000 wafers of overflow to Amkor and SPIL, which makes those names — and the OSAT tier broadly — the structural beneficiaries of any capacity that escapes the TSMC monopoly.

EUV lithography. ASML holds 100 per cent of EUV scanner sales globally, including the High-NA EUV systems that gate 2nm-and-below production. The lithography position has no substitute and no near-term challenger. It is the cleanest single-firm rent in the entire stack.

High-bandwidth memory. HBM is the second-tightest layer after CoWoS. SK Hynix held roughly 53 per cent of HBM revenue in Q3 2025; Samsung 35 per cent; Micron 11 per cent.² Samsung’s share is expected to recover above 30 per cent through 2026 as HBM4 qualifies into Nvidia’s Blackwell and Rubin generations. The global HBM market grew from approximately $38 billion in 2025 to $58 billion in 2026 on AI-driven demand. SK Hynix is the cleanest pure-play; Samsung carries non-HBM dilution; Micron is the US-listed proxy.

Compute infrastructure and neoclouds. Below the hyperscalers sits a separate tier of GPU-rental specialists — CoreWeave, Nebius, Lambda, Crusoe, Nscale, Fluidstack. CoreWeave ended 2025 with $66.8 billion of contracted backlog across 170 MW of active power; FY2025 cloud revenue ran roughly $5.1 billion at approximately 70 per cent gross margin (chapter V).³ Nebius announced a $17.4–19.4 billion five-year Microsoft contract and a $3 billion Meta deal in late 2025, guiding 2026 ARR to $7–9 billion. The neocloud market sits at roughly $24 billion in 2025 and $35 billion in 2026, with sell-side projections of $236 billion by 2031. Only CoreWeave and Nebius are public; the rest are private as of mid-2026. The layer-specific allocator question is whether neocloud margins survive the hyperscaler capex curve from chapter V — when Microsoft has its own sixty-plus-billion-dollar buildout, does it keep renting? — and the answer in the contract data so far is that hyperscalers are using neocloud capacity as an overflow buffer rather than substituting it for their own builds.

Power and grid. Chapter I has the buildout signal — Three Mile Island, Susquehanna, Kairos, Meta’s 6.6 GW nuclear announcement, Fervo’s geothermal, PJM clearing prices up roughly elevenfold between 2024 and 2025. The allocator-relevant cut is which exposures inside that signal compound. Utility equities serving the data-centre load (Vistra, Constellation, Talen) have rerated on the offtake demand; merchant-power exposure dominates regulated-utility exposure inside this thesis because the regulated side cannot capture the clearing-price shock. Transformer manufacturers (Hitachi Energy, Eaton, Hubbell) are the second-derivative play. Geothermal pure-plays remain private.

Data-centre real estate. Digital Realty (DLR) and Equinix (EQIX) own the wholesale and interconnection footprints into which hyperscaler capex flows. The two are not interchangeable — DLR is more wholesale-leasing, EQIX more retail-interconnection — and the relevant exposure depends on whether the allocator wants the leasing spread (DLR) or the network-effect moat (EQIX). Iron Mountain has pivoted from records storage to data-centre real estate but is the smaller and less-pure exposure.

Application and model layers. The frontier-model labs are private (OpenAI, Anthropic, xAI, Mistral); the application-layer winners are mostly private (Cursor, Anysphere, Glean, Perplexity, Harvey). Public-market access at these layers is limited to indirect exposure through the hyperscalers (Microsoft holds approximately 49 per cent of OpenAI’s economic share; Amazon and Google hold stakes in Anthropic) and to the consumer-facing product layer where it overlaps with already-public firms (Microsoft Copilot, Google Workspace AI, Meta AI). For a public-equity allocator the cleanest read is that the application layer is currently a future-flow position — wait for IPOs and acquisitions; do not pay for them through high-multiple proxies.

That is the layered map. Section II.2 takes each layer to its concrete public-equity vehicle.

II.2 Public-equity vehicles

For each layer of the substrate stack, the public-equity allocator’s choice is between an index ETF that owns the layer and a single-name overweight to the dominant firm. The trade-off is the standard one — diversification against tracking error — sharpened by the fact that, at most layers, the dominant firm is the layer.

Broad substrate. SMH (VanEck Semiconductor) and SOXX (iShares Semiconductor) are the two listed semiconductor ETFs. SMH holds 26 names with Nvidia at 16.4 per cent, TSMC 9.75 per cent, Intel 8.30 per cent, Broadcom 7.32 per cent, and AMD 7.09 per cent; the top ten holdings are 73 per cent of the fund.⁴ SOXX is more diversified across 30 names with smaller single-firm concentration. SMH carries the substrate thesis more directly; SOXX is the conservative substitute. The argument for index over single-name Nvidia is path-dependence — Nvidia’s customer-concentration shift documented in chapter V is itself a single-firm risk that an SMH overweight diffuses.

Logic and packaging. TSMC ADRs (TSM) trade liquidly in the US and capture both the leading-edge logic monopoly and the binding CoWoS bottleneck inside one position. Amkor (AMKR) and ASE Technology (ASX) are the OSAT overflow plays as TSMC outsources 240,000–270,000 wafers of CoWoS capacity in 2026. Both are smaller-cap and more cyclical than the substrate names; size accordingly.

Memory. SK Hynix is not US-listed — exposure requires the Korean line (000660 KS) or a Korea-tilted EM fund. Micron (MU) is the cleanest US-listed HBM proxy at 11 per cent of the market, with single-firm exposure to the HBM4 transition in 2026. Samsung Electronics (005930 KS) carries consumer-electronics and mobile dilution and is the diversified alternative.

Equipment and lithography. ASML (ASML) is the lithography monopoly. Applied Materials (AMAT), Lam Research (LRCX), and KLA (KLAC) are the rest of the wafer-fab equipment stack and trade as a basket within SMH/SOXX. Standalone ASML overweight is the cleanest position for the lithography rent specifically.

Compute infrastructure. CoreWeave (CRWV) and Nebius (NBIS) are the two public neoclouds. Backlog visibility is high (CoreWeave $66.8 billion contracted; Nebius guided $7–9 billion ARR for 2026), but margin durability under the hyperscaler capex curve is the open question. Position sizing should reflect that the multiples already price aggressive growth.

Power and utilities. XLU (Utilities Select Sector SPDR) is the broad utility ETF with Constellation at 7.9 per cent and Vistra at 4.3 per cent.⁵ UTES (Virtus Reaves Utilities) is the more concentrated AI-power thematic — its top four positions are NextEra, Constellation, Vistra, and Talen — and IDU (iShares US Utilities) sits between. For the merchant-power thesis specifically, single-name Vistra (VST), Constellation (CEG), and Talen (TLN) capture the offtake-demand exposure more directly than the index. Transformer and grid-equipment exposure runs through GE Vernova (GEV), Eaton (ETN), Hubbell (HUBB), and via the Swiss/Japanese listings of Hitachi Energy’s parent.

Data-centre real estate. Digital Realty (DLR) and Equinix (EQIX) are the two pure-play data-centre REITs. Equinix raised its 2026 AFFO guidance to $4.20–4.28 billion (9–11 per cent growth) and is targeting a capacity doubling to roughly 3 GW developable by 2029.⁶ Dividend yields are 2–3 per cent — low for REITs because the names are priced for growth rather than income. Iron Mountain (IRM) is the higher-beta pivot: its data-centre segment is approaching $800 million in revenue and is projected to exceed $1 billion in 2026, but the rest of IRM is records storage, which dilutes the AI exposure.

Application and model layers. Public access is indirect. Microsoft (MSFT) holds approximately 49 per cent of OpenAI’s economic share; Amazon (AMZN) and Alphabet (GOOGL) hold stakes in Anthropic. Adobe (ADBE), Salesforce (CRM), and ServiceNow (NOW) are application-layer incumbents racing to integrate AI. None of these is a pure-play. Wait for IPOs; do not pay for application-layer growth through hyperscaler multiples that already discount the model-layer optionality.

Clean-firm-power buildout. The chapter I buildout has no clean single ETF. Constellation (nuclear restart), GE Vernova (nuclear SMR and grid), Cameco (CCJ, uranium fuel cycle and the joint Westinghouse owner with Brookfield), and Nuscale (SMR, small and speculative) are the closest single-name exposures. NLR (VanEck Uranium and Nuclear Energy) is the listed thematic but holds non-US miners that introduce sovereign risk. The buildout exposure for a public-equity allocator is therefore concentrated at four-to-six names rather than an index.

The vehicles above are the means; the allocation rules are in Section V. The diffusion-versus-concentration discussion in II.3 is the standing argument against overweighting any single layer.

II.3 Diffusion versus concentration on a 1–3y horizon

The DeepSeek shock — R1 released 27 January 2025, Nvidia falling 17 per cent that day on a $589 billion single-day market-capitalisation loss (chapter V) — is the empirical test of whether capability diffusion erodes substrate rents on a 1–3y clock. The answer in the eighteen months since is no, with qualifications.

Epoch AI’s open-weight Capabilities Index puts the lag between open and closed-frontier capability at roughly three months as of mid-2026, down from more than two years in 2023. Hugging Face’s spring-2026 ecosystem report finds Chinese-origin models at 41 per cent of monthly downloads and the independent-developer share rising from 17 to 39 per cent (both cited in chapter V). Application-layer pricing has compressed accordingly. CoreWeave and Nebius are renting compute to a market that has more model choices each quarter.

What has not happened is substrate-layer margin compression. Nvidia’s customer concentration has risen rather than fallen (61 per cent of Q3 FY2026 revenue from four customers, against 36 per cent a year earlier). TSMC’s CoWoS bookings are sold out through 2026 (II.1). ASML’s order book is structural. The capability layer commoditises; the compute, the energy, the manufacturing, the talent, and the integrated product distribution do not.

The 1–3y position is therefore long substrate (II.1, II.2) and underweight — or at minimum cautious on — any application-layer name whose moat depends on model-layer exclusivity rather than data, distribution, or workflow integration. The DeepSeek-style event is a recurring risk to the latter and a transient drawdown opportunity in the former.

II.4 Cyclical rate-regime risk

The η > 1 trade in Section IV pays on a multi-year regime shift. The 1–3y rate question is different. The near-term curve is set by central-bank reaction functions and Treasury supply, not by structural EIS arithmetic. The allocator’s task in this window is risk management around three specific cyclical exposures.

US Treasury yields entered May 2026 with the 10-year at roughly 4.38 per cent and the 2-year at 3.90 per cent, putting the 2s10s spread at +48 basis points.⁷ The curve is positive, modestly steep, and consistent with a Fed that has cut three times in 2025 and is priced for one to two more cuts in 2026. The ECB has held the deposit facility at 2.00 per cent since June 2025. This is the baseline regime — easing into anchored growth — and most allocator positions in Section II are built against it.

The first risk is an inflation surprise from the AI buildout itself. Big-Five hyperscaler capex is guided at $660–725 billion in 2026 (chapter V); PJM cleared its 2026/2027 capacity auction at the FERC price cap of $329.17 per MW-day and the 2027/2028 auction at $333.44, with data-centre demand explaining nearly 5,100 MW of the year-on-year load increase.⁸ The buildout is structurally inflationary on the goods that supply it — electricity, transformers, advanced packaging, leading-edge memory — all of which are physical-capacity-constrained on the timescale of the buildout. The market has priced this partly into utility and grid-equipment equities (II.2). It has not consistently priced it into front-end inflation breakevens. A surprise on CPI or core PCE on the supply-driven leg of the build is the cleanest cyclical tail.

The second risk is a Fed reaction-function shift if AI productivity does show up in the data on a near-term clock. The Brynjolfsson-side empirical record from chapter IV — the Anthropic Economic Index augmentation/automation split, the Canaries updates, the Acemoglu-Restrepo decomposition when it lands — could move the Fed’s neutral-rate estimate in either direction. A higher r* (productivity-boom-driven) tightens financial conditions even at a constant policy rate; a lower r* (J-curve trough still in force) loosens them. The position to hold here is duration that is hedged against r* repricing — short-to-intermediate Treasuries rather than the long end.

The third risk is term-premium repricing on supply. Stargate’s $500 billion four-year compute commitment, the cumulative hyperscaler capex curve, and the US fiscal trajectory all imply Treasury issuance running heavy through the decade. Term premium has been compressed since 2008 by quantitative easing, foreign reserve demand, and pension-fund duration buying; none of those drivers is structurally durable. A 50-to-100-basis-point term-premium normalisation is the conservative tail bet on the long end. This is not the η > 1 trade — it is the boring duration risk that pays on supply, not on growth news.

Stock-bond correlation, the chapter VIII finding (0.80 in mid-2024 to 0.16 by late 2025), is the standing reason bonds are functioning as a hedge again. The correlation is conditional on no inflation surprise. If the first risk bites, the correlation flips back to its 2022 regime and the 60/40 hedge fails. Position sizing on the bond leg of a portfolio in this window should reflect that the hedge is real but contingent.

The single position that runs against all three risks simultaneously is short duration, hedged to DKK or EUR (per III.4), with merchant-power and grid-equipment overweights priced for capacity-price persistence. The single position that runs with all three risks is long-duration USD bonds with no convexity overlay — the asymmetric-loser bucket from III.5 has the section’s most exposed position in it.

The discipline for the 1–3y window is to size the duration position to survive a 100-basis-point move at the long end without forced liquidation. The η > 1 trade pays much further out and on a different mechanism; it does not substitute for cyclical risk management here.

III. Long-term (5–15y) — positioning for after the Baumol window

III.1 Humanoid manufacturing scale as the regime flip

III.3 Rent capture to macro variables

III.4 Currency exposure under ASI

For a Danish allocator the currency question reduces to one choice: EUR or USD. The krone has traded inside a ±2.25 per cent band against the euro through ERM II since 1999, around a central rate of 7.46038 DKK per EUR.⁹ Bulgaria’s euro adoption in January 2026 made Denmark the sole remaining ERM II member, and no Folketing majority has produced a referendum vote since 2000. The peg is the binding constraint, not the active position. Whatever the allocator concludes about the long-term currency map, the implementation lever is the EUR/USD split inside the portfolio, not anything denominated in DKK.

EUR/USD entered 2026 at roughly 1.18, with the Fed-ECB policy-rate differential at about 160 basis points after three Fed cuts in 2025 and an ECB deposit facility held at 2.00 per cent since June 2025.¹⁰ The narrowing is the live cyclical story; sell-side estimates put EUR/USD at 1.22–1.25 if the differential compresses 50 basis points further. That is the 1–3y currency view, and it has little to do with the AI thesis.

The AI thesis bites on the 5–15y horizon, through two channels that point in opposite directions.

The first is the rate-and-capital-flow channel. Substrate rent concentration in the US (chapter V) generates persistent equity-return premia that pull global capital into US assets. Higher prospective US returns sustain a stronger USD. In an Acemoglu world, where rent capture compounds, this channel is durable: the US is the substrate, the substrate is the rent, the USD is the substrate’s denomination. In a Brynjolfsson world the channel weakens as the gain broadens to the application layer and to non-US economies, but it does not reverse.

The second is the consumption-growth channel from chapter V’s mechanism. If real wage growth and real disposable income per capita stagnate in a capital-share-rising economy, the same economy’s long-run currency loses ground against one whose households retain a larger share of output. The EU’s labour share at 48.6 per cent (chapter VI) is structurally below the US’s 54.1 per cent in absolute terms, but the trajectories are diverging — the US figure is the lowest reading in the BLS series since 1947 and still falling, while the euro-area figure is up roughly 1.7 percentage points over twenty years. On the consumption-growth channel, what matters is the trajectory: a US economy whose labour share is collapsing produces muted consumption growth relative to one whose labour share is stable or rising, and the euro should therefore appreciate against the dollar over the decade. In the Acemoglu world this channel is the partial offset to the capital-flow channel; in the Brynjolfsson world it dominates, because the rate-differential premium compresses while the labour-share gap persists.

The two channels do not have a clean closed-form winner. The defensible reading is that USD core exposure is hard to avoid — the US is roughly 60 per cent of global equity market capitalisation and the substrate stack from Section II is overwhelmingly USD-denominated — and that fighting it on a tactical basis is more likely to cost than to pay. For equity exposure, unhedged USD dominates on transaction-cost and tracking-error grounds. For fixed-income exposure, where currency swings dominate total return on multi-year horizons, hedging into DKK or EUR is the conservative default.

The position III.5 sorts as a Brynjolfsson tilt — overweighting EUR against USD for the long horizon — is the trade if the consumption-growth channel dominates. The position implicit in the Acemoglu reading is the opposite: stay USD-heavy because the rent stays where it is captured. Both are tilts, not core positions. The robust position is the one Section II already names: hold the substrate, hold it in whatever currency it trades in, and let the FX exposure sit where it lands.

III.5 Reinstatement-robust positioning

The Brynjolfsson–Acemoglu disagreement does not need to be resolved before an allocator commits capital, but it cannot be ignored. The question that closes Section III is therefore not which side will be right — the paper that adjudicates has not been written, and Section II’s near-term substrate trade is largely indifferent to the answer — but which long-horizon positions survive being wrong about it. Sort the candidate positions into a 2×2: do they pay if reinstatement holds, do they pay if it fails, both, neither.

Position regardless

Substrate equity wins under both readings. In a reinstatement-failure outcome the AI rent compounds at the chip, fab, lithography, and hyperscaler layer because that is where the rent is structurally captured — chapter V’s mechanism. In a reinstatement-success outcome the same buildout scales with the broader productivity gain. Broad global equity (Norway-style VT exposure) belongs here too: in either world the rent ends up somewhere in the global index, and concentration only sharpens the case for owning the index rather than picking firms. Real assets with structural demand — power infrastructure on the chapter I buildout signal, housing in service-strained cities, materials tied to the energy and robotics ramp — round out the core.

Tilt if reinstatement holds

If the J-curve bends back, the gain spreads from the substrate to the rest of the equity market and the application layer rerates as augmentation gains aggregate. Public-equity exposure to the application layer (mostly private as of 2026, so this is a future-flow position waiting on IPOs and acquisitions) sits here. Currency exposure also tilts this way per III.4, though the trajectory argument is owned by that section, not this one. Equity duration beyond the substrate becomes safer in this world; bond duration becomes safer too, conditional on no inflation surprise.

Tilt if reinstatement fails

If the trough is the whole story, the rent stays concentrated and the policy fork from chapter VII becomes the live macro question. Overweighting substrate against the broad index pays here. Physical-AI and robotics (chapter III’s window) is the strongest position in this bucket. The Baumol-window wage spike is durable in this world: software productivity gains accrue to capital, the physical-care layer cannot substitute, and whoever industrialises humanoid manufacturing earns the rent that closes the window. Sovereign-wealth-style exposure — a stake in the redistributive vehicle that captures the capital share if the policy lever gets pulled — is the position that survives Acemoglu’s reading of chapter VII even if the lever takes a decade to move.

Asymmetric losers

Labour-intensive service businesses without an AI moat lose in both worlds, just by different mechanisms. In a reinstatement-failure world their wages are bid up against an AI-augmented private sector while their pricing is capped by public payors or competition; in a reinstatement-success world their margins are compressed by AI-augmented competitors who reach the same output with less labour. Long-duration bonds without convexity sit in the same bucket: exposed to the η > 1 tail if reinstatement holds, exposed to inflation surprise if it does not, with no offsetting upside on either side. The asymmetric-loser bucket is not a strong-prior call; it is the position the allocator does not need to take.

The η > 1 trade is not a regime hedge

The temptation after Section IV is to put the η trade in the position-regardless bucket because it pays on rate moves and rate moves happen in either world. It does not belong there. The trade pays specifically through the Euler-g channel — real rates rise enough to crush both equities and bonds if and only if consumption growth materialises. Chapter V’s rent-capture mechanism is the objection on the other side: in a reinstatement-failure world consumption growth stays muted even as output rises, because the rent is captured at the capital layer and does not flow through wages. The g in the Euler equation does not move. The η > 1 trade is therefore a tilt-if-reinstatement-holds position, sized within that bucket. Section IV is the trade’s own merits; this section is where it sits in the portfolio.

Sizing under ignorance

The empirical paper that decides the disagreement — an Acemoglu-Restrepo-style task-content decomposition of the post-2022 AI labour-demand growth, separating displacement from reinstatement by industry — has not been written. Chapter IX of the main essay names it as the watch metric and the rebalance trigger. Until it is written and read, the disciplined allocation weights position-regardless heavily, tilts light on either side, and holds the asymmetric-loser bucket at zero. The prior going into 2026 is slightly Acemoglu-leaning — labour share at 53.8 per cent, the Canaries displacement showing first, no reinstatement paper in print — but the prior is light. Both tilts will be wrong in different ways. The position that survives both is the position that does not need the prior to be right.

This is the section where the reader stops trying to pick the Brynjolfsson–Acemoglu winner and starts asking which positions do not require the pick.

IV. Tail hedges — the η > 1 trade

IV.1 One-line statement

IV.2 The math

IV.3 What you’re actually betting on — four joint claims

IV.4 Vehicles in order of cleanliness

IV.5 Where the trade has failed historically

IV.6 Why this is a tail hedge, not a core position

V. What you actually do

The structural argument compresses to one sentence: the AI substrate is producing a rent that is unevenly captured, the rent is currently outpacing wage income, and the policy and physical-AI mechanisms that would close the gap take a decade. The allocator’s question is what to hold against that pattern across time horizons, and which positions survive being wrong about which mechanism dominates.

1–3y core positions. Overweight substrate equity through SMH or a single-name basket of Nvidia, TSMC, ASML, Broadcom, and Micron, sized to ten to twenty per cent of the total portfolio. Add power and merchant-utility exposure through Constellation, Vistra, and Talen (or XLU/UTES as the indexed substitute), sized to five to ten per cent. Hold data-centre REITs (DLR and EQIX, weighted between them on the leasing-versus-interconnection preference) at three to five per cent. The bond leg is short-to-intermediate duration, hedged to DKK or EUR, sized to twenty-to-thirty per cent of the portfolio with explicit awareness that the stock-bond correlation hedge is conditional (II.4).

5–15y core positions. Hold the substrate-equity exposure through the window. Add physical-AI and robotics exposure as the manufacturing-scale signal arrives — Unitree, Figure, Agility, and Tesla’s Optimus line are the watch metrics (chapter III); the listed proxy is a broad robotics ETF (ROBO, BOTZ) at five per cent, with Tesla single-name overweight only against an explicit thesis about Optimus. Add capital-share-redistribution exposure if and when the policy lever is pulled — Norway-style sovereign-wealth-fund-equivalent positioning is the structural shape, operationalised in private markets and via Norwegian-listed proxies for the underlying instruments. Position the application layer through future-flow exposure: track the IPO and acquisition pipeline; do not pay for it through hyperscaler multiples now.

Tail hedges. Hold the η > 1 trade as a Brynjolfsson sub-tail at one to two per cent of the portfolio, via long-dated payer swaptions if institutional access exists, otherwise via TLT or TLH puts (III.5 and Section IV). The trade is a tail hedge, not a regime hedge; do not let it grow into a core position. Hold gold and Bitcoin at one to two per cent each as insurance against currency-system tail outcomes, not as asset-allocation backbone.

What to avoid. Labour-intensive service businesses without an AI moat. Long-duration USD bonds without convexity. High-multiple application-layer equities priced for execution they have not yet delivered. Single-firm Nvidia overweights beyond what an SMH-weight-equivalent already supplies. Currency-hedged USD equity exposure for a Danish allocator (it costs basis-point yield and protects against a peg that is itself the binding constraint).

The honest caveats. The substrate-rent thesis is the load-bearing claim; if it fails, the core positions above fail with it. The two specific failure modes are open-weight diffusion compressing substrate margins faster than scale-up sustains them (chapter V’s DeepSeek shock generalised) and humanoid manufacturing scaling earlier than chapter III’s window implies — both would compress the rent that the core positions earn. The watch metrics are in Section VI.

Sizing discipline. None of the position sizes above are precise — they are the shapes of a portfolio, not a recipe. Adjust to existing wealth, risk tolerance, age (chapter VIII), and what the rest of the portfolio is already doing. The single rule that survives every personal-circumstance adjustment is that the position-regardless bucket from III.5 dominates the portfolio; the tilts are small; the tail hedges are smaller; the asymmetric-loser bucket is empty.

The execution layer for a Danish reader sits in V.5. The falsification layer is in VI.

V.5 Execution in Danish vehicles

Section V’s positions are stated in pre-tax shapes; the Danish tax surface determines what actually reaches the portfolio. Three vehicles dominate.

Aktiesparekonto. The contribution ceiling is 174,200 DKK for 2026, up from 166,200 in 2025; the return is lagerbeskattet at 17 per cent annually.¹¹ The flat 17 per cent rate, against marginal capital-income rates of 37–42 per cent and share-income rates of 27/42 per cent, makes ASK the cheapest taxable vehicle for an active equity investor on the substrate thesis. Eligible holdings are stocks on a regulated exchange and minimum-taxed investment funds with at least 50 per cent stock allocation. Single-name positions in Nvidia, TSMC ADR, ASML, and Broadcom qualify directly; SMH and SOXX do not, because most US-domiciled ETFs are categorised as investment companies rather than minimum-taxed funds. Fill the ASK with single-name substrate equity first; treat it as the first DKK 174,200 of the substrate sleeve.

Positivlisten and the ETF question. Skat maintains a positivliste of investment funds taxed as share income (aktieindkomst: 27 per cent below the 2026 progression threshold of DKK 79,400, 42 per cent above) rather than as capital income.¹² The list contained roughly 5,000 funds in 2026, with about 100 ETFs added and 30 removed in the most recent update. Most US-domiciled ETFs — including SMH, SOXX, QQQ, VTI, and the substrate-relevant utility and REIT funds — are not on the list and are therefore taxed as capital income at 37–42 per cent on the lagerprincip. The practical implication is that the position-by-position vehicle choice runs as follows: single-name US substrate equity for the ASK; UCITS-domiciled ETFs (Sparindex, iShares Core, Xtrackers) on the positivlisten for the rest of the taxable account; US-domiciled ETFs only when no UCITS alternative exists and the substrate exposure justifies the higher tax drag.

Pension. Ratepension and aldersopsparing have different roles. Ratepension allows up to 68,700 DKK of deductible contributions in 2026 with PAL skat of 15.3 per cent on returns and ordinary income tax on withdrawal.¹³ Aldersopsparing allows 9,900 DKK if more than seven years to folkepensionsalder, or 64,200 DKK if within seven years; no deduction on contribution, 15.3 per cent PAL on returns, tax-free withdrawal. For an allocator under fifty-five the binding constraint is usually liquidity rather than tax — pension capital is locked. Use ratepension for the bond leg and long-horizon substrate exposure that the allocator does not need pre-retirement; use aldersopsparing within the cap for the same purpose, because the post-PAL after-tax return often beats the marginal-rate alternative at the household’s tax level. PAL at 15.3 per cent is structurally below the 17 per cent ASK rate, which inverts the usual ASK-first logic if contributions can be deducted at a high marginal rate.

Pool-allocation rule for the Section V portfolio. Fill ASK first with single-name substrate equity (Nvidia, TSMC ADR, ASML, Broadcom, Micron) up to the annual ceiling. Move overflow into UCITS substrate ETFs (for example iShares Core S&P 500 UCITS, Xtrackers MSCI World Information Technology UCITS) in the taxable account, sized against the positivlisten. Hold the long-horizon physical-AI and robotics exposure in ratepension or aldersopsparing because the time-to-payoff exceeds typical liquidity needs. The bond leg goes in ratepension; the 17 per cent ASK rate is a needless drag on bond yield. The η > 1 tail hedge, if executed via listed instruments (TLT puts), goes in the taxable account because pension vehicles do not permit derivative structures cleanly; if executed via OTC swaptions, this is institutional-only and outside the retail vehicle map.

What this section does not solve. The tax surface above is the current 2026 regime; it will change. The ASK ceiling has risen every year since 2019; the positivlisten composition turns over each spring; the PAL rate is politically stable but not constitutionally protected. The allocator’s job is to optimise against the regime that holds, knowing the regime is itself a variable.

VI. Where the thesis falsifies

TSMC CoWoS capacity expansion through 2026: FinancialContent, “The Great Packaging Pivot: How TSMC is Doubling CoWoS Capacity to Break the AI Supply Bottleneck through 2026,” January 2026; TrendForce, “TSMC’s CoWoS-L/S Reportedly Fully Booked, OSAT Partners Step Up,” December 2025; Fusion Worldwide, “Inside the AI Bottleneck: CoWoS, HBM, and 2–3nm Capacity Constraints Through 2027.” ↩
HBM market share Q3 2025 and 2026 outlook: Astute Group, “SK hynix holds 62% of HBM, Micron overtakes Samsung, 2026 battle pivots to HBM4,” 2025; Counterpoint Research, “Global DRAM and HBM Market Share: Quarterly,” Q3 2025 release. ↩
CoreWeave FY2025 results: company filings, summarised by Sacra. Nebius 2026 ARR guidance: Converge Digest, “The Neocloud Spending Surge,” 2026; ClusterMAX 2.0 neocloud tiering: SemiAnalysis, “ClusterMAX 2.0: The Industry Standard GPU Cloud Rating System,” 2026. Market-size figures from Mordor Intelligence neocloud-market projection. ↩
VanEck Semiconductor ETF holdings: StockAnalysis SMH holdings and VanEck SMH product page, updated through Q1 2026. Top-ten holdings concentration of 73 per cent reflects the index methodology (MVIS US Listed Semiconductor 25 Index). ↩
XLU constituent weights: State Street XLU product page, holdings as of January 2026. IDU and UTES from iShares and Virtus respectively. ↩
Equinix 2026 AFFO guidance and capacity targets: company Q3 2025 earnings release; Digital Realty 2026 outlook similar. Iron Mountain data-centre revenue trajectory: company filings and The Motley Fool, “Best Data Center REITs for 2026,” 2026. ↩
ETF Database, “Treasury Yields Snapshot: May 8, 2026,” confirmed against FRED T10Y2Y and the Federal Reserve H.15 daily release for the same date. ↩
PJM 2026/2027 Base Residual Auction Report, July 2025: PJM, “Auction Procures 134,311 MW of Generation Resources,” 22 July 2025; 2027/2028 auction cleared at $333.44 per MW-day in December 2025. Coverage: Enel North America, “PJM 2026/2027 Capacity Auction Results,” and POWER Magazine. ↩
European Central Bank, “Danish krone (DKK)”. ERM II central rate 7.46038 DKK per EUR with a ±2.25 per cent fluctuation band. After Bulgaria’s euro adoption on 1 January 2026, Denmark is the sole remaining ERM II participant. ↩
Investing.com, “EUR/USD Enters 2026 Near Key Resistance as Fed Cuts Meet ECB Hold,” April 2026; ECB and Federal Reserve policy-rate statements through Q1 2026. ↩
Aktiesparekontoloven 2026 indskudsloft og lagerbeskatningssats: Skat, “Aktiesparekonto,” opdateret januar 2026; Optimal Invest, “Aktiesparekonto skat og regler 2026,” 2026. Indskudsloft 174.200 DKK; sats 17 per cent på årets afkast efter lagerprincippet. ↩
Skats positivliste 2026: Invested.dk, “Skats positivliste over ETF’er som fra 2026 skal beskattes som aktieindkomst,” 2026; Indeksinvest.dk ETF-skattestatus oversigt for 2026. Progression threshold for aktieindkomst at 27 vs 42 per cent: DKK 79,400 for 2026. ↩
Pensionsindbetalingslofter 2026 og PAL-skat: Skat, “Fradrag for indbetalinger til pension,” 2026; Pensionsvalg.dk, “Aldersopsparing 2026: skat og regler,” 2026. Ratepension/ophørende livrente fradragsloft 68.700 DKK; aldersopsparing 9.900 / 64.200 DKK; PAL-sats 15,3 per cent. ↩