Nscale

Phase A — Understand the business

Company Overview

What it is. Nscale is a vertically-integrated "neocloud" — a GPU-cloud provider that designs/builds/operates AI data centres and rents accelerated compute (training + inference) to a small number of very large customers. It is the European answer to CoreWeave (US) and Nebius (US/NL), positioned explicitly around AI sovereignty and cheap Nordic renewable power.

Origin — the crypto-to-AI pivot. This is the single most important fact about the company. Nscale was incorporated 29 May 2024 in London and spun out of Arkon Energy, a Melbourne-based renewably-powered bitcoin-mining business run by founder Josh Payne. When crypto turned and ChatGPT landed, Payne wound Arkon down — Arkon lost $102M on $19M revenue in 2024 and was struggling to repay loans — and re-pointed the same playbook (buildings full of chips next to cheap power) at AI. The flagship site, Glomfjord, Norway, is a former Arkon crypto facility on 100% hydro with free ambient cooling. So Nscale is ~2 years old as an AI company, and its "supercomputing heritage" is really ~18 months of crypto-mining infrastructure re-badged.

Business model (vertically integrated). Five layers, per the company: (1) GPU compute for training/inference; (2) optimised networking; (3) modular / owned data centres; (4) a platform-orchestration control plane (integrates Kubernetes, Slurm, MLOps, plus VAST Data's "AI OS"); (5) managed services. The pitch vs. rivals is that Nscale controls power + shell + GPU + software end-to-end, which (in theory) lowers cost and speeds deployment.

Customers. Radically concentrated. The anchor is Microsoft — a $6.2B binding 5-year deal (Norway, from 2026) that expanded into a ~$14B / ~200,000 NVIDIA GB300 GPU multi-geography agreement. Secondary/announced: OpenAI (originally Stargate UK + Stargate Norway — see Lens 8, the relationship has since frayed), Singtel (GPU capacity, Sep 2024), and via the OpenAI unwind, Google took UK capacity. There is essentially one paying whale (Microsoft) and a halo of frontier-lab logos in various states of commitment.

Suppliers. NVIDIA is supplier, investor, AND lease guarantor simultaneously (Lens 3/10 — this is the circular-financing crux). Hardware/systems partners: Dell and Lenovo (servers, both also investors), Nokia (networking, investor), VAST Data (storage/AI-OS for 300,000 Grace-Blackwell GPUs).

Contract structure. Long-dated take-or-pay-style offtake — the Microsoft anchor is a binding 5-year term. This is the load-bearing feature of the whole capital structure: the offtake contracts, not the depreciating GPUs, are what lenders underwrite against (Lens 10).

Supply Chain

Map: cheap renewable power → land/shell → GPU + systems → orchestration software → hyperscaler / frontier-lab buyer. Named stakeholders at each node:

• Power (upstream chokepoint #1): Nordic hydro — Glomfjord & Narvik/Kvandal, Norway (100% hydroelectric); Keflavik/Blönduós, Iceland; plus grid-connected sites in Portugal (Sines / Start Campus), UK (Loughton Essex; Northumberland ex-coal site; 3 BT-owned sites, 14MW), and US (Texas ex-Ionic Digital shell; Monarch Compute Campus, West Virginia). The Aker ASA partnership (Norwegian energy/industrial conglomerate, now ~27.3% owner post-JV rollup, up from 9.3%) is the power+land access engine in Norway.
• Silicon (chokepoint #2 — single-source): NVIDIA GB300 (Blackwell Ultra) is the core; earlier the company launched on AMD MI300X but has pivoted decisively to Nvidia (which then invested and guaranteed leases). Vera Rubin (next-gen) already contracted — Microsoft's 30,000 additional GPUs at Narvik are Rubin-class.
• Systems/integration: Dell, Lenovo (rack/server), Nokia (data-centre networking), VAST Data (storage + AI OS).
• Real estate / shells: Ionic Digital (Texas lease), Start Campus (Sines, Portugal, third-party campus), BT (UK sites), acquired Monarch campus (WV).
• Buyer (chokepoint #3 — the real one): Microsoft is the demand chokepoint. Google and OpenAI are secondary/contingent. Singtel channels capacity into SE Asia.

Verdict on the chain: power is genuinely differentiated (Nordic hydro is cheap, clean, and cold), but silicon is single-sourced to Nvidia and demand is single-sourced to Microsoft. Two of the three chokepoints are counterparties Nscale does not control, and one of them (Nvidia) is simultaneously financing the purchase of its own product.

Competitive Advantages (moats)

Claimed moats and how real they are:

1. Cheap renewable power + cold climate (REAL, but shared). Nordic hydro at low PUE is a genuine cost edge for the European footprint, and the "AI sovereignty" wrapper resonates with EU regulators and buyers wanting non-US data residency. But Crusoe (stranded gas/renewables), Nebius (owns Finnish/EU DCs), and every European colo chasing the same hydro basins share this — it is a cost advantage, not a moat. And Nscale's US Texas footprint has no such edge.
2. The Nvidia relationship (REAL, but a leash not a moat). Jensen Huang publicly called Nscale a "national champion for AI infrastructure"; Nvidia is an equity investor across the last two rounds AND guaranteed up to $860.3M of lease obligations for the Texas facility in exchange for warrants. That guarantees GPU allocation — the scarcest input in the industry — which is a real edge over unfavoured neoclouds. But it makes Nscale a distribution arm of Nvidia, not an independent moat: the same favour can be extended to CoreWeave, Lambda, Nebius (Nvidia invested in all four) and withdrawn at Nvidia's discretion.
3. Vertical integration (WEAK moat). Owning power→shell→GPU→software could lower cost and speed deployment, but it is capital-intensive, unproven at scale for a 2-year-old, and every serious neocloud is converging on the same integrated model.
4. Switching costs (WEAK). Hyperscaler offtake is fundamentally fungible capacity — Microsoft moved OpenAI's Norway/UK allocation between providers and picked up Nebius's New Jersey capacity in parallel. Buyers treat neoclouds as interchangeable MW.

Bargaining power: Nscale is the weaker party on both sides. Upstream, Nvidia controls allocation and pricing. Downstream, Microsoft is a $3T buyer that can (and did) reroute demand. Nscale's leverage is that AI compute demand currently "far outweighs supply" (Payne's phrase) — a cyclical condition, not a structural moat. When supply catches up (as it did for H100s: $8/hr → $2/hr in ~12 months ), the bargaining position inverts.

Bottom line: The durable edge is thin. What Nscale actually has is privileged Nvidia allocation + a cheap-power land bank + first-mover European sovereignty branding — advantages that are real today and erodible tomorrow.

Segments

No segment financials are disclosed (private, our figures empty). Structurally, revenue can be decomposed by geography and by customer, both from `` contract data:

• By geography (contracted GPU capacity, GB300-equiv): Texas ~104,000 · Norway (Narvik) ~52,000 · UK (Loughton) ~23,000 · Portugal (Sines) ~12,600 = ~191,600 under the Microsoft umbrella alone. Total pipeline ~1.3 GW, with ~250 MW committed by Q4 2026 and >1 GW by 2029.
• By customer: overwhelmingly Microsoft (the ~$14B / 200k-GPU relationship dominates), with OpenAI/Google/Singtel as fringe. Customer concentration is effectively ~single-name — the defining risk of the whole thesis (Lens 13).

Trend: capacity is contracted and ramping (Q1 2026 Portugal → Q3 2026 Texas → Q1 2027 Loughton → late-2027 Rubin expansions), but almost none of it is revenue-generating yet. The "segment mix" is a construction schedule, not an operating P&L. n/a — not disclosed for any segment margin or income.

---

Phase B — Measure performance (+private overlay: lenses swap to funding / traction / cap table)

Funding & Valuation Trajectory

The story of Nscale is its funding velocity — one of the fastest capital ramps in European tech history. All ``, unaudited:

• Total equity raised: ~$3.7B in <2.5 years. Valuation more than doubled Sep 2025 → Mar 2026.
• Debt on top of equity: $1.4B GPU-backed loan (since start of 2026) + $790M Norway financing facility. Plus Nvidia's $860.3M lease guarantee for Texas. This is a debt-plus-vendor-financed balance sheet, not equity-funded (Lens 10/13).
• The number that matters — revenue: the only audited figure is a ~$24M loss for the 7 months to Dec 2024 (UK Companies House); one estimate puts FY2025 revenue at ~$1.27M. Nscale is, for practical purposes, pre-revenue. A $14.6B valuation on ~$0 of realised revenue implies the entire value is the contracted backlog + Nvidia halo + IPO optionality.
• Payne's claim: "every dollar is backed by real revenues," demand "far outweighs supply". Unverifiable and in tension with the disclosed accounts.

Lens (Traction & Unit Economics — added by +private overlay)

• Contracted (not realised) TCV: the Microsoft relationship is cited "up to $14B" in revenue over its life (FT estimate), of which $6.2B is the binding 5-year Norway tranche. No company-disclosed total RPO/backlog — n/a — not disclosed (contrast CoreWeave's audited $99.4B backlog and Nebius's disclosed guide — Lens 7).
• Unit economics: n/a — not disclosed. Neocloud economics generally: gross margins hinge on utilisation × rental rate − power − depreciation − interest. Nscale's inputs look favourable (cheap hydro, Nvidia allocation, take-or-pay offtake), but not one margin figure is public, and the industry's cautionary tale is H100 rates collapsing ~80% in a year.
• Capex intensity: ~1.3 GW pipeline at roughly $25–40M/MW all-in for AI-grade capacity implies tens of billions of eventual capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs.. This is why the balance sheet leans on debt + vendor finance + offtake collateral.

Founder / Management Interviews (sentiment) (+private: podcasts/press, not earnings calls)

No earnings calls exist. Reading Payne's public posture across 2025→2026 press: consistently maximalist and promotional — "national champion," "largest in European history" (x2), "every dollar backed by real revenues," IPO targeting late 2026. The tone is a founder selling into an IPO, and the messaging leans hard on the Nvidia endorsement and the sovereignty narrative. What's notably absent from the talk track: utilisation rates, realised revenue, gross margin, and any acknowledgement of counterparty concentration. The board additions — Sheryl Sandberg, Susan Decker (ex-Yahoo CFO/president), Nick Clegg (ex-Meta policy) — read as deliberate pre-IPO credibility-and-government-relations signalling more than operating firepower.

Cap Table & Secondary Marks + Peer Comps

Cap-table quality (the +private tell): The syndicate is unusually strong for a 2-year-old and skews toward crossover / public-market-adjacent capital — an IPO-proximity signal:

• Strategic/vendor: NVIDIA, Dell, Lenovo, Nokia (all suppliers-as-investors → circular-financing flag, Lens 10).
• Industrial anchor: Aker ASA (~27.3%) — Norwegian conglomerate, the largest owner; provides power/land + balance-sheet ballast.
• Crossover / hedge-fund / IPO-signal money: Fidelity, Point72, Citadel, Jane Street, Blue Owl, G Squared, 8090 Industries, Linden, Astra. A Fidelity + multi-pod-shop cap table this early is a strong "IPO is the plan" marker.
• Secondary marks: n/a — not disclosed.

Peer comps (public neoclouds — the market Nscale will IPO into). Multiples ``, dated; Nscale's own multiple is n/a — private:

Read-through: Public neoclouds trade ~4x (CRWV) to ~18x (NBIS) forward sales on real, audited, multi-billion revenue with disclosed backlog. Nscale wants a $20B+ IPO on contracted-but-unrealised revenue. If it prints even $1–2B of 2027 revenue as Texas/Norway ramp, a $20B tag is ~10–20x forward — in the peer band but at the frontier of it, and with worse revenue quality (concentration + no operating history) than either public comp. If the ramp slips, the comp set (CRWV at ~4x) is a long way below $14.6B.

Funding / Product / Catalyst Events (>5% value moves) (+private: rounds & contracts, not stock)

No public stock; the "value catalysts" are rounds, contracts, and the OpenAI unwind — the most revealing event in the file:

• Jul 2025: Nscale + Aker + OpenAI announce Stargate Norway (100,000 GPUs by end-2026, Narvik). Peak halo.
• Sep 2025: $1.1B Series B + Microsoft $6.2B binding deal. Twin de-risking events.
• Oct 2025: ~$14B / 200k-GPU Microsoft expansion; VAST Data 300k-GPU deal; IPO targeted late-2026.
• Mar 2026: $2B Series C at $14.6B; Aker → 27.3%; Sandberg/Decker/Clegg to board.
• Apr 2026 — the crack: OpenAI pulls out of Stargate Norway (failed to conclude terms with Nscale) and pauses Stargate UK citing UK energy cost + regulation; Microsoft takes over the Narvik capacity (+30,000 Vera Rubin GPUs) and Google takes the vacated UK capacity. Marketed as continuity; structurally it means the flagship "OpenAI Stargate" story evaporated within 9 months and Nscale is now even more dependent on Microsoft.
• Jun 2026: Fortune/Time run skeptical features — "must prove the hype can survive reality".
• Jul 2026 (context): Broad neocloud repricing — CoreWeave −14% / Nebius −17% in a single day on a Bloomberg report that Meta may build its own cloud and dump excess capacity. The public comps Nscale will IPO against just took a ~15% concentration-risk haircut.

Pattern: the market rewards Nvidia allegiance + hyperscaler anchor + cheap-power sovereignty, and punishes customer-concentration and monetisation doubt. Nscale sits on the good side of the first and the bad side of the second.

---

Phase C — Judge people & books

Management

• Josh Payne (Founder & CEO). Background: recruitment → capital markets → energy/crypto infrastructure (Arkon Energy). Track record is mixed-to-concerning: Arkon lost $102M on $19M revenue (2024) and struggled to service debt before being wound down. The bull read: he saw the crypto→AI pivot early and executed a $3.7B raise + Nvidia anointment + Microsoft anchor in ~2 years — extraordinary capital-markets execution. The bear read: he is a promoter/dealmaker who has not yet run a profitable operating business, and Arkon's economics are a warning about leverage into a hardware cycle.
• Tenure & skin in the game: Founder, ~2 years in; personal ownership diluted by ~$3.7B of raises + Aker's 27.3% but presumably still material as founder. Exact insider % n/a — not disclosed.
• Operating bench (the real strength): Nidhi Chappell (ex-Head of AI Infrastructure at Microsoft, ex-Intel DC director) — a genuinely heavy hire who knows the anchor customer from the inside; Alex Sharp (President, DC Group — "54 data centres" built, 30 yrs); David Power (CTO, 20 yrs HPC). This bench is more credible than the founder's operating record and is the reason to believe the build can actually be delivered.
• Capital-allocation history: Short and debt-heavy — $1.4B GPU loan + $790M Norway facility + Nvidia lease guarantee. The strategy is maximum leverage into contracted offtake. Whether that is brilliant or reckless depends entirely on Microsoft honouring 5-year terms and GPU residuals holding (Lens 13).
• Red flags: (1) Vendor-as-investor-as-guarantor (Nvidia) — related-party-adjacent circularity; (2) promotional communication with no operating metrics; (3) founder's prior company lost money and was wound down; (4) board stacked with policy/comms names (Clegg, Sandberg) ahead of an IPO — optics-forward.
• Archetype: Founder-promoter / capital-markets operator, not a systems-operator founder. Right archetype for raising and anchoring; the open question is operating and monetising.

Forensic Red Flags (web-only — no filings to audit)

Accounting/structural risks (all ``, unaudited — flagged because no audited statements exist):

1. Circular / vendor financing (the headline flag). NVIDIA is investor + supplier + lease-guarantor ($860.3M Texas guarantee for warrants). This is the exact structure critics flag across the sector — "vendors and customers exchange money," Nvidia funds purchases of its own silicon. It flatters demand signals and can mask true third-party pull-through.
2. Duration mismatch (balance-sheet time-bomb). GPUs depreciate over ~3–5 years (and economically faster — H100 rates fell ~80% in ~12 months), while the debt/leases assume long asset lives. Lenders heavily discount GPUs as collateral and underwrite the offtake contract instead. So the loans embed two fragile assumptions: GPU residual value holds and utilisation stays high. If either breaks, the collateral and the cash flow break together.
3. Revenue recognition risk. With ~$0 realised revenue and long-dated take-or-pay contracts, any future revenue-recognition and RPORemaining performance obligationsRevenue a company has already signed contracts for but has not yet delivered or booked. disclosure will need scrutiny — how much is genuinely non-cancellable, how much is optionality (e.g. Microsoft's option on the Texas 700MW phase 2, which is not committed).
4. Opacity. No audited financials, no disclosed backlog/RPO, no utilisation, no margins. The valuation rests on PR-disclosed contract headlines and a strong cap table. That is the single biggest forensic issue: you cannot verify the core claim ("every dollar backed by real revenues") from any public document.
5. "British champion" positioning vs. Australian/crypto heritage — an FT-raised optics point, minor but relevant to how much government-sovereignty tailwind is durable.

Regulatory findings (required sub-section). Per regulatory/regulatory-findings.md (Stage 1, 2026-07-06): Nscale has no SEC CIK (private, non-filer) → zero EDGAR Litigation Releases or AAERs possible. Non-SEC web search — "Nscale" (FTC OR DOJ OR FDA OR CFPB OR consent decree OR settlement OR fine OR penalty) enforcement — returned no material enforcement actions against the company as of 2026-07-06. The only adjacent public frictions are UK local-planning objections (Loughton council: power demand + grid impact) and European grid-connection queues (7–13 years) — operational/permitting, not enforcement. No Item-3 legal-proceedings disclosure exists (no filings). Conclusion: No material regulatory or legal enforcement findings — verified via SEC EDGAR EFTS (no CIK → none possible) and web search as of 2026-07-06; unaudited per public sources.

---

Phase D — Project & stress-test

IPO-Readiness & Path-to-Tradeable (+private: replaces Forward Projection)

Readiness assessment (be-early payoff lens). Nscale is a genuine pre-IPO / readiness ~4 on the 1–5 scale (late-stage, secondary-active, IPO explicitly targeted):

• Stated catalyst: CEO targets IPO late-2026; analysts float a >$20B IPO valuation.
• Milestones that unlock an S-1 / de-risk the print:

1. Texas + Norway ramp into realised revenue (Q3 2026 Texas start is the pivotal proof-point — first material revenue).
2. A disclosed, audited backlog/RPO to convert "contract PR" into a defensible number (the CoreWeave $99.4B template).
3. Utilisation + gross-margin disclosure — the metrics currently missing.
4. Customer diversification beyond Microsoft — the single biggest IPO-marketing gap (Google/OpenAI partially fill it but both are soft).
5. A supportive neocloud tape — the Jul-2026 CRWV/NBIS ~15% selloff shows the IPO window is weather-dependent.

• Path-to-tradeable: clean and near — strong crossover cap table (Fidelity/Citadel/Point72/8090), a real underwriting-grade anchor contract, and a management team pointed at the exit. This is one of the more IPO-ready privates on the frontier. The risk is not whether it can list but at what price and into what tape.

No EPS forecast is logged (pre-revenue private; per skill, no our model create in the watchlist loop and no EPS line for a +private name). The scoreable binary that matters: "Nscale files an S-1 (or completes a direct/IPO listing) by 2027-06-30." Base-case probability ~0.55. (Logged here as the tracked prediction; not written to our model per the unattended-watchlist rule.)

Update-back note: Nscale is absent from research/private-watch.json (checked — 0 hits). It should be added with beat: datacenters, stage: pre-ipo, ipo_readiness: 4, lead_investors: "Aker, 8090, NVIDIA, Dell, Nokia", catalyst: "Texas/Norway revenue ramp → IPO targeted late-2026", dossier: <this file>. Per wave boundaries I am NOT editing that file — flagging for the master session to action.

Bull vs Bear

Bull case. Nscale is the European neocloud winner in the making at the exact moment AI-infra demand outstrips supply. It has the three things that matter most: (1) Nvidia's favour (GPU allocation + capital + lease guarantees — a "national champion" blessing that peers would kill for); (2) a binding, multi-billion, 5-year Microsoft anchor that underwrites the debt; (3) a structural cost + sovereignty edge in cheap Nordic hydro that EU buyers and regulators actively want. It has raised $3.7B + billions in debt with a crossover cap table pointed at a late-2026 IPO into a market where public comps trade at 4–18x sales. If Texas and Norway ramp on schedule and even one more hyperscaler diversifies the book, a $20B+ listing is very achievable, and the early money (Series A at a fraction of $14.6B) is already multiples ahead. The bull's contrarian point: the market underrates favoured neoclouds — being the Nvidia-anointed European player is a winner-take-most position in a $180–400B-by-2030 market.

Bear case. Nscale is a 2-year-old crypto-miner in a hydro jacket, valued at $14.6B on essentially zero revenue, whose entire equity value rests on one customer (Microsoft) honouring long-dated contracts and GPU residuals defying physics. Three things can permanently impair it: (1) Demand concentration — Microsoft can reroute or renegotiate (it already moved OpenAI's capacity around like chess pieces; Meta's Jul-2026 "build our own cloud" signal just cut CRWV/NBIS ~15% on exactly this fear). (2) The duration-mismatch debt trap — GPU-collateralised loans + lease guarantees assume high utilisation and holding residuals; H100 rental rates fell ~80% in a year, and a demand air-pocket turns the leverage lethal. (3) Monetisation gap — "fast to build, slow to monetise" is the sector's epitaph, and Nscale's only audited number is a loss. Expectations baked into $14.6B (→$20B+) are priced for flawless execution of a build that is already slipping (Loughton 2026→2027).

Pre-mortem (18 months out, thesis broke — what happened?): Most likely story — a 2027 AI-capex digestion pause. Microsoft slows incremental commitments (as it paused parts of Stargate), Texas ramps into softer-than-contracted utilisation, GPU rental rates roll over, the $1.4B GPU loan + $790M facility hit refinancing at 12%+ into a risk-off neocloud tape, and the IPO is pulled or prices at a fraction of $14.6B. Nvidia, having already extracted warrants, does not double down on a favoured-child bailout. Equity holders below the Series C mark are impaired.

Are the multiples too high? On realised fundamentals, yes, extreme — $14.6B / $1.27M revenue is not an investable multiple, it's an option on the backlog + IPO. On contracted TCV ($14B Microsoft), a $14.6B equity value is <1.5x lifetime TCV, which is not crazy if the contracts are real, non-cancellable, and margin-rich — none of which is verifiable.

Contrarian view (what the market refuses to see): Both bulls and bears fixate on the Microsoft deal size; the thing hiding in plain sight is that Nscale's real product is not compute — it's a financing vehicle that converts Nvidia allocation + cheap power + a hyperscaler signature into cheap capital. That works spectacularly in an up-cycle and is structurally pro-cyclical — it has no counter-cyclical ballast. The IPO is less a "growth company going public" and more a liquidity event for a leveraged infrastructure bet, timed to the top of the capex wave.

Devil's Advocate (short-seller)

Dismantling the bull case:

• The one-customer problem is the whole ballgame. Strip out Microsoft and there is no business — OpenAI already walked, Google is a hand-me-down, Singtel is a channel. A short thesis writes itself the day Microsoft renegotiates, delays a phase, or exercises an option downward (note Texas phase-2 700MW is an option, not a commitment). The Jul-2026 tape proved the market will mark concentration risk down ~15% instantly.
• The moat is a rented phone. Every "advantage" is a counterparty's discretion: Nvidia's allocation, Microsoft's demand, Aker's power. Nscale controls none of its three chokepoints. Nvidia can anoint the next champion tomorrow.
• The balance sheet is the short. GPU-collateralised debt + vendor lease guarantees underwritten on offtake contracts whose margins and non-cancellability are undisclosed. If utilisation disappoints 20–30%, the cash flows that service $2B+ of debt compress while GPU collateral depreciates through the loan — a classic residual-value + utilisation double-discount. This is precisely the structure the sector's most credible bears say is mispriced by credit markets.
• Accounting you can't check. No audited financials, no RPO, no utilisation. The only audited figure is a loss. "Every dollar backed by real revenues" is a claim, not a filing.
• Founder risk. Prior company (Arkon) lost $102M on $19M revenue and drowned in debt into a hardware down-cycle — the same leverage-into-a-cycle pattern now runs at 100x the scale.
• What single scenario permanently impairs it? A 2027 capex-digestion air-pocket (Meta/Microsoft/Amazon collectively slow AI capex — already the market's live fear) that hits before Nscale's revenue ramps and during its refinancing window. Plausibility: moderate and rising. The Jul-2026 selloff is the market pricing exactly this.

If growth disappoints 20–30%: on a pre-revenue name, a slower ramp doesn't dent an EPS line — it breaks the IPO and the debt refinancing simultaneously, which is far more binary and dangerous than a normal miss.

Management Questions (ordered by information value)

1. What is your total non-cancellable contracted RPO/backlog (audited), and what fraction is a firm commitment vs. a customer option (e.g. the Texas 700MW phase-2)?
2. What share of contracted and realised revenue is Microsoft, and what is your concrete plan and timeline to get any single customer below 50%?
3. What are your current blended GPU utilisation rate and gross margin on live capacity (Portugal/Norway), and what utilisation do your debt covenants assume?
4. Walk me through the $1.4B GPU-backed loan and $790M Norway facility: maturities, rates, covenants, and what happens to them if 2027 utilisation comes in 25% below plan.
5. What GPU residual-value and useful-life assumptions underpin your leases and depreciation, and how do they survive the H100 precedent (~80% rate decline in a year)?
6. Exactly what did OpenAI's exit from Stargate Norway/UK cost you in contracted revenue, and why should we not read it as the anchor-tenant model failing its first stress test?
7. What is Nvidia's total economic interest — equity %, warrant terms, lease guarantees — and where does it create conflicts between "best hardware for customers" and "most hardware sold for Nvidia"?
8. What is your monthly cash burn and current liquidity runway at today's build pace, absent the IPO?
9. What are your realistic first-material-revenue and first-positive-EBITDA dates, site by site?
10. How much AMD (or non-Nvidia) capacity will you deploy, and if the honest answer is ~none, how is "single-source to Nvidia" not a fatal dependency?
11. What is the true all-in capex per MW for your Nordic vs. Texas vs. UK builds, and how do the unit economics differ across them?
12. On the IPO — what revenue, backlog, and margin milestones must print before you file, and what tape conditions would make you delay?
13. What contractual protections do you have if a hyperscaler builds its own capacity and dumps excess (the Meta scenario that just repriced your public peers)?
14. How do European grid-connection queues (7–13 years) and local planning objections (Loughton) constrain your >1GW-by-2029 plan, and what is already de-risked vs. aspirational?
15. Given Arkon lost $102M on $19M revenue into a hardware down-cycle — what specifically is structurally different this time beyond "demand is higher today"?

---