Lambda

Phase A — Understand the business

Company Overview

Lambda is a "neocloud" — a specialist GPU cloud that rents Nvidia accelerators for AI training and inference, plus a legacy hardware line (GPU servers and workstations). Founded 2012 in San Jose by twin brothers Stephen Balaban and Michael Balaban; it started life as an AI applications/hardware shop and pivoted into cloud in the 2020 GPU-scarcity window. It brands itself the "Superintelligence Cloud" / "AI Developer Cloud."

The business model in plain terms. Lambda buys Nvidia GPUs (and, increasingly, finances them with debt and vendor arrangements), racks them in leased colocation space, and sells the compute four ways:

1. On-Demand Cloud — hourly per-GPU billing, no egress fees; the developer front door (H100 PCIe at ~$2.49/hr vs CoreWeave's ~$4.25/hr).
2. 1-Click Clusters — pre-configured 16–512-GPU H100/B200 clusters launchable in minutes (the self-serve mid-market product).
3. Private Cloud — dedicated, air-gapped, multi-year reserved clusters for enterprises and labs (the big-contract, revenue-visibility product).
4. Hardware — Vector/Scalar GPU servers and workstations, being phased out (Vector/Scalar lines discontinued August 2025); the Inference API and Lambda Chat were discontinued September 2025 — a deliberate retreat from the app layer back to pure infrastructure.

Customers. ~200,000+ AI developers and teams on the platform (up from "10,000 customers" as a paying base in 2024 ). Named logos: Amazon, Apple, Microsoft, Tencent, Kaiser Permanente, MIT, Stanford, Harvard, Caltech, Raytheon, and the U.S. Department of Defense. The defining commercial facts of the last year are two anchor contracts: a multibillion-dollar, multi-year Microsoft deal (Nov 2025) to deploy tens of thousands of Nvidia GPUs incl. GB300 NVL72 systems, and Nvidia itself leasing back 18,000 GPUs for $1.5B over four years — making its own chip supplier its largest customer (Lens 2/5/13).

Suppliers: Nvidia (binding single source for silicon + an equity investor + now a customer), plus systems integrators (Supermicro "AI factories" partnership, Aug 2025) and colocation landlords (EdgeConneX, Prime Data Centers, Aligned, Cologix). Competitors: CoreWeave (the public benchmark), Nebius, Crusoe, WhiteFiber, together with the hyperscalers (AWS/Azure/GCP) whose H100 list prices ($4–11/hr) Lambda materially undercuts.

Contract structure: a barbell — a long self-serve/on-demand tail (usage, no lock-in, price-competitive) and a growing reserved/committed head (multi-year Private Cloud + the Microsoft/Nvidia anchors). No take-or-pay disclosure exists (private), but the anchor deals are the visibility base; the on-demand book is spot-priced and exposed to the sector-wide rate collapse (Lens 13).

Supply Chain

Upstream → Lambda → end customer, every named node:

• Silicon (the binding input): Nvidia — H100, H200, A100, B200, GH200, GB300/Blackwell Ultra; RTX 6000 Ada for workstations. Single-source dependency for frontier capacity, and Nvidia is simultaneously supplier + Series D equity investor + the $1.5B leaseback customer — a triple entanglement deeper than a normal vendor relationship. [chokepoint — the single hardest dependency]
• Systems / integration: Supermicro (Blackwell "AI factory" reference designs, Aug 2025) and standard OEM server assembly. [chokepoint — assembly + lead times of 9–12 months on Nvidia allocation].
• Colocation / real estate (Lambda is asset-lighter than CoreWeave here): Lambda leases rather than owns most sites — ~15 data centers across the US, including EdgeConneX (Chicago, Atlanta), Prime Data Centers (Vernon/LAX01, CA — 21MW), Aligned (Plano, TX), Cologix (Columbus, OH), plus its owned/anchored Kansas City AI factory (24MW at launch, scaling to 100MW+). [differentiator — colocation model shifts real-estate/power capex risk onto landlords vs CoreWeave's owned mega-sites]
• Power: the sector's true bottleneck. Lambda's stated target is 3 GW of liquid-cooled capacity and 1M+ GPUs by 2030 — but current deployed power is far smaller (Kansas City 24MW, Vernon 21MW, etc.), so the 3GW figure is an ambition, not an asset. Unlike Crusoe (which owns generation) or CoreWeave (which contracts gigawatts directly), Lambda's power is landlord-supplied — lower control, lower capex, but also less of a moat.
• Capital (increasingly part of the "supply chain"): ~$2.3–2.4B equity raised to date, ~$500M in GPU-collateralized debt, and a pending ~$350M convertible/mezzanine (Mubadala Capital). In a business where revenue lags capex ~2:1, continued capital access is itself an input — a supply-chain node that can choke.
• Downstream / distribution: self-serve web signup (the 200k-developer funnel), plus enterprise sales for Private Cloud, plus the two anchor tenants (Microsoft, Nvidia). Lambda has NO owned distribution surface (no social network, no car fleet, no model of its own) — the inverse of xAI. It is distribution-poor and must win on price, availability, and developer UX.

Chain verdict: compute-and-power-gated at the top (Nvidia + landlord power), developer-funnel-rich but distribution-thin at the bottom. The differentiated node is the developer experience + price (a genuine self-serve funnel most neoclouds lack), NOT the physical stack, which is bought/leased like everyone else's.

Competitive Advantages (moats)

Real (or semi-real) moats:

1. The developer funnel. ~200k AI developers who signed up self-serve is the one genuinely hard-to-replicate asset — a bottoms-up acquisition channel with near-zero marginal CAC that CoreWeave (enterprise-sales-led) and Crusoe (utility-led) simply do not have. It is a distribution and a data moat (usage telemetry, workload familiarity). [the strongest point]
2. Price position. H100 at ~$2.49/hr vs hyperscaler $4–11/hr and CoreWeave ~$4.25/hr. Being the low-cost credible provider in a commoditizing input is a share-taking wedge — but it is a double-edged moat (Lens 13: low price = thin margin in a rate-collapse).
3. Nvidia proximity. Priority GPU allocation during shortages, early access to new silicon (SHARP integration July 2025; among the first with GB300), and Nvidia equity + the leaseback = a supply and demand tailwind rivals must compete for. [real, but see proximity is also a dependency]
4. 8+ year Microsoft relationship predating the 2025 anchor deal — switching-cost and trust capital with the single most important buyer of merchant AI capacity.

Weak / contested moats:

• The core product is a commodity. GPU-hours are fungible; the underlying silicon is identical to every rival's. There is no proprietary chip (unlike Groq/SambaNova), no owned power (unlike Crusoe), no owned model (unlike xAI). The moat is operational (funnel + price + UX + Nvidia access), not structural.
• Hyperscaler vertical integration is the existential overhang: Microsoft's own Azure, AWS, and GCP can — and do — build the same capacity, and Microsoft's ~$13B OpenAI stake shows the incumbents integrating the whole stack. Lambda is a supplier to the entity most able to disintermediate it.

Bargaining power: weak over Nvidia (needs the chips far more than Nvidia needs Lambda); weak-to-improving over anchor customers (Microsoft/Nvidia are giants; the leaseback means Nvidia is buying its own demand); stronger over the long tail of 200k developers (no single one matters). Net: the moat is a developer funnel + a price wedge + Nvidia access — genuine but narrower and shallower than CoreWeave's $98.8B contracted backlog or Crusoe's owned power.

Segments

No audited segment disclosure exists (private). The best public decomposition:

Geography: US-centric (all 15 DCs domestic); not otherwise disclosed.

Margin by segment (the key tell): Sacra reports ~50% overall gross margin in H1 2025, but ~61% excluding the non-cloud (hardware) segments. That gap is the whole segment story: the cloud business is structurally higher-margin (~61%), and management is actively pruning the low-margin hardware/app lines to converge the blended number up — the strategic logic behind killing Vector/Scalar and the Inference API. The trend is toward a cleaner, cloud-only, higher-margin P&L ahead of the IPO. Watch whether blended GM actually climbs toward 60%+ as hardware runs off, or whether the on-demand rate collapse (Lens 13) drags cloud GM back down first.

---

Phase B — Measure performance (+private overlay: funding & traction in place of earnings)

Funding & Valuation Trajectory

The valuation ladder is the story:

Total equity raised: ~$2.3–2.4B. Plus ~$500M GPU-collateralized debt.

Read: a ~4× valuation step-up in ~18 months (Series C $1.5B → secondary ~$9B), driven by the anchor contracts and the AI-capex tape, not by a step-change in audited profitability (Lambda is still loss-making — TTM net loss ~$175M, Lens 10). The move from ~3.5× (Series C) to ~5.9× (Series D) revenue multiple, and the secondary now implying ~12× the $760M run-rate, means each round is pricing IPO optionality and multiple-expansion, not just growth. The syndicate quality (see Lens 7) is the tell that this is a genuine pre-IPO name, not a promotional one.

Founder / Management Commentary (sentiment trend) (earnings-call swap → interviews, blog, positioning)

No earnings calls exist (private). The public "sentiment surface" is the leadership blog, press, and the founders' posture:

• The narrative escalated from "GPU cloud" to "Superintelligence Cloud" over 2024→2026 — a deliberate up-branding to justify gigawatt ambitions and an IPO multiple. The 3GW/1M-GPU-by-2030 target is the anchor claim.
• The single loudest signal is a leadership overhaul (May 2026): founder Stephen Balaban stepped back from CEO to CTO; Michel Combes (ex-Sprint/SoftBank International/Alcatel-Lucent CEO) took the CEO seat; brother Michael Balaban → CPO; John Donovan (ex-AT&T Communications CEO) → Chairman. The messaging is "assembling a leadership team to power gigawatt-scale infrastructure for the superintelligence era." The subtext is unmistakable: the founders are installing telco/infrastructure-scale operators and a public-company-grade board ahead of an IPO — a classic "hire the adult" pre-listing move.
• What they stopped saying: the app-layer story (Inference API, Lambda Chat, killed Sept 2025) — a narrowing to "we are pure infrastructure," which is both more honest and more commoditized.

Tone trend: increasingly institutional and IPO-oriented, shifting from founder-scrappy to infrastructure-incumbent. That is bullish for governance/credibility and bearish for the "founder-obsession moat" — Combes is a capable operator but not a product visionary; the vision now rests with Balaban-as-CTO.

Cap Table, Syndicate & Secondary Marks (comps swap → who owns it + how it's marked)

Syndicate quality (the IPO-proximity tell): the register combines strategics (Nvidia — supplier + investor), crossover/institutional (ARK Invest, US Innovative Technology Fund, TWG Global, Mubadala Capital), sovereign (Mubadala; QIA-adjacent via the sector), venture (Andra Capital, SGW, B Capital, Gradient Ventures, G Squared), and notable individuals (Andrej Karpathy). The presence of TWG Global leading the last primary and Mubadala leading the pre-IPO convertible is exactly the crossover/sovereign signature that precedes an S-1. Banks already hired: Morgan Stanley, J.P. Morgan, Citi.

Secondary marks: Forge $49.25/sh @ ~$9B (2026-06-13); NPM $43.16/sh (2026-06-17); PM Insights notes secondary shares +12% over 90 days on IPO speculation. So the private market is marking Lambda ~55% above its last primary ($5.9B → ~$9B) — bullish demand, but also a warning that a disappointing IPO price could re-rate secondaries down.

Valuation-multiple comps table — Lambda's ~$9B secondary vs the public/private neocloud peer set, using hard-filed numbers where they exist:

Read: Lambda at ~12× run-rate sits above CoreWeave (~8–11× on a vastly larger, contracted backlog) but below Nebius (~21–24×). The honest framing: Lambda is not cheap relative to CoreWeave — CoreWeave has ~7× the revenue and a $98.8B take-or-pay backlog Lambda cannot match — but it is cheaper than the most expensive name in the group (Nebius) and carries a better organic-funnel story. The multiple is defensible only if the ~80% growth holds through the IPO and the margin-mix cleanup (Lens 4) delivers. Multiples labeled; where not sourceable → n/a.

Funding / Product Catalysts (the events that move the mark)

For a private, "what moved the valuation" replaces "what moved the stock":

• Feb 2024 — Series C $1.5B (first unicorn mark).
• Feb 2025 — Series D $480M @ $2.5B; Nvidia + ARK + In-Q-Tel join (strategic validation; +67% mark).
• Sep 2025 — Nvidia $1.5B / 18,000-GPU leaseback + IPO-bank hire (MS/JPM/Citi) (the IPO starting gun; the round-trip demand backstop).
• Nov 2025 — Microsoft multibillion GB300 deal + Series E $1.5B+ @ ~$5.9B (TWG Global) (the anchor-tenant + mega-round double; the biggest single re-rate).
• Nov 2025 — Kansas City 24MW→100MW AI factory + Prime/EdgeConneX/Aligned expansions (capacity build-out narrative).
• May 2026 — Combes CEO / Donovan Chairman leadership overhaul (IPO-governance signal).
• Jun–Jul 2026 — secondaries to ~$9–9.5B; Mubadala ~$350M pre-IPO convertible (private-market demand confirming the IPO thesis).

Pattern: the mark reacts to (a) anchor contracts, (b) Nvidia entanglement events, and (c) governance/IPO-prep milestones — not to organic profitability (there is none yet). This is a valuation driven by narrative + backstop demand + IPO proximity, which is exactly the profile the circular-financing skeptics attack (Lens 13).

---

Phase C — Judge people & books

Management

• Michel Combes — CEO (since May 2026). A career infrastructure operator: former CEO of Sprint, SoftBank International, and Alcatel-Lucent. Track record: ran carriers that "built the physical foundation the internet runs on" — genuinely relevant for a gigawatt-scale buildout and for the debt-and-capital-markets execution an IPO+datacenter roll-up demands. Archetype: professional turnaround/scale operator, not a founder-visionary; his Alcatel-Lucent and Sprint tenures were about cost, integration, and balance-sheet management — the exact skills a levering neocloud needs, but also a signal that the business is entering its industrial/financial-engineering phase.
• Stephen Balaban — Co-founder, now CTO (stepped down from CEO May 2026). Ran Lambda from 2012; the technical/product vision now sits with him. Skin in the game: substantial founder equity (twins co-founded and controlled it through Series E); the move to CTO keeps them technically central while handing operations to a public-company operator — a mature, non-ego founder transition. [positive governance signal]
• Michael Balaban — Co-founder, now CPO. Product execution.
• John Donovan — Chairman (since May 2026). Ex-CEO AT&T Communications — a public-company-grade board anchor. [IPO-readiness signal]
• Charles Fisher — CFO; Leonard Speiser — COO.

Capital-allocation history: aggressive but directionally disciplined — raised ~$2.4B equity + ~$500M debt, poured it into GPUs and colocation, and pruned the low-margin/low-strategic lines (hardware, Inference API, Chat) to focus capital on the cloud. That pruning is the single best capital-allocation signal here: they are willing to kill revenue lines to protect margin and focus — the opposite of empire-building. Red flags: the Nvidia triple-entanglement (investor + supplier + customer) is a related-party-adjacent structure that will get IPO-prospectus scrutiny (Lens 13); and a brand-new CEO 6–12 months before a planned IPO is execution risk (Combes must learn the business and sell the story simultaneously).

Net: founder-built, now professionally-operated, with a credible IPO-grade board — a stronger governance setup than a typical private at this stage. The risk is not competence; it is that the new operator's skill set (infrastructure/finance) confirms the business is now a capital-markets and buildout execution story, where the swing factor is debt cost and demand durability, not product genius.

Forensic Red Flags (no filings → structural/sector risks, web-only)

Without audited statements, this lens flags structure, not line items — and labels everything unaudited:

• Loss-making at scale. TTM net loss ~$175M; H1 2025 net loss ~$24M. Growth is real but not yet profitable — the entire bull case rests on operating leverage that has not yet arrived.
• Revenue lags capex ~2:1 across neoclouds — meaning Lambda must keep raising (equity + debt) to fund the buildout; a capital-markets freeze is an existential risk, not a growth speed-bump.
• GPU depreciation — the sector's hidden fault line. Peers disagree violently on useful life: CoreWeave 6 years, Nebius 4 years, Scaleway ~3 years. If the true economic life is 2–3 years (as Blackwell/Rubin efficiency gains and the H100 rate collapse suggest), then any neocloud carrying GPUs at 6 years is overstating profit today and facing write-downs tomorrow. Lambda's depreciation policy is undisclosed (private) — a must-ask in the S-1 (Lens 14). [the single most important accounting unknown]
• ~$500M GPU-collateralized debt against a depreciating, price-collapsing asset — smaller than CoreWeave's ~$25B, but the same structure: debt secured by GPUs whose collateral value is falling as rental rates fall ($7–10/hr in 2024 → $2–4/hr late 2025).
• The Nvidia round-trip. The $1.5B leaseback means Lambda's largest customer is its chip supplier and equity holder — revenue that "originates from within the same small circle". Not fraud, but a quality-of-revenue flag: how much of Lambda's demand is genuine third-party vs. vendor-backstopped? (Lens 13.)
• Accelerating hardware run-off could create one-time inventory/asset charges as Vector/Scalar wind down.

Regulatory findings (required sub-section). Per regulatory/regulatory-findings.md (read this pass): Lambda has no CIK — it is private and not required to file with the SEC, so no EDGAR enforcement search (Litigation Releases / AAERs) is possible; total_sec_findings: 0. Non-SEC web search ("Lambda" (FTC OR DOJ OR consent decree OR settlement OR fine OR penalty) enforcement) and a governance/litigation search returned no material regulatory actions, lawsuits, layoffs, or SEC complaints against Lambda as of 2026-07-06. There is no 10-K Item 3 to quote (no filings). Conclusion: No material regulatory or legal findings — verified via the absence of an SEC CIK (no EDGAR record possible), a non-SEC web enforcement search, and a governance/litigation web search as of 2026-07-06. All findings unaudited per public sources; this will only be truly testable when the S-1 discloses legal proceedings and related-party transactions.

---

Phase D — Project & stress-test

IPO-Readiness & Path-to-Tradeable (forward EPS swap → the be-early lens)

private-watch.json grounding: stage=late, ipo_readiness=4/5, lead investors USIT + Nvidia, catalyst "neocloud GPU capacity; pre-IPO chatter." This dossier upgrades the color:

• Milestones already cleared for an S-1: banks hired (MS/JPM/Citi, Sep 2025) ✓; public-company-grade board + CEO (Combes/Donovan, May 2026) ✓; anchor contracts for revenue visibility (Microsoft, Nvidia) ✓; a ~$760M run-rate at ~80% growth ✓; crossover/sovereign syndicate (TWG, Mubadala, ARK) ✓; a pre-IPO convertible bridging to the listing ✓.
• Milestones still open / gating: (1) a clean, cloud-only, higher-margin P&L — the hardware/app run-off must show blended GM converging toward ~60% and losses narrowing; (2) a path to profitability the market will underwrite given CoreWeave still prints losses; (3) the depreciation-policy and Nvidia-related-party disclosures the S-1 will force into the open; (4) a supportive IPO window — the "2026 IPO boom off to a bad start" backdrop and CoreWeave's own ~15% Jul-1 selloff show the tape is treating levered neoclouds harshly.
• Estimated window: H2 2026 targeted, slipping — originally H1 2026, pushed to H2 2026, and the Mubadala convertible is explicitly a bridge to "next year's" IPO. Base case: a 2026-H2-to-2027 listing, contingent on the tape.
• IPO-readiness call: 4/5 confirmed — arguably the most IPO-ready private neocloud after CoreWeave already listed. The instrument to watch is the S-1 filing itself (the first look at audited financials, depreciation policy, customer concentration, and Nvidia related-party terms) and the IPO price vs the ~$9B secondary mark.

Write-back: this dossier's path will be recorded against the private-watch.jsonlambda entry (dossier set to this file) so our model shows Lambda dossier-warm. No Brier forecast is logged — per the --watchlist rule, forecasts are only logged on genuine committed calls, and there is no binary EPS/readout to score for a pre-S-1 private.

Bull vs Bear

Bull case. Lambda is the disciplined neocloud — the one with a real bottoms-up developer funnel (~200k), the lowest credible price ($2.49/hr), a higher underlying cloud gross margin (~61%) than the blended CoreWeave trajectory, and a management team that kills low-margin lines rather than chasing vanity revenue. It has the two anchor tenants that matter (Microsoft, Nvidia), Nvidia's supply priority and equity, an IPO-grade board, and a top-tier bank/syndicate line-up. If the AI-capex cycle has years to run and merchant GPU demand stays supply-constrained, Lambda IPOs into that demand at a premium and compounds as the "cleaner, funnel-driven CoreWeave." The colocation-light model means it carries less balance-sheet risk than CoreWeave's $25B-debt owned-site empire — a feature, not a bug, if rates bite.

Bear case (2–3 permanent-impairment risks).

1. The whole category is round-trip-financed. Nvidia is Lambda's supplier, investor, and now largest customer via the $1.5B leaseback; the sector's revenue and funding "originate from within the same small circle". If end-user AI demand disappoints, the backstop unwinds and Lambda is left with depreciating GPUs, GPU-backed debt, and a rate that has already collapsed to ~$2/hr — the telecom-vendor-financing bust, replayed.
2. GPU depreciation + rate collapse = margin mirage. If the economic life of an H100 is 2–3 years (not 6), today's ~50–61% margins are overstated and write-downs are coming; simultaneously, on-demand rates fell ~60–75% (2024→2025) and keep falling as Blackwell/Rubin flood in. A low-cost provider in a price war with a mis-depreciated asset base is the worst place to be.
3. Hyperscaler disintermediation. Lambda's biggest customer (Microsoft) is also the entity most able to build the same capacity in-house. Merchant neoclouds exist to absorb overflow demand; when hyperscalers catch up on their own buildouts, the overflow shrinks first.

Pre-mortem (18 months out, thesis broke — what happened?). The IPO priced below the $9B secondary mark (or was pulled) into a risk-off AI-capex-digestion tape; the S-1 revealed a 6-year depreciation policy and heavier-than-expected Nvidia-related-party revenue, which the market discounted as low-quality; on-demand rates kept falling and blended gross margin fell instead of converging to 60%; and the ~$500M GPU debt + capex-2:1 dynamic forced a dilutiveDilutionIssuing new shares, so each existing share owns a smaller slice of the same company. down-round-equivalent. The funnel was real but not enough to offset a commodity in oversupply.

Are the multiples too high? At ~12× run-rate vs CoreWeave's ~8–11× on 7× the revenue and a $98.8B backlog, Lambda is priced for a successful IPO and continued ~80% growth — richly, not absurdly. The secondary premium (+55% over last primary) is the fragile part.

Contrarian view — what the market is refusing to see: the colocation-light, funnel-first model is a feature the CoreWeave-comparison obscures. Everyone benchmarks Lambda against CoreWeave's backlog and finds it short — but Lambda deliberately carries less debt, less owned real estate, and a lower-CAC demand engine. In a scenario where the AI-capex cycle cools (the bear case for the whole sector), Lambda's lighter balance sheet and organic funnel may prove more survivable than CoreWeave's levered take-or-pay empire — the market is pricing Lambda as "CoreWeave-minus," when in a downturn it might be "CoreWeave-safer." That is the non-consensus read.

Devil's Advocate (short-seller)

Dismantling the bull case:

• Where is revenue concentrated, and what breaks it? In two anchor entities (Microsoft, Nvidia) plus a spot-priced long tail. Nvidia is the largest customer and the supplier and an investor — if you strip out vendor-backstopped demand (the $1.5B leaseback), how much genuinely third-party, organically-demanded revenue is left? The S-1 will show it; the bull case assumes it's most of the book. If it isn't, the ~80% growth is partly a financing artifact, not organic demand — the exact critique leveled at the whole neocloud complex.
• Why is the moat weaker than bulls think? The product is a commodity (identical Nvidia silicon), the price wedge is a margin liability in a rate war (Lambda's low price is the thing bulls cite — but it means thin margin as rates collapse to ~$2/hr), the power is landlord-supplied (no owned-generation moat), and there's no proprietary chip or model. The funnel is the only real moat, and funnels can be out-marketed or made irrelevant if hyperscalers bundle GPU credits.
• Most dangerous competitor bulls underestimate: the hyperscalers themselves, especially Microsoft — Lambda's own anchor customer, which is building Azure GB300 capacity for OpenAI directly. Lambda is a supplier to its own future disintermediator.
• Worst capital-allocation / governance flag: the Nvidia triple-entanglement is a related-party structure that flatters demand; and a brand-new CEO (Combes) 6–12 months before an IPO is betting the listing on an operator still learning the business.
• What must hold for ~$9B? ~80% growth continuing, blended GM rising to ~60%, the IPO window staying open, depreciation being defensible at worse than 4 years, and the Nvidia backstop not being read as low-quality revenue. That's five things, several outside Lambda's control.
• Valuation if growth disappoints 20–30%: at ~5–6× (the Series D multiple) on a decelerated ~$550–600M revenue, the equity is worth ~$3–3.5B — a ~60% haircut from the $9B secondary mark. The downside is large precisely because the current mark prices IPO perfection.
• Single scenario that permanently impairs the business: a genuine AI-capex digestion / demand air-pocket that (a) collapses on-demand rates below cash-cost, (b) triggers GPU write-downs against the ~$500M GPU-backed debt, and (c) shuts the IPO window — forcing a dilutive rescue. Plausibility: moderate — not the base case in a still-supply-constrained 2026, but the CoreWeave Jul-1 selloff shows the market is already stress-testing exactly this failure mode across the sector.

Management Questions (ordered by information value)

1. What is your GPU depreciation policy — useful life in years — and how does it compare to the 2–3-year economic life implied by the H100 rate collapse and Blackwell/Rubin efficiency gains? (The single number that determines whether current margins are real.)
2. Of 2025 revenue, what share is genuinely third-party arm's-length demand vs. Nvidia (the $1.5B leaseback) and other vendor/investor-linked revenue? (Quality-of-revenue — the round-trip question.)
3. Show the disaggregated economics of the Nvidia leaseback: is it take-or-pay, at market rates, and how are the investor/supplier/customer relationships walled off for related-party purposes?
4. What is the exact contracted/committed backlog (RPO-equivalent) and its duration — the number that lets us compare you to CoreWeave's $98.8B?
5. What is blended gross margin today, and the bridge to your target as hardware/app lines run off — and where does on-demand cloud GM sit after the rate collapse?
6. How much debt (GPU-collateralized and other) is on the balance sheet, at what rates and maturities, and what is the covenant exposure if GPU collateral values keep falling?
7. What is the current deployed power (MW) vs. the 3GW-by-2030 target, and how much of that 3GW is contracted vs. aspirational?
8. Path to profitability: at what revenue scale and by when does the business turn FCF-positive, and what does capex-intensity (the ~2:1 revenue-lag) look like on that path?
9. Customer concentration: what % of revenue is Microsoft + Nvidia + the top 5, and what are the term/renewal/cancellation provisions on the anchor contracts?
10. Why the colocation-light (lease vs own) model — and how does it change your unit economics and downside resilience vs CoreWeave's owned sites?
11. What is the IPO timing and the rationale for the Mubadala pre-IPO convertible's ~20% discount — what does that discount say about expected IPO pricing vs the ~$9B secondary?
12. With Combes as a new CEO 6–12 months before an IPO, how is the founder→professional-operator handoff structured, and where does product/vision authority sit?
13. How defensible is the ~$2.49/hr price against hyperscalers bundling free GPU credits and against rivals matching on price — is this a durable wedge or a race to the bottom?
14. What is your inference vs training revenue mix, and why did you exit the Inference API/Chat — does that cede the higher-margin, stickier inference layer to rivals?
15. What single event would most damage the business over the next 24 months, and what is the plan if the IPO window stays shut into 2027?

---