SambaNova Systems

Phase A — Understand the business

Lens 1 · Company Overview

What it is. SambaNova Systems designs custom AI silicon and full rack-scale systems purpose-built for enterprise AI inference (and, historically, training). It was founded in 2017 in Palo Alto by three people out of Stanford: Kunle Olukotun (Stanford EE/CS professor, "father of the multicore processor"), Rodrigo Liang (CEO; ex-Sun/Oracle silicon executive), and Christopher Ré (Stanford ML/databases professor, MacArthur fellow). The name is Portuguese for "new dance" — Liang was raised in Brazil.

The product stack (vertically integrated — this is the key structural fact). Unlike Nvidia (sells chips + CUDA) or Groq/Cerebras (sell chips + cloud), SambaNova sells the entire stack as one appliance:

• The chip — the SN40L Reconfigurable Dataflow Unit (RDU), and now the SN50 (launched Feb 2026). A dataflow architecture, not a GPU: instead of fixed cores executing a kernel schedule, the RDU reconfigures its compute+memory fabric to the shape of the model's dataflow graph, fusing operations into single kernel calls.
• The system — DataScale rack-scale integrated systems (the on-prem/national-lab product line).
• The software — SambaFlow compiler + the Composition of Experts (CoE) serving model, which lets one system hold many models in memory and switch between them cheaply.
• The cloud — SambaNova Cloud (launched Sept 2024), an inference API business — the pivot vehicle.

How it makes money. Three motions: (1) on-prem systems sold/leased to enterprises, governments, national labs (DataScale); (2) sovereign/managed cloud deployments (the stc Saudi Arabia platform, SoftBank Japan); (3) API inference (SambaNova Cloud, per-token). The 2024–25 strategic pivot was explicitly away from training and toward inference + cloud services.

Customers (named). National labs — Argonne, Lawrence Livermore (LLNL), Los Alamos (LANL), Oak Ridge (ORNL), plus the DOE NNSA and Texas Advanced Computing Center (TACC). Enterprise/financial — Accenture (generative-AI deployment), NetApp, OTP Bank (Hungary), Blackbox. Sovereign/telco — stc Group (Saudi "world's largest sovereign AI cloud" for Llama 405B), Aramco / Aramco Digital ("Metabrain" internal AI), SoftBank Corp (first SN50 customer, Japan), RIKEN R-CCS (Japan), OVHcloud and Infercom (Europe).

Contract structure. Not disclosed in detail — n/a — private, not disclosed. Structurally, DataScale is a high-ACV capital sale/lease (single deployments replace hundreds of GPUs); sovereign deals are multi-year build-outs (stc, SoftBank); the cloud is consumption/per-token. Reported >50 enterprise deployments by end-2025. Customer concentration is high and government/sovereign-weighted — a defining risk (see Lens 13).

Lens 2 · Supply Chain

Map: upstream inputs → SambaNova → end customer.

• Chip fabrication: TSMC — the SN40L die is fabbed on TSMC 5nm (N5), 2× 600mm² dies per socket, 102B transistors. Single-source foundry dependency — the industry-wide TSMC chokepoint. ``
• Advanced packaging: TSMC CoWoS-S (2.5D interposer joining the two accelerator dies + HBM stacks). CoWoS capacity is the binding constraint across the entire AI-accelerator industry — SambaNova competes for the same allocation as Nvidia, AMD, Broadcom, Cerebras. As a small-volume buyer, it sits at the back of the CoWoS queue — a structural disadvantage a strategic (Intel/foundry) relationship is partly meant to solve.
• Memory: HBM (64 GiB on-package per socket) — sourced from the SK Hynix / Samsung / Micron oligopoly; plus commodity DDR DRAM (up to 1.5 TiB off-package). HBM allocation is itself a chokepoint; SambaNova's three-tier design reduces HBM dependence relative to Groq (SRAM-only, no HBM) and buys cheaper DRAM capacity — a genuine supply-chain hedge.
• The company: designs the RDU, integrates DataScale racks, writes SambaFlow, operates SambaNova Cloud.
• Compute/host partner: Intel — the Feb 2026 alliance pairs SambaNova RDUs (decode) with Intel Xeon 6 CPUs (host + "action" CPUs) and GPUs (prefill) in a heterogeneous inference blueprint. This is both a supply relationship and a go-to-market channel.
• Downstream buyers: national labs, sovereign clouds (stc, SoftBank), enterprises (Accenture, OTP), and API developers.

Chokepoints (named): (1) TSMC N5 wafer allocation; (2) TSMC CoWoS-S packaging — the hardest single dependency for a low-volume player; (3) HBM supply. All three are shared with every Nvidia challenger; none is SambaNova-specific, but SambaNova's small scale makes each worse for it than for Nvidia.

Lens 3 · Competitive Advantages (moats)

The architectural moat is real and specific — three-tier memory. The RDU carries 520 MiB SRAM + 64 GiB HBM + up to 1.5 TiB DDR in a single hierarchy. This is the differentiated bet: it lets one SN40L socket hold far larger models and many more models simultaneously than rivals. The concrete claims:

• 16 SN40L chips replace ~320 GPUs to serve a 671B-parameter model — a ~19× footprint reduction, 15–31× faster model-switching, 3.7–6.6× overall speedup vs a DGX-H100 for Composition-of-Experts inference.
• The Composition of Experts (CoE) software makes hundreds of models resident and hot-swappable — the natural substrate for agentic AI (many specialized models, low switch cost), which is exactly how SambaNova now positions the SN50.

Where SambaNova wins vs the other two challengers: Groq's SRAM-only LPU is blisteringly fast but can't fit large models on-chip — mid-size models already need large multi-system deployments. Cerebras's wafer-scale maximizes throughput but at extreme power/space cost. SambaNova's memory-capacity angle is the "run the biggest models on the fewest racks" lane. That is a defensible, differentiated position within the challenger set.

Moat durability — the honest verdict: weak-to-moderate, and shrinking on one axis.

• The CUDA problem is the whole game. Nvidia has a decade+ software-ecosystem lead; any challenger "initially lacks this breadth of software maturity". SambaNova claims HuggingFace-pipeline compatibility, but in practice application code needs modification per architecture. Switching costs run against SambaNova, not for it. The consensus outlook has Nvidia + AMD + CSP-built ASICs at 80–95% share.
• The vertical-integration lock-in cuts both ways — a customer who buys the SambaNova stack is locked in (good), but that same all-or-nothing commitment is exactly why enterprises hesitate to adopt (bad). It is a moat you have to sell past before you can benefit from it.
• IP moat: genuine — the RDU/dataflow patents and the Olukotun/Ré academic pedigree are real and hard to replicate quickly. But IP ≠ market share when the incumbent owns the toolchain developers already know.

Bargaining power. Weak over suppliers (small CoWoS/HBM buyer, back of the queue). Weak over customers in commercial markets (they have Nvidia as the default). Strongest in sovereign/national-security channels, where "not-Nvidia," on-prem control, and full-stack sovereignty are features worth paying for — which is precisely where SambaNova's real traction sits.

Lens 4 · Segments

No segment financials are disclosed (private). Structural revenue mix, `` from the public deployment record:

Geography: heavily US (national labs) + Middle East (Saudi sovereign) + Japan (SoftBank/RIKEN) + selective Europe (OTP, OVHcloud). The Gulf and Japan are doing the growth lifting — a sovereign-demand-led company, not a US-hyperscaler-led one. That is the single most important thing to understand about the revenue: it is lumpy, relationship-driven, and politically exposed, not broad commercial SaaS.

Trend & cause: 2024 delivered "virtually no revenue" on the old training model; the pivot + sovereign wins drove a claimed ~$100M ARR / record 2025 bookings. Accelerating off a near-zero base — the derivative is up, the absolute level is still tiny for the valuation asked.

---

Phase B — Measure performance (+private overlay: Lens 5 → funding/valuation, Lens 7 → cap table, + traction/unit-economics)

Lens 5 · Funding & Valuation Trajectory (replaces Earnings Result)

The whole story is in this table. It is a boom → near-death → resurrection arc, and the valuation whiplash is the thesis.

Total raised to date: ~$1.48–1.5B across 5 completed rounds.

⚠️ Source conflict, surfaced not resolved-away: the Feb 2026 Series E post-money is reported as ~$2.2B by TSG Invest / PM Insights (well-sourced, consistent with the collapsed $1.6B floor + $350M new capital), but one TechCrunch/CNBC framing states the Feb 2026 round was "at a $5.1B valuation". The $5.1B figure appears to be a conflation with the 2021 Series D peak and is contradicted by the arithmetic (a $1.6B rescue floor + $350M ≠ $5.1B) and by the June "5x in four months to $10B" framing (5× of ~$2B = ~$10B; 5× of $5.1B ≠ $10B). I weight ~$2.2B post-money as the correct Series E mark and flag the $5.1B mention as an error. ``

Read on the trajectory: SambaNova is the only major AI-chip challenger to have taken a ~57% down-round from peak and to have been shopped for sale (Oct 2025). Then, in four months, the SN50 launch + Intel alliance flipped the narrative and the rumored mark quintupled to $10B. That is not a company compounding steadily — it is a company whose valuation is a call option on the Intel relationship and the AI-chip funding mania, marked to sentiment. Burn: 2024 was near-zero-revenue with "relentless engineering pressure" — a high-burn period that forced the layoffs and the sale process.

Lens 6 · Founder/Exec Signal Trend (calls → founder interviews/positioning)

No earnings calls (private). Tracking management's public positioning shift:

• 2021–2023 (peak era): "Reimagine computing," full-stack training+inference, "buy the outcome not the chip." Ambition maximal.
• Sept 2024 – Apr 2025 (the pivot / near-death): tone turns defensive-pragmatic. SambaNova Cloud launches; then the 77-person / 15% layoff framed as "align with today's market conditions and the transition… from model training to fine-tuning and inference". This is management conceding the training market to Nvidia and retreating to a defensible inference niche.
• Feb 2026 (resurrection): Liang calls the Series E "grossly, grossly oversubscribed" with "tremendous market momentum". Positioning re-centers on agentic AI ("fastest chip for agentic AI," SN50 "3× lower TCO," "5× faster") and the Intel alliance as validation.
• What they stopped saying: "training." The word has been surgically removed. The company that once sold end-to-end now sells inference-and-agentic only.

The sentiment arc — euphoria → survival → vindication — is genuine, but note the vindication is investor-driven (a funding round + a strategic partner), not customer-revenue-driven. The tape here is the cap table, not the P&L.

Lens 7 · Cap Table & Secondary Marks (comps → syndicate quality + peer marks)

Syndicate quality — genuinely tier-1, and that is the bull's best card. Investors across rounds: SoftBank Vision Fund 2, BlackRock, Intel Capital, GV (Google), Temasek, GIC, Vista Equity Partners, Cambium Capital, QIA (Qatar sovereign), T. Rowe Price, Battery Ventures, Walden International. The presence of crossover/public-market funds (T. Rowe, BlackRock) and sovereigns (QIA, Temasek, GIC) is an IPO-proximity tell — these are the money types that show up before a listing. Intel Capital's continued participation after the acquisition fell through is the loudest signal in the cap table.

The governance bombshell: Lip-Bu Tan — Intel's CEO — is executive chairman of SambaNova. Intel is simultaneously (a) a shareholder, (b) the failed acquirer, (c) the strategic GTM/host-CPU partner, and (d) chaired by SambaNova's own board chair. This is an extraordinary, conflict-laden web — analyzed as a red flag in Lens 10 and a catalyst in Lens 12.

Public/private peer marks (the comps that matter now):

The comp read is brutal and clarifying: in the same cycle, Cerebras IPO'd at $56B and Nvidia paid $20B for Groq — while SambaNova was being shopped at $1.6B and is now asking $10B. On a pure "what did the market pay for a peer" basis, a $10B SambaNova is (a) a fraction of Cerebras's public cap and (b) half of what Nvidia paid to remove Groq — not obviously rich if you believe SambaNova is a genuine top-3 inference architecture. The bear read: SambaNova is asking $10B on ~$100M ARR (~100× ARR), a multiple only defensible on the strategic-scarcity argument, not fundamentals.

Lens 8 · Funding & Product Catalysts (the events that re-rated it)

The private-market equivalent of "moves >5%." The events that moved SambaNova's perceived value:

1. Apr 2021 — SoftBank Series D → $5.1B. Peak euphoria. ``
2. Sept 2024 — SambaNova Cloud launch. The pivot begins. ``
3. Apr 2025 — 15% layoffs, training→inference pivot. The market's "this might not work" moment. ``
4. Oct 2025 — sale exploration / Intel talks at ~$1.6B. Near-death; the floor. ``
5. Dec 2025 – Feb 2026 — Intel acquisition COLLAPSES, then $350M+ Series E + SN50 launch + Intel strategic alliance. The pivotal re-rate from ~$1.6B → ~$2.2B and a total narrative reversal. ``
6. May 2026 — Cerebras IPO (+108%) & Nvidia buys Groq ($20B). The sector re-rates violently; SambaNova is the last independent standing → scarcity bid. ``
7. Jun 2026 — reported $800M–1B raise at ~$10B. The 5x-in-4-months mark. ``

Pattern: SambaNova does not re-rate on customer revenue or benchmark wins — it re-rates on (a) marquee investor validation, (b) the Intel relationship, and (c) sector-wide scarcity events (peers exiting). The market is trading SambaNova as "the last independent inference architecture Intel is standing behind," not as a cash-flow story.

---

Phase C — Judge people & books

Lens 9 · Management

• Rodrigo Liang (CEO, co-founder). Two decades of shipping silicon at scale — led processor teams at Sun Microsystems and Oracle. This is the rare AI-hardware CEO who is a career silicon-engineering executive, not a researcher or a salesman. Track record: took SambaNova from stealth to $5.1B, through a near-death, to a resurrection — survived where most would have sold at the bottom. The decision to reject Intel's $1.6B bid and raise instead is a high-conviction, arguably vindicated capital-allocation call (the mark is now 5x higher on paper).
• Kunle Olukotun (co-founder, Chief Technologist). The credential — Stanford professor, pioneer of the multicore processor (ubiquitous by the early 2000s), director of Stanford's Pervasive Parallelism Lab. His name is the technical moat's provenance.
• Christopher Ré (co-founder). Stanford ML/databases professor, MacArthur "genius" grant, deep in the modern ML-systems lineage (the DAWN lab).
• Lip-Bu Tan (Executive Chairman). Legendary semiconductor investor (ex-Cadence CEO, ex-Walden) — and, as of 2026, Intel's CEO. His chairmanship is the connective tissue of the entire Intel relationship. Skin in the game: enormous — but see Lens 10 for the conflict.
• Capital-allocation history: raised ~$1.5B, burned hard on training (2024 near-zero revenue), corrected decisively (layoffs + pivot), then raised again on better terms. The pivot was late and costly but executed. No public buybacks/dividends (private, growth-stage).
• Founder vs professional manager: a founder-CEO with a rare deep-silicon operating spine + two academic-heavyweight co-founders. For a hard-tech company at this stage, that is close to the ideal archetype — the risk is not competence, it is the market (CUDA) and the conflicts (Intel).

Red flags at the management layer: the 2024 "relentless pressure / virtually no revenue / then layoffs" period shows a leadership that over-committed to the wrong market (training) for too long before pivoting. And the Intel/Tan entanglement (below) is a real governance concern, not a footnote.

Lens 10 · Forensic Red Flags

Accounting: n/a — private, unaudited. No public financial statements exist; there is no way to independently verify the ~$100M ARR, the "record bookings," or the burn. Every financial figure in this dossier is a company-sourced or press-sourced claim, not an audited number. That opacity is itself the first red flag — the $10B ask rests on numbers no outside party can check.

Structural / governance red flags (these are the real ones):

1. The Intel conflict web. Intel's sitting CEO (Lip-Bu Tan) is SambaNova's executive chairman; Intel is a shareholder (Intel Capital, multiple rounds); Intel was the failed acquirer at $1.6B; Intel is now the strategic GTM/host-CPU partner. Questions this raises: Was the $1.6B bid a fair-value floor or a related-party lowball that Tan sat on both sides of? Is the "oversubscribed" Series E and the Intel partnership (announced same day) a genuine arm's-length validation or a chairman engineering an outcome? Is the rumored $10B mark real third-party demand or Intel-orbit signaling? None of this is illegal or even necessarily improper — but it is exactly the kind of related-party density a forensic analyst flags, and it means the loudest "validation" signals are not independent of the chairman.
2. Valuation marked to sentiment, not cash flows. A 5x move in 4 months on a company with ~$100M unaudited ARR is a sentiment mark. Down-rounds (2024–25) and up-rounds (2026) here track narrative and sector events, not fundamentals.
3. Customer concentration in sovereign/government channels (Saudi stc/Aramco, SoftBank, national labs) — lumpy, politically exposed, and hard to underwrite (see Lens 13).
4. "Virtually no revenue" in 2024 on the prior strategy — the pivot is unproven at scale; the inference/agentic revenue base is early.

Regulatory findings. Per regulatory/regulatory-findings.md: SambaNova has no CIK and is not an SEC registrant — no EDGAR enforcement (LR/AAER) search is possible; total SEC findings: 0. Non-SEC web search ("SambaNova Systems" (FTC OR DOJ OR consent decree OR settlement OR fine OR penalty) enforcement) surfaced no material regulatory or legal enforcement actions as of 2026-07-07. One adjacent exposure worth noting for a company selling advanced AI compute into the Gulf: US export-control (BIS/EAR) sensitivity on advanced AI accelerators shipped to Middle East sovereigns — not an enforcement finding, but a live regulatory risk surface for the sovereign-cloud strategy. Conclusion: No material regulatory or legal findings — verified via SEC EDGAR EFTS (no CIK / not applicable), web search, and (no 10-K exists) as of 2026-07-07. All financials unaudited per public sources.

---

Phase D — Project & stress-test

Lens 11 · IPO-Readiness & Path-to-Tradeable (replaces Forward Projection; +private)

Grounding: stage: late, ipo_readiness: 3/5, catalyst: "AI systems; well-funded; inference pivot". On the file's scale (3 = late-stage; 4 = pre-IPO/secondary-active; 5 = S-1 filed), SambaNova is a strong 3, arguably now a 4 given the 2026 events — the private-watch entry (last updated 2026-06-30) predates full digestion of the $10B round chatter and should be nudged toward 4.

The path to tradeable — three routes:

1. IPO (base case, 12–30 months out). The Cerebras IPO (May 2026, +108%, $56B) blew the window wide open and set the template; SambaNova is the obvious next inference-chip candidate. Milestones that unlock an S-1: (a) two-to-four quarters of auditable revenue growth off the inference/sovereign base — the market will demand real numbers post-Cerebras; (b) SN50 shipping at volume (H2 2026) with SoftBank Japan live as a lighthouse; (c) diversification beyond a handful of sovereign whales into repeatable enterprise/cloud logos. A completed ~$10B private round would push readiness to 4 and set a ~2027 IPO window if the SN50 ramp delivers. ``
2. Strategic acquisition (live alternative). Intel already tried at $1.6B; a re-approach at a higher (or the alliance-implied) price is plausible, though the failed first attempt and the chairman-conflict make a clean deal harder. Other acquirers (a hyperscaler, another foundry-aligned buyer) are conceivable given Nvidia just paid $20B for Groq — M&A demand for inference specialists is proven.
3. Stay private, sovereign-funded. QIA/Gulf capital could keep it private and well-funded for years as a sovereign-AI supplier — the least-liquid outcome for outside investors.

Base call (tracked as a judgment, NOT logged via forecast.ts per --watchlist rule): SambaNova closes a raise at ≥$8B pre-money by end-2026 (~60–65% likely), and files/IPOs within ~18–30 months conditional on the SN50 ramp and auditable revenue (~40% by YE2027). The dominant value-inflection is the next funding mark, and it is a call option on the Intel alliance + sector scarcity, not on discounted cash flows. ``

Note: no forecast.ts create — this is the unattended --watchlist loop; the projection is recorded as analysis only.

Lens 12 · Bull vs Bear

Bull case. SambaNova is the last independent, full-stack, differentiated inference architecture in a market that just proved (Cerebras $56B IPO; Nvidia/$20B Groq) it will pay enormous sums for exactly that. The three-tier memory RDU is genuinely the best design for "biggest models on fewest racks" and for agentic AI (many resident models, cheap switching) — the fastest-growing inference workload. It has a founder-CEO who ships silicon, an academic-heavyweight bench, a tier-1 + sovereign cap table, and an Intel alliance giving it Xeon co-selling and (implicitly) foundry/packaging goodwill. Real, sticky sovereign demand (Saudi stc/Aramco, SoftBank Japan, Gulf capital) is a durable, less-competed lane where "not-Nvidia + full sovereignty" is a paid feature. Off a near-zero 2024, revenue is inflecting (~$100M ARR claim, record 2025 bookings). If the SN50 ramps and one or two sovereign deals scale, a $10B mark looks cheap vs a $56B Cerebras.

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

1. CUDA is destiny. The moat that matters is software, and Nvidia's decade+ lead is widening as it now also owns Groq's inference IP. Consensus puts Nvidia+AMD+CSP-ASICs at 80–95%. SambaNova can be architecturally superior on a slide and still lose because developers won't leave CUDA. This is the risk that permanently impairs the business.
2. The revenue is a mirage until audited. ~$100M ARR is unverifiable, sovereign-concentrated, and lumpy. If the "record bookings" are a few Gulf/Japan whales rather than a repeatable motion, an IPO exposes it and the mark collapses (as it did to $1.6B in 2024–25 — this company has already round-tripped once).
3. The Intel entanglement is a single point of failure. The bull thesis leans heavily on Intel — but Intel is itself a distressed turnaround, and the chairman-conflict means the "validation" isn't independent. If Intel's own crisis forces it to pull back, SambaNova loses its loudest backer and its GTM channel at once.

Pre-mortem (it's Jan 2028, the thesis broke — what happened?): The SN50 ramp slipped on CoWoS allocation; the sovereign pipeline stalled on Gulf budget cuts / US export-control friction; the $10B round either didn't close or closed then couldn't be defended by revenue; Nvidia's Groq-powered inference stack + CUDA made the RDU's speed edge economically irrelevant for anyone already on GPUs; the IPO window that Cerebras opened closed before SambaNova could reach auditable-growth escape velocity. It gets re-shopped — and this time there's no $1.6B floor, because Intel is preoccupied and the scarcity premium has faded.

Are the multiples too high? At ~$10B on ~$100M unaudited ARR (~100× ARR), yes on fundamentals — defensible only on strategic scarcity (half of what Nvidia paid for Groq; a fifth of Cerebras's public cap). The valuation is an option premium, not a cash-flow multiple.

Contrarian view (what the market is refusing to see): The bull consensus treats the Intel alliance as SambaNova's greatest asset. The contrarian read: it is the greatest hidden risk. SambaNova has effectively become a leveraged bet on Intel's own turnaround — its chairman, its GTM, and its scarcity narrative all route through the single most troubled large-cap in semiconductors. The market is pricing the Intel relationship as validation; it should price it as concentration.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case.

• What structurally breaks the business model: the customer never has to leave CUDA. Every enterprise buyer's default is Nvidia; SambaNova's full-stack lock-in is a reason not to adopt, not a reason to. The addressable market is "customers willing to re-architect around a proprietary non-GPU stack" — a small, shrinking slice as Nvidia's software lead compounds and it absorbs Groq.
• Where revenue is concentrated / what happens if it shifts: a handful of sovereign whales (Saudi stc/Aramco, SoftBank Japan) and government labs. If Gulf AI budgets tighten, if a sovereign switches to Groq/Nvidia (Groq already has a 19,000-LPU Dammam build with Aramco Digital — SambaNova's flagship sovereign region is contested), or if US export controls choke Gulf shipments, the "record bookings" evaporate. This is not diversified SaaS ARR; it's project revenue dressed as recurring.
• Why the moat is weaker than bulls think: the architectural edge is a point-in-time benchmark advantage; Nvidia ships new silicon annually and now owns Groq's inference tricks. A dataflow-vs-GPU slide win doesn't survive contact with CUDA's ecosystem and Nvidia's cadence.
• Most dangerous competitor bulls underestimate: Nvidia's Groq business unit — Nvidia specifically bought Groq to deny inference specialists escape velocity; it is now weaponizing SRAM-speed inference inside the CUDA moat. That is the exact combination (speed + ecosystem) SambaNova cannot match.
• Worst capital-allocation / governance moves: over-committing to training through 2024 (near-zero revenue, then layoffs); and the Intel-chairman related-party web — a CEO-of-the-acquirer sitting as your chairman across a failed buyout, a same-day financing, and a strategic partnership is a governance structure that should make any outside investor underwrite the "validation" at a discount.
• What must hold for a $10B price: (a) SN50 ships at volume on time despite CoWoS scarcity; (b) sovereign + enterprise revenue compounds to audited numbers that justify ~100× ARR; (c) Intel stays a committed backer; (d) the Cerebras-opened IPO window stays open. Break any one and the mark is indefensible.
• Valuation if growth disappoints 20–30%: on a sentiment mark, a growth miss doesn't dent it 25% — it round-trips toward the $1.6B–2.2B floor, because the entire premium above that is scarcity/option value that a miss erases. The 2024–25 down-round is the proof of concept for exactly this.
• The single permanent-impairment scenario, and its plausibility: Nvidia's Groq-integrated CUDA inference stack makes SambaNova's architecture economically irrelevant for the ~90% of the market already on GPUs, while the sovereign niche proves too small and too contested to sustain an independent chipmaker. Plausibility: moderate-to-high — it is the base-rate outcome for sub-scale merchant AI-silicon challengers against an incumbent that owns the toolchain.

Lens 14 · Management Questions (ordered by information value)

1. Show us the audited numbers: what were FY2025 GAAP revenue, gross margin, and cash burn, and what is current net cash / runway at today's burn?
2. What share of 2025 "record bookings" came from your top 3 customers, and how much is recurring/consumption vs one-time systems/project revenue?
3. Given Intel CEO Lip-Bu Tan is your executive chairman, how were the failed $1.6B acquisition, the Series E pricing, and the Intel partnership negotiated at arm's length — who represented minority shareholders?
4. The ~$10B rumored mark is ~5x the Feb 2026 round on flat-ish fundamentals — what changed operationally, or is this a sentiment re-rate you'd caution investors about?
5. What is your realistic, honest answer to the CUDA problem — why will a developer team re-architect off GPUs for SambaNova, and what's your evidence of that happening at scale?
6. Nvidia now owns Groq and its SRAM inference IP inside CUDA. How does the RDU's speed/TCO edge survive that combination over the next 24 months?
7. Groq has a 19,000-LPU build with Aramco Digital in your flagship Saudi region — how contested is your sovereign pipeline, and what's your win/loss record there?
8. What is SN50's realistic volume-shipment timeline, and how have you secured TSMC CoWoS-S packaging allocation against Nvidia/AMD/Cerebras for it?
9. What is your path to IPO — what auditable metrics do you need to hit, and what window are you targeting post-Cerebras?
10. If the current raise doesn't close on your terms, what is Plan B — down-round, strategic sale, or sovereign-funded independence?
11. How exposed is the Gulf/Japan revenue to US export controls (BIS/EAR) on advanced AI accelerators, and what's your mitigation?
12. What are the actual per-token economics of SambaNova Cloud vs a comparable GPU cloud, at real utilization — not peak-benchmark TCO slides?
13. You conceded the training market; is inference-and-agentic a large enough TAM to build a standalone public company on, or is this a niche that gets acquired?
14. What is insider/founder ownership post-Series E, and how much dilution does the ~$10B round imply for existing holders?
15. Which single metric or milestone in the next 12 months would you stake the thesis on, and what would falsify it?

---