Nova Compute

• The 60/100 feasibility score is well-calibrated, arguably slightly generous. This is applied distributed-systems engineering (spot-capacity arbitrage, SLA-aware scheduling, cross-cloud failover), not a research bet — there is no unsolved scientific problem here. The 99.95% measured uptime over 90 days and 41% average cost reduction across 22 teams demonstrate the core loop already works in product
• The technical frontier they lean on (inference cost curves, MoE/sparse serving) actually cuts against them, not for them: per-token inference cost has been falling ~4-10x/year (cf. GPT-4-class pricing declines 2023-2025), and each foundation-model price cut shrinks the absolute dollars their 41% arbitrage captures. Their savings are a percentage of a rapidly deflating base — a structural headwind
• Latency-SLA routing across five clouds' spot pools is genuinely hard at the tail: spot preemption (AWS gives 2-min warning, GCP 30s), cold-start on evicted GPU pools, and cross-region egress fees can silently erode the advertised 41% net-of-egress. The claimed uptime suggests they handle this, but 90 days / 22 teams is thin evidence for the p99.9 behavior enterprises underwrite.
• Defensibility is the crux the feasibility score can't rescue: the routing logic is replicable, and the disclosed data-scale moat (61/100) requires request-volume network effects they haven't yet demonstrated at $48k MRR. Note the traction/plan mismatch — description says $48k MRR (~$576k ARR) while the model reads $576k as MRR; the lower figure is the honest read and reframes this as very early.

• Commoditization from above and below: cloud providers (Azure AI, Bedrock) and inference platforms (Together, Fireworks, OpenRouter) are shipping native multi-model routing; a cross-cloud router is a feature the hyperscalers can absorb, consistent with the 37/100 competitive-headroom score at 90% intensity.
• Deflating value pool: if per-token prices keep falling 4x+/year, a 41% relative saving on a shrinking bill yields declining absolute customer value and pricing power — the arbitrage spread that funds their take rate compresses toward zero.
• Unverified net economics: no LTV/CAC, no bottom-up TAM, and cost-reduction claims are gross of cross-cloud egress and their own margin. The stress test and TAM triangulation could not run — until net-of-egress savings and retention are disclosed, the 41% headline number is unaudited.

• Push back on 55/100 as too kind, not too harsh: the ~65% mature GM is a sector reference, not observed. A router that captures a slice of a 41% cost saving faces structural GM pressure — if Nova prices as % of savings, its margin depends on a spread (spot GPU price vs SLA-met floor) that compresses as clouds add their own spot-optimization. Real GM could sit 40-55%, not 65%.
• Traction is real but thin: $48k MRR / 22 teams = ~$2.2k ARPA/mo (~$26k ARPA/yr). No CAC, no churn, no gross retention disclosed — so LTV/CAC and payback are unknowable. STRESS TEST was not run for exactly this reason. At seed, 22 logos over 90 days is a signal, not a business.
• Note the model's $576k 'MRR' is an annualized artifact ($48k x 12); do not double-count. On $576k ARR, a $6M raise implies ~10x forward ARR entry — reasonable for infra IF net revenue retention >120% and GM proves durable, expensive if churn is high or margin is thin.
• TAM triangulation not run — the $190B TAM is the whole AI-infra market, not Nova's serviceable wedge. Bottom-up: inference spend routed through third-party optimizers is a fraction; SAM is more plausibly single-digit $B. 41% cost reduction is the strongest datapoint — it's a quantified, defensible value prop if it holds across workloads.

• Disintermediation / commoditization (competitive headroom 37, intensity 90): the five clouds Nova arbitrages can each ship native cross-pool spot routing, collapsing the arbitrage spread and Nova's GM to near zero. This is the thesis-killer and it's structurally disclosed.
• Unit economics are undisclosed — no CAC, churn, NRR, or contribution margin. Without these the 55 score is a placeholder. Kill signal: gross churn >20% or blended GM <45% at scale.
• Value-capture fragility: if Nova prices on savings, customers renegotiate as savings shrink; if flat-fee, it competes with free/native tooling. Either path caps ARPA expansion.

• The 62/100 market score is defensible but flatters this business. The ~$190B AI-infra TAM at 19% CAGR is real, yet Nova's serviceable slice is narrow: it monetizes the delta between GPU pools, not compute itself. Take-rate on a 41% cost cut across five clouds' spot markets is a fraction of a fraction — I'd model SAM closer to single-digit billions, so treat 62 as a TAM ceiling the company can't ca
• Demand is genuinely elastic here — inference spend is a top-3 cost line for LLM-heavy teams, so a verified 41% reduction sells itself, evidenced by 22 paying teams and 99.95% uptime. That's the bull case: the product works and the pain is quantified. Elastic demand cuts both ways though — the same buyers will switch the instant a cheaper router (or a free cloud-native one) appears.
• Moat at 61 is too generous for what's disclosed. The routing logic is replicable; the only durable asset is placement-optimization data at scale, and 22 teams / $48k MRR is nowhere near the flywheel. Multi-cloud spot integration is a 6-9 month engineering lift, not a structural barrier — this reads closer to the 35 'undemonstrated defensibility' floor than to 61.
• Note the traction discrepancy: brief states $48k MRR (~$576k ARR) while the model's team line reads '$576k MRR' — a 12x conflation. At $48k MRR the business is pre-scale; the 64 team score built partly on a misread ARR figure should be discounted.

• STRUCTURAL — margin compression from both ends: the disclosed 65% gross margin assumes clouds keep discrete spot pricing gaps. AWS/GCP/Azure have every incentive to ship native cross-region/spot routing (or foundation vendors bundle inference optimization free), collapsing the arbitrage Nova monetizes. Competitive headroom at 37 (90% intensity) is the single most honest number in this model.
• Capital intensity flagged at 70% plus undisclosed LTV/CAC (stress test not run) is a real hole. Selling a switching-cost-light cost-savings tool implies high churn risk and expensive land-and-expand; without CAC payback I cannot verify the unit economics survive contact with enterprise sales cycles.
• Key-man dependency on continued multi-cloud API access — if a cloud rate-limits or contractually restricts spot arbitrage resellers, Nova's five-cloud failover premise degrades to a subset, eroding both the 41% savings claim and the failover value prop.

• The 77/100 is directionally right — Nova is sectoral-only (FTC §5, no AI statute, no financial-services license), so no gating regulator can block operations. But 77 undersells contractual and export risk that lives outside 'regulatory intensity': the real legal surface here is B2B and cross-border, not agency licensing.
• Core structural exposure is the five-cloud dependency. Nova routes across spot capacity from AWS/GCP/Azure/etc — most hyperscaler ToS restrict resale/brokerage of compute and prohibit circumventing pricing. A single provider revoking API access or invoking anti-resale clauses breaks the '41% cost reduction' thesis overnight. Diligence must pull every underlying cloud agreement and confirm Nova is
• Data/privacy is live because Nova sits in the inference request path: prompts/completions transit its router. Absent a federal omnibus law, CCPA/CPRA plus HIPAA/GLBA apply if any of the 22 teams push regulated data. Nova needs signed DPAs, sub-processor disclosures, and a no-retention/no-training posture in writing, or it inherits liability as a processor. Export controls (EAR advanced-compute rul
• Round structure is clean and standard: 506(b) or 506(c) Reg D, Form D within 15 days, 1x non-participating pref, broad-based weighted-average anti-dilution, pro-rata + information rights, board observer. At $6M seed a post-money SAFE or priced round is market; insist on IP assignment reps and confirmation all founder/contractor code is assigned to the company.

• Anti-resale / ToS violation: if Nova's spot-arbitrage across five clouds breaches hyperscaler terms, providers can terminate access and the product's cost-savings moat and 99.95% uptime claim collapse — this is a going-concern legal risk masked by the benign 77 score.
• Uncapped downstream liability: an SLA-driven router promising latency guarantees and failover invites breach-of-SLA and consequential-damages claims from paying teams if a misroute causes an outage; without liability caps and clear data-processing terms Nova carries disproportionate exposure for a seed-stage balance sheet.
• IP thinness: the routing logic is likely trade-secret/orchestration know-how, not patented — commoditization risk (quant flags this, competitive headroom 37) means defensibility is contractual and operational, not legal; confirm no lapsed filings and that key algorithms are protected as trade secrets with enforceable employee NDAs.