# TensorWave Financial Model

AMD-backed GPU cloud provider offering AI compute infrastructure as an alternative to NVIDIA-dominated hyperscalers.

- Canonical: https://finamodel.com/startups/tensorwave
- Excel download: https://finamodel.com/startup-models/tensorwave.xlsx
- Category: AI/ML
- Model type: Marketplace / GMV
- Funding round: Seed
- Funding: $43M
- Founded: 2024
- Geography: US (team based in Nevada; cloud targeting US market implied)
- Customer: B2B

## About the company

TensorWave provides AI compute infrastructure for customers running GPU-intensive workloads. Its commercial model is driven by available capacity, contracted cloud usage, and workload consumption, making utilisation and infrastructure economics more important than a conventional seat-based SaaS metric.

Customers can commit capacity for predictable workloads or expand usage as training and inference demand grows. Revenue therefore depends on GPU supply, deployed capacity, pricing, workload mix, contract duration, and the ability to maintain service performance as demand scales.

The model forecasts GPU capacity, utilisation, customer workloads, contracted and on-demand revenue, and churn. It connects cloud and infrastructure COGS, equipment investment, depreciation or lease costs, gross margin, sales capacity, working capital, cash burn, funding needs, and runway.

## What's included

- 5-year monthly revenue build with stage-appropriate growth assumptions
- Full P&L, headcount plan, and operating-expense schedule
- Cash-flow statement, runway, and burn-rate tracking
- Valuation via exit multiple with a DCF cross-check
- Returns analysis with MOIC and IRR
- Unit economics including CAC, LTV, payback, and cohort retention

## Product & value proposition

- AMD MI300X GPU cloud - the first RoCE v2 clusters in production
- Official AMD MI300X Launch Partner, fully supported by AMD
- Proprietary exclusive Inference Engine for dominant inference performance
- Building Reference Architectures for RoCE v2 clusters alongside AMD and Edgecore
- No code changes needed to migrate from NVIDIA: PyTorch/TensorFlow workloads run natively on AMD ROCm
- Strategic hardware partner: Edgecore
- AMD MI300X specs vs NVIDIA H100 SXM:
  - FP32: 120 TFLOPS vs 67 TFLOPS (1.8x)
  - GPU Memory: 192 GB HBM3 vs 80 GB (2.4x)
  - Memory Bandwidth: 5.2 TB/s vs 3.35 TB/s (1.6x)
  - Interconnect: Infinity Fabric vs NVLink/NVSwitch
- 1.4x better performance vs H100 on Meta Llama 2 70B; up to 2.1x FP16 latency improvement

## Market

- Global AI market forecast to cross $1 trillion by 2028, reaching $1.8T by 2030; CAGR +36.6%
  - 2021: ~$0.0T, 2022: $0.1T, 2023: $0.2T, 2024: $0.3T, 2025: $0.4T, 2026: $0.6T, 2027: $0.8T, 2028: $1.0T, 2029: $1.4T, 2030: $1.8T
- 76% of enterprises plan to increase AI spend next fiscal year; only 3% plan to decrease
- AI Processors and Data Center market revenue:
  - 2019: $2B, 2020: $4B, 2021: $7B, 2022: $11B, 2023: $16B, 2024: $24B, 2025: $29B, 2026: $38B
- 92% of models on HuggingFace are PyTorch-exclusive
- 74% of ML papers in 2023 use PyTorch/TensorFlow

## Revenue model

- GPU-as-a-Service: rent AMD MI300X GPU capacity to AI/ML customers (inference and training workloads)
- Channels: direct enterprise sales implied (B2B); AMD partnership drives referrals

## Traction & metrics

- Official AMD MI300X Launch Partner (Dec 2023)
- First RoCE v2 cluster operational
- Cloud providers described as "booking for 2025" with "long lead times" - framing unmet demand

## Competition / moat

- Primary competitor: NVIDIA HGX ecosystem (H100 SXM)
- Broader hyperscaler competition implied (AWS, Azure, GCP)
- Moat claims:
  - First-mover as AMD MI300X launch partner with official AMD support
  - Exclusive Inference Engine
  - First RoCE v2 cluster (vs NVLink/NVSwitch lock-in)
  - Superior hardware specs on memory and bandwidth
  - CUDA lock-in is weakening - PyTorch/ROCm portability removes switching cost for customers

## Team & funding ask / use of funds

- Darrick Horton - Co-Founder, CEO; infrastructure & cloud executive; ex-Lockheed Martin Skunkworks engineer
- Piotr Tomasik - Co-Founder, COO; multiple-time technical co-founder; 3x exits (Influential, LetsRolo, ActiveSide)
- Jeff Tatarchuk - Co-Founder, CGO; serial entrepreneur; multiple companies sold/acquired
- Team credentials: 6 companies founded; $65M+ raised across previous ventures; decades delivering Cloud, AI/ML, B2B SaaS, FinTech to Fortune 500
- Previous employer logos: Meta, PayPal, Venmo, IBM, Lockheed Martin
- Described as "Built the Largest Tech Startup in Nevada By Valuation" (source: PitchBook)

---

## Recommended financial model

- **Archetype + why:** GPU-as-a-Service capacity revenue model - revenue driven by GPU count × utilization rate × hourly rate. Similar structure to colocation/cloud IaaS. Not SaaS (no subscription/seat pricing shown). Not marketplace. 3-statement with capex-heavy balance sheet and debt/lease schedule is appropriate given hardware-intensive build-out.

- **Forecast horizon & granularity:** Monthly for Year 1–2 (capacity ramp is lumpy; need to track CapEx tranches and utilization build), quarterly for Year 3–5. Five-year horizon appropriate for hardware investment return cycle.

- **Key drivers & assumptions:**

| Driver | Value / Tag |
| -- | -- |
| GPU cluster size at launch (# MI300X GPUs) | 512–1,024 GPUs; typical bare-metal GPU cloud starting cluster size |
| GPU acquisition cost per MI300X | ~$15,000–$20,000/GPU list price at time of deck (early 2024) |
| Utilization rate (reserved + spot blended) | 60% Year 1, ramping to 85% by Year 3; enterprise GPU cloud benchmarks |
| Avg hourly rate per GPU | $3.50–$5.00/hr; AMD MI300X positioned at discount to H100 (~$4–6/hr market) |
| Revenue mix: inference vs training | 70% inference / 30% training; deck emphasizes inference superiority |
| Customer contract type: reserved vs spot | 60% reserved (6–12 month commits), 40% spot; enterprise-oriented positioning |
| Gross margin | 40–55%; GPU cloud peers CoreWeave/Lambda at ~45–55% before D&A |
| Power / data center opex per GPU/month | ~$300–400/GPU/month (power, cooling, colo); industry standard |
| Headcount ramp | 15–20 FTEs Year 1 rising to 50+ by Year 3; B2B enterprise sales + infra ops |
| CapEx: GPU cluster build-outs | Modeled as discrete tranches tied to fundraising / demand triggers |
| Depreciation: GPU useful life | 3–4 years straight-line |
| AMD partnership / launch partner benefits | Official launch partner status - may imply favorable pricing or allocation priority; quantum not disclosed |
| Edgecore hardware partner | Strategic partner for RoCE v2 cluster hardware |

- **Scenarios (Base / Bull / Bear - which variables flex):**
  - **Base:** 800 GPUs at launch, 70% utilization by end Y1, $4/hr avg rate, 45% gross margin
  - **Bull:** Faster cluster expansion (2,000+ GPUs by Y2), 85%+ utilization driven by AMD allocation advantage, premium rate ($5/hr+) from inference engine differentiation
  - **Bear:** Demand slower to convert (50% utilization Y1), NVIDIA supply unlocks reducing scarcity premium, AMD MI300X adoption slower than projected, rate pressure to $3/hr

- **Required sheets / outputs:**
  1. Assumptions - all drivers, toggle for scenario
  2. GPU Capacity & Revenue - cluster ramp, utilization, blended rate, revenue by workload type
  3. Income Statement - revenue, COGS (power, colo, bandwidth), gross profit, opex, EBITDA, D&A, EBIT
  4. CapEx & Depreciation Schedule - GPU tranches, useful life, accumulated D&A
  5. Balance Sheet - CapEx-heavy; debt/lease obligations if financed
  6. Cash Flow Statement - operating CF, CapEx, financing (equity raises)
  7. Unit Economics - revenue per GPU, gross profit per GPU, payback on GPU cost
  8. Runway / Funding Bridge - cash runway to next raise given CapEx burn

## Frequently asked questions

### Is the TensorWave financial model free?

Yes. The TensorWave model is a free Excel download with live formulas.

### Can I change the assumptions?

Yes. The workbook is editable and its live formulas recalculate when assumptions change.