What this research found
A 63-page venture capital due diligence report on CoreWeave, the GPU-as-a-Service provider valued at more than $19 billion whose customers include OpenAI and Microsoft. It builds a per-GPU financial model for H100 capacity, sizes the market out to 2030, applies Porter's Five Forces, and scores ten risks on likelihood and impact. The recommendation is a conditional go with a probability-weighted return of two to three times over five years, contingent on a documented NVIDIA allocation agreement and cutting top-customer concentration below 25%.
- The per-GPU model anchors the case: $32,500 of capital expenditure per H100 against a blended $1.54 per hour — a 70/30 mix of $1.25 reserved and $2.23 on-demand pricing — yields a 76.4% gross margin and $9,468 of annual revenue at 70% utilisation.
- Utilisation drives the economics. Gross margin rises from 68.8% at 50% utilisation to 80.6% at 90%, and cash payback shortens from 83.9 months to 39.8 months across the same range.
- Price leadership is quantified rather than asserted: CoreWeave lists H100 capacity at $2.23 per hour against $3.35 at Google Cloud, $3.67 at Azure, and $4.09 at AWS, putting it 35% to 83% below the hyperscalers, with Lambda Labs and Together AI closest at roughly $2.49 to $2.50.
- Supplier concentration is the one critical risk, scored 20 out of 25 on a likelihood of 4 and impact of 5. NVIDIA supplies over 95% of the fleet and controls allocation, and the Porter scoring rates supplier power at 4.5 out of 5 against competitive rivalry at 4.0.
- The market tailwind is strong but the share assumption is aggressive. The addressable market is modelled from $62 billion in 2024 to $350 billion in 2030, a 33.4% compound annual growth rate, while CoreWeave's projected capture of the serviceable market climbs from 11.4% to 26.5%.
- Customer concentration compounds the exposure: the model assumes the largest customer at 30% of revenue and the top three at 60%, so losing the largest alone would remove 30% of revenue.
How it was done
A per-GPU model for H100 capacity was built from $32,500 of capital cost, 850 W draw at a power usage effectiveness of 1.2 and $0.07 per kilowatt-hour, a $150 monthly colocation fee, and five-year straight-line depreciation, then swept across 88 utilisation and pricing scenarios. Published on-demand rates for seven providers set the competitive pricing baseline, market sizing was projected year by year to 2030 across compute, networking, and services, and ten risks were scored on likelihood and impact. Bull, base, and bear cases weighted 30%, 55%, and 15% produced the return range, alongside a staged 40/35/25 investment structure tied to milestones. The output runs 63 pages with 12 figures.
Data sources
- Published on-demand H100 pricing for CoreWeave, Lambda Labs, Together AI, Oracle Cloud, Google Cloud, Azure, and AWS
- GPU-as-a-Service market sizing projections, 2024 through 2030
- Porter's Five Forces framework (Porter, 1979)
Limitations
Several load-bearing inputs are assumptions rather than disclosed figures: customer concentration, utilisation, and revenue projections are all modelled, and the report's own investment conditions call for verifying actual utilisation above 70% before committing. The unit economics also assume the $1.54 blended rate holds; a 20% price compression is judged survivable only while utilisation stays above 65%.
Figures from this analysis
How this research was produced
K-Dense Web planned and ran this venture capital investigation end to end — gathering the sources, carrying out the analysis, producing the figures, and drafting the report. The full session transcript, including every intermediate step, is available to view.
