20VC

Harry Stebbings × Thomas Sohmers

Positron co-founder Thomas Sohmers frames inference as a memory-wall problem: GPU FLOPS rose 120x from 2014–2024, versus 17x for memory bandwidth.Positron targets cutting a 500-megawatt Nvidia deployment to 100 megawatts, while cached tokens cost roughly one-thousandth of recomputation and support reported 80-point Anthropic API gross margins.Planned capacity may shift locations rather than disappear, but debt, regulation, and concentration remain risks to monitor.

Harry Stebbings × Cliff Weitzman

Speechify’s owned-GPU strategy compares $35,000–50,000 annual H100 rental with roughly $30,000 to buy, potentially lowering compute costs.NVIDIA’s deal with Blackstone, BlackRock, Apollo and Goldman Sachs could create a liquid secondary market and price floor through buybacks up to 25% of GPU value.Weitzman calls missing B2B his biggest strategic mistake; Simba 3.2 costs $10 per million characters versus ElevenLabs’ $100, but execution against an “unstoppable” rival remains unresolved.

Harry Stebbings

NVIDIA delivered a record $96.2B quarter and guided to 70% growth for the fiscal year ending January 2028, with end-user demand and share loss the key risks.The reported near-$12.9B Hugging Face deal, not finalized or confirmed, would reinforce open-source compute economics and NVIDIA’s push to win every segment.Agents are becoming software buyers, supporting the cases for Clay at $7B and Linear at $2.5B, while cyber exposure and reward hacking remain unresolved.

Harry Stebbings

NVIDIA’s $6B Poolside licensing deal, $1B investment at a $12B pre-money valuation, and 109-engineer transfer show frontier-model capital intensity pushing capable startups toward strategic “failure” exits.OpenAI’s 2027 IPO appears compelled by Anthropic’s $60B mid-year run rate versus under $30B this year, while enterprises’ $200–400B AI-spending capacity remains the key test.

Harry Stebbings

Jensen Huang’s 50-company open-weights letter, signed by Microsoft, Meta, IBM and Sam at OpenAI but not Anthropic, signals that frontier labs now face competition from their own customers.Open weights threaten Nvidia’s CUDA margins while LLM agents create a likely security catalyst, and Google’s negative free cash flow plus upcoming CIO budget ceilings could amplify AI volatility.

Harry Stebbings

Kimi and Qwen are accelerating open-weight competition: China-created models already drive half of OpenRouter traffic, while Kimi K3 requires frontier-scale compute.A blanket US ban is “massive overkill,” but data-export concerns could bring tighter restrictions and pressure OpenAI and Anthropic growth within one to two years.OpenRouter’s routing layer is becoming embedded infrastructure, making a sale before commoditization a live strategic question.

Harry Stebbings × Lin Qiao

Fireworks argues that private enterprise data favors millions of specialized models over one AGI, with open weights offering control and customization once models cross the quality threshold.It processes 40+ trillion tokens daily and expects 10x lower token costs to drive 100x usage, while $800M in AR and 30-40% gross margins reflect rapid scaling and leave infrastructure bottlenecks in energy, chips, and transistors.

Harry Stebbings

OpenAI’s proposed 5% stake for the US government exposes a new permission economy, as Fable 5’s 19-day ban ended with pre-approval discussions still unresolved.The debate sits alongside late-cycle signals: Nvidia recognizes hardware revenue with put-back rights while Meta sells excess compute, leaving demand diffusion and downside management as key risks.

Harry Stebbings

DeepSeek’s $7.4B Series A at a $50B valuation, with the founder committing roughly 40% and investors receiving no rights, frames open source as state-backed competition.Open source is about 5x cheaper and targets the number-three closed model, while $750B of annual capex requires replacing roughly 7-8% of US labor with tokens; 2027 is the ROI test.

Harry Stebbings × Aravind Srinivas

Aravind’s headline call is that Micron could surpass Meta in value within 6-12 months because memory remains the bottleneck after prices rose 5x.His broader thesis shifts value from commoditized search and models to orchestration and token value per watt, while power constraints, China’s vertical integration, and a 20-30% DeepSeek-moment risk remain key watchpoints.