Jul 22, 2026, 11:50 AM ETFireside Chat, Interview

AI Sticker Shock: Open Source as the Answer to Broken Tokenomics | Vipul Ved Prakash | RAISE 2026

TogetherAI (Vipple) crossed $1 billion in annual revenue earlier this year, growing from $100 million.
Current cash on hand exceeds $1 billion.
The company's most recent valuation stands at $8.3 billion.
Inference volume surged from 30 billion tokens per month to 400 trillion tokens per month on their platform.
This represents a 10,000x growth rate, significantly outpacing Google's Gemini (7x growth) during comparable periods.

The company attributes a "stampede" toward open models over the last nine months to economics, geopolitics, and model performance parity.
Economic drivers: Companies are prioritizing token optimization over "token maxing" to manage massive inference costs.
Geopolitical drivers: Sovereignty and control over model weights have become critical for national and corporate security.
Performance parity: Chinese open-source models (e.g., GLM, Kimmy, Minimax, Nemotron, DeepSeq) are approaching the performance of frontier closed models like Claude 3.5/4.7/4.8.
Competitive threat: Major labs (e.g., Anthropic, OpenAI) expanding into adjacent verticals (e.g., pharma, design) are viewed as existential threats, prompting companies to seek data sovereignty to prevent IP loss.
Security architecture: TogetherAI operates on U.S. or European infrastructure to ensure data residency, guaranteeing they do not train on customer data to build their own frontier models.

The company is addressing global compute and power constraints through a "global building" strategy, partnering with other clouds, neo-clouds, and sovereign states.
They are underwriting sovereign infrastructure investments by bringing workloads to local sites to ensure immediate investment positivity.
Sovereign partnerships are being established across two continents, with specific country announcements pending.
The strategy explicitly targets regions facing U.S. municipal moratoriums on new data center builds.

Evolving consensus: Co-founder Vinod Khosla previously advocated for closed models, while current leadership argues open weights are essential for a distributed, non-monopolistic AI future.
Historical analogy: The market is expected to follow a pattern of "centralization followed by disaggregation" similar to the mainframe-to-PC or OS-to-App shifts.
Risk of concentration: Closed API dominance poses a risk of "industrial capture," where a few labs could control entire industries (pharma, manufacturing) via data and know-how.
Security misconceptions: The founders argue that "open weights" are merely downloadable files subject to intense public scrutiny, countering fears of security risks associated with local deployment.
Sovereign equity proposal: The speaker suggested a hypothetical model where global governments (GDP-weighted) hold equity in AI infrastructure to ensure global access and prevent single-nation capture, though no specific deal was made.

The market is characterized as an "early moment" within a massive five-year arc, requiring continuous updates to business priors.
The primary challenge for growth remains access to compute, power, and the operational complexity of deploying open models.
Future growth is predicted to rely on a "collaborative commons" approach where infrastructure providers handle the complexity of open weights, allowing customers to focus on application and data.