Dave McClure warns open source threatens frontier AI margins
- Open-source models now process nearly 80 percent of tracked token volume.
- Venture capitalists are shifting capital from closed AI labs to open weights.
- Enterprise buyers deploy local models to protect trade secrets from central APIs.
The high-margin monopoly of frontier AI labs is collapsing.
On This Week in Startups, Uncork Capital founder Jeff Clavier characterized recent 50 percent price cuts by OpenAI and Anthropic as a venture-subsidized race to the bottom. Proprietary labs spend billions training massive architectures, only to sell token access at razor-thin margins. Meanwhile, open-weight alternatives command 78 percent of tracked token volume on routing platforms, according to investor Dave McClure.
Venture capital managing partner Jenny Fielding noted on the show that early-stage founders care strictly about price and performance. Building on closed APIs leaves startups exposed to platform risk and direct competition from their model vendors. Consequently, portfolio companies across early-stage funds are systematically shifting toward open-source architectures.
Five days later on TFTC, Nous Research strategist Tommy Eastman explained that open-source models have reached a capability threshold where 90 percent of corporate users can handle 90 percent of routine workloads without relying on proprietary frontier endpoints.
This technical parity coincides with a sharp pivot in enterprise data governance. Law firms, financial institutions, and corporate buyers are pulling data out of centralized API pipelines over fears that feeding core intellectual property into black-box models violates fiduciary duties.
To resolve corporate privacy risks, teams are pairing open-source models with dedicated execution harnesses like Hermes, designed by Nous Research to maintain self-referential memory and reproducible task paths. Host Marty Bent outlined how using such a harness with local storage allows businesses to replace manual sales and editorial workflows while keeping data within secure cloud enclaves.
The resulting surge in local model execution has inverted hardware dynamics. Eastman noted that data center sales teams are turning enterprise buyers away as inference demand completely overwhelms available GPU server capacity, shifting value away from closed labs and toward raw compute providers.
Proprietary labs built the frontier, but open weights are winning the market.
Source Intelligence
- Deep dive into what was said in the episodes
Marty Bent
#796: Own Your Intelligence Stack with Tommy Eastman • Sep 28
- Tommy Eastman explains that Noose Research designed the Hermes Agent to solve reproducibility issues in AI agents by storing and recalling the exact procedural primitives of successful task paths.
- Marty Bent outlines a three-legged architecture for business AI integration comprising the file system or second brain, the agentic harness, and the underlying AI models.
- Tommy Eastman claims that open-source models have reached a performance threshold where 90 percent of users can complete 90 percent of their workloads without relying on frontier closed models.
- Tommy Eastman emphasizes that utilizing open-source Chinese AI models does not compromise data privacy if the weights are run locally or hosted on domestic US cloud infrastructure.
Also discussed on this episode: (6)
Safety (1)
- Tommy Eastman warns that feeding trade secrets and intellectual property to closed AI models like OpenAI and Anthropic poses a massive fiduciary risk for enterprises.
AI Infrastructure (1)
- Tommy Eastman reports that the high-performance computing market has inverted, leaving graphics processing unit compute clusters so highly squeezed that buyers must pitch internal sales teams to secure contracts.
Open Source (1)
- Tommy Eastman argues that corporate lobbying against open-source AI is a disingenuous attempt to capture profits rather than a genuine effort to mitigate existential technology risks.
Payments (1)
- The Aven Bitcoin Visa card allows users to access a line of credit up to 1 million dollars backed by their bitcoin without selling assets or triggering taxable events.
Mining (1)
- Simple Mining operates more than 40,000 machines in Iowa, utilizing low industrial power rates to mine bitcoin at a discount to spot price.
Health (1)
- Marty Bent notes that his family of five pays 700 dollars per month for crowdfunded healthcare through CrowdHealth, avoiding traditional health insurance.
VCs Would Bet on Open-Source AI Over OpenAI and Anthropic | E2341 • Sep 23
- Dave McClure highlights a massive market shift toward open-source artificial intelligence. Data shows open-weight models now command 78% of tracked token volume on routing platforms, leaving closed-weight frontier models with just 21% of the market.
Also discussed on this episode: (12)
VC (8)
- Jeff Clavier raised one of the first dedicated seed funds in 2007, a $15 million vehicle. At the time, Fitbit's 2008 seed round of $2 million at a $5 million pre-money valuation was considered massive.
- Instinct reached a rumored $2.5 billion valuation in under five months and is now negotiating a $1 billion raise at a $10 billion valuation. The unreleased iMessage assistant has accumulated 100,000 waitlisted users while burning massive cash on compute.
- Jason Calacanis warns that extreme valuations like Instinct's mimic the Clubhouse and Fab.com cycles. Historically, Weblogs Inc was sold to AOL for $30 million, yielding a $5 million return on Mark Cuban's initial $300,000 seed investment.
- Jeff Clavier reports that Uncork Capital secured $1.5 billion in exit distributions over a recent three to four month window. This represents a significant liquidity event following a prolonged five-year drought in the technology acquisition market.
- Venture capital seed stage valuations have surged dramatically over the last decade. Median seed valuations rose from $8 million in 2017 to $28 million, while elite deals in the 95th percentile spiked to $208 million.
- Dave McClure pitches single asset continuity SPVs as a tool for early stage managers to generate DPI. Managers can sell a portion of a winner to the SPV at a post-money Series B or C price, satisfying LP liquidity demands.
- Dave McClure advises startup employees and GPs to donate illiquid equity or carried interest directly to donor-advised funds. This strategy secures immediate tax deductions and avoids capital gains liabilities ahead of end-of-year tax deadlines.
- Jeff Clavier warns that mid-market companies valued at cheap 2x revenue multiples face deep writedowns. Private acquirers like Bending Spoons are buying these struggling VC-backed startups, leaving historical cap tables with minimal returns.
Regulation (1)
- Dave McClure predicts that governments will establish an AI sovereign wealth fund within 12 months. This fund will function as a regulatory shakedown, granting companies a product liability shield in exchange for a 10% to 20% equity stake.
Safety (1)
- Jeff Clavier argues that AI safety requires direct CEO accountability rather than regulatory shields. He highlights that 26% of Anthropic's team focuses on safety alignment, yet preventing unpredictable autonomous agent behavior remains extremely difficult.
Startups (1)
- Dave McClure suggests Stripe acquired Open Router to act as an anti-fraud layer. This move allows Stripe to leverage its fraud-prevention capabilities to monitor and secure untrusted AI token usage rather than simply offering low-cost API routing.
Space (1)
- Jeff Clavier details Loft Orbital's new $1 billion Abu Dhabi investment to construct a space-based AI compute constellation. The constellation uses onboard GPUs to process satellite data on orbit, eliminating the latency of streaming massive raw images.