Half of the 35 startups in Axiom Partners' $52 million fund are expected to fail. Founder and managing partner Sandhya Venkatachalam, who led early institutional investments in Groq at Social Capital, says the model needs one outstanding company to return the fund. The fund backs AI that performs work in construction, industrials and insurance rather than selling another software tool. The stance matters because customers are already paying for such systems from labor budgets in contracts worth hundreds of thousands of dollars.
How Axiom picks founders and real-world jobs
Venkatachalam built her operating background in infrastructure and communications, leading product at a data center hardware company later sold to Cisco and serving as a product executive at Skype before its sale to Microsoft. After Social Capital and a period investing at Khosla Ventures, she founded Axiom Partners to back startups producing outcomes in underserved industries. The portfolio spans software-only products as well as systems involving hardware, sensors or robotics. The common filter is work that matters to customers in construction, industrials and insurance, not a conventional enterprise tool.
Axiom is organized around people who build and sell AI in their other roles. Those practitioners work with the firm part time, have dedicated time for investing and receive carry, acting as working partners. They stay close to productizing, pricing and taking AI to market, which keeps investment thinking current. The firm also runs what it calls the Axiom Brain to track market trends, surface people and companies, and accelerate diligence. Risk assessment focuses on the next milestones: whether the team can deliver and whether success could be large.
The thesis traces back to a 2016 investment in Groq, when AI inference was far from an established category. Venkatachalam had examined why Google built its own networking switches and then chips, which led her to Jonathan Ross, who left to start Groq. Ross argued that training attracted attention, but the larger future market would be inference to support applications built on models. She barely understood inference then, but wider model use pointed to infrastructure demand. The lesson was to move a little early, with patience, before consensus forms.
What this means for companies buying AI
For buyers, the distinction is between software licenses and payment for completed work. Axiom requires evidence, even at alpha or design-partner stage, that customers will buy from labor budgets, and reports such behavior across most portfolio companies. Contract values often reach the hundreds of thousands of dollars, above a typical midmarket software tool. For a small contractor or insurer, one such system can cover a scarce function without hiring. For a large industrial firm, the same capability must integrate with existing systems, data and workflows to justify the larger commitment.
Durability comes from that last mile, according to Venkatachalam. A defensible company integrates deeply, learns customer data, trains on it, understands the workflows that matter and stands behind the result, rather than placing an interface on a model. Such ties are harder for another startup to replicate and may hold little interest for large model providers. Still, faster building and easier imitation pressure any product without embedded deployment. Buyers should verify integration effort, data handling, responsibility for outcomes and replacement cost before treating a pilot as proven.
The marker to watch is whether design partners convert to labor-budget contracts at the hundreds-of-thousands level and expand across sites. If one of Axiom's 35 companies scales that pattern in construction, industrials or insurance, the model of tolerating half failing will read as discipline. Without that conversion, outcome-based AI remains a sales narrative rather than a procurement shift.
