Rayan Krishnan is 25 years old and has a $400 million valuation for a company that tells trillion-dollar tech giants their AI models are not as good as they think. His startup, Vals AI, sells independent evaluation and benchmarking for large language models [TechCrunch, September 2026]. The bet is that public leaderboards are broken, and the real market for AI needs a private, trusted scorekeeper.
Vals AI runs 32 independent benchmarks across domains like law, finance, healthcare, and coding [Vals AI, retrieved 2026]. Its enterprise product, Vals Smith, lets customers determine which model is best for a specific use case, even building custom coding benchmarks from their own GitHub repositories [Perplexity Sonar Pro Brief, retrieved 2026] [TechTimes, August 2026]. The company’s reported $40 million Series A, led by Andreessen Horowitz last August, values the business at $400 million [TechCrunch, September 2026] [Vals job listing, August 2026]. For a16z and other backers like 8VC and Bloomberg Beta, the check is a wager on a new infrastructure layer for the AI economy.
The Wedge Against Public Benchmarks
Vals AI’s core argument is that public benchmarks like MMLU or HumanEval are gamed. Model developers train on the test data, leading to performance inflation that doesn’t translate to real professional work. The startup’s wedge is private datasets and domain-specific leaderboards designed to prevent test leakage [Perplexity Sonar Pro Brief, retrieved 2026]. It works with domain experts to turn actual workflows in law, tax, and medicine into rigorous evaluations [Calista Reyes - Building and Investing, retrieved 2026].
The result is a different performance picture. For instance, Meta’s Llama 4 reportedly underperformed on Vals’s held-out private benchmarks, despite strong public scores [BigGo Finance, retrieved 2026]. The company’s public leaderboard shows Llama 4 Maverick with 67.0% accuracy and a latency of 7.72 seconds on its tests [Vals AI, retrieved 2026]. For enterprise buyers at financial institutions or hospital systems, this kind of apples-to-apples comparison on cost, latency, and accuracy is the purchasing decision.
The Team and Traction
Krishnan, the CEO, interned at Palantir and worked at Microsoft and Stanford’s AI lab as an undergraduate [TechCrunch, September 2026]. He co-founded the company with Langston Nashold, another Stanford computer science graduate [Perplexity Sonar Pro Brief, retrieved 2026]. The team has recruited from Palantir, Microsoft, NVIDIA, Meta, and quantitative trading firm Hudson River Trading [Perplexity Sonar Pro Brief, retrieved 2026]. Chief of Staff Udani Satarasinghe joined from the investing team at private equity firm Alpine Investors [Alpine Investors, retrieved 2026].
Headcount has scaled quickly. The company grew from eight employees to roughly 30 over the course of 2026, and reported 35 employees as of late September [CNBC, October 2026] [TipRanks.com, retrieved 2026]. Six open roles were listed in late 2026, including Founding Product Manager and Founding Head of Partnerships, signaling a commercial push [Perplexity Sonar Pro Brief, retrieved 2026].
The Investor Consensus
The funding history shows a rapid consensus forming around Vals AI’s thesis. Early believers 8VC and Bloomberg Beta led a seed round in 2025 [TechCrunch, September 2026]. The Series A syndicate expanded to include Andreessen Horowitz as the lead, with participation from existing investors and new money from Hudson River Trading and NextLadder Ventures [TechCrunch, September 2026]. The table below outlines the disclosed capital stack.
| Round | Date | Amount | Lead Investor(s) | Reported Valuation |
|---|---|---|---|---|
| Seed | 2025 | Undisclosed | 8VC, Bloomberg Beta | Not disclosed |
| Series A | August 2026 | $40 million | Andreessen Horowitz (a16z) | $400 million |
Sources: [TechCrunch, September 2026], [Vals job listing, August 2026]
The $400 million valuation, reported in a company job listing, implies significant confidence in the startup’s potential to become a standard [Wilson Sonsini, retrieved 2026]. For a16z, the investment is a policy play as much as a commercial one. Firm co-founder Ben Horowitz hosted Krishnan on a podcast titled “Who Grades the AI Models?” discussing the need for independent evaluation [a16z, September 2026].
The Commercial and Competitive Landscape
Vals AI sells to a broad target list: startups, enterprises, research labs, foundation-model developers, financial institutions, and hospital systems [Perplexity Sonar Pro Brief, retrieved 2026]. The enterprise product, Vals Smith, is the primary commercial vehicle. The company claims to work with “all the major foundation model labs,” though specific customer names are not publicly verified [Perplexity Sonar Pro Brief, retrieved 2026].
The competitive field for independent AI evaluation is still forming. No direct competitors are named in available sources, which suggests Vals is early in defining the category. The real competition is the status quo: internal evaluation teams at large enterprises and the continued reliance on free, public benchmarks. Vals must convince buyers that its proprietary testing is worth the SaaS premium.
The risks to the model are straightforward and significant.
- Customer concentration. Success depends on landing and retaining large contracts from model labs and regulated industries. A few key defections could stall growth.
- Benchmark moat. The company’s advantage rests on the quality and exclusivity of its private datasets. If those leak or are reverse-engineered, the product becomes commoditized.
- Commercial execution. The team’s background is strong on technical and investment sides, but scaling an enterprise sales motion from zero is a different discipline. The hiring of a Founding Head of Partnerships is a direct response to this [Perplexity Sonar Pro Brief, retrieved 2026].
The company’s answer is likely a blend of relentless dataset curation and embedding its tools deep into customer workflows, like the GitHub integration for coding benchmarks [TechTimes, August 2026].
The Next Twelve Months
The $40 million war chest from Andreessen Horowitz, 8VC, Pear VC, Bloomberg Beta, Hudson River Trading, and NextLadder Ventures buys runway for aggressive hiring and product development [TechCrunch, September 2026]. The next milestones will be commercial. Watch for named enterprise customer announcements, likely in finance or healthcare, to validate the sales motion. The expansion of the Vals Smith platform and potential international regulatory partnerships, hinted at in an a16z Policy episode, are also on the roadmap [a16z Policy, August 2026].
The fundamental question for Krishnan and his team is whether they can transition from a well-funded research project to the de facto grading system for a multi-trillion-dollar industry. If they can, the $400 million valuation will look cheap. If the enterprise buyers decide to build their own tests, it will look ambitious. For now, the checkwriters have placed their bet. Can Vals AI become the S&P 500 rating for AI models, or will it remain a niche tool for the labs? The next year of customer logos will provide the answer.
Sources
- [TechCrunch, September 2026] Vals, backed by Andreessen Horowitz, is looking to become the gold standard for AI benchmarking | https://techcrunch.com/2026/09/19/vals-backed-by-andreessen-horowitz-is-looking-to-become-the-gold-standard-for-ai-benchmarking/
- [Perplexity Sonar Pro Brief, retrieved 2026] Vals AI company and product overview
- [Vals AI, retrieved 2026] AI Benchmarks for Real-World Tasks | Vals AI | https://www.vals.ai/home
- [Vals job listing, August 2026] Series A valuation reference
- [Wilson Sonsini, retrieved 2026] Vals AI valuation reference
- [CNBC, October 2026] Reference to employee growth
- [TipRanks.com, retrieved 2026] Employee count reference
- [BigGo Finance, retrieved 2026] Reference to Llama 4 performance
- [Calista Reyes - Building and Investing, retrieved 2026] Essay on Vals AI's approach
- [TechTimes, August 2026] Article on Vals Smith GitHub integration
- [Alpine Investors, retrieved 2026] Background on Udani Satarasinghe
- [a16z, September 2026] Who Grades the AI Models? Ben Horowitz & Rayan Krishnan podcast | https://a16z.com/podcast/who-grades-the-ai-models-ben-horowitz-rayan-krishnan/
- [a16z Policy, August 2026] Measuring AI: What Policymakers Can Learn from Benchmarks | https://a16zpolicy.substack.com/p/measuring-ai-what-policymakers-can