A $40 million Series A for Vals AI surfaces a genuine structural tension at the center of the AI evaluation industry: a benchmark can be auditable or it can be uncontaminated, but it is very hard to be both at once.
TechCrunch reported on September 19, 2026 that Vals AI raised a $40 million Series A led by Andreessen Horowitz, following an earlier seed from 8VC and Bloomberg Beta. The company builds benchmarks that test modelperformance on real-world professional tasks โ across law, finance, coding, cybersecurity, and biosecurity โ and keeps its test materials confidential so that models cannot be trained against them. No valuation was disclosed.
Co-founder Rayan Krishnan, 25, previously interned at Palantir and worked at Microsoft and Stanford’s AI lab. The company’s origin thesis is stated plainly by Krishnan himself: “We were seeing a bunch of new, very capable models come to market quickly, and the academic benchmarks [were] not keeping up with that frontier advance.” His argument is that once a test is public, a model developer can train against it โ what he characterises as “arguably cheating on their exam.” The business model is straightforward: companies pay Vals to test their models.
Headcount grew from eight employees at the start of 2026 to twenty-five. Revenue is reported as eight times higher than the prior year; no base was disclosed, so no absolute revenue level is estimated here. The twenty-five-person headcount is the more informative figure precisely because it is a level rather than a ratio.
The key insight: The confidentiality that makes Vals’s benchmarks harder to game is the same property that makes them harder to audit. That is a genuine design trade-off with real costs on each side โ not a flaw in the approach, and not unique to Vals. Every serious benchmark designer faces it. The consequence worth naming is what it asks of the reader: trust the examiner rather than inspect the exam.
The Structural Read
Publishing a test lets anyone verify that it measures what it claims to measure โ that the questions are fair, that the scoring is sound, that the thing being scored resembles the thing being advertised. Publishing it also lets anyone optimise directly against it. Those two properties are not in tension by accident. They are the same property viewed from different sides: openness is what enables both verification and contamination simultaneously. Removing one removes the other.
The consequence of confidential benchmarking is not that it is wrong โ it is that it moves the locus of trust. A score from a public benchmark is checkable in principle by anyone. A score from a confidential benchmark is checkable only by the organisation that ran the test. That shifts the question a score-reader must answer from is this test well-designed? to is this organisation trustworthy? Nothing here says whether Vals is trustworthy, and nothing here implies that it is not. The point is structural: the information regime changes what the number means to whoever relies on it.
FDE Framework โ Enabler Position
Vals sits in the Enabler layer of the AI stack โ the infrastructure that makes other infrastructure legible.
In the FDE lens, Enablers do not build the models (Founders) and do not deploy them at scale to end users (Distributors). They build the instruments โ the measurement rails, the trust scaffolding, the verification layer โ that the rest of the stack depends on to function credibly. The business risk for an Enabler in this position is not competition from peers; it is the question of whose authority backs the instrument. A standardised test’s authority derives from its institutional position as much as from its methodology. Vals is, at this stage, building both simultaneously.
Krishnan’s SAT comparison is his own, and it earns more examination than it typically receives. The comparison is sound on the surface most people reach for first and dismiss: a test-taker paying for a test is entirely ordinary, and it is not in itself compromising. The structural difference worth adding is not an accusation but a geometry of interested parties. In the SAT case the score consumer is a university โ a third party with its own strong interest in the score being honest and the practical ability to notice if scores stop predicting anything. Where a model developer pays for a test and then circulates the result, the party actually relying on that score is further from the transaction, and has less ability to detect drift. That is an observation about where parties sit relative to each other, not a suggestion that anyone misreports anything.
The deeper pattern is not specific to benchmarking. Independent verification costs money. Somebody has to pay for it. Every arrangement for paying creates a relationship between checker and checked. The answer is never “remove the relationship” โ that option does not exist. The answer is to make the relationship visible enough that a reader can price it into the score they are reading. Vals’s model makes some relationships visible (who pays, what domains are tested) while necessarily leaving others opaque (what the questions are, how scoring is calibrated). That is the design, and it is a coherent one.
Three Implications
IMPLICATION 1 โ THE EVALUATION LAYER IS BECOMING INFRASTRUCTURE
As model capabilities compound, the gap between what a model scores on a public academic benchmark and what it actually does in a regulated professional context widens. That gap is a market. Vals is building in that gap. The Enabler position in the AI stack โ the measurement and trust layer โ has historically been underbuilt relative to the model and application layers above and below it. A16z’s participation signals that at least one major allocator thinks that is changing.
IMPLICATION 2 โ CONFIDENTIALITY CREATES AN INSTITUTIONAL AUTHORITY PROBLEM
The value of a confidential benchmark is only as durable as the credibility of the institution running it. Public benchmarks accumulate authority through reproducibility and community scrutiny over time. Confidential benchmarks accumulate authority through track record, customer loyalty, and the perceived independence of the examiner. Vals is 25 people and a few years old. The authority-building process is not automatic, and it is not fast. That is not a critique of the model; it is the central strategic challenge the model creates.
IMPLICATION 3 โ THE SCORE-READER BEARS MORE OF THE EPISTEMIC BURDEN
When a confidential benchmark score circulates โ in a product announcement, a procurement decision, a regulatory filing โ the downstream reader cannot verify the underlying methodology. They are relying on the chain of custody between examiner and exam being intact and honest. In high-stakes domains like biosecurity or legal advice, that chain of custody matters more, not less. The evaluation layer becoming proprietary raises the question of what disclosure norms, if any, should govern how confidential scores are presented to parties who did not pay for the test.
Business Engineer Framework
The Map of AI โ Where Vals Sits in the Stack
Vals is an Enabler play in the AI stack โ not building models, not deploying applications, but building the measurement rails that make the rest of the stack legible and trustworthy. The Map of AI framework maps 200+ companies across 9 layers of the AI value chain. Understanding which layer a company occupies โ and which adjacent layers it might need to control โ is the structural read that separates a funding announcement from a business-model insight.
Not investment advice. This article is analytical commentary on publicly reported information. Nothing here constitutes a recommendation to buy, sell, or hold any security or asset. No valuation is estimated for Vals AI.
The Bottom Line
Vals AI’s $40 million Series A is less interesting as a funding event than as a crystallisation of a structural question the AI industry has not answered: when the benchmark is secret, the score is only as good as the trust you extend to the examiner, and trust of that kind is built slowly, institutionally, and not by press release. That is the business Vals is actually in โ not just evaluation, but the slower, harder work of becoming the kind of organisation whose confidential judgment a regulated industry will accept. The money buys runway. The authority has to be earned.
Get structural reads like this every week
Business Engineer covers AI strategy, infrastructure, and competitive dynamics โ no hype, no summaries.
This is not investment advice.No valuation was disclosed for the Series A, and none is estimated above. The report that revenue is eight times last year’s carries no disclosed base, so it says nothing about absolute size and no revenue figure is inferred above. The comparison to students paying the College Board for the SAT is Rayan Krishnan’s own analogy, reported as his and not as something uncovered. The trade-off described between an auditable benchmark and an uncontaminated one is a genuine design choice faced by every benchmark designer, not a flaw in this company’s approach; nothing above says whether Vals is trustworthy, and nothing above implies that it is not. “Gold standard” describes an ambition reported in the source headline; nothing above states that Vals is or will become any kind of standard. No customer, model score, price or competitor figure appears above, no laboratory is named as having trained on a test set or gamed a benchmark, and no company is claimed to have done so. Nothing is predicted.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
