AI Startups to Watch (Sep 20): Nscale, Vals AI & Jev
AI startups to watch on Sep 20, 2026: Nscale files for a $30B IPO, Vals AI raises $40M to grade models, and TypeSafe ships Jev to developers.
The most interesting AI startups this week are not selling bigger models. They are selling judgment. Nscale is testing public markets with a loss-heavy cloud balance sheet. Vals AI wants to be the referee that decides which models actually work. TypeSafe is shipping a model that refuses to write prose. And Vantora is betting that factories, not chatbots, are where the next wave of value gets built.
Nscale: renting the shovels, at a price
Nscale is a London-based AI cloud provider, founded in 2023, that rents compute, power, data center capacity, and the software to train and run models. It competes with CoreWeave, Nebius, Crusoe, and Lambda, operates in 14 regions, and claims a power pipeline above 10 gigawatts.
The move: on September 18 the company filed its S-1 to list on the New York Stock Exchange under the ticker NSCL, targeting a valuation near $30 billion. That is roughly double the $14.6 billion it was worth after a March round. Goldman Sachs, J.P. Morgan, and Morgan Stanley are lead underwriters, and Nvidia bought $1 billion of a $3.1 billion convertible bond issue this week. Nvidia also holds more than 5% through non-voting shares. Source: Reuters
The numbers: $140.6 million of revenue in the six months to June 30, up 1,252% from $10.4 million a year earlier, against a net loss of $1.02 billion, up from $368.9 million. Source: Financial News
Who it hurts: the listing tests whether public investors will fund capital-hungry infrastructure after calls from AI lab CEOs for a slower development pace rattled the sector. It also piles pressure on CoreWeave and Nebius to keep pricing power. The concentration risk is the catch: one unnamed customer produced 52% of first-half revenue, and Anthropic's $45 billion contract for West Virginia compute dominates the $103 billion backlog.
Vals AI: grading the graders
Vals AI is a San Francisco startup formed in 2024 by Rayan Krishnan, who interned at Palantir and worked at Microsoft and Stanford's AI lab before dropping into the evaluation business. It raised a seed round led by 8VC and Bloomberg Beta, then $40 million in a Series A led by Andreessen Horowitz last month. Source: TechCrunch
The angle: legacy benchmarks are public, which means labs can train against the test. Vals keeps its materials private and evaluates models on work that resembles real jobs in law, finance, and coding, along with recursive self improvement, mental health, cybersecurity, biosecurity, and the law of armed conflict. Companies pay to be evaluated, a model Krishnan compares to paying the College Board for the SAT.
Why it is disruptive: revenue is now eight times last year's level, the team has tripled from eight people to 25, and Vals recently launched a program to evaluate models for federal agencies. If benchmark results start appearing in IPO filings, whoever controls the tests controls the marketing.
TypeSafe AI: a model that says "I don't know"
TypeSafe AI was founded by Diogo Almeida, who left OpenAI two years ago after helping build the instruction-following methods that ended up behind ChatGPT. His critique: language models are optimized for human talk, and strings are a terrible interface for software. This week the company released Jev, the first of what it calls System One Models. Jev is not an LLM. It outputs type-safe values with calibrated probabilities, so it cannot produce text, and therefore cannot hallucinate. Source: TypeSafe AI
The economics are the pitch: input tokens at $0.042 per million, output tokens free, and response times of 70 to 500 milliseconds versus the 3 to 329 seconds typical of frontier models. Pranit Sharma, a software engineer at Vercel, said his team swapped OpenAI's ChatGPT Luna 5.6 for Jev on a command-safety classifier and got results five to 18 times faster with better accuracy. Source: TechCrunch
Demand briefly outran capacity and the API had trouble serving users. Armin Ronacher, CTO of Earendil, noted the tradeoff: confidence scores "delegate the hallucination problem a little bit to the user," because a 50% answer should be treated as a coin toss. No funding round has been disclosed for TypeSafe.
Vantora: startups that build startups
Vantora is a Boston company, founded by Vikram Ravichandran, Madhavan Seshadri, and Manpreet Singh Khurana, that builds and operates AI-native businesses rather than selling one product. Its focus is physical AI: software that touches cars, robots, and industrial systems instead of chat windows. Source: Vantora
The move: Vantora raised $100 million from Silversmith Capital Partners. Ravichandran's framing is blunt: lots of AI startups in this space are really product companies with a demo and no operations, and the winners will be the ones that run the business, not just build the software.
Why it matters: it is the studio model applied to hardware-adjacent AI, where capital intensity and field operations are the moat. If a company can spin up operative ventures faster than founders can hire, it changes the default path for physical AI talent.
Startup Snapshot
| Startup | What They Do | Latest Round | Valuation | Key Differentiator |
| --- | --- | --- | --- | --- |
| Nscale | Rents GPUs, power and data centers | IPO filing (Sep 18) | ~$30B target, $14.6B in March | $103B contracted value, 10GW pipeline |
| Vals AI | Independent AI benchmarking | $40M Series A (a16z) | Not disclosed | Private tests labs cannot train on |
| TypeSafe AI | Calibrated decision models | Not disclosed | Not disclosed | Jev cannot hallucinate, free output tokens |
| Vantora | Builds physical AI companies | $100M (Silversmith) | Not disclosed | Operates ventures, not just products |
Frequently Asked Questions
What are the top AI startups to watch today?
Nscale, Vals AI, TypeSafe AI, and Vantora are the AI startups to watch today. Nscale filed for a $30 billion listing, Vals raised $40 million to grade models, TypeSafe shipped a non-language model, and Vantora raised $100 million for physical AI. Together they sell judgment rather than benchmark scores. (~50 words)
Why is Nscale's IPO risky?
Nscale's IPO is risky because its loss is about seven times its revenue and one customer provided 52% of sales. Revenue reached $140.6 million in the first half while the net loss hit $1.02 billion. The $103 billion of contracted value depends on data centers that still have to be built and powered. (~52 words)
How is Jev different from ChatGPT?
Jev is different because it does not generate text at all, only structured values with calibrated confidence scores. That design makes hallucination structurally impossible and pushes cost down to $0.042 per million input tokens, with output tokens free. It is built for software decisions, not conversation. (~48 words)
The Bottom Line
The pattern across all four companies is the same: the money is moving toward infrastructure, measurement, and reliability, not toward another chat interface. For the policy and industry side of the same day, see the General AI News digest from Sep 19, and browse the AI Startups tag page for earlier rounds.
---
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch →
Related reading
Deeper context from the archive on the companies and technologies in this story:
