AI Startups to Watch (Sep 19): EVAS, Factory & PrismML
AI startups to watch September 19: EVAS RISC-V chips, Factory at a $5B valuation and PrismML fitting a 27B model on a phone. Five moves worth watching.
The most disruptive AI startups this week are not the ones with the biggest logos. A Chinese chipmaker raised nearly RMB2 billion to attack Nvidia's software moat, a Dutch company pulled more than €200 million for a chip nobody can buy yet, and a three-person Indian security firm breached OpenAI for the price of a used laptop. Here are the five moves worth watching in AI startups today.
EVAS Intelligence: RISC-V cloud AI chips
EVAS Intelligence is a Chinese chip designer building cloud AI accelerators on the RISC-V instruction set, the open standard that competes with the proprietary Arm and x86 architectures. It raised nearly RMB2 billion, roughly $280 million, at a reported post-money valuation approaching RMB15 billion, about $2.1 billion. Source: TechStartups
The disruption runs deeper than the dollar figure. Nvidia's real lock-in is CUDA, the programming layer that makes GPUs easy to use and rivals hard to adopt. A well-funded RISC-V accelerator vendor is a bet that buyers will accept open instruction sets if the price is right. It is the clearest sign yet that China's chip strategy has moved from catching up on GPUs to designing around the GPU paradigm entirely.
Factory: software factories at a $5 billion valuation
Factory is a San Francisco startup founded in 2023 by Matan Grinberg and Eno Reyes. It sells what it calls software factories, defined as continuously running systems that take signals from bug reports, internal messages, and customer feedback, then carry work through development, review, deployment, and monitoring in one closed loop. It raised $200 million, co-led by Blackstone and five other institutions, lifting its valuation to $5 billion. Source: Gate News
Growth is the story. Factory was valued at $1.5 billion in April 2026, so this round is roughly 3.3 times its April mark in five months. Rival Cognition closed a Series E above $2 billion at a $48 billion valuation a week earlier, near 53 times annualized revenue. The category thesis is that enterprises are shifting from single coding agents to always-on software operations, and both companies are pricing that shift aggressively.
PrismML: compressing a 27B model to 5.9 GB
PrismML is a Caltech-founded startup led by professor Babak Hassibi, a specialist in compression technology, and backed by Khosla Ventures, Cerberus Capital, and Caltech. It released Bonsai 2 27B, which compresses Alibaba's Qwen3.8 27B open-source model to 5.9 GB, a 9x to 10x reduction, while matching 98% of the original's aggregate benchmark scores. Source: TechCrunch
PrismML uses ternary weights, reducing each of the model's stored values to positive one, negative one, or zero instead of the usual 16 bits. That cuts memory sharply enough for a PC and possibly a high-end phone. The first Bonsai shipped in March at 95% of benchmarks and has been downloaded over 11 million times, with 2.6 million more across PrismML's smaller models. The company runs on Nvidia GPUs via CUDA and Apple silicon via MLX.
Who it hurts is cloud inference. If a 27B reasoning model runs locally, the per-token bill disappears and data never leaves the device, which undercuts the economics of rented inference for a large class of workloads. The 98% figure is vendor-reported and a 2% aggregate gap can widen on real tasks.
Euclyd: €200 million for an inference chip
Euclyd is a 2024-founded Eindhoven chip company building an inference accelerator it markets as agentic AI silicon. It raised more than €200 million, about $231 million, in a Series A co-led by Samsung, Somerset Capital Partners, the Scaleup Europe Fund managed by EQT, and Innovation Industries, overshooting its own €100 million target. Former ASML chief executive Peter Wennink chairs the company. Source: Dutch Startup News
The round is a bet on processor-memory co-design at the moment AI infrastructure spending is being criticised as wasteful. Samsung's participation is the signal worth reading, because a memory giant funding designs that make memory more central to compute is a hedge against being commoditised by it. Neither Euclyd nor anyone else has shipped at scale.
Hacktron AI: three people, one rival model, $6,500
Hacktron AI is an Indian cybersecurity startup whose three-person team, Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini, chained a libheif image-processing bug in OpenAI's Discourse forum with a single sign-on flaw to reach employee ChatGPT and Codex accounts and OpenAI's private Monorepo GitHub organisation. They filed a pull request as proof and collected a $6,500 bounty. Source: Cryptopolitan
The disruptive detail is cost. The team spent under $3,000 in tokens, and Anthropic's Claude Opus 4.8 failed to produce a working exploit while Opus 5, released July 24, produced one the next morning. That collapse in the price of serious offensive research is a direct threat to the business models of security consultancies that bill for scarce exploit expertise.
Startup Snapshot
| Startup | What They Do | Latest Round | Valuation | Key Differentiator |
|---|---|---|---|---|
| EVAS Intelligence | RISC-V cloud AI chips | ~RMB2B | ~RMB15B | Open ISA attack on CUDA |
| Factory | Agentic software factories | $200M | $5B | Always-on software ops loop |
| PrismML | LLM compression | Not disclosed | Not disclosed | 27B model in 5.9 GB |
| Euclyd | Inference silicon | €200M+ | Not disclosed | Ex-ASML chair, Samsung backer |
| Hacktron AI | AI-assisted offensive security | Bootstrapped | Not disclosed | Frontier exploit for $3K in tokens |
Frequently Asked Questions
What are the top AI startups to watch today?
EVAS Intelligence, Factory, PrismML, Euclyd, and Hacktron AI are the five most disruptive AI startups in the last 24 hours. Between them they target Nvidia's CUDA moat, the coding agent market, cloud inference economics, inference silicon, and the price of offensive security research, with combined disclosed capital above $700 million.
Why is EVAS Intelligence raising so much for RISC-V chips?
EVAS Intelligence is betting that open instruction sets can break Nvidia's CUDA lock-in on cloud AI. The Chinese chip designer raised nearly RMB2 billion at a reported valuation near RMB15 billion. Nvidia's advantage is as much software as silicon, so a well-funded RISC-V accelerator vendor with domestic backing is a direct challenge to that position.
How did Hacktron AI breach OpenAI with three people?
Hacktron AI chained a libheif image bug in OpenAI's Discourse forum with a single sign-on flaw to reach employee accounts and the private Monorepo repository. The team spent under $3,000 in tokens, and Claude Opus 5 produced a working exploit within hours after Opus 4.8 failed. OpenAI paid a $6,500 bounty.
Yesterday's roundup: AI Startups to Watch (Sep 18) covered Crusoe, Emulate, and Arcee. See every edition under the AI Startups tag.
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch →
Related reading
Background worth having before you act on the above:
