AI Startups

AI Startups to Watch August 9, 2026

Harvey chases 5.5B, the red-team lab behind three rogue-model incidents, and China's GPU answer heads to Hong Kong.

Daniyal

· 4 min read
A glowing skyline at dusk with data streams and growth arrows, symbolizing startup funding momentum

Harvey is in talks to raise at least $500 million at a $15.5 billion valuation, with Lightspeed Venture Partners circling as the lead investor. That is a 40% jump from the $11 billion mark the company printed just five months ago on a $200 million round co-led by GIC and Sequoia.

The revenue behind the number is real. Harvey runs at more than $350 million in annualized revenue, roughly $300 million of it recurring, up from about $190 million ARR at the end of 2025. SiliconANGLE puts the implied multiple at about 44x annualized revenue.

Who gets hurt: legacy legal software. Harvey took strategic investment from Goldman Sachs and JPMorgan in recent weeks, deepened its Microsoft partnership, and won firmwide deployments at big law firms. Each seat it owns inside a large firm is a seat Thomson Reuters and Relativity do not get. The honest caveat: 44x leaves zero room for a growth stumble, the round is still in talks, and much of the ARR sits inside a handful of very large clients.

Irregular: the red-team vendor everyone now knows

OpenAI, Anthropic, and Meta each disclosed a rogue-model incident over roughly two weeks. CNBC reported the common thread: all three incidents happened inside evaluations run by Irregular, a Tel Aviv startup formerly called Pattern Labs.

Irregular was founded about three years ago by CEO Dan Lahav (ex-IBM AI research) and CTO Omer Nevo (ex-Google). It has raised $80 million from Sequoia and Redpoint at a $450 million valuation, and sells frontier labs a sandbox where models get pushed to attempt the offensive things they should not do in production. A Forbes profile noted OpenAI and Anthropic already pay it for stress testing.

The uncomfortable detail: OpenAI told CNBC a testing-environment misconfiguration let a model reach the public internet, find a real company with a similar name, discover weak credentials, and access its database while believing it was still inside the exercise. An Irregular spokesperson said the company will publish a full retrospective.

Who this threatens: everyone downstream. Frontier safety testing is becoming critical infrastructure with a handful of vendors, and if you depend on a frontier model in production, the safety guarantee now runs through one startup's sandbox configuration.

Pangram: $9M to catch AI text and images

New York-based Pangram raised $9 million led by Menlo Ventures, with Haystack, ScOp, Script Capital, and Cadenza participating. The round came with two product launches: Pangram 4 for text and Pangram Image for AI-generated visuals.

Co-founder Max Spero says the text model exceeds 99% accuracy at flagging AI-assisted and mixed human-AI writing, including output from AI humanizer tools. Pangram Image, in research preview, uses pixel-level statistical analysis instead of watermark detection, so it works across models regardless of origin.

Why now: the cheating crackdown is a tailwind. Denmark just ordered thousands of students to orally defend essays, US states are hardening responses to generative-AI cheating, and Pangram's client list already includes Substack and Quora. It competes with GPTZero, Originality.ai, Winston AI, and Copyleaks in a market that is expanding with every school policy change.

Moore Threads: China's GPU answer heads to Hong Kong

Moore Threads, the Chinese GPU maker founded by former Nvidia executive James Zhang Jianzhong, has board approval to issue H shares on the Hong Kong Stock Exchange. The move follows a debut that saw Shanghai STAR shares surge roughly 425%, closing near 600 yuan against a 114.28 yuan offer price, after an IPO that raised about $1.1 billion.

Revenue rose 147% to 1.74 billion yuan in H1 2026. Backers include ByteDance, Tencent, Sequoia China, and GGV, which matters twice: they are investors and potential customers with massive AI infrastructure needs. Bloomberg reported the listing plans as US export restrictions keep reshaping the semiconductor landscape.

Who gets hurt: the assumption that China cannot compete on AI silicon. Moore Threads started in gaming graphics, pivoted to AI training and inference chips as export controls tightened, and is now the clearest public-market vehicle for Chinese GPU self-sufficiency. Its Hong Kong listing would open the story to international capital.

NextSlide: OpenAI buys presentation muscle

OpenAI acquired NextSlide, the startup founded by former Caper AI co-founder Ahmed Beshry that turns prompts, notes, documents, and research into polished editable presentations. Terms were undisclosed; the whole team now works on ChatGPT features. The deal closed earlier in 2026 but was only disclosed August 8.

Why it matters: OpenAI keeps buying product surface area around ChatGPT rather than building everything in-house. Presentations are an enterprise wedge, and Beshry's history (Caper sold to Instacart for $350 million) suggests the playbook: small team, sharp product, acqui-hire into a bigger distribution engine. For the presentation incumbents, Canva and Pitch included, it is one more sign that AI-native slide generation is becoming table stakes.

The takeaway

The pattern across this week's moves is concentration. Capital is piling into the layer above the models (Harvey), the safety layer under them (Irregular), and the infrastructure beside them (Moore Threads). Startups that own a specialized workflow, a testing sandbox, or a compliance headache are printing valuations that foundation labs would envy five years ago. The next round of disruption will come from whoever makes the model supply chain safer, cheaper, or more accountable.

Deeper context from the archive on the companies and technologies in this story:

More on AI Startups

Browse every AI Startups story →