Only 12 of the 23 new startups attacking the 2022–24 AI wave publish a number that beats the incumbent they name — and the cheapest of them is priced 6,000× below its target

Across 19,307 result rows from 101 web queries, 56 startups founded in 2025 or 2026 name an established AI company they intend to displace. Twenty-three attack an AI startup from the 2022–2024 wave rather than a frontier lab, and only twelve publish a quantified speed, cost or accuracy claim against it. Four were founded by someone who built the incumbent’s product. The widest gap between evidence and price is Polar, built by a former Perplexity Comet engineer, which beat Comet-class agents 98.0 to 75.5 on a browser-agent benchmark and raised a $5.7M seed against Perplexity’s ~$20B.

The three highest-conviction bets

What you pay versus what they attack

What the data says

1. The insider defection is the strongest single signal, and it is rare. Four of the 23 challengers were founded by someone who built the thing they now attack: Polar (ex-Perplexity Comet), Sail Research (ex-Together AI), UMA Robots (Rémi Cadéne, who led Tesla Optimus and then Hugging Face’s LeRobot) and River AI (xAI co-founder Igor Babuschkin). Three of the four also carry a published head-to-head number. Two near-misses are worth the same attention: Gradium’s Neil Zeghidour built the audio models at Kyutai, Google Brain and DeepMind rather than at ElevenLabs, and Entire’s Thomas Dohmke is the former CEO of GitHub, now attacking GitHub.
2. The most exposed incumbents carry a valuation with no deployment — or a deployment that already churned. Figure AI holds $39B with zero external customers; Skild AI holds $14B with no publicly disclosed commercial deployment; Physical Intelligence holds $11B and discloses no revenue. At the other end, 11x reported roughly $14M ARR of which only about $3M survived, on reply rates 55% below human SDRs. Walden Robotics is in a live Toyota plant, Enigma is on 100+ real robots, and Monaco has ten named design partners — all at a fraction of the price.
3. Cost structure, not model quality, is where the 2022–2024 cohort is soft. Cursor sits at $60B and is unprofitable because compute consumes the whole of its $1B+ revenue; Perplexity trades near 100× ARR on ~$200M; Runway is the most expensive video generator in its category. So the attacks are priced, not featured: 1/6 of ElevenLabs (Fish Audio), sub-50ms against a 400ms norm (Reactor), 10× lower cost (Sail), 50× lower compute cost (OXMIQ), and OpenBuilder selling unlimited use on open weights at a flat subscription against Lovable’s $13.3B and ~$600M run-rate.
4. Two challengers attack the incumbent’s customer instead of the incumbent. Moritz built a law firm where AI does 80% of the work with 50 contracted co-counsel, and has run $2B of transactions for 100+ companies with a seven-person core team. It never has to win a bake-off against Harvey’s $11B distribution, because it competes with the firms that buy Harvey. Mike OSS did the same to the software: one former Latham & Watkins associate rebuilt the core of Harvey and Legora in 14 days, and the repository took 2,200 stars in 72 hours.
5. The largest 2026 rounds are the weakest disruption trades. Ineffable Intelligence ($1.1B seed at $5.1B), AMI Labs ($1.03B at $3.5B) and Unconventional AI ($475M at $4.5B) price a research thesis before a product. Unconventional AI is a two-month-old chip company with no production silicon and no revenue, priced at a 1.5× discount to Groq — strictly more risk for almost no discount. OXMIQ attacks the same incumbent by licensing GPU IP instead of taping out, for $60M against Groq’s $6.9B. Same thesis, 77× cheaper entry.
6. Watch who sells to whom. Trajectory, which retrains open models weekly on logged failures, already counts Harvey, Decagon and Rogo as customers. Engram, whose compressed KV-cache “Cartridges” cut token use up to 100×, counts Notion, Microsoft and Harvey. Callosum has Cerebras integrated and its CEO on the record. These are suppliers to the 2023 cohort rather than attackers of it — a lower-variance way to own the same thesis, and the first place a real architectural shift shows up in someone else’s revenue.

The shortlist

A blue rule at the left marks the five candidates added in this revision.

Priced as a thesis, not a disruption

• Ineffable Intelligence — $1.1B seed at $5.1B. David Silver’s pure-RL “superlearner” against every LLM lab. Sequoia and Lightspeed co-led with Nvidia, Google and the UK Sovereign AI Fund. No product, no benchmark, no customer.
• AMI Labs — $1.03B seed at $3.5B. Yann LeCun’s JEPA world models against transformers, backed by Nvidia, Bezos Expeditions, Toyota and Samsung. A research bet on an architecture, with the paper as the only artefact.
• Mirendil ($200M at $1B), Recursive Superintelligence (~$500M), Discovery Loop (raising $1B at ~$10B, Jeff Dean and Oriol Vinyals), Thinking Machines Lab ($2B at $12B) — all attack frontier labs, none attack the 2022–2024 startup wave, and all price the team ahead of the wedge.
• River AI ($1.1B at ~$5B) is the closest of this group to a real wedge: LoRA and RL fine-tuning on open weights, a full RL run in 15–20 minutes at 2–4× lower cost, billed per token. Igor Babuschkin left xAI to build it. The price already assumes it works.
• Also screened and set aside: Standard Template Labs ($49M seed, CRV and ICONIQ, a continuously updated digital twin of a company, versus Serval), Perceptic ($12M, ex-Palantir, connective tissue across drug discovery), NeoCognition ($40M, Yu Su; angels Lip-Bu Tan, Ion Stoica, Dawn Song) and Poetiq ($45.8M, ex-DeepMind, topped Gemini 3 Deep Think on ARC-AGI-2 at half the cost). Strong teams, but each names an incumbent outside the 2022–24 startup wave.

Method: 101 web queries across eleven sweeps returned 19,307 unique result rows, read on 21–22 August 2026. Founding years, valuations, funding, benchmarks, customers and risks were extracted per page, then merged on a normalised company key. “Quantified head-to-head” counts only a numeric speed, cost or accuracy claim stated against the named incumbent or its direct category norm; a build-time or turnaround claim does not count, and Kilo Code’s figure counts against it. Every such figure is the challenger’s own claim unless a third party is named, and none were independently reproduced here. Incumbent valuations are the latest reported private marks and move fast; Lovable is shown at its August 2026 mark of $13.3B, up from $6.6B earlier in the year. Tier A means an insider or category-expert founder, a shipping product, a named customer and a published number; B means two or three of those; C means one. Entry price is the disclosed post-money valuation where one exists, otherwise the round size — the two are not comparable, and the chart marks which is which. Demo links are as reported by the sources and were not opened.