20 AI startups to watch in H2 2026 — backed by real data, not hype
H1 2026 brought in $510 billion in startup funding, and four companies took 65 percent of it. The startups worth watching in H2 aren't the ones landing mega-rounds. They're the ones converting real revenue, real deployments, and structural advantages into durable positions.
Most "startups to watch" lists are built from press releases. A company announces a funding round, the round gets covered, and the name lands on twenty watchlists whose authors never looked past the valuation to ask whether a business sits underneath it.
This list starts from different inputs. Every pick comes with at least one concrete data point: a disclosed ARR figure, a published user count, a documented enterprise deployment, or a measurable market advantage. Where a company is earlier-stage and the numbers are thinner, I say so openly and lay out the structural argument instead.
The picks fall into five categories that I think produce lasting AI businesses: foundation model challengers, vertical AI replacing legacy software, developer infrastructure at scale, physical AI moving from demo to deployment, and creative AI finding enterprise revenue. OpenAI, Google, Meta, and Microsoft are left out, because at their scale "to watch" is an odd way to describe them. So is any company whose only public number is a valuation with no revenue behind it.
All figures reflect publicly disclosed data as of July 2026.
Category 1: foundation model challengers
The thesis: OpenAI and Anthropic together took 43 percent of all startup funding in H1 2026. That concentration opens structural room for alternatives that position themselves differently: open-weight models, European data sovereignty compliance, or deep expertise in a specific domain.
1. Mistral AI
Based in Paris, with $1.1 billion raised at a $6 billion valuation. Its open-weight strategy won developer adoption faster than any European AI company before it, and the EU AI Act is turning into a structural tailwind instead of a regulatory burden. European enterprises that need data sovereignty have a short list of credible options, and Mistral's compliance position is the strongest on it. Its coding model, Codestral, has found a niche among developers who want an open alternative to proprietary coding tools. H2 2026 catalyst: As EU enterprises shift from AI pilots to production, Mistral's regulatory edge stops being a marketing story and becomes a procurement requirement. Risk: It has limited penetration in the US market, and its compute access trails the US-based frontier labs.
2. Cohere
Based in Toronto, with $1.6 billion raised, a $6.8 billion valuation, and an H2 2026 IPO target. Cohere is enterprise-only by design, offering private deployment, data residency guarantees, and fine-tuning on proprietary data without exposing it to a shared cloud. That focus has concentrated its customers in financial services, healthcare, and government, where data control is non-negotiable and contract cycles run long. If conditions hold, an H2 IPO would make Cohere one of the first foundation model companies to go public, and the clearest read yet on what enterprise AI multiples look like in public markets. Risk: If the IPO window closes or public appetite for AI cools, Cohere has to grow into its private valuation on private capital alone.
3. Thinking Machines Lab
Founded in 2025, with $2 billion raised and led by Mira Murati, formerly OpenAI's CTO. It's the most closely watched new AI lab of the year for a simple reason: Murati ran OpenAI's technical and product organization through its most consequential stretch of growth. Thinking Machines is pursuing frontier research and applied AI products together, a broader mandate than most new labs try. A $2 billion seed round says as much about the team as the product, and the investor base reflects it: this is frontier-pedigree capital. There's no disclosed revenue or product yet, so the whole thesis rests on execution and team quality. Risk: It can't be evaluated as a business at this point. Watch H2 product releases for the first signs of direction.
4. Safe Superintelligence (SSI)
Founded in 2024, with $3 billion raised. Co-founder Ilya Sutskever set SSI a single stated goal: building AI that stays reliably safe at AGI scale, free from the commercial pressure to ship products every quarter. Depending on your priors, that mandate is either the most important research agenda in AI or a very expensive philosophical exercise. The H2 signal to watch is whether SSI publishes technical research showing real progress on the alignment problems it has set itself, instead of staying opaque. Capital isn't the constraint here. Credibility is what gets built or lost in H2 2026. Risk: A company that neither ships products nor publishes research can't be evaluated. Watch for output.
Category 2: vertical AI replacing legacy software
The thesis: Both Forbes AI 50 and CB Insights data show vertically focused AI startups growing revenue faster than generalists. Harvey in legal and Ambience in healthcare are the most-cited examples, but the pattern runs through every major enterprise software category.
5. Harvey AI
$600 million raised across two rounds in H1 2025, at a valuation in the multiple billions. Harvey handles legal research, contract review, and case analysis for top firms including Allen and Overy, PwC, and A&O Shearman. Legal software has resisted disruption for twenty years, because accuracy demands are high and the cost of errors is severe. Harvey is winning because it meets both constraints: it outperforms generic LLMs on legal tasks, and it supplies the audit trail firms need for liability reasons. H2 2026 catalyst: Contract renewals and expansion inside existing firm relationships, which tend to follow 12-24 month land-and-expand cycles in professional services. Risk: Larger legal software vendors are building AI features of their own, and incumbents with existing relationships are a genuine competitive threat.
6. Ambience Healthcare
$250 million Series D in January 2026, a $4.1 billion valuation, and more than 700,000 physicians using the product. Ambience automatically generates clinical documentation from physician-patient conversations, removing one of the most complained-about administrative burdens in medicine. The 700,000 figure is the metric that matters. It shows a penetration rate that validates both the product and the sales motion in one of the most compliance-sensitive industries anywhere. According to 2026 data, healthcare AI is the fastest-growing vertical by deal count, and Ambience leads that wave with documented scale. Risk: EHR integration complexity is a continuing cost, and the major EHR vendors (Epic, Oracle Health) are building competing AI documentation features.
7. Sierra AI
Founded in 2023, valued at $4.5 billion, and led by Bret Taylor (former Salesforce co-CEO and Twitter board chair) and Clay Bavor (former Google VP). Sierra builds AI customer service agents that handle complex resolution workflows, not just FAQ lookups. The founding team is the first signal: Taylor and Bavor have each worked where enterprise software meets consumer-scale products, across two major companies apiece. Enterprise AI customer service is among the most immediate ROI use cases in the market, since it replaces or augments costly human agent operations, and Sierra enters the category with one of the strongest leadership lineups in AI startups. Risk: Customer service AI is crowded. Intercom, Zendesk, and Salesforce itself all have agent products running at scale.
8. Decagon
AI customer support agents with documented deployments at GitHub, Notion, and Rippling. Decagon grew from nothing to significant ARR without the large funding rounds its competitors raised, so its business model has been validated by customer revenue rather than investor capital. Its current customers, developer tools and growth-stage SaaS, belong to the early-adopter tier for AI support agents, and that tier tends to expand heavily once an initial deployment proves itself. H2 2026 catalyst: Moving beyond developer tools into broader enterprise accounts, which calls for a different sales motion but opens a far larger addressable market. Risk: Without the capital cushion of better-funded rivals, Decagon has to grow efficiently rather than at any cost.
Category 3: developer infrastructure at scale
The thesis: In 2026, the quickest route to durable AI revenue is through tools that developers adopt first and then push their organizations to mandate. The coding tool category is producing some of the most dependable ARR figures in AI.
9. Cursor (Anysphere)
$3.3 billion raised, a $29.3 billion valuation, estimated ARR of $500 million to $1 billion, and use at 64 percent of Fortune 500 companies. By the time it reaches $1 billion ARR, Cursor may be the fastest-growing software company in history. That isn't hype. It's the number multiple analysts derive from disclosed growth trajectories. Building as a VS Code fork gave Cursor native multi-file context and an agent mode (Composer) that competing extensions couldn't replicate by design. H2 2026 catalyst: Growth from individual developer adoption to team-wide and company-wide licensing, the natural next stage once developer penetration inside an organization passes a threshold. Risk: GitHub Copilot, Windsurf, and Claude Code are all credible at the same price point. How well Cursor retains customers at scale is the question H2 will begin to answer.
10. Lovable
Based in Stockholm, with $400 million ARR. Lovable lets anyone build and deploy full-stack web applications from a browser tab by describing what they want in plain language. The $400 million ARR stands out: it makes Lovable one of the highest-revenue AI application companies that many enterprise observers haven't yet registered. Its addressable market is everyone with a software idea and no engineering resources, a population far larger than the developer tool market. H2 2026 catalyst: Expanding from individual builders and early-stage founders to mid-market teams that want to prototype quickly. Risk: Complex applications still need engineering oversight. Product-market fit is strong at the prototype layer, while durability in production deployment is less proven.
11. Scale AI
$13.8 billion valuation and more than $1 billion in ARR. Scale supplies the data labeling, evaluation, and fine-tuning infrastructure that AI models rely on for quality. The business now extends beyond labeling to government contracts for AI in national security, giving it a revenue base most AI application companies can't reach. Alexandr Wang is 28 years old and runs what is effectively a critical piece of AI infrastructure for both commercial and government development. H2 2026 catalyst: Government AI adoption is accelerating, and Scale's existing relationships position it to expand as federal agencies deploy more AI. Risk: Synthetic data generation threatens the labeling business model at the lower end of the market.
12. ElevenLabs
$500 million ARR and fast enterprise adoption. ElevenLabs has gone from "that impressive voice demo" to production infrastructure for enterprise voice applications faster than most observers expected. Reaching $500 million ARR with a relatively small team points to exceptionally efficient monetization. Its expansion from text-to-speech into voice cloning, dubbing, and conversational AI widens the product surface and the range of enterprise use cases. H2 2026 catalyst: Voice-driven conversational AI agent deployments are growing, and ElevenLabs is well placed as the voice layer for those systems. Risk: Google, Microsoft, and Amazon all run voice infrastructure at scale. Being the independent premium alternative is a sustainable position until it isn't.
The companies in this category share a pattern: they built significant revenue before raising the rounds that made them famous. That sequence, revenue before reputation, is one of the clearest signs of product-market fit in the AI market right now.
Category 4: physical AI crossing from demo to deployment
The thesis: Between 2025 and 2026, autonomous vehicles, humanoid robotics, and robotic foundation models crossed from research demonstrations into commercial deployment at scale. The companies to watch in H2 are those whose unit economics are becoming visible.
13. Waymo
$16 billion raised in February 2026 at a $126 billion valuation, 15 million paid trips in 2025 (3x year over year), commercial operations in 6 US cities, and plans to expand to more than 20 new cities plus London in 2026. In 2025, Waymo completed more paid autonomous trips than every other commercial robotaxi program combined. Its 82-90 percent reduction in serious-injury crashes versus human drivers turns the safety case from theory into documented fact. Goldman Sachs projects the US robotaxi market will reach $48 billion by 2035. H2 2026 catalyst: The London expansion is the first major non-US market test, and the international regulatory playbook it creates matters for global scale. Risk: Tesla's robotaxi timeline remains unclear, but the competitive threat is real if execution follows the announcements.
14. Physical Intelligence (pi)
Commercial deployments at Amazon, Mercedes-Benz, and BMW, plus an order pipeline of more than 70,000 units. Physical Intelligence is building the robotic counterpart to foundation models: general-purpose robot learning systems that adapt to new tasks without task-specific programming. Commercial deployments at major manufacturers take it past the "impressive robot video" phase that defines most robotics startups. The Amazon context is especially meaningful, given the scale of warehouse operations and the efficiency pressure they create. H2 2026 catalyst: Conversion of the order pipeline into delivered units is the metric to watch. 70,000 units ordered is a signal, and 70,000 units shipped is proof. Risk: Manufacturing hardware at scale is a different discipline from AI research, so execution risk in physical deployment exceeds that in software.
15. Figure AI
$2.6 billion raised, a BMW partnership, and humanoid robots working in production at BMW manufacturing facilities. Figure competes with 1X, Agility Robotics, Boston Dynamics, and Tesla's Optimus for the labor-substitution opportunity in physical work settings. The BMW deployment sets it apart: that's a production line, not a demo. The global labor shortage is estimated as a $1.5 trillion addressable market, the kind of TAM that justifies large capital rounds, but only if the robots perform reliably in industrial environments. H2 2026 catalyst: Unit economics from the BMW deployment, and whether more automotive or logistics partners follow. Risk: Tesla's Optimus has manufacturing-scale advantages that no pure-play robotics startup can currently match.
16. Isomorphic Labs (formerly DeepMind spinout)
Ranked #2 on Seedtable's 2026 AI startup index, this Google-backed spinout focuses on AI-accelerated drug discovery. AlphaFold permanently changed protein structure prediction, and Isomorphic is building a commercial drug discovery platform on that foundation. Its 2024 partnership deals with Eli Lilly and Novartis gave disclosed validation of the approach at pharmaceutical scale. The drug discovery TAM is large and the regulatory moat deep enough that even an early position in the category carries substantial long-term value. H2 2026 catalyst: Pipeline results from the pharma partnerships are what investors are waiting to see. Risk: Drug discovery has a notoriously long timeline from discovery to revenue, and the business model demands patience that tech investors haven't typically shown.
Category 5: creative AI finding enterprise revenue
The thesis: Creative AI tools reached consumer scale first, yet the enterprise revenue opportunity in content production at scale is much larger. The H2 question is which creative AI companies make the leap from consumers to enterprises.
17. Runway ML
An AI video generation platform with a significant partnership pipeline, at Series C stage. Runway's video quality is the benchmark that competing tools are measured against in professional production settings. The enterprise opportunity lies in marketing content, film and television pipelines, and corporate video, each a far bigger market than the creator and consumer tier. H2 2026 catalyst: Enterprise partnership announcements and production studio integrations are the signals to watch. Risk: OpenAI's Sora, Google's Veo, and Kuaishou's Kling are all credible video generation rivals with larger infrastructure behind them.
18. HeyGen
The fastest-growing AI video startup by user count in its category, offering AI avatar video generation plus translation and dubbing. HeyGen's core capability is generating talking-head video from text and translating it into multiple languages, which solves the localization problem at a price that puts global video production within reach of small companies. Dubbing, where real video is translated and lip-synced into other languages with the speaker's voice characteristics preserved, is a clear enterprise workflow with ROI that content teams operating globally can document. H2 2026 catalyst: Expanding the enterprise tier with API access for content production pipelines. Risk: Open-source alternatives are commoditizing the avatar and dubbing categories quickly.
19. Perplexity AI
A $22.6 billion valuation, roughly $450 million ARR, 45 million users, and a February 2026 decision to drop advertising. Moving to a subscription-only model was a significant strategic bet: it gives up one revenue stream but aligns product quality more cleanly with the business model. Its ARR trajectory, from $80 million at the end of 2024 to $450 million in H1 2026, is among the fastest growth rates in AI applications. H2 2026 catalyst: The subscription-only model proves itself over the rest of 2026. If ARR keeps growing without ad revenue, the model is validated. If growth slows, the bet will look different in hindsight. Risk: Google's AI Mode and AI Overviews compete directly for the search-replacement use case, and 45 million users against Google's billions remains a structural disadvantage.
20. Suno
Consumer and enterprise music generation, based in Cambridge, and the most widely used AI music platform by active user count. Suno reached 10 million users in its first year, one of the fastest consumer adoption curves in generative AI. The commercial licensing question, whether AI music can be used commercially without settling IP claims, is both its main risk and its H2 watchpoint. If Suno reaches licensing agreements with the major labels or wins the pending legal cases, its commercial opportunity expands dramatically. If not, the consumer tier stays strong but enterprise deployment is constrained. H2 2026 catalyst: A legal resolution on copyright is the single most important event for AI music in the second half. Risk: Udio, Google MusicLM, and Meta's AudioCraft are all building in the same category, and first-mover advantage in generative AI hasn't proven durable in other content categories.
What separates H2 watchlist companies from the broader AI funding surge isn't the size of the round. It's at least one specific, falsifiable signal: a disclosed revenue figure, a documented deployment, a public user count, or a named enterprise customer. Watchlists built on valuation alone mistake momentum for evidence.
The selection framework, spelled out
Every company above was tested against the same set of questions, not a vibes-based read of the market.
Is there a specific data point? The company needed at least one of: a disclosed ARR or revenue run-rate, a named enterprise customer, a documented deployment at scale, or a verified user count. Companies whose only public figure is a valuation were excluded.
Is there a structural advantage? Something that makes the position defensible beyond "first to market" or "best team": EU regulatory positioning, proprietary deployment data, customer relationships in industries with high switching costs, or unique access to training data.
Is there a specific H2 2026 catalyst? A named event or decision point that makes H2 especially important for the company, such as an IPO filing, a legal resolution, a market expansion, a product launch, or an enterprise partnership announcement.
What is the honest risk? Not a generic "competition is increasing," but a named, specific risk that could undermine the position.
Three kinds of company were excluded: those where funding is the only signal, those building undifferentiated "AI wrappers" on foundation models with no clear distribution or data advantage, and those on watchlists mainly because they issued a press release.
Several companies in the $1-5 billion valuation tier also missed the cut because their latest public data was more than six months old. Fast-moving markets need recent evidence, and a strong 2024 metric says little about H2 2026.
H1 2026 was defined by concentration: 4 companies, $188 billion, 65 percent of global VC. H2 2026 will be defined by something harder to track in real time. Of the companies building on top of that concentration of infrastructure, which will turn it into durable revenue, and which will find their market position was weaker than their funding round suggested?
The 20 companies above are the ones where the data gives a reason to watch instead of a reason to wait.
Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.
I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com