Ilya Sutskever’s company has a website with six sentences on it. No product page, no pricing, no blog. Safe Superintelligence Inc. has roughly fifty employees, has published no papers, and has never demoed anything in public. It’s also worth $32 billion, and in July, Nvidia added another $5 billion to that pile anyway.
That fact does more to explain 2026 than any funding chart could. Somewhere in the last two years, “we have no product” stopped being a disqualifying admission for a certain class of AI startup and became the entire pitch. Sutskever said as much when he founded the company: its first product would be safe superintelligence itself, “and it will not do anything else up until then.”
When Meta reportedly tried to buy SSI in 2025 and got turned down, it settled for hiring away one of Sutskever’s own co-founders, Daniel Gross, instead. SSI kept going regardless. That’s either the most disciplined bet in the history of venture capital, or a very expensive way to avoid ever being tested against a product review — and from the outside, there’s genuinely no way to tell which yet.
That’s one half of 2026. The other half is almost the opposite story: a set of startups growing revenue faster than any software companies in recorded history, not on vibes but on contracts a CFO signed. Both things are true in the same twelve months, and figuring out which is which is basically the whole game this year.
The money is genuinely deranged, even by AI standards
Some context, because the scale here is easy to undersell. Global venture funding hit $510 billion in the first half of 2026 alone — more than the entirety of 2025 — and AI ate more than 70% of it. OpenAI and Anthropic pulled in $217 billion between them in six months, which works out to 43 cents of every venture dollar deployed anywhere in the world, in any industry, going to two companies.
Anthropic is now the most valuable private company on the planet (SpaceX stopped counting once it went public), sitting at $965 billion after a $65 billion round in May, with $47 billion in annualized revenue that’s pulled ahead of OpenAI’s. OpenAI is still bigger on raw mindshare — 900 million-plus weekly ChatGPT users — and filed confidentially for an IPO in June, reportedly targeting a $1 trillion valuation its CEO has called a floor, not a ceiling. Whichever way that gap closes is probably the biggest story in tech for the rest of the decade. But at this point it’s a story about two companies the size of oil majors — not really a startup story anymore. The more interesting stuff is happening one rung down.
The neolabs: enormous checks, almost nothing to point at
Sutskever isn’t the only one getting funded largely on reputation, just the most extreme case. Call it the “neolab” pattern: senior researchers who left one of the big labs, taking enough trust and enough of a Rolodex with them that investors treat the founder’s name as the product, at least for now.
Mira Murati is the more self-aware version of this. She left OpenAI as CTO in 2024, in the same messy stretch of boardroom drama that OpenAI staff reportedly still call “the blip” internally, and then went almost completely dark for eighteen months — no interviews, one shipped product, a solid-but-unglamorous fine-tuning API called Tinker. In June she finally resurfaced at a Bloomberg conference to describe what her company, Thinking Machines Lab, is actually building: “interaction models” meant to process conversation continuously, in 200-millisecond windows, instead of the stop-and-wait rhythm of a normal chatbot. It’s a genuinely interesting idea. It’s also, as of this writing, one open-weight model — Inkling, shipped in July, solid rather than best-in-class by Murati’s own description — propping up funding talks reportedly around a $50 billion valuation, five times where the company started sixteen months ago.
Fei-Fei Li has a stronger claim than almost anyone on this list to have earned the speculative treatment. She built ImageNet, the dataset that helped set off the deep learning boom in the first place. Her company, World Labs, is betting the next real gap in AI isn’t language but spatial reasoning — giving a model an actual, physically consistent sense of three-dimensional space, the kind you’d need before letting a robot loose in a warehouse. Her first product, Marble, generates navigable 3D environments and picked up Autodesk as a strategic investor, which is a meaningfully less abstract signal than most of this list gets, since Autodesk sells software to people who build physical things for a living. Reported valuation: around $5 billion, on $1.2 billion raised.
Whether any of this converts into a durable business is an open question that even the people writing the checks can’t answer with a straight face. Ask again in eighteen months.
The part of 2026 that’s actually boring, in a good way
Here’s the story that gets less attention, mostly because it doesn’t involve a company with a six-sentence website: a handful of startups are hitting $100 million in annual revenue faster than Salesforce, Slack, or Zoom ever did, and they’re doing it by selling something a CFO can point to and defend in a budget review.
Sierra, co-founded by former Salesforce co-CEO Bret Taylor, builds customer-service agents that don’t assist a human rep — they replace the interaction outright. It went from zero to $200 million in annual recurring revenue in 26 months, which sounds aggressive until you notice that Harvey, now used across most of the AmLaw 100 law firms, doubled its own revenue from $100 million to $190 million in five months flat. The same shape shows up across half a dozen adjacent categories: Abridge turning doctor-patient conversations directly into clinical notes, OpenEvidence doing roughly Perplexity’s job but for clinical literature doctors can act on, Decagon chasing Sierra’s own opportunity for mid-market companies that move faster than the Fortune 500, Mercor running engineering hires through AI screening on its way to $1.5 billion in gross run-rate after seventeen months. Different professions, identical bet: find one expensive task a human does today, do it faster and cheaper, sell it to whoever already owns that line item in the budget.
The coding tools are the most extreme version of the same pattern — and the one most worth a little skepticism. Cursor, built by four MIT grads with no prior startup experience under a company called Anysphere, went from nothing to $2 billion in ARR in under three years and is reportedly in talks around a $60 billion valuation. Lovable, a Swedish company that lets non-programmers build full apps by describing what they want in plain English, hit $500 million ARR about fourteen months after launch with a headcount of 146 — revenue per employee in the same range as Nvidia’s. Cognition is the case worth watching most closely: it’s valued at $26 billion largely on the strength of Devin, marketed as the first AI software engineer, capable of picking up a real GitHub issue, writing the fix, running the tests, and opening the pull request unsupervised. In Cognition’s own early benchmark, Devin resolved about 14% of real-world issues completely on its own. That’s a genuinely impressive number for an autonomous system, and also, plainly, a long way from what “first AI software engineer” implies if you’re reading the marketing copy instead of the fine print.
Underneath all of it sits the least glamorous, most reliably real layer: infrastructure. Databricks, the data platform most serious AI deployments eventually run through, is worth $134 billion on an actual $4.8 billion revenue run-rate growing faster than 55% a year. Cyera, which handles the specifically-2026 problem of governing what an autonomous AI agent is allowed to touch inside a company’s data, has tripled its revenue three years running. Neither company will trend on social media. Both are probably the safest bets on this entire list.
The one that won’t sort into either bucket
Not everything fits neatly into “real revenue” or “pure reputation bet,” and Perplexity is the best argument for why that split is too clean. Its valuation sits around $22–23 billion on somewhere between $450 and $500 million in annualized revenue — a multiple that would look aggressive for almost any other kind of software company, but is actually one of the more restrained ones on this list. And yet the number that should worry its investors most isn’t the valuation, it’s traffic: Perplexity’s global web traffic share slipped from roughly 2.0% in March to 1.3% by late May, even as revenue kept climbing. The company’s answer has been to stop competing purely as a search engine and push its Comet browser agent directly into Microsoft 365 — Word, Excel, PowerPoint, Outlook, Teams — betting that the win condition isn’t owning the search box anymore, it’s owning the workflow around it. Whether that’s a shrewd pivot or a tell that the original product hit a ceiling is exactly the kind of question you can’t answer yet. Which is what makes it worth watching more than most names on this list, not less.
The one nobody can actually describe
If you want the purest distillation of how strange 2026 got, it isn’t Sutskever’s company, it’s Jeff Bezos’s. Prometheus is an industrial AI startup that raised a $12 billion round this year at a roughly $41 billion valuation, pushing total funding past $18 billion, in under a year of existing. It hasn’t launched. Reporters covering it can say it has something to do with manufacturing and logistics and not much else, because the company has said almost nothing on the record. The valuation isn’t pricing a product. It’s pricing Bezos, and whatever his team has told a small number of investors in a closed room that the rest of us haven’t heard yet. It might be the most consequential company on this entire list. It might also just be a very well-funded rumor. Both are live possibilities right now, which is an uncomfortable place for $18 billion to sit.
Physical AI more broadly finally started pulling in serious institutional money this year, for the first time since the last robotics hype cycle fizzled out around 2015. Figure AI (humanoid robots, BMW as a customer) is valued at $48 billion on the bet that general-purpose robots become viable for factory work within two to five years. Physical Intelligence is trying to build something like a foundation model for robot manipulation — a system that transfers what it learns across tasks the way a language model transfers what it learns across topics — backed by Bezos and OpenAI at $11 billion. Waymo, well past the speculative stage and running commercial robotaxis in multiple US cities, is valued at $126 billion. The distance between “Waymo is obviously real” and “Prometheus is an expensive rumor” is basically a map of how mature this category actually is, company by company.
What I’d actually watch
If I had to guess which of these hold up and which get quietly written down in two years, I’d split it about like this: the vertical AI companies and the infrastructure layer underneath them are the closest thing on this list to boring, established-business economics, and boring is a compliment this year. They have customers who’d notice immediately if the product vanished tomorrow. The neolabs and Prometheus are venture capital doing what it has always done at its most extreme — betting on people instead of products — except at a valuation and a scale with no real precedent, even inside an industry that’s used to having no precedent for anything.
The actually interesting question for the rest of 2026 isn’t which company on this list is “the best.” It’s whether this turns out to be the year the market got genuinely good at pricing a founder’s reputation years ahead of any product — or the year it got embarrassingly, publicly wrong about a handful of very expensive bets, all at once, in front of everyone. Right now there’s real evidence for both stories happening in the same twelve months. That tension is the actual story, not any single company on it.