5 AI Startups Everyone Is Talking About
From autonomous agents reshaping the workplace to biotech companies that are rewriting how doctors diagnose disease, these five companies are not just building software. They are building the next decade.
There is a strange, electric feeling in the air around artificial intelligence right now. It is the kind of tension that practitioners in Silicon Valley, London's Tech City, and Singapore's one-north district all seem to recognize, even if they cannot quite name it. Funding rounds that would have seemed absurd three years ago are now treated as routine. Founders who barely finished their second year of university are appearing on the covers of business magazines. And somewhere beneath all the hype, something genuinely important is happening. Real companies, solving real problems, are beginning to prove that the promises of AI are not just a pitch-deck fantasy.
To separate the signal from the noise, this publication spent the past several weeks speaking to investors, researchers, enterprise buyers, and founders across three continents. What emerged was a clear picture of five artificial intelligence startups that are attracting serious attention for serious reasons. These are not companies chasing trends. They are, in many cases, the companies setting them. Here is what you need to know about each one.
Cognition AI
When Cognition AI unveiled Devin, its AI software engineer, the reaction from the developer community was somewhere between disbelief and mild existential panic. Devin is not a code completion tool or a glorified chatbot. It is a fully autonomous agent capable of reading a brief, spinning up its own development environment, writing code, running tests, finding its own bugs, and deploying a finished product. Backed by a formidable roster of investors including Peter Thiel's Founders Fund and with a valuation that crossed the billion-dollar threshold faster than almost any AI company before it, Cognition has become the reference point for what "AI agents" actually means in practice. The company's founders, several of whom came out of competitive programming at the international level, believe that software development is only the starting point. Their longer-term thesis is that autonomous AI agents will take on any task that currently requires a human sitting at a computer. That is either a thrilling prospect or an unsettling one, depending on who you ask.
The debate around autonomous agents has intensified precisely because companies like Cognition are moving so quickly. Critics warn that the benchmarks used to showcase Devin's abilities have been disputed and that real-world performance remains patchy in complex codebases. Supporters counter that even a version that works 40 percent of the time dramatically changes the economics of software development. Both sides agree on one thing: the category itself is not going away.
Perplexity AI
Google has ruled the internet's information layer for so long that it can be easy to forget that search is, at its core, a deeply unsatisfying experience. You type a question. You get a list of links. You click, you skim, you click again, you lose twenty minutes and often still feel uncertain. Perplexity AI set out to fix exactly that. Its AI-powered answer engine reads across the live web and synthesizes a direct, cited, conversational response to whatever you ask. The product has attracted tens of millions of users who describe it simply as "the search engine that actually works." Led by former OpenAI researcher Aravind Srinivas, the company has raised hundreds of millions of dollars and forged distribution partnerships with device manufacturers and telecom providers. The real fight, of course, is with Google, which has now deployed its own AI Overviews in search results. But Perplexity has something Google often lacks in its answer product: speed, clarity, and trust from a user base that has grown genuinely attached to the experience.
"The companies rewriting the rules are not the ones asking for permission. They are the ones shipping products while everyone else is still writing memos."
Cohere
While much of the AI conversation is dominated by consumer-facing chatbots and science fiction-tinged visions of general intelligence, Cohere has quietly built one of the most commercially serious AI companies on the planet. Founded by former Google Brain researchers, including Aidan Gomez, who was a co-author on the original "Attention Is All You Need" paper that launched the transformer era, Cohere focuses entirely on enterprise language models. Its platform allows large organizations to deploy AI within their own secure environments, meaning sensitive corporate data never needs to touch an external server. For banks, hospitals, law firms, and government agencies, that is not a nice-to-have. It is a hard requirement. Cohere has structured its entire business around that reality. With a valuation now exceeding five billion dollars and revenue growth that has impressed even skeptical analysts, it is proof that the B2B AI market is not a consolation prize for companies that could not build a consumer product. It may, in fact, be the more durable business.
What makes Cohere worth watching beyond the financials is the philosophy at its center. The company has consistently argued that most enterprises do not need frontier general intelligence. They need highly capable, reliable, controllable models that do specific jobs well and can be audited when something goes wrong. That is a different product vision than OpenAI or Google DeepMind, and in regulated industries, it is landing extremely well.
Imbue
Ask most AI researchers what they think is missing from current language models and a significant portion will say the same thing: genuine, reliable reasoning. Today's AI systems are extraordinary at pattern matching, language generation, and retrieval. They are far less reliable when a task requires multi-step logical deduction, causal inference, or the kind of careful thinking that a chess player or a scientist would recognize. Imbue, formerly known as Generally Intelligent, has made reasoning its singular obsession. The San Francisco-based research lab, funded at over two hundred million dollars by investors including Astera Institute, believes that until AI systems can actually reason, everything else is a parlor trick. Their approach involves training models in ways that prioritize coherent thought chains over surface-level fluency. The work is slower, harder, and less immediately flashy than releasing a new chatbot. But the researchers there are among the most cited in the field, and the long bets in AI have historically gone to the people who were right about something nobody else was ready to take seriously yet.
Nabla
Of all the sectors that AI has the potential to improve, healthcare stands apart, both in the scale of its opportunity and in the severity of the consequences when things go wrong. Nabla has chosen to enter this space with a specific, practically urgent product: an AI copilot for physicians that listens to patient consultations, understands the clinical context, and automatically generates structured clinical documentation. Anyone who has spent time in a doctor's office in recent years will know that the physician often seems to be typing more than they are listening. Burnout rates among clinicians are at historic highs, and administrative burden is consistently cited as a leading cause. Nabla's product directly addresses that reality. Already deployed in major hospital systems across Europe and North America, the company has received extraordinary reviews from the physicians who use it. They report spending less time on paperwork and more time actually caring for patients. With the global healthcare AI market projected to reach well into the hundreds of billions of dollars over the coming decade, and with Nabla building trust in one of the most difficult regulatory environments in the world, this is a company that investors and policy-watchers alike are monitoring very closely.
What these five companies have in common is not just funding or media coverage. It is a quality that is rarer than either of those things: a genuinely clear-eyed understanding of what problem they are solving and why their approach to solving it is meaningfully different from what came before. That is the thing that tends to separate the AI companies that matter from the ones that will quietly disappear once the current wave of enthusiasm retreats. The wave always retreats. What remains after it does is what counts.
The broader lesson for anyone paying attention to the AI landscape right now is that the most important companies of the next decade will not necessarily be the ones with the largest models or the most talked-about launches. They will be the ones that found a real problem, built something that genuinely works, and stayed disciplined enough to grow into it. By that measure, all five of the companies profiled here deserve to be on your radar. Whether they live up to the attention they are getting is something only time, and real customers, will determine.
Your Questions, Answered
What makes an AI startup worth paying attention to in 2025?
The most credible AI startups in 2025 share a few characteristics: they have a specific problem they are solving rather than chasing a vague idea of "general intelligence," they have real customers paying real money, and they have a technical advantage that is difficult for a well-funded competitor to replicate overnight. Funding and press coverage matter less than those fundamentals, even though both tend to follow companies that get the fundamentals right.
Is Cognition AI's Devin actually as capable as the demos suggest?
The honest answer is: yes and no. Devin genuinely can complete end-to-end software engineering tasks that no AI tool was capable of handling even eighteen months ago. However, independent evaluations have found that the conditions under which it performs best are more constrained than the original demos implied. That is not unusual for early versions of transformative technology. The trajectory matters more than any single benchmark, and the trajectory is pointing steeply upward.
How is Perplexity AI different from just using ChatGPT with web browsing enabled?
The core difference is product design philosophy. Perplexity is built from the ground up around the question-and-answer use case, with a strong emphasis on citations, source transparency, and speed. ChatGPT's browsing capability is a feature added onto a conversational assistant. Users who spend time with both consistently describe Perplexity as feeling more like a research tool and less like a chatbot that happens to look things up. The distinction matters for anyone who uses search heavily as part of their professional workflow.
Why does Cohere focus only on enterprise clients instead of releasing a consumer product?
Cohere made a deliberate strategic choice early on that enterprise is a fundamentally more defensible and sustainable business than consumer AI. Consumer AI products are extremely expensive to run and tend to attract users who are sensitive to price. Enterprise clients, particularly in regulated industries, are willing to pay significant premiums for reliability, privacy, customization, and support. Cohere also recognized that many large organizations needed AI they could deploy within their own infrastructure, which is a capability that consumer-first companies rarely prioritize.
What is Imbue's theory of AI reasoning and why does it matter?
Imbue operates from the premise that the current generation of language models, however impressive their outputs look, are not actually reasoning in any deep sense. They are extremely sophisticated pattern matchers. Imbue's research aims to produce models that can follow multi-step logical chains reliably and explain why they reached a conclusion. This matters enormously in high-stakes domains like medicine, law, and scientific research, where "the model said so" is not an acceptable explanation and where errors in reasoning can have serious consequences.
Is it safe for hospitals to use AI tools like Nabla in clinical settings?
Safety is exactly the question that Nabla has had to answer before every major deployment, and the answer, based on its track record so far, appears to be yes, provided the tool is used appropriately. Nabla does not make clinical decisions. It assists with documentation, which is then reviewed and approved by the physician. That distinction matters both ethically and regulatorily. The company has been through the necessary approval processes in the European Union and works within HIPAA frameworks in the United States. That level of regulatory engagement is one of the reasons healthcare systems trust it.
Which of these five AI startups is most likely to become a household name?
If consumer reach is the measure, Perplexity AI has the clearest path to broad name recognition simply because it competes in the search space that billions of people use every day. If impact is the measure, Nabla arguably has the highest potential to touch people's lives in a direct and meaningful way, even if most patients never know the company's name. Cognition AI is the most likely to generate continued media coverage given how viscerally the idea of an AI software engineer resonates with both enthusiasts and skeptics. All five deserve a seat at that table.
0 Comments