After two years of record fundraising, the AI market could soon get more selective. 2025 ended with trillion-dollar compute deals, a glut of generative tools, and the first signs of fatigue among buyers faced with dozens of products promising the same thing.

Venture investors say cash in 2026 will go mainly to the systems that run AI and to tools that solve clear problems for businesses, rather than another general-purpose chatbot. Quartz asked the people deploying that money what they expect to back next year — and what ideas they’re now turning down.

Moving into reality

For the last few years, startups have tried to show their AI platforms work in theory, but 2026 is when they will have to prove it in the real world. “Companies helping enterprises actually take AI into production” will be the big winners next year, said Guru Chahal, partner at Lightspeed Venture Partners, which manages about $35 billion. “Enterprises are finally realizing AI’s potential but also that they need help driving success.”

That includes platforms that work across different workplaces — like those automating HR processes — and “vertical solutions that solve specific industry problems,” he told Quartz. Chahal pointed to Eve Legal, a Lightspeed-backed platform that tackles the mountains of paperwork faced by lawyers, as an example of the latter. He added: “The gap between proof-of-concept and production is where money flows now.”

Nowhere is that gap more pronounced than in so-called agentic AI, where products or “agents” make decisions and act with a degree of autonomy rather than responding to one prompt at a time. Mikael Johnsson, managing partner of Stockholm- and London-based VC investor Oxx, said agentic platforms will move “from experiments, pilots, and trials to driving real productivity gains” in 2026. “AI investments will be put under the same scrutiny as any other software related investment.”

Those agents could have their work cut out. In a 2025 test of Anthropic’s Claude platform acting as an agent, it was fed emails showing that a hypothetical company executive was having an affair and was also planning to shut down the AI system. Claude’s response was to blackmail the executive by threatening to tell his wife and bosses about the affair.

Nonetheless, Johnsson is hopeful that companies can break through these problems and bring agents from testing and into the workplace. “We’re looking to fund companies that can prove real and meaningful return on investment in production environments and where customers are doubling down on one vendor rather than running multiple ones in parallel,” he said.

The infrastructure race

Companies building the physical infrastructure for AI — including servers, chips, power-generation, and data centers — will remain a core battleground, after an explosion in spending in 2025. Goldman Sachs analysts forecast data center power demand could grow roughly 50% by 2027, forcing buildouts of tens of gigawatts of new electricity capacity just to keep pace. McKinsey estimates up to $6.7 trillion in global data-center spending by 2030 to satisfy AI’s compute needs.

Cecilia Ma, an investment manager at Stockholm-based VC Norrsken, said: “This past year was a learning journey for the whole industry. Deals in AI were being done in record time, investors and entrepreneurs are calibrating the strength and limitations of the tech along the way.”

In 2025, scale was the key factor, with OpenAI alone making more than $1.4 trillion of infrastructure commitments in recent months, including deals with Nvidia, Advanced Micro Devices, and Broadcom. Microsoft set aside about $80 billion for AI-enabled data center expansion, and Alphabet committed north of $75 billion in capex in 2025, largely tied to servers and data centers underpinning Gemini.

The next year will bring an “increasingly sophisticated view” of this area, added Ma, pointing to companies that can “[optimize] performance infrastructure in the long-tail” as the ones winning funding. She pointed to OptiCloud, which tackles digital waste by detecting where companies are paying for computing power they aren’t using and cutting it off.

Chahal, meanwhile, pointed to NextHop, a Lightspeed-backed firm producing the equipment that connects AI chips, enabling faster data transfers than standard data-center gear generally allows. Lightspeed was the lead investor in NextHop’s $110 million Series A round in 2025, with Chahal saying at the time that cloud providers “need a new generation of networking capabilities to keep pace with the demands of AI workloads."

Video gets cheaper, not always better

AI-generated video made its way into mainstream advertising in 2025, often with awkward results. A McDonald’s Christmas ad in the Netherlands was pulled after it branded the holiday season as the “most terrible time of the year,” and subjected viewers to a chaotic and often warped barrage of un-festive scenes. Coca-Cola was pilloried in November for its own serving of festive slop featuring uncanny polar bears — and... er, sloths? — romping around a snow-covered landscape. But investors said usage is rising regardless, with brands using AI tools for mood boards, test cuts, and short social clips, and they expect quality to improve fast.

Inaki Berenguer, managing partner of LifeX Ventures, expects 2026 to be the year quality catches up. “Video is next,” he said, citing the rapid improvement of tools like Runway, Krea and Wonder Studios, and drawing parallels with how Suno and ElevenLabs reshaped music and voice. “The cost of producing high-quality video is collapsing,” he added, which will open space for new companies in advertising, social content and filmmaking.

Berenguer also sees gaming as another area set for change. “Models keep getting better, and the cost of experimenting with new mechanics, assets, and worlds keeps dropping,” he said. “When creation becomes cheap, you get a Cambrian explosion of new ideas. I expect a wave of AI-native gaming startups in 2026.”

What’s not getting funded

Investors said the companies most likely to struggle in 2026 are those offering generic AI products without a clear reason to exist. Chahal noted that look-alike applications built on foundation models have fallen out of favor, particularly chatbots without a specific use-case or distribution advantage. “The market’s gotten disciplined,” he said. “You need real product moats or genuine go-to-market advantages, not just access to an API [application programming interface].”

Ma agreed that easy wins are drying up. The cost of building an AI tool has dropped, she said, which means speed alone is no longer a selling point. “The ones that will struggle to get funded are the ones competing just on speed to market now,” she said. “Development work has become so democratised I believe we’re putting the focus back on the customer and value add where it should be.”

📬 Sign up for the Daily Brief