Since the beginning of 2025, Microsoft for Startups team members have traveled to three continents and multiple countries (USA, UK, Barcelona, Eastern Europe, and India), served on more than 10 panels, and consumed more airport coffee than I care to admit. But it’s all been worth it, because I have the opportunity to share these insights with you. Okay, maybe not the coffee.
With Microsoft Build 2025 coming up, I’ve spent some time reflecting on the events I’ve already had the opportunity to attend in 2025—and the insights startups can garner from these industry gatherings. This year, the Microsoft for Startups team has been hopping from city to city—London to New York to Bengaluru—to take the opportunity to meet with and listen to early‑stage AI startup founders. And I mean truly listen.
Pro tip: As a startup founder, treat these face-to-face moments as golden opportunities for customer discovery. There’s nothing quite like an in-person chat to get unfiltered feedback on your product or idea. No survey or call can replicate the spontaneous insight from a hallway conversation.
After the keynotes and clapping, we actively sought out individual founders and corporate contacts, grabbing time in quiet corners to ask two simple and fluff‑free questions:
- What is actually working?
- What is still painfully broken?
Through their responses, we picked up on five consistent themes and challenges. And if you’re hearing similar issues or objections in your organization, we’ve also provided some options to help you resolve them.
Here are the themes we identified:
1. Beyond agents: A lot of AI problems remain unresolved
“AI agents” are dominating the headlines. Every major keynote and panel seem to talk about them. The startups leaning into agentic AI are getting attention—and often massive fundraising rounds. From a corporate perspective, however, enterprises still have several problems that they’re sharing with us. For example:
- Coding assistants are a top area for AI adoption, but enterprises are burdened with a glut of legacy code that needs to be refactored before it’s intelligible. Hint: Moderne helps with that.
- Enterprise leaders worry about proving return on investment (ROI). Hint: This is where Pay-i and Faros AI are helping teams track the ROI of their AI initiatives.
- Data teams still struggle with the basics: data privacy, security, and quality. They worry their data isn’t “AI-ready”—a reminder that garbage data leads to garbage AI. Hint: Techniques like Federated Learning help unlock the sensitive data for AI use, and companies like Integrate AI help with that.
For founders asking if they should pivot to agentic AI:
- Skip the hype cycle.
- Stay close to your customers and learn about the challenges they’re facing today.
- Listen for opportunities to solve real problems and capture value.
2. The increasing need for a cost control playbook (such as controlling compute costs)
Building AI products requires massive computing power—both to train models and to run them in production. Unlike a typical software startup, an AI startup might need clusters of GPUs. Shortages of that computing power have made capacity scarce and pricey. Many founders shared stories of how they start by prototyping with a powerful hosted model like GPT-4o, and before they know it, the API bill becomes their biggest expense. Ouch.
While GPU hours are cheaper than two years ago, it’s important to keep a pulse on the model releases. Once you’ve proven value and retention, explore downgrading to cheaper models (for example, drop from a GPT-4o to a Phi-4). Depending on your use case, some optimize their models for efficiency, even if it sacrifices a bit of accuracy—a slightly less complex model might cut cloud costs by five times.
Many AI teams are also exploring other ways to reduce costs such as prompt caching, batch processing, and other creative approaches. We’re also seeing creative hybrid strategies that involve orchestrating across multiple models to utilize cheaper models for most traffic and use premium models for only the tasks where quality truly matters.
Pro tip: For startup founders, the generative boom holds an unpleasant truth: compute expenses can outstrip revenue and threaten your startup’s survival if left unchecked. So make sure you start thinking through your cost-control playbook.
3. “Vibe coding” lowers the barrier on building software—but raises the bar on moats
Since Andrej Karpathy’s February tweet, investors can’t stop asking about “vibe coding”—the prompt‑driven, chat‑with‑your‑IDE (Integrated Development Environment) workflow that lets non‑technical founders ship production apps in days. This boom has led to thousands of AI startups springing up, many with overlapping ideas. And that brings up its own challenge: how do you stand out?
If the technical barriers to building AI applications are falling, it’s more important than ever to focus on building a defensible company. Differentiation is more critical to your startup’s survival than ever. And to differentiate in a crowded market, you need to be thinking long and hard about your “moat.”
Need some inspiration? Here are a few examples:
- Proprietary data is one way to safeguard your position and ensure defensibility.
- Superior UX and workflow integration is another—an AI tool embedded seamlessly in a customer’s existing software might win over a slightly smarter tool that requires them to change platforms Hint: this is why startups are also building Copilot extensions.
- Network effects or community, if designed well into the product, also help since AI apps can get smarter with more users or data. Or many open-source projects that built a strong community grew to become bigger startups.
- Speed to market and brand—becoming the known solution—in a niche before others catch up. For instance, OpenAI’s first-mover advantage with ChatGPT gave it brand-name recognition, which many smaller competitors now envy.
The big takeaway here is embrace the new “vibe coding” world but plan for defensibility. Leverage AI doors to move fast—but don’t skip the strategy. Ask yourself at every step: “If it’s this easy for me, couldn’t someone else do the same thing? How will I stay ahead?”
4. The AI talent wars
AI job postings grew 38% between 2020 and 2024 and are still climbing.¹ But every founder we met said the same thing: “I can find prompt engineers but I cannot find skilled AI engineers.”
Behind every AI product is human talent—data scientists, machine learning engineers, research scientists, and the like. Hiring these specialists is a major challenge, especially for early-stage startups. There’s fierce competition for anyone with machine learning (ML) and AI skills, as tech giants and well-funded later-stage startups can offer hefty salaries and stock packages. Moreover, new sub-domains like AI safety, LLMOps (large language model operations) and MLOps (machine learning operations), and data engineering for AI require experience that few people have.
Founders shared stories about how long it took to fill key positions. In some cases, critical roles stay open for more than six months because the talent pool is so limited. The strategies we discussed included:
- Looking worldwide (and remote). Expand your search beyond the usual hubs. Look in locations with strong talent pipelines like Eastern Europe, India, and China to tap into AI graduates there.
- Hiring domain experts and upskilling them in AI. For instance, one startup hired physicists and trained them on prompt engineering skills.
- Fostering an attractive mission and learning culture to lure AI practitioners who might otherwise join a big tech.
- Pitching a compelling product vision to potential hires.
- Compensating for talent gaps through partnerships or advisory networks (for example, academic collaborators, incubators that provide mentorship in AI).
5. Face‑to‑face still wins in an LLM world
Yes, we all live on Microsoft Teams meetings, Viva Connect communities, Loop pages, and email threads, but every pivotal insight in this post originated from a hallway track conversation: seeing a founder’s whiteboard or back-of-the-napkin sketch, overhearing a complaint about data lineage, sharing cloud bill horror stories over dinner, discussing the differences between Model Context Protocol (MCP) and Agent2Agent (A2A). The list goes on and on, but they all came through in-person conversations.
That’s why our next big stop is Microsoft Build (May 19 through 22, 2025, in Seattle, WA)—the perfect chance to pressure‑test these ideas with more than 5,000 builders, swap notes on Copilot roadmaps, and maybe settle the “vibe‑coding vesus real‑coding” debate once and for all.
I’m looking forward to engaging in even more of these conversations—and testing these insights—at Microsoft Build 2025, our annual developer conference where developers, founders, and tech enthusiasts gather to learn about the latest innovations and updates from Microsoft. I look forward to asking those same questions of enterprise customers and founders in attendance.
And I also look forward to catching up with our Pegasus startups in attendance including Bria, Coactive, D-iD, Dataloop, Factory AI, Faros AI, Galileo AI, Howso, Kubiya AI, Moderne AI, Nimble, and Qodo.
We’ll be on the expo floor trading stories, showcasing some of our top AI startup solutions that will revolutionize the way you build AI applications, and hunting for the next hallway inflection. If any of the lessons above resonated with you, find Microsoft for Startups while you’re there and tell us more about what you’re building.
¹Growth in AI Job Postings Over Time: 2025 Statistics and Data, January 5, 2025. Michael Bernzweig.
