Rounak Adhikary will tell you, if you ask, that the desktop metaphor is dying. Not gradually—it's already dead, he'd argue, we just haven't noticed yet.
His startup, ProjectX, is a six-person outfit that emerged from Y Combinator with a product called Infinity. The pitch? It's billed as the first operating system purpose-built for a world where humans and AI agents work together, not just side by side but intertwined. Every application runs in its own isolated environment with dedicated GPU resources. All of it accessible through a single URL, no downloads required.
Audacious doesn't quite capture it. Maybe audacious undersells it.
But if you squint at where AI is headed—agents that book your travel, negotiate contracts, manage workflows—the question of where they'll actually run starts to matter. And ProjectX is betting everything on having an answer.
Containers, Not Virtual Machines
The technical architecture diverges from typical virtual desktop infrastructure in ways that matter, at least on paper. Rather than spinning up full virtual machines for each user, Infinity isolates applications in what the company describes as containerized microservices. Each gets its own compute and GPU allocation, then streams to users via WebRTC.
The real bet centers on GPU partitioning—specifically NVIDIA's Multi-Instance GPU technology, which theoretically lets the system slice GPU resources on demand instead of locking entire GPUs to single users. ProjectX claims sub-three-second cold starts, though no independent verification or technical benchmarks are available. They also say you can run Windows and Linux applications simultaneously in the same session, which would be notable if it holds up.
Whether it does remains an open question. The company hasn't published technical benchmarks or let independent testers put the system through its paces under real load. That's not unusual for an early-stage startup, though it does leave the claims in the "trust us" category for now.
A Long Road From India to San Francisco
The founding trio—Adhikary as CEO, Bishal Karmakar as CTO, and Sourya Majumder as COO—started ProjectX back in 2022. That timing matters: they were building this before ChatGPT made "AI agents" a dinner party topic. By mid-2024, they'd announced a public beta in India, though adoption figures from that launch remain murky.
The company's acceptance into Y Combinator brought a relocation to San Francisco. Their YC directory profile describes a team now grown to six people, working on "the first cloud-native OS built for a multi-agent, GPU-native future." The phrasing has the unmistakable polish of pitch-deck language.
Funding details are somewhat opaque. The company secured $200,000 in Google Cloud credits in early 2024—over two years ago now—and received grants from the Indian government around the same period. A LinkedIn post mentions backing from Pontaq, though that hasn't turned up in press releases or standard funding databases. For a company trying to reimagine computing infrastructure, the capital story seems notably quiet.
The Multi-Agent Question
Here's where it gets interesting, and where the gaps become harder to ignore. The "multi-agent operating system" positioning appears throughout ProjectX's materials. It's in the YC directory. It's on the website. It's clearly central to how the company sees itself.
What's missing is the technical substance behind it. ProjectX hasn't published a whitepaper explaining how Infinity handles multi-agent orchestration at the OS level. How do agents coordinate? What security model governs their interactions? How does the system manage conflicting agent requests for the same resources?
The company isn't working in a vacuum, exactly. Academic researchers have explored multi-agent OS architectures, though one notable paper (KAOS) was withdrawn from arXiv in 2024 for reasons unclear. A more recent preprint called Qualixar OS surfaced in April 2026, claiming similar territory. But none of this provides clarity on what ProjectX has actually built versus what it's marketing.
Perhaps "multi-agent" means something simpler here—multiple applications with GPU access running simultaneously. Which would be useful, certainly. But it wouldn't quite be the paradigm shift the positioning suggests.
Crowded Territory
The market ProjectX is entering isn't empty. AWS AppStream 2.0 already streams Windows applications through HTML5 browsers with GPU-backed instances. Microsoft's Azure Virtual Desktop does this. So does Citrix's Virtual Apps and Desktops. A newer player, Neverinstall, offers browser-based access to Linux applications.
ProjectX's differentiation seems to rest on three things: the per-app isolation model instead of full VMs, those claimed sub-three-second cold starts, and the multi-agent narrative. A GITEX exhibitor listing for the company mentions over ₹20 crore (roughly $2.4 million) in signed letters of intent and a community exceeding 30,000 users. Neither figure has been corroborated by third-party sources, and letters of intent have a way of not converting to revenue.
The product is live now at infinity.projectx.cloud, with signup flows for both individuals and enterprises. Pricing? Not publicly disclosed.
The Steep Climb Ahead
Six people trying to reimagine computing infrastructure—it sounds almost quaint until you remember that's how many people Larry Page and Sergey Brin started with. Or the WhatsApp founding team. Scale isn't everything at the beginning.
But building an operating system is different from building an app. The coordination required, the security considerations, the performance optimization across wildly different use cases. It's a grind that has broken better-funded teams.
Still, there's something to the central premise. AI agents will need somewhere to run. They'll need environments that can spin up quickly, allocate resources efficiently, and maybe—this is the interesting part—coordinate with each other in ways current operating systems weren't designed for.
Whether ProjectX has cracked that problem, or just identified it, we'll know soon enough. The computing industry has a way of separating ambitious marketing from actual innovation fairly quickly. Usually within eighteen months, sometimes less.
For now, Adhikary and his team have a product in market, a Y Combinator pedigree, and a thesis about where computing is headed. What they don't have yet is proof. But then again, proof is expensive. Maybe impossibly so for a six-person startup taking on the operating system itself.
The product is out there, accessible through a browser. That's something. Whether it's enough is the question they'll be answering throughout 2026 and beyond.
