Parallel Works

The thing about high-performance computing is that everyone agrees it is important and almost nobody agrees on where it should live. For decades it lived in a room - a big, humming, refrigerated room owned by a university or a national lab, full of machines with names, running jobs submitted through a scheduler that a graduate student learned to placate through trial and superstition. Then the cloud showed up and said: you don't need the room. And it was right, sort of, except now you have three clouds, each with its own console, its own billing model, its own idea of what a "cluster" is, and a compliance officer asking whether any of this is allowed. This is progress. It does not feel like progress.

Parallel Works exists in the gap between those two sentences. Its product, ACTIVATE, is what the industry calls a "control plane," which is a nicely modest term for something fairly ambitious: a single interface that sits on top of your on-premises supercomputer and your AWS account and your Azure tenancy and your Google Cloud project and, increasingly, the newer "neocloud" GPU providers, and lets you provision, run, orchestrate and - critically - budget your compute as if it were one resource instead of five. You log into one place. You spin up a cluster. You don't especially care where it physically is. The point of the tool is that you shouldn't have to.

The company was spun out of Argonne National Laboratory in 2015, and its DNA is unusually academic even by infrastructure-startup standards. Its technical roots run back to Swift, an open-source parallel scripting language that co-founder Michael Wilde helped guide at the University of Chicago and Argonne around 2005-2006 - which is to say, back when "running a workflow across many machines" was a research problem rather than a product category. The founding cast includes Wilde, Michela Wilde, the engineer Mihael Hategan-Marandiuc, and a civil engineer named Matthew Shaxted, who is now the CEO and whose backstory is the good kind of unlikely.

Shaxted trained at Northwestern in civil engineering with a concentration in architecture, and then went to work at Skidmore, Owings & Merrill - the firm responsible for a meaningful chunk of the Chicago skyline - where he ran sustainability and daylight simulations for large buildings. The way you get pulled into HPC, it turns out, is that your simulations keep getting bigger, and the machines you run them on keep getting bigger, until one day you look up and you are standing inside Argonne National Laboratory wondering how you got there. He met Wilde. The company happened.

What ACTIVATE actually does, feature by feature, is less romantic and more useful: cluster provisioning, Kubernetes support, workflow orchestration, interactive applications, storage management, GPU utilization tuning, role-based access control, and - the part enterprise buyers get quietly emotional about - cost control and budget enforcement. That last one deserves a moment. GPUs are expensive, AI workloads are gluttonous, and the standard cloud experience is that you find out what you spent after you've spent it. ACTIVATE's pitch is that you can set a budget, enforce it, and watch the meter before the invoice, not during the argument about the invoice. Governing the spend before it happens is a boring feature that saves genuinely un-boring amounts of money.

Then there is the defense story, which is where a small company did something structurally large. In 2025, ACTIVATE's High Security Platform became the first hybrid multi-cloud computing solution to receive a U.S. Department of Defense IL5 Authority to Operate - an ATO, in the acronym-dense language of federal compliance. If you have never priced out what it costs, in time and money and paperwork and sheer institutional patience, to earn a DoD authorization at that level, the short version is: a lot, usually undertaken by contractors with more lawyers than Parallel Works has employees. That a 22-person shop got there is the kind of detail that tells you the company decided, early, that compliance was not a cost center to be minimized but a moat to be dug.

The authorization was not theoretical. The DoD's High Performance Computing Modernization Program awarded Parallel Works a contract to give defense scientists, engineers and acquisition professionals a single interface spanning on-prem supercomputers and secure commercial cloud. The Naval Research Laboratory has already put the platform to work running weather-forecasting workloads - which is a nice, legible example, because weather is one of those problems that is simultaneously a science project, a national-security concern, and an enormous consumer of compute. And it is not only the military: NOAA uses Parallel Works to manage cloud HPC across AWS, GCP and Azure through its research portal, which means some meaningful fraction of American environmental modeling is being orchestrated through this company's software.

The strategic read here is the one Matt Levine readers will recognize as the sensible-boring bet: during a gold rush, the reliable money is frequently in the shovels, the maps, and the toll roads rather than the gold. The AI and HPC world is loud with companies training enormous models and buying enormous quantities of GPUs. Parallel Works is not, particularly, in that business. It is in the business of the layer underneath - the connective tissue that decides where a job runs, what it costs, who is allowed to touch it, and whether it complies. The compute market is fracturing into hyperscalers and neoclouds and on-prem holdouts, and Parallel Works has bet that the fracture itself is the opportunity. When the landscape splinters, sell the thing that makes the splinters behave like a whole.

None of this is to oversell a seed-stage company. Parallel Works has raised comparatively little - public records point to modest early funding rather than the nine-figure rounds thrown around in AI infrastructure - and it remains small. But small and consequential are not opposites, and the roster of institutions running real workloads through the platform suggests the product does the unglamorous thing well: it works, quietly, across systems that were never designed to cooperate. Argonne and Parallel Works jointly won the Federal Laboratory Consortium's top technology-transfer honor, which is the research world's way of certifying that a lab idea successfully became a real product - a transition that fails far more often than it succeeds.

What can you actually do with it? If you are a climate scientist, you can run a model across three clouds without learning three consoles. If you are a genomics lab, you can burst onto GPUs when you need them and turn them off when you don't. If you are building a digital twin, or analyzing a hurricane, or simulating space weather, you can treat compute as a portable, governable resource rather than a series of separate procurement headaches. And if you are a defense researcher, you can do all of that inside a security envelope that has already cleared the DoD's bar. The company's own framing, from Shaxted, is admirably free of hype for a category this hype-prone.

It is worth dwelling on why "workload mobility" - the ability to move a job from one environment to another without rewriting it - is harder than it sounds. Each cloud speaks a slightly different dialect. An on-prem cluster expects a scheduler; a hyperscaler expects an API; a Kubernetes deployment expects containers and manifests. Historically, bridging those meant employing a small priesthood of engineers who knew all the dialects and could translate between them, which is expensive and fragile and does not scale. A control plane like ACTIVATE is essentially an attempt to encode that priesthood into software, so that the translation happens automatically and the scientist submitting the job never has to become a cloud-infrastructure specialist as a precondition of doing science. That is the whole game: keep the domain expert in their domain.

There is also a governance argument that has grown sharper as GPUs became a strategic resource rather than a line item. When compute is scarce and expensive, someone has to decide who gets it, for how long, and at what cost - and in most organizations that decision is made badly, through a mixture of email, seniority, and whoever shouts loudest in the standup. Role-based access control and automated budget enforcement are, underneath the dry naming, an attempt to make that allocation legible and fair: a research director can see who is spending what, set limits that actually bind, and reallocate capacity without a spreadsheet archaeology project. For a large institution running mixed workloads across secure and commercial environments, that visibility is arguably as valuable as the raw performance.