The Compute Crisis Is Coming for Orbit. The Answer Isn't More Satellites - It's Modularity.

PERSPECTIVE · ORBITAL INFRASTRUCTURE

By Merry Walker, Co-Founder & CEO, Symphony Space, Inc.

May 27th, 2026

In a single week this spring, the orbital data center race stopped being a thought experiment and started being a supply chain. The question for the rest of us is no longer whether compute is going to space. It's what kind of infrastructure will be waiting for it when it gets there.

Over the past several months, the most powerful companies on Earth have made the same bet in public. Starcloud - which put the first NVIDIA H100 in orbit and trained the first AI model in space - just signed a deal with SpaceX's Starlink to wire its satellites together with laser crosslinks, the connective tissue that turns a scatter of spacecraft into a functioning distributed data center. Google's Project Suncatcher is in talks with SpaceX for launch, with Planet building the prototype satellites, aiming to fly TPUs in orbit and eventually fan out to a cluster of dozens. SpaceX, having absorbed xAI, is now selling orbital compute as a core part of its growth story to investors. Anthropic has signaled interest in space-based capacity measured in gigawatts.

The orbital data center market has grown from concept to a $1.44 billion industry in 2026, projected to reach $3.81 billion by 2034 at a compound annual growth rate of 12.96%. Alternative projections tracking more aggressive constellation buildouts estimate the market reaching $39.09 billion by 2035 at a 67.4% CAGR. At least eight companies now actively pursue orbital data center infrastructure, with three already operating hardware in orbit. SpaceX and xAI merged into a $1.25 trillion entity and filed for one million data center satellites, while Starcloud filed separately for 88,000.

Each of these is a serious commitment from a serious organization. Together they tell us something important: the people closest to the compute problem have concluded that the grid, the water, the land, and the permitting on Earth will not scale fast enough for what AI is about to demand. Orbit offers near-continuous solar power, no cooling water, and no zoning board. The logic is sound. But it is not yet settled. The skeptics - and there are credible ones - are right that the hard physics is still being argued: thermal rejection in a vacuum, radiation durability over years, and high-bandwidth data movement all get harder, not easier, as you scale. Google's own Suncatcher research names these as open problems. Small-scale feasibility has been shown. Scale economics have not. The honest framing is the most exciting one: the opportunity is enormous, and the physics is still open.

The open question is whether the industry is answering that uncertainty with a new kind of architecture, or simply scaling up the old one.

The trap of the purpose-built satellite

Almost every approach in the market today shares one assumption: that an orbital data center is a thing you build, launch, and operate as a fixed unit. You design a satellite around a specific generation of chips, you launch it, and that configuration is what you have for the life of the spacecraft. If the workload changes, if the chip generation changes, if the customer changes - you launch another one.

This is how you end up with constellation filings for tens of thousands, even hundreds of thousands, of satellites. It is a fixed-architecture answer to a problem that keeps changing shape. And it carries the same liabilities that terrestrial data centers carry, just with worse access: hardware obsolescence on an 18-month cadence, capacity that is stranded the moment demand shifts, and an upgrade path that runs exclusively through the launch manifest. Unlike conventional data center chips with a five-year economic lifespan, AI data center chips face accelerated obsolescence cycles. A data center takes 18-36 months to develop, but AI is driving massive change every few months. A purpose-built compute satellite is a bet that the future will look like the present. In AI infrastructure, in most tech stacks, that bet has rarely paid off.

A different premise: the platform, not the payload

Symphony Space was not founded to build orbital data centers. We were founded to deliver infrastructure in orbit as a service and bring down the barrier of entry to space. We understood the market demand and back-engineered to our first technological solution. Our inaugural space station is far more than a satellite. We think of it more as a rack: a standardized, power-managed, thermally-regulated, attitude-controlled platform with swappable payload modules that can be reconfigured, serviced, and upgraded on orbit rather than discarded and replaced.

We did not design Adagio for compute. We designed it for modularity - and that is precisely why it is one of the strongest answers to the compute question now sitting in front of the industry. Orbital data centers are, fundamentally, a power, thermal, pointing, and connectivity problem wrapped around a box of chips. Those four things are exactly what our platform was architected to provide and to manage dynamically. The compute is a module. So is everything else.

Why reconfigurability wins the decade, not the demo

Here is the difference that matters. When a chip generation turns over - and it will, repeatedly, over any platform's operational life - a purpose-built satellite becomes a depreciating asset in a place you cannot easily reach. An Adagio platform gets a new module. When a customer's workload shifts from training to inference, or from compute to sensing, or from one sovereign mission to another, the platform reconfigures rather than retires. Capacity that would otherwise be stranded gets redeployed. The same physical infrastructure serves compute this year and something entirely different next year, because the platform was never married to a single use in the first place.

This is a different optimization than the one the current race is running, and it is the one we believe wins over time. The leaders are optimizing for the most powerful demonstration - the first model trained, the first cluster flown. Those are real milestones with real and significant long-term implications. Demonstrations are won by purpose-built hardware. Decades are won by infrastructure that can absorb change without being rebuilt. Modular construction delivers time savings of 25-50% and cost savings up to 25% of a typical construction budget compared to traditional fixed approaches. When applied to orbital infrastructure, the ability to reconfigure without replacement transforms the economics of long-duration space assets. The history of computing on the ground is unambiguous on this point: the durable winners were never the people with the best single machine. They were the people who built the platform everything else could plug into.

Speed is the other half of the argument

Reconfigurability solves what happens after you are in orbit. The second advantage solves how fast you get there. Modular, standardized platforms are not bespoke spacecraft that take years to design around a customer. They are a known quantity that flies on the cadence the launch market already supports - rideshare today, dedicated missions as the fleet scales. When demand for orbital capacity is doubling on the timeline AI is setting, the ability to deploy quickly and then adapt on orbit is not a nicety. It is the entire game. A faster path to orbit and a faster path to reconfiguration, stacked together, is a fundamentally different cost curve than launch-and-replace.

Where this leaves us

The infrastructure layer this entire race depends on - power, thermal, pointing, connectivity, and the ability to change your mind after launch - is the layer we have been building from the beginning. As the giants commit to putting compute in orbit, the market is going to need platforms that can host it without locking customers into a single chip, a single workload, or a single moment in time. Adagio is built for exactly that, and it is built to do it for compute and for everything else orbit is about to be asked to carry.

The companies making headlines this spring have proven the demand, and they have demonstrated that compute can run in orbit at all - no small thing. What no one has yet proven is that it works at data-center scale, where thermal rejection in vacuum, radiation durability over years, and moving data fast enough stop being engineering details and start being the whole problem. And even once those are solved, the obsolescence problem that fixed architecture guarantees will still be waiting. We think the answer to both was never going to be more satellites. It was always going to be smarter ones - platforms designed, from the first sketch, to manage exactly those hard problems, and to change.