Space data center plans usually arrive with enormous numbers attached: gigawatts, million-satellite constellations and multi-billion-dollar valuations. Google’s first one arrives with small numbers. On 1 October 2026, a satellite called MVP is scheduled to carry four of Google’s Tensor Processing Units (TPUs) into orbit on a SpaceX Falcon 9, running on about one kilowatt of solar power. Engadget called it a “teensy-tiny AI data center”, and the description is fair. The whole payload has roughly the computing power of one server.
That smallness is the story worth examining. This article measures Google’s first space data center against the things it will one day have to compete with: one of Google’s own TPU pods, the chips Google rents on the ground, the gigawatt campuses now under construction and the far larger orbital plans from Starcloud and SpaceX. It works through the arithmetic of going from one kilowatt to one gigawatt, explains why starting small is a sensible choice, and sets out what a single satellite can and cannot prove.
We covered the launch, the radiation and vibration tests and the history of the programme in our report on Google’s first test of AI chips in space under Project Suncatcher. Here the focus is scale.
Table of contents
- What Google Is Actually Sending to Orbit
- How Small Is Google’s First Space Data Center?
- The Per-Chip Arithmetic Behind a Tiny Space Data Center
- Why Google Started Its Space Data Center So Small
- From One Kilowatt to One Gigawatt: The Space Data Center Scaling Ladder
- What Breaks When a Space Data Center Grows
- How the Space Data Center Race Compares on Scale
- What a Tiny Space Data Center Can and Cannot Prove
- What the Space Data Center Test Means for Businesses
- Space Data Center FAQ
- References
What Google Is Actually Sending to Orbit
The satellite belongs to Project Suncatcher, the Google research moonshot announced in November 2025 that explores whether fleets of solar-powered satellites could one day run AI computing. MVP is its first piece of hardware in orbit, and every figure attached to it is modest.
Four chips and one kilowatt
According to The New York Times, MVP carries four TPUs, the custom chips Google designs to train and run its Gemini models. Together they offer about the computing power of a single server in a data center. Its solar panels supply about one kilowatt. Engadget compared that to the energy needed to run a hair dryer. On the ground, a single Google data center runs thousands of these chips.
What “space data center” means here
Calling MVP a space data center stretches the phrase. It has no racks, no rows and no cooling plant. It is a refrigerator-sized satellite built by Planet, fitted with Google’s chips, a layered cooling system and software. Google itself describes the mission as an experiment. The phrase is shorthand for what MVP is meant to lead to, not what it is.
Fifteen minutes on, then a rest
The most telling figure is time. Travis Beals, who leads product management for Project Suncatcher, told The New York Times that the chips can run for about 15 minutes before they must shut down to cool. In those bursts MVP will answer short Gemini queries. Google plans to operate it for about a year, and it could stay in orbit for up to six years before it re-enters the atmosphere and burns up.
| Measure | Google MVP satellite | What it compares with |
|---|---|---|
| Chips | Four TPUs | About one data center server |
| Power | About 1 kilowatt from solar panels | A household hair dryer |
| Duty cycle | About 15 minutes of work, then a cool-down | Ground data centers run around the clock |
| Size | About the size of a refrigerator | A single appliance, not a building |
| Service life | About a year of operation, up to six years in orbit | Ground servers are repaired and refreshed on site |
How Small Is Google's First Space Data Center?
The size of a space data center only means something against a reference point. Three are useful here: Google’s own largest TPU system, the capacity Google rents on the ground, and the power the AI industry expects to need.
Against one of Google’s own TPU pods
In April 2025, Google announced Ironwood, its seventh-generation TPU. In its largest configuration, Ironwood scales up to 9,216 liquid-cooled chips linked by Google’s Inter-Chip Interconnect network, “spanning nearly 10 MW”, according to Google. Dividing the chip count by four gives 2,304 MVP-sized payloads in one pod. Dividing the power by one kilowatt gives 10,000. Either way, a single Google pod on the ground is thousands of times larger than the first space data center.
Against the chips Google rents on Earth
Set Google’s first space data center against the capacity Google buys from others and the contrast is sharper still. Barchart reported that Google agreed, before SpaceX’s stock market listing, to lease roughly 110,000 GPUs from SpaceX’s AI infrastructure, from October 2026 to June 2029, at about $920 million a month. Barron’s put the figure at almost $1 billion a month. That is 27,500 times as many chips as MVP carries, and all of that capacity sits on the ground, not in orbit.
Against the power the industry expects to need
Barron’s cited Morgan Stanley analyst Stephen Byrd, who sees AI power demand growing to almost 100 gigawatts by 2028, against total US generating capacity of roughly 1,300 gigawatts. At one kilowatt each, matching 100 gigawatts would take 100 million MVP-sized satellites. Nobody is proposing that. The number simply shows how far a first space data center sits from the scale at which it would matter to the grid.
The chart below puts these figures on one scale. The gaps are so large that a normal bar chart cannot show them, so the bars use a logarithmic scale from 100 watts to 100 gigawatts: every extra ninth of bar length means ten times more power.
The Starlink figures come from a TNW report in March, which noted that SpaceX’s network of about 10,000 satellites generates around 200 megawatts. Dividing one by the other gives an average of about 20 kilowatts per satellite, 20 times what MVP draws. The same report found more than 25 gigawatts of data centers under construction in the US alone.
The Per-Chip Arithmetic Behind a Tiny Space Data Center
There is a second, less obvious way in which this space data center is small. It is not only a few chips; each chip seems to get less power than it would in a data center.
What each chip gets on the ground
Google’s own Ironwood figures work out at just over a kilowatt per chip: nearly 10 megawatts divided by 9,216 chips is about 1,085 watts. That figure covers more than the chip, because the pod power includes the network and the liquid cooling that link and serve the chips. For another reference point, Nvidia lists its H100 graphics processor, the chip Starcloud flew in 2025, at up to 700 watts in its SXM form.
What each chip can get in orbit
MVP’s whole budget is about one kilowatt, and that kilowatt also has to run the rest of the satellite: its radios, its computers and its pointing systems. Four chips sharing it leaves 250 watts each at the very most. Google has not said which TPU generation flies on MVP or how hard the chips will be pushed, so this is an upper bound, not a measurement.
Why the gap matters
A space data center is limited by the heat it can shed, not only by the power it can collect. On MVP, even a modest power budget forces a 15-minute limit on work. That suggests the cooling, not the chips, sets the pace. Any future design has to close both gaps at once: more watts per chip and far more hours per day.
Why Google Started Its Space Data Center So Small
Starting a space data center with four chips and a borrowed satellite can look timid next to rivals promising constellations. It is the opposite. It is a deliberate way to learn cheaply and quickly.
Learning before scaling
Beals wrote on Google’s blog that “exploring space as a viable location for scalable AI compute won’t happen all at once.” He added: “It takes methodical engineering, starting with proving our hardware can handle the physical and unpredictable realities of operating in orbit. This first launch is about seeing what works, identifying points of failure, and applying those findings to future missions.” A small space data center gives answers to the basic questions without betting a large satellite on them.
Ready-made hardware saves a year
The original plan, announced in November 2025, was for Planet to launch two prototype satellites by early 2027. Google wanted to be in orbit this year. Eric Stevens, a director of systems engineering at Planet, told The New York Times that Google was willing to take risks to get there, so it installed its chips in a satellite Planet had already built. Benzinga noted that this puts the first hardware in orbit months ahead of schedule.
Small is cheaper to launch
MVP rides on Transporter-18, one of SpaceX’s rideshare missions, which carries dozens of satellites from different customers and shares the cost of the rocket. A dedicated Falcon 9 launch costs around $74 million, Engadget noted. A refrigerator-sized experiment sharing a ride is a far cheaper way to learn than a custom spacecraft on its own rocket.
Failure is part of the plan
Beals told The New York Times that a perfect record would be a warning sign. “If five years from now, everything we’ve done has worked perfectly, it probably means we’ve not taken enough risk and we’ve not learned as much as we could,” he said. A tiny first space data center is an affordable place to fail, which is exactly what an early experiment should be.
From One Kilowatt to One Gigawatt: The Space Data Center Scaling Ladder
Google has sketched the steps between MVP and anything resembling a full-size space data center. Each one is an order of magnitude or more beyond the last.
Step one: survive
MVP’s job is to show that TPUs survive the shaking of launch, months of radiation and cooling in a vacuum, and that they can answer real queries from orbit. That is a test of the hardware, not of the economics.
Step two: talk to each other
In 2027, Google and Planet plan to fly two satellites to test laser links between them. A cluster only works as one computer if the satellites can share data very fast. Google’s research blog said the links would need tens of terabits per second, and its bench-scale demonstrator has already reached 800 gigabits per second in each direction, 1.6 terabits per second in total.
Step three: fly in formation
Google’s preprint paper modelled an illustrative cluster of 81 satellites at a mean altitude of 650 kilometres, packed into a radius of 1 kilometre with neighbours roughly 100 to 200 metres apart. The New York Times reported that Google has designs for fleets of more than 80 satellites flying in close formation. Google has said each future satellite would carry dozens of TPUs.
Step four: grow the satellites
Google is also talking to designers about a much larger custom satellite, about the length of a football field, according to The New York Times. Engadget noted that Google has not said how much money it is spending on the effort. Google’s own estimate is that orbital computing could reach rough cost parity with ground facilities in the mid-2030s.
| Stage | Hardware | What it proves | Timing |
|---|---|---|---|
| MVP | One satellite, four TPUs, about 1 kW | Chips survive launch, radiation and vacuum cooling | Launch scheduled for 1 October 2026 |
| Laser pair | Two satellites with optical links | High-bandwidth links between satellites | 2027 |
| Cluster model | 81 satellites within a 1 km radius | Formation flying and shared AI workloads | Research design, no date |
| Large satellite | Custom craft about the length of a football field | Bigger power and cooling per spacecraft | In discussion with designers |
| Cost parity | Launch below $200 per kg | Orbit competes with ground energy costs | Google estimate, mid-2030s |
What Breaks When a Space Data Center Grows
A four-chip space data center hides problems that thousands of chips expose. Physics does not scale kindly, and a few rough calculations show why.
Heat needs radiators, and radiators need area
In a vacuum there is no air to carry heat away, so a satellite can only shed it by radiating infrared light from its surfaces. The Stefan–Boltzmann law sets the rate. An ideal radiator at about 27°C emits roughly 460 watts per square metre from each face. A panel radiating from both sides therefore sheds about 920 watts per square metre, ignoring the sunlight and Earth’s infrared that fall on it and make the real figure worse. On that basis one kilowatt needs about a square metre of radiator, and one gigawatt needs over a square kilometre.
Power needs panels, and panels need area
Sunlight above the atmosphere delivers about 1,361 watts per square metre, the solar constant. Assuming solar cells that convert 30% of it, each square metre of panel yields roughly 410 watts. That is 2.4 square metres for MVP’s kilowatt, but about 2.4 square kilometres for a gigawatt. Google’s argument is that in the right orbit panels can be up to eight times more productive than on Earth, which is true, but the area still has to be launched and unfolded.
| Space data center size | MVP-sized satellites | Solar panel area | Two-sided radiator area |
|---|---|---|---|
| 1 kW (MVP) | 1 | About 2.4 square metres | About 1.1 square metres |
| 1 MW | 1,000 | About 2,450 square metres | About 1,090 square metres |
| 10 MW (one Ironwood pod) | 10,000 | About 24,500 square metres | About 10,900 square metres |
| 1 GW (a large ground campus) | 1,000,000 | About 2.4 square kilometres | About 1.1 square kilometres |
These are idealised figures, based on 1,361 watts per square metre of sunlight, 30% efficient cells and a perfect radiator at 27°C with nothing falling on it. Real hardware will need more area, not less. A standard football pitch covers about 7,140 square metres, so a one-gigawatt space data center would need solar panels covering well over 300 pitches.
Bit flips multiply with chip count
Radiation causes bit flips, where a stray particle turns a stored one into a zero or the reverse. Google found in particle-beam tests that restarting the chips could usually clear them. That is manageable with four chips. With thousands of chips working on one job, a restart on any one of them can stall the rest, so error handling becomes a system-design problem rather than a reset button.
Launch bills scale with mass
Google’s paper argues that launch costs may fall below $200 per kilogram by the mid-2030s, at which point the cost of launching and running a space data center could become roughly comparable to the energy costs of a ground facility. Until then, every square metre of panel and radiator is paid for by the kilogram, which is why Google’s first space data center is so light.
How the Space Data Center Race Compares on Scale
Google is not the only company planning a space data center, and it is the most cautious on size. Its rivals have gone bigger sooner, at least on paper.
Starcloud
Starcloud, a start-up from Redmond, Washington, flew the first Nvidia H100 in orbit on Starcloud-1 in November 2025, a 60 kg satellite. Its next satellite, Starcloud-2, carries an Nvidia Blackwell chip and what TNW described as the largest deployable radiator ever flown on a private satellite. Starcloud-3 is designed as a 200-kilowatt, three-tonne spacecraft. DatacenterDynamics reported that Starcloud wants a constellation of 88,000 satellites delivering 20 gigawatts, which works out at roughly 230 kilowatts per satellite, about 230 times MVP.
SpaceX Starmind
SpaceX, which acquired Elon Musk’s AI company xAI in February 2026, calls its orbital computing plan Starmind and has asked the Federal Communications Commission for permission to operate up to a million satellites. In a letter to the FCC reported by PCMag, SpaceX said each Starmind satellite could weigh just under 4,000 kilograms, against about 575 kilograms for a Starlink V2 Mini, with a design life of five years. Musk has said the first Starmind satellite, built with Nvidia, is targeted for late 2027.
Where Google’s space data center sits
Google’s first space data center is the smallest of the three first steps, and the only one built around the company’s own chips rather than Nvidia’s. That gives Google control from silicon to software, and it matches the research-first tone of the programme. It also means Google has the furthest to climb.
| Programme | First computing hardware | Per-satellite scale | Stated ambition |
|---|---|---|---|
| Google Project Suncatcher | Four TPUs on MVP, launching 1 October 2026 | About 1 kW | Fleets of more than 80 satellites, cost parity mid-2030s |
| Starcloud | One Nvidia H100 on Starcloud-1, November 2025 | 60 kg today, 200 kW for Starcloud-3 | 88,000 satellites delivering 20 GW |
| SpaceX Starmind | First Nvidia-based satellite targeted for late 2027 | Just under 4,000 kg | Up to 1 million satellites |
Mass tells the same story as power. MVP’s mass has not been published beyond its refrigerator size, but the planned rivals are heavy.
What a Tiny Space Data Center Can and Cannot Prove
A single-satellite space data center is a good instrument for some questions and a useless one for others. Being clear about which is which is the best way to read the results when they arrive.
What one satellite can tell Google
MVP can show whether TPUs survive a real launch and months of real radiation, how often bit flips happen in orbit and whether restarts clear them, and whether the layered cooling system behaves as it did in a thermal vacuum chamber. It can also show that a Gemini query can be answered from orbit end to end, which is a useful proof that the software stack works.
What it cannot tell Google
It cannot say whether a space data center will be cheap, because one satellite on a rideshare has none of the economics of a fleet. It cannot test laser links, which need two satellites. It cannot show how thousands of chips cooperate, how radiators behave at megawatt scale or how a failed satellite is replaced within a cluster.
| Question | Can MVP answer it? | What would |
|---|---|---|
| Do TPUs survive launch? | Yes | This flight |
| How often do bit flips occur in orbit? | Yes, for four chips | A year of operating data |
| Does vacuum cooling work? | Partly, at 1 kW with 15-minute bursts | Larger satellites running continuously |
| Can satellites share AI work? | No | The 2027 laser-link pair and later clusters |
| Is orbit cheaper than the ground? | No | Fleet-scale operation and launch below $200 per kg |
The sceptics’ arithmetic
Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon University, told The New York Times that expanding from one satellite to a vast network that operates like a giant data center will take years and enormous funds. “If you do this on a large scale, there are additional engineering problems,” he said. Engadget made the same point more bluntly: a cooling system that manages 15 minutes with four chips has a long way to go before it can serve thousands.
Google’s own expectations
Google is not overselling it. “We don’t expect, to be perfectly frank, that we’ll have anything usefully operational in the next few years,” James Manyika, Google’s senior vice president for research, told The New York Times. He compared the effort to Google’s long research into driverless cars before any product appeared. On that timeline, a teensy-tiny first space data center is exactly what the first year should look like.
What the Space Data Center Test Means for Businesses
For organisations that buy computing, Google’s space data center test changes nothing this year or next. It is still worth understanding, because it shows where the largest AI buyers think the limits are.
Capacity planning stays on the ground
No provider sells orbital computing to ordinary customers, and Google does not expect anything usefully operational for years. Plans for data center operations and cloud capacity in 2026 and 2027 should rest on today’s ground prices and constraints. Small is having a moment on the ground too: Crusoe has raised money for both giant campuses and small modular AI factories that can be placed near spare power.
Read the per-kilowatt numbers, not the headlines
When space data center announcements arrive, the useful questions are simple. How many kilowatts does each satellite deliver, for how many hours a day? What does each kilowatt cost to launch, and how long does the hardware last? The figures in this article give a baseline: MVP at about 1 kW, Starcloud-3 designed for 200 kW, and ground pods at around 10 MW.
Where orbit could matter first
If a space data center proves useful, it will probably start with data that is already in space. Earth-observation satellites collect far more imagery than they can send down, and running computer vision models in orbit to pick out ships, floods or damaged crops before downlink would save bandwidth and time. That market needs kilowatts, not gigawatts, which is closer to where MVP sits today.
Debris and e-waste belong in the sums
Satellites that burn up after five or six years are hardware that is thrown away on a schedule. We have written about the growing e-waste problem of AI data centers on the ground, and orbit adds re-entry debris to it. Astronomers questioned SpaceX’s debris estimates for Starmind in PCMag’s report, and any sustainability claim for a space data center should say how it counts both.
Space Data Center FAQ
How big is Google’s first space data center?
It is one refrigerator-sized satellite, called MVP, carrying four Google TPUs with about the computing power of one server. Its solar panels supply about one kilowatt, roughly what a hair dryer uses.
Why is Google starting its space data center so small?
A small payload on a shared rideshare flight is a cheap, fast way to test whether TPUs survive launch, radiation and vacuum cooling. Google used a ready-made Planet satellite to reach orbit in 2026 rather than wait for its 2027 prototypes.
How many MVP-sized satellites would match a ground data center?
At about one kilowatt each, matching one of Google’s Ironwood TPU pods, at nearly 10 megawatts, would take about 10,000. Matching a one-gigawatt campus would take about a million.
When could a space data center be useful?
Google estimates rough cost parity with ground facilities in the mid-2030s if launch prices fall below $200 per kilogram. Its research lead does not expect anything usefully operational in the next few years.
Who else is building a space data center?
Starcloud has flown an Nvidia H100 since November 2025 and plans 88,000 satellites delivering 20 gigawatts. SpaceX’s Starmind plan seeks approval for up to a million satellites, with its first Nvidia-based satellite targeted for late 2027.
References
Google is sending a teensy-tiny AI data center to space
Google takes the AI data center race to outer space (The New York Times)
Behind Project Suncatcher, our moonshot to put AI in space
Towards a future space-based, highly scalable AI infrastructure system design
Ironwood: The first Google TPU for the age of inference
Starcloud raises $170M at a $1.1B valuation to build data centres in orbit
Starcloud closes $250m Series A extension at $2.3bn valuation
Will SpaceX’s orbiting data centers rain debris over Earth?
SpaceX intends to launch its first Nvidia-powered AI satellites in late 2027 (Barchart)
Why Google and SpaceX are testing AI in orbit: The earthly grid can’t keep up (Barron’s)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
