Google sent AI chips into orbit | Future Intelligence #29

Four TPU accelerators from Google orbited Earth on Thursday aboard the Project Suncatcher prototype - the first test of the thesis that the ceiling on computing capacity need not be determined by the availability of electricity on Earth. Amazon is meanwhile tackling the same problem on the ground: with a twenty-year contract, it locked in 690 megawatts from the Calvert Cliffs nuclear plant. And Broadcom will lend Anthropic up to $42 billion for a five-year lease of its own chips, making the supplier the creditor of purchases of its own goods - the same logic Tesla chose with a $30 billion credit facility for AI and robotaxis. Financing of computing capacity is moving from balance sheets into debt structures, and that is a shift that will play out over years, not quarters.

Future Intelligence is an exclusive analytical report published once a week and available only to Bulios Black members. Members receive it automatically every Sunday morning by email - in full, including specific scenarios and market implications. Permanent access to the report is obtained with Bulios Black membership.

Project Suncatcher: four Google TPUs orbited Earth

On Thursday, October 1, a Falcon 9 rocket launched from Vandenberg Space Force Base carrying a prototype satellite with four Google Trillium-generation TPU accelerators. Liftoff occurred at 11:18 a.m. local time as part of the rideshare mission Transporter-18, which carried a total of 130 payloads, according to reports. This is the first orbital test of the Project Suncatcher initiative.

The satellite, nicknamed MVP, was built by Google in collaboration with Planet Labs, is about the size of a refrigerator, and is powered by solar panels with an output of approximately 1 kilowatt. After launch, Google confirmed that it had established contact with the satellite and that it is operating as expected. The mission was announced on its blog on September 24.

Why space, and not another turbine in Texas

The project's logic rests not on space romance but on energy. Data centers on Earth face constraints from electricity availability, grid connections, and cooling; in orbit, sunlight is nearly uninterrupted and heat can be dissipated by radiation into the vacuum. Suncatcher is an attempt to move computing power to where the energy ceiling does not apply the same way as on Earth.

The target architecture is not a single satellite but a swarm. According to published plans, future satellites are to fly in clusters of 81, within one kilometer of each other, at an altitude of about 650 kilometers, and connect to one another via free-space optical links. Each is to carry on the order of tens of TPU chips.

The optical link is the hardest part. Google describes the aiming precision as hitting a coin-sized target from several miles away, while both ends are moving relative to each other. Without optical interconnectivity, the swarm is just a collection of small computers; only shared throughput turns it into a single computing node.

What is actually being tested

The mission is not validating performance but survival. The Trillium TPUs were exposed to a proton beam at the Crocker Nuclear Laboratory at UC Davis before launch, and Google reports they withstood a total ionizing dose above the level of a five-year mission in space. But the laboratory cannot simultaneously combine radiation, launch vibration, and the thermal cycling of low Earth orbit - that can only be learned by flying.

Public data is sparse: Google has not disclosed the orbital parameters or what workloads the chips will compute in orbit. The scale gap is brutal. Four chips at one kilowatt stand against ground clusters whose consumption is measured in hundreds of megawatts - a difference of six orders of magnitude. This is not a product; it is a data point.

Who profits and who pays

For Alphabet $GOOGL, Suncatcher is a cheap option against its own capital program, which according to reports is approaching a value of up to $205 billion. The cost of four chips is a rounding error in that budget, but it changes the company's position toward energy suppliers and grid operators: an operator with a proven path outside the distribution grid negotiates differently.

The second name is Planet Labs $PL. The company has so far mainly sold satellite imagery; the role of platform manufacturer for a hyperscaler's computing payload shifts its story toward an infrastructure supplier with recurring orders. For a stock of this size, the type of customer determines an acceptable earnings multiple, not any single contract.

A competing trajectory is shown by SpaceX, which plans to launch the first Starmind satellites with Nvidia $NVDA chips in the fourth quarter of 2027 and is building them around the Vera Rubin NVL72 rack system.

Both Suncatcher and Starmind lead to the same question: if orbital computing proves scalable, demand for accelerators will cease to be tied to the pace of construction of ground-based power plants and grid connections.

The horizon remains years, not quarters. But for the valuation of energy and data-center suppliers, it is relevant already now: every additional flight test that does not end in hardware failure weakens the assumption that the availability of electricity on Earth is a permanent ceiling on computing capacity growth. The nearest measurable milestone is the telemetry results from MVP.

Tesla: 486,532 vehicles and a $30 billion credit facility for AI and robotaxi

On October 2, Tesla announced 486,532 vehicles delivered in the third quarter, roughly 25 thousand units above the Wall Street consensus of about 461,100. It produced 464,391 vehicles, fewer than it delivered - the company reduced inventory. Year over year, this is still a decline of approximately 2%, but compared with the second quarter, a slight increase.