Google Is Putting TPUs in Orbit
Project Suncatcher launches two satellites in early 2027, built with Planet Labs and carrying Trillium-generation TPUs. The theoretical end state is an 81-satellite cluster inside a 1km radius at 650km altitude - flying, in Google's own words, in tighter formation than any current or previous satellite constellation.
Here is the part the coverage keeps leaving out. The design requires 10 Tbps of aggregate bandwidth per link. Current inter-satellite connectivity runs 1 to 100 Gbps. That is a hundredfold gap, and Google's proposed route across it is commercial off-the-shelf DWDM transceivers.
The economics have a similar gap. Suncatcher needs launch costs near $200/kg. Today they run $1,500 to $2,900/kg. Google's own projection for reaching $200 is 2035, via Starship.
And Google lists the physics problems itself rather than burying them: thermal management needing advanced interface materials, space radiation producing uncorrectable errors in High Bandwidth Memory, Single Event Effects on training runs that "require further studying", and no possibility of human maintenance - the answer to which is redundant provisioning.
That is an unusually honest set of caveats for a moonshot announcement. Two satellites in 2027 is a real experiment. The 81-satellite orbital data centre is a 2035 proposition with at least two unsolved engineering problems in front of it.
Anthropic Is Renting $45 Billion of xAI Capacity
SpaceX's IPO filing discloses Anthropic paying $1.25 billion a month through May 2029 for 300MW inside xAI's Colossus facilities in Memphis - about $45 billion, with a 90-day termination clause for either side. Google pays roughly $920 million a month on a separate deal.
The same filing concedes that SpaceX's widely repeated 1GW "nameplate compute draw" figure "does not represent actual power consumption or utilization", and references 330MW operational. Full breakdown in our piece on the filing.
Microsoft Named Two Models in a Threat Report
The 2026 Digital Defense Report cites Anthropic's Mythos and OpenAI's GPT-5.5 orchestrating a 32-step attack chain to complete domain control, puts median vulnerability-to-weaponisation well below 24 hours, and measures open-weight models as seven months behind closed ones at attack orchestration. Phishing as an entry vector tripled year on year, from 7% to 23%. The full findings.
FieldAI Raises $700M at $10B
The robotics foundation-model company raised $700 million at a $10 billion valuation - a 5x step up from its $2 billion mark in 2025. It reports $135 million in revenue plus contracts across more than 30 customers. Backers include Bezos Expeditions, Emerson Collective, Khosla, Nvidia and Samsung.
A 5x valuation increase in roughly a year against $135 million of revenue puts the multiple around 74x. Read that as a statement about where robotics capital is rather than about FieldAI specifically.
Brussels Moves on Azure and AWS
The European Commission issued a preliminary gatekeeper determination under the Digital Markets Act covering Microsoft Azure and Amazon AWS. DMA penalties run to 10% of global revenue, 20% for repeat breaches.
Cloud has been outside the DMA perimeter until now. If the determination holds, the two providers hosting most AI workloads in Europe acquire a set of interoperability and self-preferencing obligations they have not previously had.
Oracle's Wisconsin Campus Slips
Project Lighthouse - 902MW of compute against 1.3GW of total demand - now expects partial power in December 2027 and full supply in October 2028, with a worst case of April 2029.
Worth holding beside the Suncatcher story. One company is solving for compute by launching TPUs into orbit; another cannot get a transmission line energised in Wisconsin before 2028. Both are the same constraint, which is power, and only one of them is a physics problem.
The Thread
Four of today's six items are about electricity and the buildings it goes into. Google is going to orbit for it. Anthropic is paying $15 billion a year to rent it from a competitor. Oracle is waiting until 2028 for it. SpaceX is buying $2.8 billion of gas turbines to make it.
The model race has been a compute race for a while. As of this week it is openly a power race, and the filings say so more plainly than the announcements do.