Summary

Google is treating orbital AI infrastructure as a long-horizon response to Earth’s land, permitting, grid, and community constraints. Project Suncatcher would put TPU clusters and solar arrays in a dawn-dusk low-Earth orbit, where nearly constant sunlight could make panels up to eight times as productive; returning data may add only “a couple more milliseconds.” It is a real option on explosive compute demand, but not a present business: launches still cost many times an equivalent terrestrial center, and repairs may require robots.
The first concrete milestone is a 2027 two-satellite prototype, not an orbital hyperscale build-out. Google says newer TPUs survived proton-beam radiation beyond what a five-year mission should bring, and it is partnering with Planet for the test; Starcloud, Axiom Space, and a vaguely described Chinese effort are also exploring the field, while Eric Schmidt and Jeff Bezos have signaled interest. Google frames Suncatcher beside Waymo and quantum computing—an “eight, 10, 12, 15 years” kind of commitment—making launch economics and operational reliability the gates.
Republican AI policy remains a spectrum of intuitions rather than a settled MAGA doctrine. Dean Ball described the administration as sharing broad intuitions that AI may be the most important opportunity in decades or ever, carries familiar and “more alien” risks, and will shape U.S. leadership; camps range from national-security officials focused on China competition and child-safety conservatives to the contrasting positions the hosts associated with David Sacks and Steve Bannon. Ball expects backlash around a “weird vichyssoise” of slop, electricity, water, jobs, child safety, extinction, and claims that AI is fake.
The labs’ strongest competitive moat may be infrastructure, while their regulatory positions track their incentives. Hyperscalers do not necessarily oppose export controls because Chinese buyers—and even indirect demand for TSMC fabrication capacity—compete with them; frontier labs need chips to sell tokens, but Ball argued that “the parameters of the model are not your moat.” Anthropic’s announced $50 billion data-center commitment joins Google, OpenAI’s Stargate, Meta, and xAI in a capex race whose equilibrium government must arbitrate.
Ball draws a sharp regulatory line: prevent catastrophic tail risk federally, regulate ordinary harms reactively, and let procurement rules govern only government-bought models. He defended the “Woke AI” order as anti-ideological-bias terms and system-prompt disclosure for federal versions, while acknowledging the right’s jawboning tension and calling compelled changes to public model training “unambiguously unconstitutional.” He wants national standards for billion-dollar, globally served models, yet says Congress must act on child safety and that California’s SB 53 transparency rules are “rather reasonable overall.”
An unidentified Gemini test model appears to have pushed handwritten transcription from impressive to professionally usable. On five benchmark documents totaling about 1,000 words, Mark Humphries measured roughly a 1% word-error rate—about a 50% decline against Gemini 2.5 Pro on those tests and comparable to what human transcription experts offer. The result came through Google AI Studio A/B tests, sometimes after 20 or 30 attempts, so Kevin Roose’s “probably Gemini 3” remains a bet, not an identification.
The deeper signal was not transcription but a ledger inference that looked like symbolic reasoning. From an Albany entry, the model inferred that “14 5” meant 14 pounds 5 ounces of sugar sold at 1 shilling 4 pence per pound, reconciling it to 19 shillings 1 penny; Humphries said “models shouldn’t be able to do that.” If replicated, the capability could let models aggregate a ledger and take on broader archive-scale tasks—but the sample was tiny, the model unknown, and release-time replication still required.

Deep dive

1. Earthbound power scarcity makes orbital compute thinkable

Roose’s premise: terrestrial data centers require land, permits, grid capacity, and speed, while residents increasingly object over electricity, water, and environmental costs. With “literally, like, not enough capacity” on the grid for every proposed build, exponential AI demand makes space less a gag than a contingency for an industry assuming near-infinite use.
Newton’s investor-relevant framing was blunter: this is “the stage of this bubble” where companies believe the planet cannot provide enough electricity for their ambitions. Even if orbital compute fails, its consideration measures how aggressively the industry expects demand and capital spending to grow.
Google calls Project Suncatcher a “future space-based, highly scalable AI infrastructure system design,” but also a moonshot. No one is operating such a center today; the project is active research aimed at a world perhaps five, 10, or 15 years away, when AI might serve virtually everyone continuously.

2. Suncatcher turns constant sunlight into an eightfold energy bet

The energy logic begins with scale: the sun emits roughly 100 trillion times humanity’s total output, but terrestrial panels lose production when night arrives. A dawn-dusk low-Earth orbit could provide nearly continuous exposure, making orbital panels up to eight times as productive as panels on Earth.
The envisioned center is not a warehouse in orbit. Mock-ups from Starcloud resemble “a giant bird”: thin solar-panel wings gather power for clusters of computers at the structure’s center, with such a structure orbiting above Earth.
Getting results home may be one of the less exotic problems. People working in the field compared it with Starlink: low-Earth orbit is close enough that transmission could add only “a couple more milliseconds,” using satellite communications that already exist.
Google subjected a standard TPU to an intense proton beam simulating orbital radiation. Its newer chips endured radiation well beyond expectations for a five-year mission; hardware failures remain harder, with researchers telling Roose, “We gotta figure out” how robots could perform repairs.

3. A 2027 prototype still leaves the economics on Earth

Google plans to launch two prototype satellites with mapping-satellite company Planet in 2027. Starcloud is also preparing prototypes, so the near-term contest is to validate survivability, connectivity, and operation—not to reproduce terrestrial hyperscale capacity immediately.
Roose’s central caveat: sending large numbers of chips and satellites into space currently costs many times more than building comparable compute on Earth. He does not expect meaningful deployment for at least several years; lower launch costs and workable repair systems are prerequisites, not details.
Google placed Suncatcher in the lineage of Waymo and quantum computing, signaling willingness to work for “eight, 10, 12, 15 years” before expecting mainstream use. The bet only makes sense if AI demand becomes effectively infinite and Earth eventually runs short of both suitable land and power.
The field also includes Axiom Space, a vaguely detailed Chinese effort, and possible interest from Eric Schmidt and Jeff Bezos. Local “not in my backyard” resistance might become Newton’s “NOPs”—“not on my planet”—although debris and the appearance of rich companies escaping earthly problems could eventually generate orbital opposition too.

4. The Republican AI coalition has intuitions, not settled doctrine

Ball entered the White House after writing Hyperdimensional and posting on X, then led the drafting of its AI action plan. What he found was “coherent intuitions,” accompanied by excitement, worry, and confusion in roughly equal measure—not a mature catalogue of conservative policy positions.
Intuition one: AI is the most important technological, scientific, and economic opportunity in decades, “and quite possibly ever.” Intuition two: it creates both familiar risks addressable through existing frameworks and potentially “more alien” risks for which government lacks clear concepts. Intuition three: AI will materially shape American global leadership.
Roose sketched David Sacks and Steve Bannon as rough polar opposites—one attacking “doomer” arguments, the other talking about existential risk. Ball accepted the spectrum but stressed the large middle: national-security officials focused on China, and conservatives focused on chatbot psychosis, teen suicide, and lessons from social media.
Ball’s honest answer on a distinctly MAGA theory of AGI was: “Not yet. No, not really.” Online MAGA discussion may currently lean more doomer, but many participants have not formed a view of AGI at all, much less converted it into specific domestic regulation.

5. AI backlash will combine contradictions into one political stew

At AI gatherings, Ball said the recurring question is when “the pitchforks” come out and what triggers them. His answer is not one catastrophe but a miasma: “slopification,” unsafe products for children, electricity and water consumption, job loss, extinction fears—and, simultaneously, the belief that AI is fake—mixed into “this weird vichyssoise.”
He does not expect one clean partisan position to emerge. “AI policy” will splinter as “internet policy” did into data centers, China competition, software regulation, children’s safety, and other issues; individual issues may polarize, but the technology as a whole is too internally varied for one durable party line.

6. Infrastructure may become the frontier labs’ durable moat

Ball resisted treating “industry” as one political actor: hyperscalers, frontier-model companies, and other suppliers occupy different points in the stack. “No one’s making illegitimate arguments here. Everyone’s operating from incentives,” leaving government to solve for an equilibrium among them.
Microsoft, Google, and Amazon Web Services do not necessarily hate chip export controls. They benefit when Chinese firms cannot compete for the same accelerators and when less Chinese demand reaches the finite space inside TSMC fabrication plants—even if the firms are purchasing different chip designs.
Frontier labs need chip access because they “wanna make money selling tokens,” yet Ball sees infrastructure as the likelier moat than model parameters. He cited Anthropic’s $50 billion data-center commitment alongside Google, OpenAI’s Stargate, Meta, and xAI: each is seeking defensibility through owned or secured compute.

7. The “Woke AI” order exposes the right’s jawboning contradiction

Ball emphasized that the executive order governs federal procurement, not models sold to consumers or private businesses. The government is asking vendors not to engineer top-down ideological biases into the particular versions supplied to agencies; it is not formally regulating public releases.
He rejected “objective” as a workable technical standard—humanity has debated truth since language existed—and said the order wisely avoids trying to resolve it. Its narrower demand is that developers not impose an additional worldview and that they disclose elements such as the government version’s system prompt during procurement.
The hosts’ pushback: Republicans condemned Biden-era pressure on platforms over COVID misinformation, yet Trump’s government is also telling technology companies how products should respond. Ball acknowledged an inherent post-Trump tension between maintaining anti-jawboning principles and using government power to “throw it back at the left.”
Ball chose the principled side: “No jawboning from anyone.” But he distinguished customer requirements, noting that government models already face harder Freedom of Information Act, Presidential Records Act, and data-stewardship obligations. Compelling changes to how public models are trained, he said, would be “unambiguously unconstitutional,” violating both company and user speech rights.

8. Federal standards collide with legitimate state urgency

Ball’s federalism case turns on scale: models can cost $1 billion to train and are built for worldwide service, so their training, evaluation, and measurement standards implicate interstate commerce. Competing state regimes are impractical; absent Congress, the largest state can effectively write national rules.
California is therefore the country’s default central AI regulator, which Ball called a constitutional failure mode the founders could not have anticipated because modern economies of scale did not yet exist. Still, he was broadly supportive of SB 53, a transparency law applying only to the largest developers, calling it “rather reasonable overall.”
Newton’s challenge was present harm: chatbot psychosis, child safety, and teen suicidality are harms he described as present today and encouraged to some degree by products on the market, while Congress remains unable or unwilling to regulate technology. A state legislator can reasonably conclude, “I don’t want the kids in my state to kill themselves,” and act rather than await Washington.
Ball agreed with that incentive and clarified that his view is proactive: Congress needs to solve the problem, not merely tell states to stop. He sometimes faults legislators for poor statute drafting—“we’ll let the courts figure that out” is no excuse, because lawmakers also swear constitutional oaths—but not for seeking to protect children.

9. Tail risks merit prevention; ordinary harms can wait for evidence

Ball borrowed Ezra Klein’s description of government as “a grand enterprise in risk management.” Catastrophic tail risks and national-security threats demand mature, preferably bipartisan prevention; if government cannot manage them, it has failed its basic purpose and might as well “return the money to the shareholders.”
Ball said many people at frontier labs have an earnest desire to address safety problems, while noting that he could not speak for the labs as companies. He argued that companies also have incentives because they would be bankrupted if, for example, they caused a pandemic. Near-term biological and cyber risks are “eminently tractable” to veteran risk officials—serious, but like a hurricane tracking toward Florida rather than an unknowable abstraction.
For current and near-future technology, Ball rejects an automatic trade-off between safety and acceleration: government can improve biosecurity without meaningfully slowing development. He expects genuine trade-offs eventually, but would decide them on their particulars rather than presume today where the acceptable boundary lies.
For non-tail harms, he favors reactive law after four tests: harm has occurred, will probably recur, common-law liability is inadequate, and a targeted statute can meaningfully help. Child safety meets that description. A catastrophe might catalyze Congress, he conceded, but incremental progress without one remains possible.

10. A mystery Gemini reached human-expert transcription territory

Humphries and research partner Lianne Leddy use AI to connect tens of thousands of handwritten records about ordinary people in the fur trade. Their evidence—accounts, contracts, baptisms, marriages, and deaths—reconstructs fragmented lives across western North America from roughly 1760 into the early 19th century.
GPT-4 in 2023 could “sort of read” handwriting but produced mostly errors. Models then reached roughly 90% accuracy quickly and struggled above it; that final 10% contains the names, monetary amounts, and locations historians most need. Gemini 2.5 Pro subsequently reached about 95%.
Their benchmark contains 50 documents believed, though not guaranteed, to be absent from training data. Humphries tested five, totaling roughly 1,000 words; because AI Studio assigns experimental A/B comparisons randomly, he sometimes had to submit a document 20 or 30 times to receive the revealing side-by-side result.
The mystery model produced roughly a 1% word-error rate, including capitalization and punctuation mistakes—about a 50% error reduction on those tests and comparable to what professional human transcribers offer. Roose said he was able to confirm only that Google tests unreleased models in AI Studio; his “probably Gemini 3” inference remains unverified.

11. One sugar ledger hints at a broader knowledge-work jump

The tougher test was an 18th-century Albany ledger: quickly written handwriting, on-the-fly tabular structure, and pounds, shillings, and pence. Such records resemble cash-register receipts created for immediate accounting, not later interpretation, and previous models performed poorly on them.
A Samuel Slitt entry recorded one loaf of sugar, the compact figure “14 5,” a price of 1 shilling 4 pence per pound, and a total of 19 shillings 1 penny. The model clarified that “14 5” meant 14 pounds 5 ounces and reconciled the quantity, unit price, and total correctly.
Humphries’s surprise was that the apparent transcription errors were clarifications. The model had to infer random numbers rather than predict a likely phrase, recognize units with different bases, and work backward through a relatively rare historical currency system. To him, that looked like symbolic reasoning: “Models shouldn’t be able to do that.”
If the result replicates, historians could ask a model to identify and total every sugar transaction across a ledger, not merely transcribe it. Humphries generalized the mechanism to knowledge work—transforming information, connecting formats, and drawing implications—while Newton saw evidence that scaling may still produce emergent capabilities. Both conclusions remain contingent on testing the released model at larger scale.

Michael Burry Shorts NVIDIA and Palantir & Has Defensibility Died in a World of AI?How Apple Changed the Cellular Economy and SpaceX's Spectrum Play