On July 16th at 1PM ET, beehiiv is going live with a look at the future of publishing, audience growth, and digital business.

What started as a newsletter platform has evolved into something much bigger: a place where creators and brands can grow, monetize, and own their audiences without stitching together half the internet to make it work.

The next chapter starts live at the Summer Release Event. 

Join us to see what’s coming next.

RSVP now.

Hi {{first_name}},

The pitch is almost too clean: if AI data centers are straining our grids, drying up our aquifers, and provoking backlash in towns from Virginia to Chile, then just move them off the planet. Unlimited solar, the vacuum of space as a free heat sink, and nobody's water supply to fight over. In the past year that pitch went from whitepaper to hardware in orbit, and billions of dollars are now chasing it. I find the whole thing genuinely fascinating and also can't shake a nagging question, which is the subject of today's issue. Then, as ever, the jobs.

Craig

Here is where things actually stand, because it is wilder than most people realize right now.

In November 2025, a startup called Starcloud put an NVIDIA H100 GPU into orbit. By December it had trained a small language model in space and run a version of Google's Gemma there, the first time either had ever been done. Demis Hassabis called it "first LLM contact from space." Eric Schmidt called it "a seriously cool achievement." Andrej Karpathy, whose open-source nanoGPT was the model trained on orbit, agreed.

By March 2026, Starcloud was a unicorn, valued at $1.1 billion after a $170 million Series A led by Benchmark and EQT Ventures. It is the fastest company in Y Combinator's history to reach that status. Seventeen months from demo day to billion-dollar valuation.

Then the filings started, and things got genuinely strange.

Starcloud asked the FCC for permission to operate a constellation of up to 88,000 satellites. SpaceX, which had just merged with xAI into a $1.25 trillion company, the largest private merger ever recorded, filed for up to one million orbital data centers. NVIDIA announced space-grade chips. Google published radiation tests on its TPUs and revealed plans for 1.6-terabit-per-second optical links. Blue Origin, Axiom, and a Chinese plan for a 200,000-satellite constellation all joined the race. Eight separate organizations made serious moves within ninety days. A Robinhood co-founder raised $50 million to beam power from orbit and run AI workloads on the same satellites.

Elon Musk has predicted that within five years, more AI compute will launch into space annually than currently exists in cumulative total on Earth. "A few hundred gigawatts per year of AI in space and rising," he said on the Dwarkesh Podcast in February. That quote was one of the stated motivations for the SpaceX-xAI merger itself.

So, what is actually going on here, and should you believe any of it?

The underlying physics is genuinely attractive. Before poking holes, it is worth being saying there is a possibility this could work.

In low Earth orbit, you get near-constant exposure to sunlight. Solar panels work about 90% of the time up there, compared to far lower capacity factors on Earth's surface. Power is abundant, and over time, the cost per watt falls substantially once the capital expenditure of launch is amortized. Starcloud's CEO has claimed ten times lower energy costs and ten times lower lifetime carbon than an equivalent terrestrial facility. That claim is aggressive, but it is grounded in something real.

Heat dissipation is the other half of the argument. Terrestrial data centers consume enormous quantities of freshwater to cool their servers. In 2024, the sector was already drawing on the water supplies of drought-stressed communities across the American Southwest and Southeast. In orbit, you radiate waste heat directly into the vacuum of space. No cooling towers. No water. No fights with local governments about aquifer depletion.

And then there is the zoning problem. Building a large data center on Earth now takes years of permitting, community opposition, grid interconnection queues, and political negotiation. Orbit has none of that. No neighbors. No county commissioners. No NIMBY campaigns. You just need a rocket and an FCC license, and the FCC process, while slow, is at least predictable.

Then you read the people who study this for a living, and it becomes more complex.

Writing in The Conversation, two professors of data-center and space-systems design lay out constraints that terrestrial facilities simply never face. Radiation is the first. Low Earth orbit exposes hardware to cosmic rays and charged particles from the sun. These corrupt data and degrade chips over time in ways that ground-level shielding prevents. Radiation-hardened chips exist, but they trade performance for durability, and the tradeoff is painful at today's margins.

Thermal management is the second, and it is the one that stops most engineers cold. Data centers generate enormous quantities of waste heat. Radiating that heat into space sounds elegant until you do the numbers: dissipating ten megawatts of waste heat requires radiator panels roughly the size of two football fields. Modern hyperscale data center campuses can run at hundreds of megawatts. The math gets uncomfortable fast.

The GAO flagged the collision and debris risk in a spotlight report this April. Its key takeaway: a significant increase in the number of orbital structures could be difficult to manage and would raise the probability of collisions. Orbital data centers, unlike satellites optimized for longevity, may be decommissioned frequently as hardware generations turn over. Each decommissioning is a reentry event. Many decommissionings create debris fields. The report does not say orbital data centers are impossible. It says the debris math needs to be solved before the constellation math makes sense.

MIT Technology Review asked four hard questions about what would actually need to be true for this to work at scale: radiation-tolerant hardware that does not sacrifice too much compute, thermal rejection systems far beyond current technology, maintenance regimes for hardware that cannot be easily serviced, and orbital traffic management for constellations measured in the tens or hundreds of thousands. None of those are solved problems today.

And then there is the question of connectivity. Data centers work because they are close to the networks and users they serve. Low Earth orbit satellites are moving. They are overhead for minutes at a time before another takes their place. Latency and handoff protocols for that kind of compute are a genuine engineering frontier, not something you paper over with a press release.

SemiAnalysis published a detailed total cost of ownership analysis in June 2026 that is the most rigorous independent assessment published so far. Their headline number: space compute runs at roughly $8.64 per GPU-hour today versus $2.37 on the ground. That is a 3.6x premium, driven primarily by launch costs and the gap in hardware lifetime. Orbital hardware lasts about five years before radiation damage and component degradation make replacement necessary. Terrestrial hardware runs 15 years with maintenance.

I do very small scale angel investing in startups, which means I see these proposals constantly. The "put it in space" pitch has become a recurring character over the past eighteen months.

Terrestrial power is tightening, land is scarce, communities are pushing back. Space sidesteps all of it. The slide deck looks clean. And then you ask specific engineering questions, and the conversation gets vague in very familiar ways.

I am not against exploring space, I want to be clear, because this critique gets misread. GPS, weather satellites, communications infrastructure: the return on sixty years of orbital science is genuinely staggering. However, when we talk about putting AI compute into orbit, what problem are we actually solving, and for whom?

From where I sit, a lot of these proposals are not really about solving a problem. They are about moving a problem somewhere it cannot be easily seen. The energy costs are real. The water stress is real. The community opposition to large data center buildouts is real. Space does not eliminate those costs. It relocates the visual of them. The launch emissions go up, the orbital debris accumulates, the hardware replacement cycle churns on a timeline of years instead of decades, and the bill lands somewhere. The bill always lands somewhere.

The strongest argument for orbital data centers is that they allow AI's energy appetite to keep growing without paying the local costs. No water stress in Arizona. No grid strain in Virginia. No community opposition in Georgia. The problems disappear from the balance sheet because they move to orbit, where there is no balance sheet yet.

We are contemplating launching a million satellites at four times the current terrestrial cost to avoid doing work we already know how to do on the ground. We know how to build solar and storage. We know how to modernize grids. We know how to site data centers near abundant water or in climates that do not require intensive cooling. We know how to schedule heavy compute for off-peak hours. We know how to use models that are appropriately sized for the task instead of always defaulting to the largest one available.

On that last point: Sasha Luccioni has argued in Time that a significant share of the AI energy problem is a model selection problem. Her research has shown that models up to sixty times smaller can match larger ones on many common tasks. The infrastructure conversation focuses almost entirely on supply: more power, more cooling, more compute, eventually more rockets. The demand side, running leaner models when lean models are sufficient, gets far less attention. The cheapest gigawatt is the one you never needed.

None of this makes orbital compute a scam. Some workloads genuinely belong in space. Processing Earth observation data where it is generated rather than beaming it down in bulk is a real use case. Satellites that watch for wildfires, track floods, and monitor shipping lanes already have compute on board. The question is whether AI workloads that currently live in terrestrial hyperscale facilities should migrate to orbit, and the honest answer is: some of them, eventually, maybe, for premium applications where the physics advantages outweigh the cost premium.

The worry is different. The worry is that "just launch it into space" becomes a story we tell ourselves to avoid the unglamorous, achievable fixes on the planet we actually inhabit. Water-stressed communities need relief this decade. Grid instability is a present-tense problem in Texas, the UK, and South Asia. Carbon from the data center sector is climbing now. 2040 parity with a space solution does not help those problems in 2026.

Fund the moonshot. Seriously. Starcloud's H100-in-orbit demonstration is genuinely impressive engineering. If launch costs keep falling and radiation-hardening advances on schedule, orbital compute will be a real option for real workloads within the decade. Some missions, training models on Earth-observation data, running compute aboard climate monitoring constellations, powering low-latency edge compute in places terrestrial infrastructure cannot reach, are genuinely better served from orbit. The research should continue.

But do not let the moonshot launder the present.

The real challenges are solvable now with technology that already exists. Energy efficiency standards for AI training runs. Better scheduling to shift workloads to times and places where renewable power is abundant. Mandatory water reporting for the data center sector. Model efficiency as an engineering discipline, not an afterthought. Grid investments that move faster than the interconnection queue currently allows.

The most radical thing this field could do in 2026 is not escape the planet. It is fix the data center down the road. And then ask the harder question: what is all of this compute actually for?

We debate the energy source, the cooling method, whether the satellite goes up this year or next. We rarely stop to ask what the output of all this infrastructure is supposed to do for people who are not already well-served by existing technology. Billions of people still lack reliable electricity, clean water, and basic healthcare. We are about to spend trillions accelerating AI systems whose primary beneficiaries are concentrated in a pretty narrow band of the global population.

AI Operations Manager
Bees & Bears · Berlin, Germany
Build AI-driven workflows at a climate FinTech financing solar, storage, and heat pumps for households, with €500 million committed to greening 25,000 homes.

Data Scientist, Supply Chain Analytics
Amplify · Remote (US)
Build machine learning models that forecast demand and cut waste for an education company reaching more than 18 million students across all 50 states.

Senior Advisor, Pollution & Public Health Campaigns
Climate Power · Remote (US) · $175,100 to $209,090
Run hard-hitting campaigns holding polluting industries accountable for their health impacts, at a strategic communications shop focused on the politics of climate.

Head of Francophone Africa
LemFi · Remote (Europe-based team)
Own the P&L for remittance corridors serving Francophone diaspora communities, at a fintech moving over $1 billion a month to underserved markets across the Global South.

Project Manager, Blue Economy
FUNDES, IDB Lab project · Sinaloa, Sonora, or Baja California, Mexico
Lead an IDB Lab-funded blue economy project in the Gulf of California, strengthening coastal communities and small enterprises around a fragile marine ecosystem.

Senior Strategist
Future Currents · Remote (US)
Design scenario-planning exercises that help movement organizations prepare for authoritarian threats, economic shifts, and the crises that actually keep organizers up at night.

Investment Operations Associate
AngelList · New York or San Francisco · $150,000+
Build the operations behind a fund broadening public access to venture capital, in a role that explicitly wants someone applying AI to build better systems.

Infectious Disease Technology Lead
Wellcome Trust · London, UK (hybrid)
Shape trustworthy data science to fight infectious disease in the world's most affected communities, part of a £16 billion, ten-year health research commitment.

Energy Data Officer (several positions)
International Energy Agency (OECD) · Paris, France · from 4,556 EUR/month, tax-exempt
Build the energy data that underpins the IEA's work on secure energy transitions, central to the agency's modernization and its use of new data methods.

Data Architect, Pathogen
Ellison Institute of Technology · Oxford, UK
Design the data backbone of a global pathogen metagenomics system, using whole-genome sequencing and cloud technology to catch outbreaks earlier.

Program Director, AI for Education
Team4Tech · Remote (US) · see featured pick above
Also worth a direct look if you skipped the top: leading AI capacity-building for education nonprofits in low-resource settings worldwide.

Fellow, The Upshot
The New York Times · New York or remote
Report data-driven stories for the Times' Upshot desk, turning statistics, elections, and economics into journalism people actually read.

Technology Communities Coordinator
GSMA · London or remote
Coordinate the mobile-industry communities working on digital inclusion, climate, and connectivity for the world's least-connected populations.

Research & Methods Lead, Partnership for Battery Action
Global Development Incubator · Remote · closes July 17
Build the evidence base for cleaning up battery supply chains, from mines to recycling, across the Global South.

Emerging Voices Fellowship
Center for Humane Technology · Remote · $30,000 · closes July 12
Six months and a stipend to develop your public voice on how technology should serve humanity, with mentorship from the team behind The Social Dilemma.

Developer Grants Program (up to $200,000)
Supercell · Game studios based in Africa · closes August 9
Equity-free grants of $20,000 to $200,000 for African game development studios, backing local teams building the continent's games ecosystem.

Innovation Professorships (Associate Professor / Professor)
Aarhus University · Denmark · closes September 10
Senior posts for researchers commercializing science for the public good, with themes including water technology and catchment-scale water challenges.

Deal Analyst
New Majority Capital · Remote (US)
Help underrepresented entrepreneurs buy and grow small businesses, using data and financial modeling to close the ownership gap.

Impact joke of the day. A startup founder told me his data center would run in orbit to save water. I asked how he'd fix a broken server 500 kilometers up. He said that was a Series B problem.

Starcloud trains the first AI model in space (CNBC). The moment the race turned real: an H100 in orbit, a small model trained on the works of Shakespeare, and a version of Google's Gemma running 500 kilometers up. The optimistic case in the founders' own words.

Intriguing on paper, brutal in reality (The Conversation). Two engineering professors on the radiation, the football-field-sized radiators, the repair costs, and the debris risk. The clearest sober counterweight to the hype.

What the US government is watching (US GAO). A short, plain spotlight on the policy questions: crowded orbits, radiation corrupting data, and data centers that may be decommissioned often enough to worsen space debris and reentry risk.

Google plans to put data centers in space (The Guardian). How a mainstream hyperscaler, not just a startup, is now treating orbital compute as a real answer to AI's energy demand, and why that shift matters.

Jobs, jobs, jobs. The roles above are a handful of this week's picks. For the full, continually updated set of social-impact and tech-for-good openings, the PCDN job board adds new postings daily, and our own AI for Impact Career Platform tracks roles at the AI-and-impact intersection.

Person to follow: Sasha Luccioni, AI and former Climate Lead at Hugging Face and co-founder of the new Sustainable AI Group. She built the tools that measure AI's carbon and water footprint and makes the case, better than almost anyone, that smaller models and smarter siting beat brute-force scale.

A role we missed, feedback on our career platform, or a strong disagreement with my take on orbital data centers: reply to this email, or use the form below. A human reads every message.

Know someone weighing the hype against the real work? Forward this along. They can subscribe at impactai.beehiiv.com.