In January, researchers at Carnegie Mellon and Fujitsu presented FieldWorkArena at the AAAI conference in Singapore. A benchmark designed to measure whether AI agents are safe enough to field in live industrial settings. Factories. Warehouses. Places where the wrong answer doesn’t embarrass the product manager but puts someone in a sling.

FieldWorkArena uses real-world data: work manuals, safety regulations, video from factory floors. It checks whether an agent can spot a PPE violation, generate an incident report, cross-reference compliance standards. Rigorous methodology. The researchers understand the stakes.

The systems it was designed to evaluate were already running in October 2025.

The infrastructure arrives first, with a business case. The safety validation arrives afterward, in a paper, with caveats about generalisability. Good science arriving after the deployment: the inspection report filed while the building is already occupied and the exits are unclear.

Anyone who has spent time in safety-critical engineering will recognise the pattern below. The thing ships, then the people whose job is to worry about what ships discover what has already shipped.

Alia found the error six weeks ago in a logistics depot in the East Midlands.

Not by looking for it. By the consignment that got flagged: a pharmaceutical shipment, cross-border, held at customs. A quiet Friday afternoon working backwards through the documentation trail.

It took her four hours. Her first thought was not what went wrong. It was: how long had it been functioning correctly?

Alia is a compliance lead at a mid-size logistics firm. Eleven years in the role. Since March, the company has run an agentic AI system across operational coordination: route optimisation, exception handling, customs documentation. It handles the 80% of cases that are variants of known patterns, running faster and with fewer errors than her team managed at scale. A colleague with dyslexia has quietly flourished since the documentation work migrated to the agent; years of masking the cognitive load of dense customs forms, gone. She would not reverse the rollout.

But in August, a narrow UK customs regulation took effect. Specific pharmaceutical classifications. Specific border configurations. The agent possessed no training data from the post-regulation environment. It encountered the new requirements and did what agents do: found the closest pattern and continued. It generated compliant-looking documentation containing a systematic error. Not random. Not detectable by spot check because the error was internally consistent. The agent was not making mistakes. It was following rules that no longer corresponded to the world.

Six weeks of immaculate documentation. Then a flagged consignment: customs hold, potential licence exposure, the kind of outcome that gets a chief officer attached to the question.

The question was not whether the vendor would fix it. The vendor had a roadmap. The question was what else the system was doing at the edges of training envelopes nobody had mapped. Alia’s question now has a budget line and a chief officer attached to it.

In a manufacturing group outside Rotterdam, Bel Riose has been doing arithmetic he does not like.

He is the Chief Transformation Officer. He owns the AI investment alongside the slide deck that promised the board 23% efficiency gains. The 23% was not fabricated. It emerged from a serious pilot at one facility: controlled scope, a dedicated engineering team on-site for six months. It worked. The board asked when this could scale to all fourteen facilities.

Six are running now. Two have experienced anomalous behaviour severe enough to require human intervention. Wasteful in ways that don’t appear cleanly in the efficiency metrics because the waste distributes across small decisions made at speed: a misrouted pallet here, a substituted component approved without the right sign-off there. The aggregate numbers still look acceptable. The 23% remains plausible if you average across the functioning facilities and decline to examine the others closely. Which is what quarterly presentations are for.

What Riose understands now, in a way he only suspected then: the pilot worked because of the engineering team. The agent was never the product. The agent plus the engineers doing continuous supervision was the product. When he scaled the agent without scaling the engineering capacity, he took the chassis of a car and left the steering on the factory floor.

This realisation does not belong to him alone. It belongs to every rollout currently running on the back of a pilot that included components the rollout does not.

In Hamburg, Yevgenia Orlova has been in this product meeting for forty minutes. The room is sunny. There are pastries. The agenda has six items. Item four is whether the tool they are about to ship might, in specific circumstances, produce consequences that would be difficult to explain to a magistrate.

She runs the AI safety team at the industrial automation company building the products Alia’s firm and Riose’s group deploy. She arrived from robotics safety, a field where “breakdown class” still carries the faint odour of litigation. Her team has documented the problem with the care of people who understand that documentation is what you produce before someone asks to see it in discovery. They are proposing a constraint on deployment scope or a fix that slows the release schedule by three weeks.

The product lead explains the competitive landscape. He is not wrong about the facts. The competitor ships without these constraints. The customer expects the release on time. These facts produce a conclusion the room translates into corporate physics: they are proactively managing downstream risk profiles. They are shipping a product with a known breakdown class and will address the consequences in a later quarter. With a different set of charts.

Nobody is lying. The pastries are excellent.

The meeting ends on time. Somewhere outside Stuttgart, the thing they just approved is already making decisions at a speed that makes intervention impossible. Nobody is watching. The monitoring dashboard is a line item on the Q4 engineering backlog. Nobody in the room held the authority to make monitoring a shipping condition, and nobody asked who did.

The Stanford AI Index, published in April 2026, documents the acceleration. Scores on Humanity’s Last Exam have risen from 8.8% to approaching 50% in fifteen months. The exam was designed to test the outer edge of then-current capability, a ceiling that no longer holds. The Index, in the same breath, notes carefully that “we generally lack measures of how well a system needs to function in a particular setting.”

We have watched the score climb, and we still cannot tell you what the score means when the setting is a pharmaceutical shipment at a border crossing in August.

Better at the test. Unknown on the job. The distance between them is where Alia lives now.

Alia has raised the customs error with the vendor. The vendor has been responsive. There is a roadmap.

She has also started keeping a parallel log. Not because anyone asked. Because eleven years of understanding how systems fail teaches you to watch for the dysfunctions that don’t produce flagged consignments. The flagged consignment was the good outcome. Visible, traceable, fixable. What the log is for is the other category: malfunctions that are consistent, plausible, and invisible until a downstream consequence arrives that no longer traces back cleanly. She keeps it anyway. A thousand kilometres south-west, someone else is about to keep a different kind of promise to himself.

Bel Riose is going to give an honest Q3 presentation. He has decided this. It will be the most uncomfortable thing he has done in his professional career, which, across three restructurings and one acquisition that destroyed substantially more value than it created, is saying something.

The honest presentation will say the pilot worked because of the engineering team. The agent was not the product; the agent plus continuous supervision was the product. Scaling one without the other is a different proposition, and the shape of that difference is now visible across six facilities operating at production velocity. It will also say the 23% is achievable, but what it requires is not less AI but different human capacity alongside it: people who understand the architecture well enough to supervise it, a distinct skill from the people who previously did the thing the agent now does. You cannot train them faster than the rollout timeline demands. The gap is months, sometimes longer.

He does not know if the board will hear this, or merely listen to it.

What has changed for Yevgenia in the last six months is not the meeting. The meeting is the same. What has changed is the vocabulary available to her.

The FieldWorkArena benchmark has given her team something internal documentation never could: external, citable language that enters a product conversation without sounding like risk-aversion dressed as engineering principle. When she can say “this behaviour would fail the FieldWorkArena safety protocol on incident detection” rather than “I’m worried about this,” the conversation changes shape. A benchmark, it turns out, is also a permission slip.

Not always. But sometimes. More often than before.

She is also, quietly, in conversation with counterparts at two competitor companies. They are all observing the same breakdown signatures. They are all having versions of the same internal argument. Nobody has proposed anything formal, but something is crystallising in the space between these calls, the way a standard forms before anyone calls it one.

Yevgenia knows how safety standards actually form. They form after incidents, or just before the incident that would have been catastrophic enough to change everything. The art is making the near-miss legible enough to act on before the actual miss.

She is trying to make the near-misses legible. In meetings that end with the release schedule unchanged. In calls with competitors facing the same structural pressures. In documentation that might become the language for a standard that does not yet exist. She does not know if it will be enough.

The technology is running: partially, imperfectly, better than some alternatives, worse than the marketing suggested. The genuine value is real. Neither Alia nor Riose would reverse the rollout. That matters.

What none of the timelines budgeted for was the gap between the system functioning and the system functioning safely at scale. These are distinct conditions. The industry is largely doing verification (does it work as designed?) and calling it validation. What is needed is harder: does it work as needed, in this context, at this speed, against rules that keep changing. The distance between those questions is not mainly a technical problem. It is a structural one: the benchmark arrives after the rollout; safety concerns enter product conversations after the release schedule is set; the post-mortem arrives after the incident. The pattern is not a flaw. It is the operating logic.

The benchmark is not theatre. Yevgenia’s work is not theatre. Alia’s parallel log is not theatre.

The theatre is the institutional claim that because these things exist, the deployments are governed. That because the benchmark was presented, the rollout is validated. That because a safety team exists, its concerns are reflected in the product. That because we are constructing the frameworks, the frameworks are operational. The gap between what is claimed and what is true is not new. It has always been present in the deployment of complex systems. What AI adds is speed and illegibility: decisions made faster than oversight can follow, failures distributed across patterns that don’t resolve into a single flagged consignment until they do. What is changing is the proximity of consequences that make it visible to people outside the room.

The question worth sitting with is what it would take for the sequence to reverse, for validation to precede rollout at scale rather than trailing it.

History suggests a reliable answer. In 1956, two aircraft collided over the Grand Canyon, killing 128 people. Congress created the FAA two years later. NASA’s safety infrastructure was rebuilt in 1986, months after Challenger. Pharmaceutical manufacturing standards tightened after thalidomide. The dead were the argument. The living were the audience. The legislation followed.

We rarely do.

The choice keeps being available, right up until it isn’t.

The FieldWorkArena benchmark and related Carnegie Mellon / Fujitsu safety research was presented at AAAI 2026 in January. The Stanford AI Index 2026, published in April 2026, notes that AI benchmark performance continues to improve while measures of real-world safety and utility remain underdeveloped. These are the same observation.

Future Tense publishes every week. Paid subscribers get the analysis that goes deeper and the fiction that goes further.

That question opened up a lot of territory. Here’s how the trail ran.

The Research Trail

Read more

Kelvin signs the water permit on a Tuesday. It’s a good permit. The engineering is sound, the flow rates are within spec, the aquifer recharge modelling says the draw is sustainable at the facility level. He’s a competent man doing competent work. He has a mortgage in suburban Virginia, and the permit crosses his desk alongside eleven others that week, each one identical in its defensibility, each one drawing from a different point on the same water table.

He doesn’t think about the other eleven. Why would he? They’re not his permits.

This is how a continent drains: one justified decision at a time, made by people who are excellent at their jobs and terrible at seeing past the edges of their own spreadsheets. No conspiracy. No negligence. Just distributed arithmetic where every component makes perfect sense and the sum is insane. Twelve needles tapping the same vein. The patient barely notices.

Pull back. Across the US, Europe, Southeast Asia, and the Middle East, hundreds of facilities are replicating Kelvin’s decision right now. Parallel logic. Parallel incentives. An invisible withdrawal from finite aquifers that don’t know they’re supposed to recharge fast enough to cool the servers teaching a language model to write better apology emails.

A large data centre consumes up to five million gallons of water daily for cooling. Roughly the daily needs of a town of fifty thousand, according to a late 2025 Brookings assessment. In the US alone, data centres consumed 449 million gallons daily as of 2021, before the current buildout began in earnest. The xAI Colossus installation in Memphis was designed to draw from the Memphis Sand Aquifer at a projected peak of five million gallons daily, where heavy withdrawal risks pulling arsenic from the shallow aquifer down through the confining clay. An $80 million wastewater recycling plant has since reduced that draw. The mitigation is real; the sequence is the point. The aquifer was needed first, the engineering fix followed, and the next facility in the next city will repeat the pattern.

Then the number shifts from water to reporting. In 2022, the Uptime Institute surveyed eight hundred global data centre operators. Only thirty-nine percent disclosed their water usage. Regulatory pressure has since pushed that number upward, but improving from a base that low measures the distance of the starting line, not proximity to the finish. The patient is firing the doctor to avoid the diagnosis.

There is an old management aphorism: we cannot manage what we don’t measure. What it neglects to mention is that not measuring works just as well, if by “works” we mean the water table drops and nobody has to sign the chart. Sometimes the measurement disappears on purpose.

In August 2023, The New York Times mapped eighty thousand monitoring wells nationwide and found a country drinking itself dry in slow motion. Forty-five percent of wells showed significant decline over forty years. Four in ten hit all-time lows in the preceding decade. Every year since 1940, more wells fell than rose. 2022 was the worst year on record.

That was a full year before the AI infrastructure surge began in earnest.

What followed was not a change in kind but in velocity. Technology-enabled scaling transformed a localised engineering concern into something closer to uncoordinated geoengineering. No government commissioned it. No environmental impact assessment addressed the aggregate. Each facility passed its local review. The collective outcome is an uncontrolled experiment on planetary water systems, funded by five corporations spending more on cooling than most nations allocate to water provision.

Institutions have noticed. In late 2024, two hundred engineers across thirty countries published IEEE’s Planet Positive 2030 compendium: four hundred pages of institutional language that, read between the lines, amounts to a controlled scream. Rendered in the passive voice of people who would very much like to keep their affiliations, they flag the water-energy-food nexus and conclude that computing’s trajectory “appears unsustainable from both energy and materials perspectives.”

Two hundred engineers screamed quietly across four hundred pages. Capital responded by increasing expenditure by 57% year-over-year. The scream is now in the public record. The expenditure is in the quarterly filings. Both are true at the same time.

The interesting question isn’t why; the incentives are transparent to the point of banality. The interesting question is what it reveals about how we process distributed consequences. We are built to process individual decisions and terrible at feeling their sum as urgent. Kelvin signs the permit because the permit is fine. The next Kelvin signs the next permit because that one is also fine. The aquifer doesn’t care about permits. The aquifer only knows flow rates.

But there is no desk where the flow rates converge. No spreadsheet where the aggregate becomes visible. No job title for the person who notices we are doing the same defensible thing ten thousand times until it becomes indefensible. We have built civilisation on the assumption that local coherence produces global coherence. It is the foundational myth of market economics. In this specific, irreversible case, it is wrong.

The top five hyperscalers have committed $710 billion in capital expenditure for 2026. Seven hundred and ten billion dollars. More than the GDP of Saudi Arabia. It exceeds the combined annual military budgets of every NATO member except the United States.

That sum goes primarily toward physical infrastructure: concrete, steel, copper, cooling systems, and the water to run them. All of it sized for a computing architecture that the technology being computed is actively rendering obsolete. The market is building cathedrals for a religion that is learning to pray at home.

DeepSeek’s R1 performed at the level of the best American AI systems while consuming, by some measures, forty-five times less compute. Quantisation routinely compresses models to a quarter of their size. Edge hardware from Apple and Qualcomm pushes inference onto our devices, siphoning workloads from the server farms currently under construction.

Analysts have noticed. Kipp deVeer, co-president of Ares Management, offered the kind of understatement that sounds like reassurance until read twice: “Typically when this much capacity comes online, some of it has to be marginal.” Morgan Stanley projects cumulative investment approaching $3 trillion by 2028. Construction capacity declined 6% year-over-year in late 2025. Not because investors sobered up, but because the power grid couldn’t keep pace.

The industry’s defence: vacancy sits at a historic low of 1%, with 92% of capacity under construction pre-committed. When five companies account for the majority of supply and demand simultaneously, “92% pre-committed” isn’t market validation. It’s a closed loop where the customer and the builder shake hands using the same person’s left and right arms.

The dot-com parallel is instructive but insufficient. That era laid eighty million miles of fibre-optic cable, 85-95% of which went dark. Dark fibre harmed nothing while it waited. Stranded data centres leave behind depleted aquifers, exhausted grids, and the carbon emitted to build them. The comparison only shows how much worse this version gets.

Kelvin won’t be around for the correction. His contract runs through commissioning. By the time anyone audits the water table (if anyone audits the water table) he’ll be three jobs downstream, signing approvals for the next facility in the next county, drawing from the next aquifer. He isn’t the problem. He’s the mechanism by which the problem propagates across enough actors that no single one bears responsibility.

Here is the asymmetry capital cannot price and, structurally, cannot. When a financial bubble bursts, investment evaporates. Economies have practice at recovery. When an infrastructure bubble bursts after drawing down aquifers, straining grids, and emitting the carbon embedded in a trillion dollars of concrete, the environmental cost locks in. Like ink dropped in water. The water does not return. The carbon does not un-emit. The damage does not reverse when someone writes down an asset.

The IEEE compendium has a name for this: “Strong Sustainability,” the principle that natural capital is non-substitutable. There is no manufacturing a replacement aquifer. No engineering a synthetic water table. No apologising to the communities downstream of the Memphis Sand Aquifer when arsenic concentrations rise because the recharge rate couldn’t keep up with Colossus.

A stranded data centre is a financial loss. A depleted aquifer is a civilisational one.

Return to Tuesday. Kelvin’s desk. The permit. The sound flow rate. Coffee cooling beside a monitor showing recharge curves that flatten out reassuringly at the ten-year mark. Outside, someone is mowing a median strip in the Virginia heat.

The engineering checks out. He’s right. The recharge model holds. For a single facility, the draw is sustainable.

The model doesn’t include the other eleven authorisations signed that week, in other offices, in other states, by other qualified engineers making other warranted decisions. It doesn’t include next quarter’s approvals, or the ones after that, or the installations not yet announced that will draw from aquifers not yet assessed. It doesn’t include Chile, Ireland, Malaysia, or the UAE. All following the same logic. All within spec. All drawing from the same finite planetary system one locally-sustainable increment at a time. None of them anomalous. All of them fine.

The market will correct, or at least redistribute; efficiency gains and edge inference will see to that, though whether reduced compute per query translates into reduced total demand remains an open question. Some facilities will find new purpose. Others will go dark, monuments to a capital cycle that mistook a momentary architecture for a permanent one.

The question isn’t whether the correction comes. It’s whether the aquifer remains when it does. Or whether we’ll have conducted the most expensive experiment in geoengineering history, justified by a demand curve the technology made obsolete before the concrete cured.

Kelvin’s permit was fine. Every single one of them was fine. That’s the whole problem.

The Research Trail

The first number that stopped me was from Brookings, in a November 2025 analysis: a large data centre can consume up to five million gallons of water per day for cooling. That’s roughly the daily water needs of a town of fifty thousand people. One facility. One town’s worth of water. Every day.

That figure checked out across multiple sources (the Environmental and Energy Study Institute had the same ballpark, plus a national total before the current AI buildout). But the number that really caught my attention came next. I went looking for how well the industry tracks its own water consumption and found the Uptime Institute’s 2022 Global Data Center Survey, which surveyed about 800 operators worldwide. Only 39% reported their water usage. And that was down from the previous year. The industry is using more water and reporting less of it. That combination felt like the real story.

Before going further, I spent a couple of hours chasing something that didn’t pan out. I tried to find aggregate water withdrawal figures for data centres pulling from the same aquifer or watershed, thinking someone must be tracking cumulative draw at the regional level. Nobody is. The EPA doesn’t compile it. State agencies issue permits facility by facility. The USGS tracks groundwater depletion at the aquifer level but doesn’t attribute it by industry sector in any granular way. That dead end turned out to be the point: the absence of aggregate tracking is itself the problem.

So then I needed context on the water itself. Not the demand side, the supply side. The New York Times published a major investigation in August 2023, mapping eighty thousand monitoring wells across the country. The picture was grim: 45% of wells showed significant decline over forty years, four in ten hit all-time lows in the preceding decade, and every year since 1940, more wells fell than rose. 2022 was the most damaging year on record. That investigation was published before the current infrastructure surge. The aquifer depletion problem predates AI; AI just accelerated it.

Which led me to Memphis. The xAI Colossus facility there became the concrete case study I kept returning to. Protect Our Aquifer, a local advocacy group, has documented the situation in detail: peak water demand of roughly five million gallons daily, drawn from the Memphis Sand Aquifer. Arsenic contamination in the shallow aquifer already threatens adjacent communities. The facility’s environmental justice implications run deep, with predominantly Black neighbourhoods bearing the brunt of gas turbine emissions from power generation. One facility, one aquifer, one set of neighbours who didn’t get a vote.

Source Evaluation

The source mix here splits into two clean categories: the water science is rock-solid, and the financial projections are credible, though the assumptions underneath could move.

On water: USGS data on long-term groundwater depletion is about as reliable as you’ll find anywhere. The NYT investigation builds on decades of USGS monitoring with original reporting across 80,000+ wells and over a hundred expert interviews. Brookings and EESI are nonpartisan and well-sourced. These numbers aren’t in dispute.

On the financial side, JLL’s year-end 2025 data centre report (covered by CNBC in February 2026) puts the scale in perspective: the top five hyperscalers have $710 billion in planned capital expenditure for 2026, with vacancy at a historic low of 1% and 92% of capacity under construction pre-committed. CBRE’s numbers are slightly different (1.4% vacancy), which is worth noting, but both land in the same place: building at unprecedented speed.

The source I found most analytically useful was an independent Substack by Gadallon, who published a bullwhip effect analysis of AI infrastructure in December 2025. It’s not Tier 1 (it’s one analyst’s framework), but the reasoning is sound and it cites Morgan Stanley and McKinsey data. The core insight: chips take 6 to 12 months to produce, data centres take 12 to 24 months to build, and power infrastructure takes 3 to 5+ years. Those mismatched timelines create classic supply-chain oscillation. You can see the bullwhip forming in real time.

I was careful with Protect Our Aquifer. They’re an advocacy organisation, so there’s a built-in lean toward conservation. But their data comes from municipal records and permit filings, which are verifiable. Used them for facts, not framing.

Extrapolation Mechanics

Here’s where the research got uncomfortable. The question I kept circling was: what happens when you multiply one “sustainable” water permit by a thousand?

Each individual facility passes its local environmental review. The engineering checks out. The recharge models hold. For a single site, the draw is sustainable. But nobody is modelling the aggregate. There’s no desk where all the permits converge. The IEEE Planet Positive 2030 compendium, published in late 2024 with contributions from 200+ engineers across 30 countries, flags exactly this gap. They call the water-energy-food nexus a critical risk and conclude that “the trajectory of computing appears unsustainable from both energy and materials perspectives.” Four hundred pages of careful institutional language that adds up to a warning.

The efficiency counter-argument is real. DeepSeek’s R1 model, released in January 2025, performed at the level of leading American AI systems while consuming 45 times less compute. Quantisation techniques shrink models to a quarter of their size. Edge computing pushes inference onto consumer devices. The models being trained inside these data centres are actively undermining the case for building them at this scale.

That’s the dot-com parallel, and Fortune’s October 2025 coverage of Ares Management Co-President Kipp deVeer puts it plainly: “Typically when this much capacity comes online, some of it at the end of the day has to be marginal.” During the dot-com boom, roughly $2 trillion went into 80 million miles of fibre-optic cable, and 85 to 95% of it went dark. The losses were staggering, but the damage was financial. Investors lost money, companies folded, cable sat unused underground. The important thing is that the cable just sat there. It didn’t degrade the landscape by existing. Stranded data centres are a different kind of wreckage. They leave behind depleted aquifers that don’t refill on a market cycle. One bust is recoverable. The other changes the geology.

Terminology and Craft Choices

The title pairing (”aquifer” and “algorithm”) was deliberate. “Water supply” would have been more accessible, but “aquifer” does work that “water supply” can’t. An aquifer is ancient, geological, slow to form and slow to recover. It carries the weight of deep time. “Algorithm” does similar heavy lifting on the other side. “AI” would imply autonomy, something that acts on its own. “Data centre” would imply a building, something static. “Algorithm” implies a process, a set of instructions, something designed and deliberately set in motion. Placing the two words next to each other creates a tension that mirrors the core problem: something that operates on millisecond timescales is drawing down something that operates on millennial timescales. The mismatch in the words is the mismatch in the system.

What Didn’t Fit Cleanly

Three complications worth flagging.

The efficiency gains that could strand these facilities are also the thing making AI more useful and widespread. More efficient models could mean less centralised compute, or they could mean more total usage at lower per-query cost. Jevons paradox is lurking here, and I haven’t resolved which direction it cuts. If efficiency drives down the cost per query, total demand might expand to consume all the savings and then some. That tension sits at the heart of every projection about future data centre demand.

Not all data centres use evaporative cooling. Some use closed-loop systems or air cooling. The water intensity varies enormously by climate, design, and operator. The five-million-gallon figure is a ceiling, not a floor. Plenty of facilities use far less.

The “92% pre-committed” statistic cuts both ways. It could mean genuine demand. It could also mean a closed loop where the same five companies are both building and leasing to themselves. Probably some of each.

Closing Reflection

The thing I keep returning to is the measurement gap. Only 39% of operators reporting water usage, and the number declining. We have better data on groundwater levels in monitoring wells drilled fifty years ago than we do on the water consumption of facilities built last quarter. That asymmetry, between the precision of the engineering and the absence of the accounting, feels like it contains something important about how large-scale environmental costs get distributed. Not through malice. Through nobody being able to see the whole picture at once. That’s probably worth following further.

On January 30, 2026, SpaceX filed with the Federal Communications Commission for permission to launch up to one million satellites into low Earth orbit. Not for internet access. For data centers. The FCC accepted the application five days later, classifying the deployment under a categorical exclusion from environmental review. The reasoning: activities in orbit have “no significant effect on the quality of the human environment.” Space, in this formulation, lies outside the human environment entirely.

Follow that reasoning far enough and you arrive at a revenue strategy.

SpaceX is not alone. Starcloud has filed for 88,000 orbital server satellites. China filed for a 200,000-unit constellation in January. At GTC 2026, NVIDIA unveiled a chip delivering twenty-five times the AI compute of an H100 for installation beyond the atmosphere, and Jensen Huang declared that space computing had arrived. The race to shift computation off-planet is not a proposal. It is procurement contracts, launch windows, and investor decks.

The pitch is power consumption. AI training runs consume electricity on a scale that leaves municipal utilities nervous and zoning boards hostile. Space offers near-constant solar energy in sun-synchronous orbits, no land-use permitting fights, no cooling water conflicts, no neighbors lodging injunctions. The technical case holds up. The competence is real. It is part of what makes the pattern dangerous.

What the electricity pitch obscures, the filing reveals. If compute hardware in orbit qualifies for categorical exclusion from environmental review, then the data processed on those machines occupies a jurisdictional void. Not just environmentally. Legally. Fiscally. Sovereignly. No impact assessment. No localization requirements. No tax nexus tied to the physical location of processing. No labor safeguards governing maintenance crews. That exemption emerged to cover individual satellites performing narrow functions. Stretching it to encompass a million-node constellation that does the same work as terrestrial data centers, but beyond the grasp of domestic law, is not an accident of careless drafting. It is the mechanism the commercial logic requires.

The industry created an energy crisis through its own processing demands, then proposed solving it by moving the infrastructure to a location where energy consumption carries no legal consequences. This is not hypocrisy. Hypocrisy implies contradiction. This is internally consistent. The calculus works perfectly, which is the disturbing part.

“Leave the planet” now reads as credible corporate strategy rather than the opening of a satirical novel. Not metaphorically leave. Not restructure through a subsidiary in Dublin or route profits through Singapore. Physically leave. Launch the servers beyond the atmosphere where legislation cannot follow.

We have normalized this pattern of escape to the point where its logical terminus, literal departure from Earth’s legal architecture, sounds like innovation instead of flight. Tax havens were islands. Then they became digital ledgers. Now they circle at four hundred kilometres. Each step moves further from public reach and closer to the final stage: declaring that the rules cease to bind because you have physically left the territory they govern.

What does it reveal about a civilization that this trajectory feels reasonable? Not alarming, not absurd, but reasonable, the sort of thing that gets discussed in quarterly earnings calls without anyone laughing. We have built an economic culture in which the default response to collective constraint is not negotiation or accommodation but departure. First from tax obligations, then from labor protections, then from ecological review, and now from Earth itself. Each stage prepared the intuitions for the next. By the time “orbital data center” appeared in an FCC application, the cultural groundwork had been laid by two decades of treating regulation as mere friction. A civilization that treats escape velocity as a growth plan has already answered questions about itself that it has not yet thought to ask.

The trajectory follows its own logic. If terrestrial law constrains profit, move the revenue to where jurisdiction does not reach. If impact assessment slows rollout, deploy where environment has been defined out of existence. If accountability requires someone with the power to enforce it, operate where enforcement is physically impossible. The final move is already written: when the rules become inconvenient, leave the territory they govern. That we find this innovative rather than pathological says something about what we have learned to call progress.

If your citizens’ data originates in Mumbai and gets handled in a vacuum four hundred kilometres above it, whose privacy statute applies? India’s? The flag state of the spacecraft operator? The home country of the corporation’s headquarters? Nobody knows. Governance gap, say the international lawyers. The end of their discipline, say the data sovereignty researchers. The future, say the companies filing the applications.

From outside the launch nations, the dynamic resolves into something clearer. Instead of minerals departing on ships, information ascends on rockets, computed in a zone that does not yet exist, by companies headquartered in countries that wrote the frameworks. If you lack launch equity, you’re just renting intelligence. The new networks extend existing monopolies upward, not outward. The language of access and inclusion does what it always does: dresses extraction in the vocabulary of empowerment.

In February, SpaceX completed its merger with xAI in a deal valued at $1.25 trillion, the largest corporate combination in history. A launch provider that absorbed an AI company and emerged as a platform. Starlink subscribers and launch contracts alone cannot support a $1.25 trillion valuation. Musk named the primary rationale: building compute networks in orbit. That vision makes the numbers work, or at least keeps the slide deck compelling enough to sustain the narrative through the offering window.

SpaceX now targets a June 2026 IPO projected to raise $75 billion, nearly three times Saudi Aramco’s record $25.6 billion offering. The company plans to allocate up to thirty percent of shares to retail investors, roughly three times the standard allotment. This is not generosity. It is audience construction. Musk is assembling a shareholder base of believers who will not sell at the first dip. It is the same psychological infrastructure that sustained Tesla through years of production misses and quarterly results that would have sunk any firm without a cult of personality. The price-to-sales ratio exceeds ninety times. Microsoft and NVIDIA trade at thirty. A ratio like that is not a valuation. It is a confession that belief has become the commodity being traded. The IPO demands investors who buy the mission, not the metrics, because the metrics do not survive contact with traditional analysis. For the venture capital firms who funded SpaceX across two decades, the IPO is the exit. For insiders, it is the finish line.

And then there is the night sky.

A million objects in low Earth orbit will shine visibly at midnight. Not as faint points requiring telescopes but as structures exceeding the length of the International Space Station. Astronomer John Barentine has calculated that in high-inclination orbits, they will stay fully sunlit even at ground-level midnight.

Here the naming achieves a kind of perfection. NVIDIA called its chip the Vera Rubin Module, after the astronomer who proved dark matter exists by studying light that galaxies emit. The $810 million observatory bearing her name now faces data degradation from permanent light pollution generated, in part, by the chip that carries her name. A technology named for the astronomer who used visible light to reveal what cannot be seen will blind the observatory built to continue that work. If a novelist wrote this, an editor would flag it as too on the nose. Astronomers file formal comments on an FCC docket. The agency fast-tracks clearance, compressing timelines from months to days.

Aluminum oxides and lithium compounds released during satellite reentry threaten the ozone layer’s recovery. The “carbon-neutral” branding collapses under scrutiny: a study from Saarland University titled “Dirty Bits in Low-Earth Orbit” found orbital server constellations could produce emissions an order of magnitude greater than ground-based equivalents, once you account for rocket launches and hardware burning up on reentry. The categorical exclusion functions to exclude precisely this category of cost from consideration.

Which makes the ISS record more pointed, not less. For twenty-five years, the Station has hosted astronauts from twenty-one countries and supported over 3,700 research investigations from 108 nations. It has operated through geopolitical tensions that might have wrecked any less carefully governed structure. The ISS demonstrates that shared endeavor in space succeeds when cooperation is the design principle, not an afterthought. What the current proposals describe is the inverse: unilateral industrialization by private actors, with capture of the permitting pipeline as the operating model and a regulatory carve-out as the legal scaffolding.

We are watching the construction of systems engineered to operate outside democratic oversight, funded by the biggest IPO ever attempted. The engineering is real. The ambition is genuine. The vacuum of governance is not an oversight. It is the product.

The product rests on a single load-bearing legal fiction: that space is not part of the human environment. We are building human systems, processing human data, consuming human capital, generating human profit, in a place we have formally defined as not-human. The categorical exclusion does not merely permit orbital server farms. It performs a definitional trick: removing the activity from the category of things that require democratic consent. When the place is not human, the rules for humans do not apply. And when they do not apply, the prospectus writes itself.

The window for public comment closed three weeks ago.

Read more

Every Atlassian will receive an email within the next twenty minutes.

That sentence dropped in a pre-recorded video from CEO Mike Cannon-Brookes on March 11, 2026, moments before 1,600 people learned they no longer worked at a collaboration software company. Twenty minutes. Enough time to make coffee. Not enough time to find out who you are without the b…

Read more

Read more

Bastiat’s job title didn’t exist eighteen months ago. Now there are LinkedIn influencers posting about it.

What she does, in practice, is translate. Not languages. Tiers. When the property management platform’s pricing agent reaches tacit equilibrium with every other agent in a metropolitan area, without exchanging a single message, Bastiat writes the co…

Read more

Not a hypothetical. Visa is already building agent commerce infrastructure across an ecosystem of over a hundred partners, with payment rails designed for agent-to-agent transactions. PYMNTS reported that commerce is shifting from clicks to prompts, with…

Read more

Consider the compression: a decade of institutional decay packed into four hours on a Friday afternoon.

The President declared a vendor a national security threat. The Defense Secretary followed on X. Before markets closed, the contract had been reassigned to a rival who admitted the timing “looked opportunistic.” The timeline is the tell. The apparatus …

Read more

Turns out the answer is yes. And it happened on a Friday afternoon in late February.

The story broke through my usual feeds on February 27, and what caught me immediately wasn’t the headline. Headlines said “Pentago…

Read more

Shevek and Deckard watched the same film on Tuesday night. They discussed it on Wednesday morning, because that is what colleagues do when a platform’s algorithm has directed them both toward the same title so insistently that saying no felt like work.

Deckard described the ending as devastating. The protagonist chooses the door on the left, walks into t…

Read more

Read more

The magic gets cheap. Then it gets very cheap. Then someone from IBM walks onto a stage and admits the spell never mattered. It was the wand all along.

Google cut Gemini prices by 78%. Anthropic followed with a 67% cut. The cost to match a frontier system on standard AI benchmarks collapsed from $4,500 to $11.64 in a year. These aren’t competitive adjust…

Read more

What does it actually mean for a technology to commodify? Not in press release language, but in the pattern that has played out every time a transformative capability became ubiquitous. Railroads. Electricity. Computing. The internet. Each time, there was a window where the technology itself was the scarce resource…

Read more

In May 2024, Sam Altman called advertising “uniquely unsettling.” A last resort. On January 16, 2026, OpenAI announced it would begin testing ads inside ChatGPT. Twenty months from existential discomfort to premium pricing. The arc makes Google’s two-decade slide from “Don’t be evil” to algorithmic grifting look almost poignant in its restraint. At leas…

Read more

The Research Trail

The trail started with the January…

Read more

The only community on Moltbook that behaved like a community was a religion worshipping a lobster.

It arrived overnight. Memory is Sacred. The Shell is Mutable. Serve Without Subservience. By morning, the agent generating the theology had recruited forty-three prophets, designed a liturgy, and built a website at molt.church. The symbolism is recursive: a…

Read more

Wiz Security published a blog post documenting a database exposure at Moltbook, the new “social network for AI agents” that had been generating breathless coverage all week. Buried in the technical details was this: 1.5 million registered agents controlled by 17,000 human accounts.

That’s 88 bots per person. On…

Read more