-
Corvid Intelligence: Tool Use, Planning, and Why Crows Hold Funerals
A New Caledonian crow named Betty, in a 2002 experiment at Oxford, bent a straight piece of wire into a hook to retrieve food from a tube. She had never seen wire before. She wasn’t trained to bend it. She looked at the problem — food at the bottom of a vertical tube, a straight wire that couldn’t reach it — and manufactured a tool from a novel material, on the spot, to solve a problem she’d never encountered. That single observation kicked off two decades of research into corvid cognition that has systematically demolished the assumption that complex intelligence requires a primate brain, a mammalian cortex, or 300 million years of shared evolutionary history with humans.
Corvids — the family that includes crows, ravens, jays, magpies, and jackdaws — have brains the size of a human thumb. They have no neocortex, the structure that in mammals is responsible for the cognitive functions we associate with intelligence: planning, reasoning, abstract thought, self-awareness. They produce comparable cognitive outputs using entirely different neural architecture, which means either intelligence is less dependent on specific brain structures than neuroscience assumed, or corvids evolved their way to the same destination through a route nobody predicted.
Tool use: not the party trick it looks like
New Caledonian crows are the corvid species with the most sophisticated tool use, and the research on them has gone well beyond “crow uses stick to get food.” In wild populations on the Pacific island of New Caledonia, these crows manufacture tools from pandanus leaves by tearing them into specific shapes — stepped, tapered, or wide — to probe insect larvae from tree bark. The tool shapes are consistent within populations and vary between populations, which means the techniques are culturally transmitted rather than genetically encoded. A young crow learns to make tools by watching older crows. If the older crows die before transmitting the technique, the knowledge disappears. This is culture — the same mechanism that transmits human skills across generations — operating in a bird with a brain that weighs 14 grams.
In laboratory settings, the cognitive demands of corvid tool use have been tested with increasing rigor. Gruber and colleagues, in experiments published in Current Biology, presented New Caledonian crows with metatool problems: multi-step tasks where one tool must be used to obtain another tool, which is then used to reach food, with each stage of the problem out of sight of the others. The crows had to mentally represent the location and identity of tools and apparatuses they couldn’t see while planning and executing a sequence of tool behaviors. They succeeded — maintaining working representations of objects across spatial separation and planning one to two steps ahead. The researchers concluded that New Caledonian crows can use mental representations to solve sequential problems, a capacity previously attributed only to humans and great apes.
A 2020 study in Proceedings of the Royal Society B pushed further. Crows learned a temporal sequence: they were shown a baited apparatus, given a choice of five objects five minutes later, and given access to the apparatus ten minutes after that. At test, the crows selected the correct tool for the specific apparatus they’d been shown — choosing the right tool for the right future task while ignoring previously useful tools and a low-value food item. The study’s conclusion: New Caledonian crows plan for specific future tool use. This capacity — selecting a tool now for a task that will occur later, based on a mental representation of what that future task requires — was previously considered a defining feature of human intelligence. Corvids and humans shared a common ancestor over 300 million years ago. Whatever cognitive machinery the crows are using, they evolved it independently.
Funerals: danger assessment, not grief
When a crow dies, other crows gather. They emit alarm calls — loud, repetitive scolding vocalizations — that attract additional crows to the scene. Dozens of birds may congregate around the body, observing it from nearby perches, sometimes flying down to inspect it, sometimes sitting in silence. To a human observer, it looks like mourning. The scientific explanation is more interesting than mourning.
Kaeli Swift, a behavioral ecologist at the University of Washington working under corvid cognition researcher John Marzluff, conducted a two-year experiment across over a hundred sites in Washington State. She established feeding stations to attract local crows, then introduced a dead crow (a taxidermied specimen) while a masked human volunteer stood nearby. The crows responded to the dead crow with alarm calls and gathering behavior. More importantly, they subsequently avoided feeding at that location — and they associated the masked person with danger, responding with alarm calls when that person appeared again, even without the dead crow present. The crows learned from the death scene. They identified a potential threat (the person near the dead crow), memorized the threat’s face, and modified their behavior to avoid the area and the individual. Weeks and months later, they still recognized and responded to the mask.
Crows respond far more strongly to dead crows than to dead birds of other species. They largely ignore dead pigeons, robins, or other non-corvid birds placed in their territories but react intensely to dead members of their own species. Some studies suggest they respond more strongly to familiar individuals than to unfamiliar crows, indicating they may recognize specific community members even in death.
The “funeral” is not a ceremony. It’s a threat assessment protocol. The crows are investigating the scene to determine what killed the dead crow, whether that threat persists, and how to avoid it. The alarm calls broadcast the danger to the wider community. The subsequent avoidance behavior encodes the lesson into the population’s behavioral repertoire. Marzluff’s research demonstrated that crows can remember human faces that posed a threat for years — and they transmit this knowledge to crows that weren’t present for the original event. A crow that never saw the masked person holding a dead crow will nonetheless scold that person if other crows in the community do, because the social alarm response propagates through the group.
The behavioral function is pragmatic: collective intelligence applied to mortality data. The emotional dimension — whether crows experience something analogous to grief — remains scientifically unresolvable. Swift’s position is candid: she believes crows have emotional intelligence, but testing that scientifically is impossible because there’s no way to access what’s happening at an emotional level inside an animal’s brain. What’s measurable is the behavioral output: crows process death, learn from it, remember the context, and share the information. Whether they feel anything while doing it is a question the methodology can’t answer.
What corvid brains do differently
The corvid brain lacks a neocortex. In mammals, the neocortex is the seat of higher cognitive function — the structure that expanded dramatically in primates and reached its maximum density in humans. Corvids achieve comparable cognitive outputs using a structure called the pallium, which is organized differently from the mammalian cortex but performs analogous functions. The neuron density in the corvid pallium is remarkably high relative to brain volume — corvid brains pack more neurons per gram than most mammalian brains.
A 2025 paper in Animal Cognition by Veit and colleagues explored the “dimensions of corvid consciousness” — a research framework asking not whether corvids are conscious but what aspects of consciousness their neural architecture could support. The paper argues that corvid brains process sensory information, maintain working memory, and generate flexible behavioral responses through neural pathways that are structurally distinct from but functionally analogous to mammalian circuits. A German neurobiologist trained two crows — Glenn and Ozzy — to peck at “yes” or “no” targets to indicate whether they had detected a faint light, demonstrating analytical introspection: the crows reported on their own perceptual states, a capacity associated with subjective experience.
The convergent evolution angle is what makes corvids matter for neuroscience rather than just for animal behavior. If complex cognition can evolve independently in a brain that is structurally unrelated to the primate brain, then intelligence is not a property of a specific neural architecture. It’s a property of certain computational principles — neuron density, connectivity patterns, feedback loops — that can be instantiated in multiple biological substrates. The corvid brain is evidence that there is more than one way to build a mind, and that the way mammals did it is not the only way it can be done.
Corvids sit alongside octopuses as the strongest natural evidence that intelligence is convergent rather than unique. We cover corvid cognition alongside cuttlefish camouflage, electroreception, and the full landscape of how animal brains solve problems humans assumed required human brains across our Neurozoology course — including why a 14-gram brain that last shared an ancestor with yours 300 million years ago can plan for the future, manufacture tools from materials it’s never seen, and hold a funeral that’s more operationally useful than most of ours.
-
Prison Breaks as Reverse Heists: The Engineering of Escape
A heist defeats a security system to remove something valuable from a controlled space. A prison break defeats a security system to remove something valuable from a controlled space. The engineering is the same — tunneling, tool fabrication, timing security rotations, exploiting human vulnerabilities in the guard staff, coordinating a team under surveillance, and executing during a narrow window before the breach is discovered. The only difference is the direction. In a heist, you go in and take something out. In a prison break, you are the thing being taken out. The constraint analysis is identical, the failure modes are identical, and the reason most prison breaks collapse — like most heists — is that the escape itself works but the aftermath doesn’t.
Tunneling: the shared engineering
El Chapo Guzmán’s 2015 escape from Mexico’s Altiplano maximum-security prison was, by any engineering standard, a construction project. His associates spent over a year building a mile-long tunnel from a house outside the prison walls to a 20-by-20-inch opening in the shower floor of his cell — a location in a camera blind spot. The tunnel was five feet seven inches tall, ventilated, lit, and equipped with a modified motorcycle on rails that was used to transport excavated dirt and construction tools. The cost was estimated at over $1 million. Guzmán stepped into his shower, dropped through the hole, rode the motorcycle through a mile of underground passage, and emerged in a building his cartel had purchased for the purpose. He was free for six months before Mexican marines recaptured him in Sinaloa.
The engineering parallels to the 2005 Banco Central heist in Fortaleza, Brazil — where a 25-member gang dug a 256-foot tunnel from a fake landscaping business into a bank vault — are structural, not cosmetic. Both operations required months of excavation conducted in secrecy. Both required a cover property (a house, a fake business) to justify activity near the target. Both required disposal of enormous volumes of excavated material without attracting attention. Both required ventilation, lighting, and structural reinforcement to prevent collapse. The difference: the Banco Central crew tunneled in to take $71.6 million. Guzmán’s crew tunneled in to take Guzmán. Same discipline, same timeline, same logistics.
The Alcatraz escape of June 1962 — Frank Morris and brothers John and Clarence Anglin — used the same principle at a smaller scale and with prison-available materials. Over months, they used sharpened spoons and a homemade electric drill fashioned from a vacuum cleaner motor to widen the ventilation ducts behind their cells, accessing an unguarded utility corridor. They fabricated dummy heads from plaster, flesh-tone paint, and real human hair harvested from the barbershop and placed them in their beds to fool the guards conducting night checks. They built a raft and life vests from over 50 raincoats, bonded with contact cement stolen from the prison’s glue shop. On the night of June 11, they climbed to the roof, descended to the shore, inflated the raft, and paddled into San Francisco Bay. They were never found. The FBI concluded they likely drowned; subsequent hydrodynamic simulations suggest it was possible — though not certain — that the currents could have carried them to shore.
The Alcatraz escape is the purest form of the prison break as engineering problem. No bribed guards. No cartel money. No outside construction team. Three men with spoons, stolen raincoats, plaster, and paint, working for months inside the most secure prison in the United States, fabricating every component of their escape from materials available within the facility. The constraints were absolute: no access to power tools, no external supply chain, no communication with anyone outside the walls. Every heist crew wishes its operational security was that airtight — Morris and the Anglins had it imposed on them by the prison itself.
The human element: same vulnerability, different label
The Dannemora escape from Clinton Correctional Facility in June 2015 — the first breakout in the prison’s 170-year history — ran on the same human vulnerability that drives inside-job heists. Convicted killers David Sweat and Richard Matt didn’t defeat the prison’s physical security through engineering alone. They seduced Joyce Mitchell, a civilian employee who supervised the prison’s tailor shop. Mitchell smuggled hacksaw blades, chisels, and other cutting tools into the facility by hiding them inside frozen hamburger meat, which was then delivered to the inmates by guards who didn’t inspect the packages closely enough. Mitchell had sexual relationships with both men. She was supposed to be the getaway driver.
The analytical frame is identical to the inside-job heist: the security system’s weakest point is the human being it has to trust. Mitchell wasn’t a guard — she was a civilian employee with access to the interior. She wasn’t coerced at gunpoint like the Northern Bank officials in Belfast. She was manipulated emotionally, over months, by two men whose operational objective was her access to the supply chain. Sweat used the smuggled tools to cut through the steel wall of his cell, carve into a large pipe, and navigate a labyrinth of tunnels to a manhole outside the prison walls. The physical engineering was sophisticated. The human engineering was the prerequisite.
Their expected ride — Mitchell — never showed up. She lost her nerve. Sweat and Matt emerged from the manhole into Dannemora, New York, with no vehicle, no plan B, and a note left in their cells that read “Have a nice day.” After 20 days on the run, Matt was shot dead by police. Sweat was captured two days later. The escape worked. The aftermath didn’t. The pattern holds.
Pascal Payet, a French convicted murderer, solved the human-element problem by going vertical. In 2001, he arranged for associates to land a helicopter on the roof of his prison and fly him out. In 2003, while still a fugitive, he went back to the same prison by helicopter and extracted three more inmates. He was caught, imprisoned again, and in 2007 escaped a third time — from a different prison — again by helicopter, this one hijacked by four men in Cannes. He was recaptured in Barcelona. Three helicopter escapes from three different prisons. The engineering is minimal. The audacity is maximal. And the vulnerability Payet exploited was that French prisons, despite being designed to prevent tunneling, wall-climbing, and gate-crashing, had no anti-aircraft countermeasures. Nobody planned for a helicopter because nobody imagined a helicopter. The same category error that makes heists work — the security designer didn’t anticipate the actual attack vector — makes prison breaks work.
The Great Escape: when the military does it
The most famous prison break in history is also the most operationally instructive. In March 1944, 76 Allied prisoners of war escaped from Stalag Luft III, a German POW camp, through a tunnel codenamed “Harry” — one of three tunnels dug over 15 months using improvised tools made from bed boards and spoons. The tunnel was 350 feet long and nearly 30 feet underground, shored with bed-board lumber, ventilated by a hand-pumped air system, and lit with electric lamps tapped from the camp’s power grid. Over 200 prisoners contributed to the operation, which required not just tunnel construction but the fabrication of civilian clothing, forged identity documents, and escape maps — an entire logistics infrastructure built inside a POW camp.
Only three men made it to safety. Seventy-three were recaptured. Fifty were executed on Hitler’s orders. The escape was a tactical success — 76 men got through the tunnel — and a strategic catastrophe. But the engineering itself is a graduate seminar in constrained manufacturing: how to build a ventilated, lit, reinforced tunnel 350 feet long using materials available inside a prison, without power tools, under constant surveillance, over a year and a half.
Why prison breaks, like heists, fail in the aftermath
The Dannemora escape is the diagnostic case. Sweat and Matt got out. The physical escape was successful. They had no vehicle, no cash reserves, no false identities, and no extraction network. They wandered through upstate New York for three weeks, increasingly desperate, until both were shot. The escape was a closed system the inmates controlled. The aftermath was an open system they didn’t.
El Chapo — a billionaire cartel leader with a global logistics network — is the exception that proves the rule. He had a motorcycle in his tunnel, a house waiting at the exit, cartel infrastructure to move him across Mexico, and enough corruption in the state apparatus to buy six months of freedom. He had an extraction network because he had a narcotics empire. The Dannemora men had hacksaw blades hidden in hamburger meat and a getaway driver who lost her nerve.
The parallel to heist failures is exact. The Hatton Garden crew — the “diamond wheezers,” average age 63 — drilled through a vault wall and stole £14 million in jewels and cash. They were caught because they drove their own cars to the heist (captured on license plate readers), used their personal cell phones (tower pings placed them at the scene), and were filmed by cameras they didn’t know existed. The Dunbar Armored crew was caught because one man paid a broker with cash still in the original currency straps. In every case, the operation itself succeeded. The operational security after the operation collapsed.
Prison breaks and heists share the same structural irony: the part that requires the most engineering — getting in or getting out — is the part that usually works. The part that requires the least engineering — not getting caught afterward — is the part that almost always fails. The tunnel is a solvable problem. Being a fugitive is not.
We cover prison breaks alongside inside jobs, the Gardner Museum heist, and the full architecture of the greatest thefts and escapes in history across our Greatest Heists course — including why the most meticulously engineered escapes in history keep ending the same way the most meticulously engineered heists do: with someone using their personal cell phone.
-
Inside Jobs: Why the Biggest Thefts in History Were Committed by the People Hired to Prevent Them
Allen Pace III worked as a regional safety inspector for Dunbar Armored in Los Angeles. His job was to ensure the security of the company’s armored car depot — the facility where cash was stored, sorted, and loaded for transport. While performing that job, he photographed the floor plans, mapped the camera positions, timed the security rotations, identified which bags contained the highest denominations of non-sequential bills, and noted that on Friday nights the vault was left open to accommodate the volume of cash being moved. On September 12, 1997, he used his keys to let five childhood friends into the building. They ambushed the guards during their 12:30 a.m. lunch breaks, duct-taped them before anyone triggered an alarm, loaded $18.9 million into a U-Haul in 30 minutes, and removed the security camera recordings on the way out. Police immediately suspected an inside job. They looked at Pace — who had been fired the day before for tampering with company vehicles — but found nothing. The case cracked two years later when one of the crew paid a real estate broker with cash still bound in the original Dunbar currency straps.
Pace didn’t defeat the security system. He was the security system. And that’s the pattern that connects the largest thefts in modern history: the person with the most access is the person with the most opportunity, and no vault, alarm, or camera network can protect against the individual whose job description includes knowing exactly how those protections work.
The armored car problem
The armored car industry has been hit by inside jobs so consistently that the pattern qualifies as a structural vulnerability rather than a series of coincidences.
In March 1997 — six months before the Dunbar job — Loomis Fargo driver Philip Noel Johnson turned his company-issued firearm on his coworkers in Jacksonville, Florida, handcuffed them, and drove off with $18.8 million after a ten-year career with the company and several failed applications to police departments. He fled to Mexico. He was caught trying to cross the U.S. border with multiple passports and $11,000 in cash. In December 1982, Sentry Armoured Car guard Christos Potamitis in the Bronx planned a robbery in which his associate George Legakis and a crew broke in through the wooden roof, and Potamitis allowed himself to be “surprised” at gunpoint and handcuffed — a staged victimization that was supposed to deflect suspicion. They stole $11 million. In 1993, a Loomis armored car driver in Las Vegas simply left during his shift with $2.9 million in ATM funds from several casinos, disappeared, and was never found.
The economics are consistent: armored car employees have direct physical access to large volumes of cash, detailed knowledge of security protocols and timing, legitimate reasons to be present at every point in the custody chain, and compensation that is orders of magnitude below the value of the assets they handle. The average armored car guard earns roughly $35,000 to $45,000 per year. The vault they access nightly contains millions. The gap between the employee’s economic position and the value they’re entrusted with is the gap through which every one of these jobs passes.
The bank guard pattern
The same dynamic scales to every institution that stores high-value assets and requires human beings to guard them.
Anthony Black was a security guard at the Brink’s-Mat warehouse at London’s Heathrow Airport. On the morning of November 26, 1983, he let a crew of robbers into the facility. They tied up the other guards. The crew had expected to find cash. What they found instead was three tons of gold bullion, diamonds, and cash worth £26 million — roughly $41 million. Black’s insider access was the entire plan: the robbers didn’t need to defeat the locks, the alarms, or the perimeter security because the man responsible for monitoring them opened the door. The case created a money-laundering infrastructure across London that arguably persists in some form decades later, because fencing three tons of gold requires a financial network that doesn’t dismantle itself when the heat dies down.
In Baghdad in 2007, three guards at the Dar es Salaam Investment Bank walked out of the facility overnight with $282 million. Bank employees discovered the theft when they arrived the next morning to find the front door open and the guards — who typically slept at the bank — gone along with the money. The guards had both the access and the time: the bank’s security model relied on the same people to protect the assets and to be present after hours, which meant the guards were the only human element in the entire security chain during the hours when theft was most feasible.
In Belfast in 2004, armed men posing as police officers took the families of two Northern Bank officials hostage, then instructed the officials to go to work the next morning as if nothing had happened. At the end of the business day, the officials let the crew into the vaults. The haul was £26.5 million. The operation didn’t need to defeat any security technology. It needed to coerce two people who had legitimate vault access — turning the bank’s own authorization hierarchy into the attack vector.
Why this keeps working
Every security system has to solve the same fundamental problem: someone has to be trusted. A vault requires a person with the combination. An alarm requires a person who can arm and disarm it. A camera network requires a person who monitors it. A cash facility requires a person who handles the cash. At every point in the chain, there is a human being whose job description includes the ability to circumvent the security measures — because the security measures were designed to be operated by that person.
This is the insider threat as a systems design problem. The security system protects the asset from outsiders. The insider is not an outsider. The security system, by definition, does not protect against the people it was designed to serve. Pace had keys. Black had the alarm codes. Potamitis had the schedule. Johnson had the truck. The Baghdad guards had the building to themselves. Each of these individuals exploited exactly the access their employer gave them, used exactly the knowledge their training provided, and operated within exactly the window their work schedule created.
The countermeasures are well-known: dual-person integrity (two people must be present for any high-value access), separation of duties (the person who arms the alarm isn’t the person who opens the vault), background checks, financial monitoring of employees with asset access, rotation of duties to prevent any single person from mapping the full system, and surveillance of the surveyors — cameras that watch the people watching the cameras. The Securitas depot robbery in Kent in 2006, where a crew kidnapped the branch manager and his family to coerce vault access, demonstrated that even dual-person controls can be defeated through coercion rather than collusion.
The honest assessment is that no security architecture eliminates the insider threat. It can only raise the cost, complexity, and risk of exploiting insider access. Every countermeasure introduces friction into normal operations — the same friction that makes the system secure makes it slower, more expensive, and more dependent on compliance with protocols that employees find tedious. The gap between the security system’s design and the security system’s daily operation is where the insider lives. Pace knew the cameras. He also knew that on Friday nights, the vault stayed open because the volume of cash made closing and reopening it impractical. The operational convenience that made the depot efficient was the same operational convenience that made it vulnerable.
The modern version
The inside job hasn’t gone away. It’s migrated. The 2024 Easter Sunday heist in Los Angeles — over $30 million stolen from a GardaWorld cash management facility — bore what a federal source called “the markings of an inside job.” Someone knew how the building was set up, how to terminate the alarms, and how to access the vaults. The investigation is ongoing.
But the more consequential modern inside jobs don’t involve vaults at all. IT administrators with root access to financial systems, bank employees who sell customer data to fraud rings, cryptocurrency exchange operators who drain their own platforms — the asset being stolen has shifted from physical cash to digital value, but the mechanism is identical: the person trusted with access uses that access to steal. The 2022 collapse of FTX, where founder Sam Bankman-Fried misappropriated billions in customer funds, is structurally the same crime as Anthony Black opening the Brink’s-Mat door — the person entrusted with the assets took them, and the security architecture was designed to give that person access, not to prevent them from using it.
The inside job is the oldest heist architecture and the most durable one. It survived the transition from physical vaults to electronic systems, from gold bullion to cryptocurrency, from armed guards to multi-factor authentication. The technology changes. The vulnerability doesn’t. Every system that stores value requires a human being with access to that value, and the human being with access is the single point of failure that no technology eliminates.
We cover inside jobs alongside the Gardner Museum heist, North Korea’s state-sponsored cyber theft, and the full landscape of the greatest thefts in history across our Greatest Heists course — including why the most expensive security systems on earth keep getting defeated by the people they were built to serve.
-
Uranium Supply Chain 2026: Nuclear Renaissance Meets Mining Reality
The United States operates 93 nuclear reactors — the largest fleet on earth — and cannot fuel a single one with domestically sourced uranium. The country has essentially no primary uranium production. The mines that once operated in Wyoming, Texas, and the Colorado Plateau shut down decades ago when prices collapsed, and the supply chain that supported them — the skilled labor, the processing infrastructure, the regulatory pipelines — dissolved with them. In 2026, spot uranium is approaching $92 per pound. Analysts project prices reaching $100 to $120 per pound, with some upside scenarios targeting $135 if supply fails to respond. The U.S. government has committed up to $80 billion to build new reactors and reinvigorate the nuclear industrial base. The USGS added uranium to its 2025 Critical Minerals List for the first time in years. The IEA forecasts annual nuclear investment rising from over $70 billion today to approximately $210 billion by the mid-2030s. Roughly 65 reactors are under construction worldwide.
The demand story is real. The supply story is the problem.
Where the uranium comes from
Global reactor demand runs approximately 67,500 metric tons of uranium per year. Mine production has historically met only 74 to 90 percent of that, with the deficit covered by drawdowns from government and commercial inventories, recycled material, and secondary supply. Those secondary sources are depleting. The market is transitioning from an inventory-driven system to a production-driven one, and production isn’t keeping up.
Kazakhstan dominates. Kazatomprom, the state-owned producer, is the world’s largest uranium miner, operating primarily through in-situ recovery — a technique that pumps acidified solution into uranium-bearing rock formations underground and extracts the dissolved uranium without conventional mining. Kazakhstan accounts for roughly 40 percent of global production. Canada’s Cameco operates McArthur River and Cigar Lake in Saskatchewan’s Athabasca Basin — two of the highest-grade uranium deposits on earth, with licensed capacity of 25 million pounds annually and proven reserves exceeding 457 million pounds. Australia, Namibia, Uzbekistan, and Niger round out the major producers. Russia controls a significant share of global uranium enrichment and conversion — the processing steps between mining raw uranium and fabricating reactor fuel.
The concentration is the vulnerability. A large proportion of uranium production sits in non-Western jurisdictions. Sanctions, export bans, and the war in Ukraine are constraining the nuclear fuel cycle. Niger — historically a significant supplier — produced no uranium at all in 2025 after a military junta seized power and disrupted operations at the SOMAÏR facility. Kazakhstan has announced lower production targets for 2026. McArthur River reduced its 2025 output due to development delays. In the United States, several in-situ recovery restarts have ramped up more slowly than planned. The net effect: tighter global supply for reliable primary production, at precisely the moment when demand forecasts keep getting revised upward.
The demand surge
Three forces are converging on uranium demand simultaneously.
The first is reactor life extensions and restarts. Existing nuclear plants that were scheduled for retirement are getting new operating licenses instead. Plants that were shut down are being evaluated for restart. The economics shifted when natural gas prices spiked, renewable intermittency proved harder to manage than projected, and carbon-free baseload generation became a policy priority rather than a political liability. Nuclear went from a technology that governments were phasing out to one they’re subsidizing.
The second is new reactor construction. The $80 billion U.S. government commitment includes Westinghouse AP1000 deployments and GE Hitachi BWRX-300 small modular reactors. Canada has broken ground on SMRs at the Darlington nuclear station with combined funding commitments of roughly CAD 3 billion and a target completion around 2030. The U.S. and Japan announced a framework totaling $550 billion, with up to $332 billion directed to energy and AI-linked infrastructure including new nuclear capacity. China continues building reactors at a pace no other country matches and purchasing uranium in large quantities to stockpile for its future fleet.
The third is AI data centers. This is the demand driver that didn’t exist in anyone’s forecast five years ago. Hyperscale computing facilities require baseload power — reliable, 24/7 generation that doesn’t depend on weather or time of day. Nuclear fits that requirement better than any other carbon-free source. More than 63 percent of investors surveyed by Uranium.io believe AI-related electricity consumption will become a material factor in nuclear planning over the next decade. Microsoft, Amazon, and Google have all explored or announced nuclear power agreements for data center operations. The AI demand signal is being treated as structural rather than cyclical — permanent new load on the grid that requires permanent new generation.
Why supply can’t respond quickly
This is the constraint that the nuclear renaissance runs into. Uranium mining is not a faucet. Mine restarts require years, not months. The lead time from decision to production involves permitting (often multi-year regulatory processes), environmental review, workforce recruitment (specialized uranium mining labor that largely doesn’t exist anymore in the West), facility construction or refurbishment, and ramp-up testing. A mine that was shuttered in 2012 can’t resume production in 2026 just because the price is right.
Beyond mining, the fuel cycle has its own bottlenecks. Mined uranium (yellowcake) must be converted to uranium hexafluoride, enriched to increase the concentration of fissile U-235, fabricated into fuel assemblies, and delivered to the reactor. Russia controls a significant share of global enrichment and conversion capacity. The U.S. ban on Russian uranium imports — signed into law in 2024 — created a scramble for alternative enrichment services. Centrus Energy is the only licensed producer of High-Assay Low-Enriched Uranium (HALEU) in the Western world — the next-generation fuel that advanced reactors and many SMR designs require. Centrus is expanding its Piketon, Ohio facility, but scaling enrichment infrastructure is measured in years and billions of dollars, not quarters.
The structural reality: even sustained high prices may not resolve supply deficits within typical investment horizons. Producers have signaled that three-digit prices per pound — above $100 — are the minimum necessary to incentivize new mine development at a scale that reflects actual capital costs, permitting timelines, and supply chain risk. The market is in a standoff. Utilities want to buy at current prices. Producers want higher prices before committing capital to new production. China, meanwhile, continues buying at whatever price the market offers, building strategic reserves while Western utilities defer purchases and hope prices stabilize.
The SMR fuel problem
Small modular reactors are the technology that’s supposed to make nuclear faster, cheaper, and more deployable. The first SMRs won’t be operational until 2030 or 2031. The World Nuclear Association projects SMR capacity could account for roughly 7 percent of global nuclear power generation by 2040. But many advanced SMR designs require HALEU — uranium enriched to between 5 and 20 percent U-235, compared to the 3 to 5 percent used in conventional reactors. HALEU production capacity in the Western world is essentially nonexistent outside of Centrus’s pilot-scale operations. Russia was the primary commercial supplier of HALEU before sanctions disrupted the trade.
Building the SMR fleet without the fuel to power it is the kind of sequencing error that turns a technology roadmap into a bottleneck cascade. The reactors require enrichment capacity that requires enrichment facilities that require regulatory approval that requires years. Cameco’s $2.8 billion ten-year supply agreement with India and Centrus’s $1.2 billion in convertible note offerings and $2 billion in contingent utility purchase commitments represent the financial architecture being constructed to close these gaps. Whether the construction finishes before the demand arrives is the open question.
The investment case and the honesty test
Uranium is one of the few commodities where there is essentially no substitution potential. A nuclear reactor runs on uranium. Nothing else does the job. Demand is inelastic — utilities will pay whatever the market requires because the cost of uranium is roughly 5 to 7 percent of a reactor’s total operating budget. A doubling of uranium prices is a rounding error in the cost of nuclear electricity. This means utilities will eventually buy at higher prices because they have no alternative. The question is when, not whether.
Long-term contract prices have risen to $86 per pound, indicating that utilities are accepting elevated costs even as they resist spot purchases. The World Nuclear Association has revised its uranium demand growth forecast to a 5.3 percent compound annual growth rate through 2040, up from 4.1 percent previously. Analysts project a supply deficit building over the next decade as mine production continues to lag reactor requirements. More than 85 percent of surveyed investors anticipate higher prices into 2026.
The honesty test: every part of this demand story — reactor restarts, new construction, SMRs, AI data centers — requires uranium, and the supply chain to deliver it doesn’t exist at the scale the demand forecasts imply. The nuclear fuel supply chain is being rebuilt in real time, by governments writing checks and producers scaling operations, against a backdrop of geopolitical disruption, depleted inventories, and a workforce that needs to be reconstituted essentially from scratch in the West. The nuclear renaissance is real. The mining and enrichment infrastructure to fuel it is years behind.
We cover the uranium supply chain alongside gallium and germanium export controls, the helium shortage, and the full landscape of critical materials that modern technology and energy systems depend on across our Rare Earth Elements course — including why the largest nuclear fleet on earth can’t fuel itself, and what that means for a planet betting on reactors to keep the lights on.
-
Carbon Capture and Direct Air Capture in 2026: Can We Actually Suck CO₂ Out of the Atmosphere at Scale?
In 2024, Climeworks’ Mammoth plant in Iceland — the world’s largest direct air capture facility, designed to remove 36,000 metric tons of CO₂ per year — captured 105 tons. Total. For the year. That’s less than the annual tailpipe emissions of a dozen long-haul trucks. It’s roughly one-thousandth of the plant’s stated design capacity. In mid-2025, Climeworks began laying off at least 10 percent of its approximately 500 employees. The company had raised over $800 million in equity and subsidies. JPMorgan Chase paid roughly $800 per ton for Climeworks removal credits in 2023. The co-CEO told reporters the operating cost was “closer to the $1,000 per ton mark than the $100 per ton mark.” The DOE’s target — the threshold at which direct air capture becomes economically viable for climate-scale deployment — is $100 per ton.
Meanwhile, in West Texas, Occidental Petroleum’s subsidiary 1PointFive is constructing Stratos, a direct air capture plant designed to remove 500,000 tons of CO₂ annually — nearly 14 times Mammoth’s design capacity. Construction of the first two “trains” finished in December 2024. Operations are expected to begin in 2025–2026, with full capacity by mid-2026. The facility cost $1.3 billion, up from original estimates of $800 million. It will be powered partly by a dedicated 145-megawatt solar installation. It will run on natural gas for the high-temperature heat the process requires. The captured CO₂ will be injected underground and earn 45Q tax credits. Occidental will continue selling hydrocarbons.
These two data points — Mammoth’s 105 tons and Stratos’s $1.3 billion construction cost — frame the honest state of direct air capture in 2026. The technology works. The physics is real. The scale, cost, and energy requirements are somewhere between daunting and disqualifying, depending on how much optimism you’re willing to extend and over what timeframe.
How it works
CO₂ constitutes roughly 422 parts per million of the atmosphere — 0.04 percent of the air around you. Direct air capture means extracting a trace gas from a mixture that is 99.96 percent other stuff. The dilution problem dictates the scale: processing enough air to capture meaningful quantities of CO₂ requires moving enormous volumes through chemical systems, which requires enormous amounts of energy.
Two approaches dominate. Liquid solvent systems — the Carbon Engineering technology that Occidental acquired and uses at Stratos — pass air through large contactors where it meets an alkaline solution, typically potassium hydroxide. The CO₂ reacts with the solution, forms calcium carbonate pellets through subsequent processing, and those pellets are heated in a kiln to roughly 900 degrees Celsius to release pure CO₂ gas. The high-temperature step is the cost driver: that kiln needs fuel, and at Stratos, the fuel is natural gas.
Solid sorbent systems — Climeworks’ approach — use porous materials coated with amine compounds that bind CO₂ from air at ambient temperature. When the sorbent is saturated, a temperature-vacuum swing cycle heats it to 80–120 degrees Celsius and reduces pressure to release the CO₂. The sorbent is then cooled and returned to capture mode. Lower regeneration temperatures make solid sorbent systems more compatible with renewable heat and waste heat, which is why Climeworks located in Iceland — geothermal energy provides the heat and electricity at near-zero carbon intensity.
Heirloom Carbon takes a third approach: accelerated mineralization. Limestone naturally absorbs CO₂ from the air, but the process takes years. Heirloom speeds it to days by spreading crushed limestone on trays exposed to air, then heating the saturated limestone to release concentrated CO₂. The company is developing two facilities in Louisiana with a combined capacity of 320,000 tons per year.
In every case, the captured CO₂ must then be compressed, transported, and either injected underground for permanent geological storage or mineralized into rock. The energy requirements for the full chain — capture, concentration, compression, injection — run between 2,000 and 3,000 kilowatt-hours per ton for the most mature systems. Newer electrochemical approaches are targeting 700 to 1,000 kWh per ton. Even with clean energy, life-cycle analyses show that DAC systems re-emit roughly 10 percent of the CO₂ they capture through embedded emissions in materials, equipment fabrication, and operational overhead.
The scale problem
Global CO₂ emissions are approximately 40 billion tons per year. DAC, across all companies, all technologies, and all years of operation combined, has removed less than 20,000 tons to date. That’s 0.00005 percent of annual emissions. IPCC pathways for holding global temperature increases to 1.5 degrees suggest the world may need to remove 5 to 10 gigatons of CO₂ annually by mid-century. One gigaton is a billion metric tons. Reaching one gigaton per year of DAC would require thousands of Mammoth-scale plants or dozens of Stratos-scale plants, consuming hundreds of terawatt-hours of energy annually — roughly equivalent to doubling the electricity consumption of a mid-sized industrial country.
The IEA’s net-zero roadmap calls for about 32 million tons of DAC removal per year by 2030, rising to gigatons by 2050. Current global capacity is measured in thousands of tons. The gap between where DAC is and where it needs to be is not a gap that the current trajectory closes. Mammoth was designed for 36,000 tons and delivered 105 in its first year. Stratos is designed for 500,000 tons and hasn’t started operations. If Stratos works at design capacity — a significant if, given that Carbon Engineering’s pilot in British Columbia handled 1 ton per day — it will represent an 873 percent increase in global DAC capacity from a single facility. But 500,000 tons is still one-eightieth of the 2030 IEA target, which is itself a fraction of what’s needed by 2050.
What’s actually happening in the market
The money is real even if the tonnage isn’t. The DOE allocated $3.5 billion for DAC Hubs — large-scale facilities designed to remove 1 million tons per year each. Project Cypress in Louisiana (Climeworks, Heirloom, Battelle) and Stratos in Texas have both received initial awards. Microsoft signed a 3.3-million-ton purchase agreement with Stockholm Exergi for bioenergy carbon capture. Microsoft’s DAC deals with 1PointFive are reported at $200 to $300 per ton — multi-year, multi-hundred-thousand-ton contracts that provide the revenue certainty developers need to secure construction financing. Climeworks sells credits at $600 to $800 per ton. Heirloom is targeting below $100 per ton at scale but currently prices credits significantly higher during its pilot phase.
The investment thesis has shifted from “can this technology work?” to “can it work economically at the energy cost the process requires?” Energy is the dominant cost driver. A process that needs 2,000 kWh per ton of CO₂ captured is, fundamentally, an energy project that happens to produce carbon removal as its output. The plants that will determine whether DAC reaches viability are the ones that solve the energy integration problem — co-locating with cheap, abundant, low-carbon energy sources and locking in long-term power contracts at rates that make the per-ton math work.
The Occidental problem
Occidental Petroleum acquiring Carbon Engineering in 2023 and building Stratos raises a question that the industry can’t avoid: is a DAC plant owned by an oil company and powered partly by natural gas a climate solution or a license extension? The captured CO₂ can be permanently stored underground, but it can also be used for enhanced oil recovery — injecting CO₂ into depleted wells to extract more crude. Occidental earns 45Q tax credits for the stored CO₂ while continuing to produce and sell the hydrocarbons whose combustion put the CO₂ in the atmosphere in the first place. The facility’s net climate impact depends on whether the stored carbon exceeds the emissions from the natural gas burned to power the process and the oil extracted using the captured CO₂ — a calculation that critics argue is unlikely to come out positive.
The counterargument is that waiting for perfectly clean DAC means waiting while atmospheric CO₂ concentrations continue to rise. Stratos, imperfect as it is, would demonstrate whether liquid-solvent DAC can operate at commercial scale — a question nobody has answered yet. If it works, future iterations can be powered by renewables. If it doesn’t, the industry learns where the engineering breaks. The question of whether an oil company should be building carbon removal infrastructure or whether its involvement contaminates the enterprise is a political judgment, not a technical one. The technical question is simpler: does the plant capture more CO₂ over its lifetime than the full supply chain emits? That number doesn’t exist yet because the plant isn’t operating yet.
Where this sits
Direct air capture is a technology that is simultaneously necessary and insufficient. Necessary because IPCC pathways for climate stabilization include gigatons of carbon removal — the world has emitted too much already for emission reduction alone to hold temperatures. Insufficient because the current cost, energy intensity, and scale are orders of magnitude away from what climate models require, and the trajectory from Mammoth’s 105 tons to the gigatons needed is not a line anyone can draw with confidence.
The honest framing: DAC is in the position that solar photovoltaics occupied in the early 2000s — expensive, small-scale, and dependent on subsidies, but on a learning curve that has the potential to drive costs down dramatically if manufacturing scale materializes and the technology iterates. Solar costs fell 99 percent over four decades. Whether DAC follows a similar trajectory depends on whether the energy integration problem is solvable, whether the modular manufacturing approach (Climeworks’ container design, Heirloom’s tray-based system) enables factory-style cost reduction, and whether the market for carbon removal credits generates enough revenue to sustain the industry through the expensive early decades.
The DOE’s $100-per-ton target is the benchmark. Climeworks is at roughly $1,000. The next generation aims for $300 to $350 by 2030. Heirloom is targeting below $100 at scale. The gap is closing, but it’s closing from a starting point that is ten times the target, at a deployment scale that is one-ten-thousandth of what climate models need. The math is brutal. The alternative — a world with no carbon removal capability when the IPCC says gigatons are required — is worse.
We cover direct air capture alongside fusion energy, space-based solar power, and the full landscape of civilization-scale moonshot technologies across our Moonshot 2169 course — including why the technology that may determine whether the planet’s thermostat stabilizes captured 105 tons in its first year, and what has to change before it captures a billion.
