In November 1989, a man with his face blurred and his voice altered sat down in front of a camera at a Las Vegas television station and told the reporter George Knapp that the United States government was hiding nine flying saucers in a hangar in the Nevada desert. He called himself Dennis. He said he was a physicist, that he had worked at a site called S-4 near Area 51, and that his job had been to reverse-engineer the propulsion systems of craft that were not built by human hands. The saucers, he said, ran on a reactor. The reactor ran on a fuel. And the fuel was an element that did not yet exist on the periodic table: number 115.
His name was Bob Lazar, and for thirty-five years he has been either the most important whistleblower of the twentieth century or its most durable fabulist, depending on who you ask. But set the saucers aside for a moment, because the interesting part of the Lazar story was never the aliens. It was the fuel. Lazar described element 115 as a stable, machinable metal, stored as small discs no larger than a coin, that produced a gravitational field as an inherent property of its nucleus. Feed it into the reactor, amplify the field, and the craft could warp the space around it and fall forward through the distortion — flight without wings, thrust without a nozzle. There was only one catch. You couldn’t make the stuff on Earth. It had to be forged somewhere superheavy elements occur naturally, out among the stars, and then recovered. The entire capability lived or died on whether you happened to possess the metal.
Then, in 2003, something happened that made a lot of people take Lazar seriously. A joint Russian and American team at the Joint Institute for Nuclear Research in Dubna bombarded americium-243 with calcium-48 and produced a few atoms of element 115. The empty slot on the periodic table was filled. In 2016 the element was officially named moscovium. Fourteen years after a man in a blurred video had named it on television, element 115 was real.
The tell
Here is what the headlines left out. What Dubna made is the opposite of what Lazar described, on the exact property the whole story depended on. Moscovium’s longest-lived isotope, moscovium-290, survives for about two-thirds of a second before it falls apart. You cannot machine it. You cannot hold it. You cannot make a coin out of it, or a reactor, or anything at all except a fleeting signature on an instrument. Lazar’s element was stable; the real one is among the most unstable things human beings have ever produced. And nothing in nuclear physics — not one calculation, not one model — predicts that any isotope of element 115 does anything whatsoever to gravity.
Naming 115 was never the hard part, and it was never prophecy. Nuclear theory had penciled that slot in since the 1960s, and physicists fully expected the region to fill in eventually, one element at a time, the way rows of the periodic table always do. Calling 115 in advance was calling next year. What Dubna did to the rest of Lazar’s story was quietly dismantle it. And when reality contradicted him — first by making 115 here, which he had said was impossible, then by making a version that was useless — the claim did not retract. It retreated. His 115, he explained, was a different, more stable isotope, the kind you could only recover from wreckage, the kind we cannot synthesize and therefore cannot check. Which is the tell. A claim engineered to survive every possible experiment has quietly stopped being a scientific claim at all. The stable isotope is always the one behind the next door.
I have spent a good part of this year thinking about that move, because I keep hearing it. Not from ufologists. From the AI industry.
The two materials
The story we are all being told is that artificial intelligence will keep getting exponentially better, more or less forever, and that the remaking of every industry, every job, and every institution follows as a matter of arithmetic. That story quietly rests on two things we do not yet have. It needs far more computing power than physics is currently willing to give us cheaply, and it needs far more electricity than the grid can currently deliver. Both walls are real. And for each wall, the industry has a rescue technology it gestures toward — a way through, arriving soon. In both cases the rescue technology is an element-115 story: an exotic, inherently stable material that dissolves the hard problem, announced long before it is demonstrated, with the confirming evidence perpetually just out of reach.
Let me be precise about one thing first, because precision is the whole game. The large language models we use today do not run on quantum computers. They run on classical silicon — on graphics chips, in racks, drawing megawatts. Quantum computing is not, right now, on the critical path to a better chatbot, and it may never be. So when I say AI’s future leans on quantum, I do not mean the models secretly require it to function. I mean something narrower and more damning: the promise that we will punch through the compute ceiling — that the exponential will not flatten into an S-curve — leans hard on quantum as the paradigm leap that rescues it. That is the escape hatch. And over the past few weeks, that escape hatch has been publicly, and peer-reviewedly, taken apart in the pages of Nature.
The stable qubit
Quantum computers are hard to build for one reason above all others: qubits are fragile. The same delicate quantum state that makes them powerful also makes them exquisitely sensitive to noise, and the faintest disturbance from the outside world collapses the computation. The phenomenon is called decoherence, and it is why the most advanced quantum machines on Earth hold, at most, a few hundred usable qubits, while the problems worth solving demand millions.
Microsoft’s answer, which it has chased for nearly twenty years, is the most elegant idea in the field. Instead of fighting decoherence with brute-force error correction, build a qubit that is stable by construction. The theory, laid out by Alexei Kitaev and Michael Freedman in 1997, is that you can store quantum information not in one fragile particle but in the topology of a system — a global, nonlocal property spread across an entire structure — so that local noise simply cannot reach it. The exotic ingredient this requires is a Majorana zero mode, a peculiar collective behavior of electrons that had never been convincingly observed in nature or the laboratory. Make and control Majoranas, and you would have a qubit with error resistance baked into the hardware itself. You would have reached, in effect, the island of stability for quantum computing: the special configuration where the fragility problem dissolves.
That is the pitch, and it is a genuinely beautiful one. It has also now produced two of the most contested announcements in modern physics.
In February 2025, Microsoft unveiled Majorana 1, a chip it called the world’s first quantum processor powered by a topological core, and declared it a clear path to a machine of a million qubits — enough, the company said, to do things all the world’s current computers working together could not, from breaking down microplastics to inventing self-healing materials. Useful quantum computing, Microsoft announced, was now a matter of years, not decades. Satya Nadella spoke of a straight road to a million qubits.
Then people read the actual paper. Published in Nature alongside the press release and carrying more than 160 authors, it did not demonstrate a working topological qubit. It described a measurement technique. More pointedly, the Nature editors took the unusual step of appending a note to the peer-review file stating that, in the reviewers’ assessment, the results did not amount to evidence that Majorana zero modes were present in the devices. The authors themselves conceded that their measurements could not, on their own, establish that the states they had detected were topological at all. The press release, meanwhile, told the public the paper was peer-reviewed confirmation that Microsoft had created Majorana particles. The distance between what the marketing said and what the science said was not a nuance. It was the whole story.
And beneath it sat a problem with a name: the Andreev mode. It happens that a trivial, useless, topologically boring state — an Andreev mode — can produce a signal in these devices almost identical to the exotic, useful Majorana one. Telling them apart is the entire ballgame, and it is genuinely difficult. Microsoft’s Majorana 1 data, its critics noted, was fully consistent with the device containing nothing but ordinary Andreev modes and not a single Majorana. If that pattern sounds familiar, it should: this exact confusion — claiming Majoranas, then finding the data fit the boring explanation just as well — is what forced the retraction of a high-profile 2018 Nature paper from researchers tied to the same program. The field has run this loop before.
The teardown
The skepticism might have subsided into the usual academic grumbling, except that on June 24, 2026, Nature published a formal, peer-reviewed critique by the physicist Henry Legg of the University of St Andrews. Its findings are specific and hard to wave off. Microsoft had validated its devices with an automated software test called the Topological Gap Protocol, whose entire purpose was to strip out human bias and rule out false positives. Legg found the protocol was itself unreliable: simply shifting the measurement window could flip its verdict, classifying the very same region of the very same device as suitable for quantum computing, or unsuitable, depending on an arbitrary choice. He found that Microsoft had shown only the favorable outcomes while contradictory results went unpublished, and that when he examined the underlying data Microsoft had not released, the supposedly successful measurements had been taken in disordered regions that looked gapless — the opposite of what the topological reading requires. One outlet reduced the coding problems to basic Python errors. Legg’s own line was that Microsoft had claimed to build the equivalent of a precision Swiss watch. His paper was the teardown.
The next door
Now watch what happened next, because this is where a physics dispute becomes an element-115 story. Nature accepted Legg’s critique in April but did not publish it until late June. In the interval, Microsoft did not wait to answer it. It announced Majorana 2.
The new chip, unveiled on June 2, swaps aluminum for lead in its superconducting stack — a change Microsoft says more than doubled the topological gap that protects the quantum states and delivered a thousandfold jump in stability, with the key lifetime leaping from a few milliseconds to a reported twenty-two seconds, occasionally a full minute. On the strength of it, the company cut its timeline for a scalable quantum computer nearly in half, from the 2033 it used to cite down to 2029. The chip, Microsoft added, was developed with the help of its own agentic AI.
Sit with that last detail. The company is using artificial intelligence to accelerate the quantum computer that is supposed to rescue artificial intelligence. The miracle is being summoned to save the miracle. It is a perfect little ouroboros, and it should make you more suspicious, not less.
Because the critics were unmoved, and their objection is the one that counts. A working qubit has to be controllable — you have to measure it along two different axes, which physicists call X and Z, and steer it between its states. The Majorana 2 paper reports only Z measurements. It does not show that the device functions as a qubit at all. The celebrated twenty-two-second figure is a parity lifetime, which, as Legg pointed out, is not the same as how long the actual quantum information survives — a number that sounds like proof of a stable qubit while quietly measuring something next to one. “Nothing in this preprint resolves the fundamental issues,” Legg said, adding that nothing in the released data, in his view, proved a topological qubit or a Majorana existed in these devices at all. And Majorana 2 is, for now, only a preprint; a formal critique cannot even be written until it clears peer review. The objections that sank Majorana 1 carry straight across. This is relocation, not resolution.
The most revealing moment in the whole affair came from Chetan Nayak, who leads Microsoft’s quantum effort, and who conceded to Science that his team had been cagey with its peers. They had told the field, in effect, to just trust them that they would make it work, he said, and they were now revealing more of how they had done it. He added that Microsoft does hold data showing it can fully control the qubits and even run algorithms on the chip — data that is, as of this writing, unpublished. Which is where we came in. The confirming evidence exists, we are assured. It is stable, it is real, it works. It simply is not something you can examine yet. The stable isotope is always behind the next door.
In fairness — and fairness is what keeps a piece like this defensible — some physicists find the new results encouraging, and the materials-science and fabrication work is, by nearly every account, genuinely impressive. Microsoft has not been caught lying, and I am not accusing it of it. But even sympathetic experts place topological quantum computing twenty to thirty years behind the trapped-ion and superconducting platforms that actually run today. And it is worth noticing that Microsoft’s own uncontested quantum results — the ones no one disputes — come from renting someone else’s trapped-ion hardware, not from its topological chips. The moonshot is the part on the magazine cover. The part that works is another company’s machine.
The clean reactor
The second wall is far less exotic and far more certain, which is precisely why the industry talks about it less. Artificial intelligence, in its present form, is a machine for turning electricity into tokens, and it is astonishingly hungry. Global data-center electricity demand jumped seventeen percent in a single year and is on track to double before 2030; were the world’s data centers a country, they would already rank among its largest consumers of power, somewhere between Japan and Russia. The five biggest technology companies poured more than four hundred billion dollars into this build-out in 2025, and plan to spend far more in 2026.
The grid cannot keep pace, and the evidence is not speculative. In the PJM territory that runs from Chicago to Washington, the capacity auction that sets the price of guaranteed power cleared at its maximum allowable level — roughly ten times what it fetched in 2022. The queue of projects waiting to connect to the American grid has swollen past 2,600 gigawatts, with waits stretching five to eight years and the large majority of projects eventually abandoned. Nearly half of the AI data centers planned for 2026 have slipped, opening a gap that is stranding hundreds of billions of dollars in committed capital behind transformers and transmission lines that do not yet exist. In Virginia, data centers already draw better than a quarter of the state’s electricity, and in the spring of 2025 the grid there came within a hair of rolling blackouts when sixty of them dropped offline at once. The bill, as always, lands on ordinary ratepayers — analysts project cumulative increases of thirty to sixty percent across PJM by 2030, with the residents of northern Virginia effectively subsidizing the power appetite of the largest corporations on the planet.
So what is the rescue technology? The same silhouette as Lazar’s coin-sized reactor: clean, compact, limitless power, arriving soon. The industry has fallen for next-generation nuclear — small modular reactors, and further out, fusion. Microsoft is paying to restart a reactor at Three Mile Island, rebranded the Crane Clean Energy Center, in a $1.6 billion refurbishment with Constellation, contracted to buy the entire output under a twenty-year agreement, targeting 2027. It has also signed with a fusion startup for power “later this decade.” Amazon, Google, and Meta have each announced nuclear commitments of their own; the pipeline of reactor offtake deals nearly doubled in a single year.
And here the pattern closes. The small modular reactors meant to solve this need five to ten years from permit to power, and cost more per kilowatt-hour than the large reactors and the renewables they are supposed to leapfrog. Fusion is, as it has been my entire life, later this decade. Roughly forty percent of the additional electricity these data centers will pull through 2030 is still projected to come from natural gas and coal. The fastest thing you can actually stand up on-site is a gas turbine, carbon and all. And the quietest, most telling number in the entire sector is that something like a third of data centers are expected to go fully off-grid by the end of the decade — “bring your own power,” the industry calls it, which is the sound of a whole industry conceding that the public grid cannot carry what it has promised. Restarting a reactor built in the 1970s and firing up gas peakers to bridge the gap is the moscovium-290 of the energy transition: what is actually on the floor, decaying in real time, while the stable clean reactor stays in the render.
A reactor with no fuel
There is a deeper point buried in the Lazar story, and it is the one worth ending on. Every real form of flight ever built is a method — an arrangement of forces that anyone with enough skill and resources can eventually reproduce. A rocket is a fuel and an oxidizer and a nozzle. An ion engine strips a gas and hurls it out the back with a field. The secret of any genuine capability is a technique, which means the rest of the world can eventually reach it, because techniques diffuse. Lazar’s story inverted that. In his telling, the secret was not a method. It was a material. You either had the metal or you didn’t, and if you didn’t, all you had was a box of parts.
The AI industry badly wants intelligence to be a material — a proprietary substance, a moat you can own, a miracle you hold and others don’t. That framing is what justifies the valuations and the trillion-dollar build-out. But the honest read, from someone who ships these systems in production on ordinary silicon every day, is that the exponential everyone is pricing in depends on two materials the industry does not possess. It needs a stable qubit, and the stable qubit is a preprint whose author’s own peers cannot yet inspect the proof. It needs a clean reactor, and the clean reactor is a decade out while the gas turbines spin up. Strip both away and look at what is actually on the floor: classical chips, improving through real but incremental engineering, running headlong into an energy ceiling the grid cannot meet.
None of which means AI is fake, or useless, or going anywhere — I would not have built a career on it if I believed that. It means the story being sold is running out ahead of the physics, the way it always does, and that the two breakthroughs meant to close the distance are both being announced rather than demonstrated. Bob Lazar’s tell, in the end, was never that he was lying. It was that his central claim had quietly retreated to a place no experiment could reach — a stable isotope no one could hold, always recovered from somewhere else, always just out of view. Listen closely to the AI roadmap and you will hear the same retreat. The stable qubit is the next chip. The clean power is the next decade. The proof is the unpublished data. Trust us.
Nobody keeps antigravity in a jar. And nobody, so far, keeps a working topological qubit or a clean gigawatt in one either. Until someone finally puts one of these materials in your hand — until it is demonstrated and not merely announced — the road to superintelligence is a beautiful reactor, exquisitely engineered, with nothing to put in it.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs and Chief Architect at Prana Entertainment, where he builds agentic AI systems in production on the Microsoft and .NET stack. Across nearly thirty years he has shipped software for Microsoft, Sony, Rockstar Games, 2K Games, Best Buy, and Allegiant Air; he was the sixth employee at Break.com and a Los Angeles Business Journal CTO of the Year finalist. His current work centers on MCP servers, retrieval-augmented generation, SQL Server 2025 vector search, and his PLogger observability framework. He is the author of the novel The Archive of Lost Suns, co-hosts the motorcycle podcast Edge Grip, and trains Brazilian jiu-jitsu in Las Vegas. He writes about the distance between what executives promise about AI and what actually ships in production.