AI Data Centers Have a Quantum Problem

[00:00:00] Speaker 1: In 2026, Amazon, Google, Microsoft, and Meta are on track to spend between $650 and $725 billion on AI data centers in a single year. Next year, Wall Street expects close to $1 trillion. All of it assumes this is how computing will work for the next 20 years. But in March 2026, Quantinuum ran calculations on 94 protected logical qubits, and Google is racing toward a commercial quantum computer by 2029. The smartest money in tech is now asking one question: What if the biggest bet in history is built on the wrong machine? Let us start with the size of this bet. Because it is easy to hear a number like $700 billion and feel nothing. So let us make it real. In 2025, the biggest cloud companies spent somewhere between $365 and $410 billion on infrastructure, depending on which analyst you ask. In 2026, that jumps by more than 70%. Amazon has committed to roughly $200 billion. Alphabet guided to between $180 and $190 billion earlier this year. Meta is guiding between $125 and $145 billion. And Microsoft spent nearly $31 billion in a single quarter, up 84% from the year before. Reuters noted that the combined 2026 spending from the big hyperscalers has already topped $700 billion, and that the number keeps rising. Gartner puts total global data center spending above $650 billion this year, up almost 32% from nearly $500 billion in 2025. Wall Street is not even sure that is enough. Analysts keep raising their forecasts. And the current consensus for 2027 sits around $934 billion, only about 7% short of the $1 trillion mark. JP Morgan expects the wider group of hyperscalers to climb from about $800 billion this year to $1 trillion next year. But investors are starting to flinch. When Meta raised its spending guidance and blamed higher component prices and extra data center costs, its stock dropped more than 9% in a single day. It was one of the first real signs that even the market wonders how long this can go on. And the question is simple: where is the money coming back from? To be fair, the revenue is real. Microsoft says its AI business has passed an annual run rate of $37 billion, up 123% in a year. That is explosive growth. But at its current quarterly pace, Microsoft alone is spending well over $120 billion a year on infrastructure. Every one of these companies is betting that demand will keep climbing fast enough to justify the concrete, the chips, and the power plants. Even Gartner has acknowledged that AI infrastructure keeps growing, despite constant concerns about an AI bubble. Because the real cost is not the buildings, it is the power. The International Energy Agency projects that electricity demand from data centers will more than double by 2030, to around 945 terawatt hours. That is slightly more than the entire electricity consumption of Japan today. Data centers built specifically for AI are expected to see their demand more than quadruple. In the United States, data centers will account for almost half of all growth in electricity demand between now and 2030. And by then, the US is expected to use more electricity to process data than to manufacture aluminum, steel, cement, and chemicals combined. Read that again: more power for data than for heavy industry. The IEA's base case for 2035 goes even further, to around 1,200 terawatt hours, with a range of anywhere between 700 and 1,700, depending on how fast AI actually grows. So here is the tension: the entire AI industry is racing to build more of the same kind of machine, faster, while the power grid struggles to keep up. And while that race is happening, a completely different kind of machine has been slowly climbing out of the lab. To understand why it matters, we need to go back more than 40 years. In 1981, the physicist Richard Feynman made a simple but radical point at a conference: "Nature does not run on classical logic. It runs on quantum mechanics." So if you want to simulate nature properly, you should build a computer that works the same way. For decades, most of the industry treated that idea as a beautiful curiosity. Qubits were fragile. A tiny bit of heat or vibration could destroy a calculation in a fraction of a second. Building a useful machine looked less like engineering and more like a dream. Here is the basic difference: A classical computer, including every GPU inside an AI data center, works with bits that are either zero or one. A quantum computer works with qubits, which can exist in a blend of both states, and can be linked together through something called entanglement. This does not mean a quantum computer is just a faster version of what we have today. It means it is extremely good at a specific family of problems, like simulating molecules, certain kinds of optimization, and breaking some types of encryption. For those problems, the speed difference can be absurd. For others, it gives you nothing at all. Keep that in mind because it matters a lot later in this video. The first big claim came in 2019, when Google said its chip had done in minutes what would take a supercomputer 10,000 years. IBM pushed back quickly, arguing that with better methods, the same task would take a few days. For years, that argument hung over the entire field. Every quantum claim came with an asterisk. Then, in October 2025, Google published a result in Nature that was different. Its Willow chip, with 105 qubits, ran an algorithm called Quantum Echoes. The key measurement took about two hours on the quantum chip. Google's team estimates that the Frontier supercomputer would need around three years for the same job. A gap of roughly 13,000 times. And here is the important part. The result is verifiable. Other quantum machines can repeat it and get the same answer, which earlier claims could not offer. Google used the method to model molecules with 15 and 28 atoms and matched real laboratory data from nuclear magnetic resonance experiments. Now, to be fair, researchers have pointed out that this was a narrow scientific task, not a general purpose win. But for the first time, the asterisk got a lot smaller. But a faster chip is not enough. The real wall in quantum computing has always been errors. Physical qubits are noisy, so engineers bundle many of them together into one reliable unit, called a logical qubit. Think of it like a team of unreliable workers, who keep checking each other's math until the group answer can be trusted. There is a famous rule behind this idea: if the error rate of each physical qubit falls below a certain threshold, then adding more qubits makes the whole system more reliable, instead of less. Google's Willow chip was one of the first machines to show that effect in real hardware. That is why the number to watch is not how many physical qubits a company announces, it is how many logical qubits it can actually run. In March 2026, Quontinuum demonstrated computations using up to 94 protected logical qubits on a trapped ion processor. And IBM has put a date on its goal. It is building a machine called Starling, planned for 2029, designed to run 100 million operations on 200 logical qubits. After that comes a system called BlueJ, targeting 2,000 logical qubits and 1 billion operations around 2033. IBM's new error correction design, called the Gross Code, packs 12 logical qubits into 288 physical qubits in total. And the company says its newer codes cut the physical qubit overhead by roughly 90%. These are roadmaps, not finished machines. And roadmaps slip. But the direction is clear. And then there is the man whose company sits at the center of the entire AI boom. In January 2025, Nvidia CEO Jensen Huang told analysts that very useful quantum computers were probably 15 to 30 years away. Quantum stocks collapsed. Rigetti, IonQ, and Quantum Computing, Inc. all fell more than 40% in a single day, wiping out billions in market value. Two months later, at Nvidia's own conference, Huang hosted a Quantum Day and joked that it was the first event where a CEO invites all the guests to explain why he was wrong. In June 2025, he said Quantum was reaching an inflection point. And by October 2025, Nvidia had launched NVQ Link, a connection that links quantum processors directly to GPU supercomputers and called it the start of the quantum GPU computing era. The biggest skeptic in the room did not just change his tone, he started building the bridge. Google's own leadership has been just as direct. Right after Huang's comments in early 2025, Google said it was optimistic that real-world quantum applications would appear within five years. Two companies, two very different timelines, and both of them have a trillion-dollar stake in being right. So is Quantum about to kill AI data centers? This is where we have to be honest, because a lot of videos on this topic skip this part. No, not in the way the headlines suggest. Training a large AI model means pushing enormous amounts of data through huge numbers of matrix calculations. And that is exactly what GPUs were built to do. To this day, nobody has shown a quantum machine that can train or run a frontier language model better than a GPU cluster. Quantum computers are specialists. They are not replacements for the general purpose muscle inside a data center. And the money tells the same story. McKinsey estimates that quantum computing companies earned between 650 and 750 million dollars in revenue in 2024 and expected to pass 1 billion dollars in 2025. Compare that with more than 700 billion dollars in AI infrastructure spending in 2026. That is a gap of roughly 700 to 1. Even McKinsey's optimistic scenario for 2035 puts the quantum computing market between 28 and 72 billion dollars. That is big. But it is still a fraction of what the AI build-out costs in a single year today. So what would actually have to happen for quantum to threaten the build-out? Three things: First, logical qubits would have to scale from the hundreds we see today to thousands and do it reliably. Second, someone would have to find quantum methods for the workloads that actually fill AI data centers, which so far does not exist at scale. And third, the cost of running those machines would have to beat a rack of GPUs on price, not just on speed. None of those three are impossible. But none of them are done either. That is why the honest answer is not that quantum is killing the boom. It is that quantum is the one variable the boom has not priced in. So where is the real quantum problem for AI data centers? There are two, and both are more real than the clickbait version. The first is that data centers are not being replaced. They are being turned into hosts. IBM is building Starling inside a data center in Poughkeepsie, New York. Google built a quantum AI campus in Santa Barbara that includes its own data center. And NVIDIA's NVQ-Link exists for one reason: to wire quantum processors into the same racks that hold GPUs. Even the startups are thinking this way. Xanadu, a photonic quantum company, has published a roadmap that includes networked quantum data centers, launching around 2029 and 2030, with up to 200 logical qubits in 2029, and more than 1000 by 2031. The most likely future is a hybrid one, where GPUs do the heavy lifting, and quantum chips handle the narrow jobs where they dominate. That means every hyperscaler has to plan for a second kind of hardware, with a different power profile, different cooling, and different software. That is a new layer of cost and complexity that the current build-out was never designed around. The second problem is the one that should keep security teams awake at night. Data centers hold the world's information, and most of it is protected by encryption, that quantum computers are expected to break one day. In 2019, Google researcher Craig Gidney estimated that breaking a standard RSA 2048 key would take about 20 million noisy qubits. In May 2025, he published a new estimate, fewer than 1 million qubits, running for less than a week. That is a 20 times drop, achieved through better algorithms rather than new hardware. Today's machines are still far from a million qubits, so nobody is saying your encryption is broken today. But there is a catch, and security experts have a name for it. Harvest now, decrypt later. Attackers can copy encrypted data today, store it, and simply wait. In August 2024, the US standards agency, NIST, finalized its first post-quantum encryption standards, and its draft guidance says vulnerable systems should be deprecated after 2030 and disallowed after 2035. Google has already started rolling out one of those new standards across Chrome and its internal systems. That means every major data center operator needs a migration plan, and soon, not a replacement of the building, a replacement of the locks on every door. Money is already moving toward that second clock. McKinsey counted about $10 billion in new public quantum investment announcements in early 2025 alone, with Japan accounting for nearly 75% of it. Private companies are raising Syria's capital, too. Zanadu reported roughly $686 million in total available capital as of the end of June 2026. That is small, next to a $700 billion buildout, but it is no longer pocket change. And it is enough to keep building real hardware for years. Governments now treat this race as a strategic priority, not just a science project. So here are the two clocks that are running right now. The first clock is the AI buildout, spending around $700 billion this year, and aiming for a trillion next year. The second clock is the quantum timeline, where IBB says 2029. Google has targeted the same year, and independent estimates generally put a truly useful, fault-tolerant machine somewhere between 2028 and 2032. If the quantum clock runs fast, companies that bet everything on one type of machine will have to adapt quickly. If it runs slow, the AI data center boom gets another decade to pay for itself. Nobody knows which clock wins. And that uncertainty is exactly why the smartest money in tech is quietly hedging. Here is the bottom line: quantum did not kill AI data centers. Anyone who tells you that is selling a headline. But something quieter is happening. The machine that the whole industry is betting on is no longer the only machine being built. The same companies writing those $700 billion checks are also funding the technology that might one day change what those checks buy. Feynman said it in 1981. Google proved a piece of it in 2025. Quantinuum showed 94 logical qubits in 2026. And IBM has circled 2029 on its calendar. The real story is not the death of the data center. It is the moment an industry that spent a trillion dollars realized it might need a second brain. If you want to keep tracking both of those clocks, subscribe to the channel. And tell me in the comments, do you think useful quantum computers arrive before 2030, or are we still a decade away?