E185｜The Quantum Race Without Consensus: Contenders Vie as Microsoft Faces Controversy
Summary
- The key investment fact in the quantum race is not that a winner has emerged, but that hardware and algorithms still lack consensus. After Nvidia brought together 12 quantum companies and quantum-computing scientists from Microsoft and Amazon at GTC, what emerged was not a single breakthrough path resembling the early GPU market, but radically different approaches. John, an investor in D-Wave and IonQ, summed it up: “There is no unified path, no consensus, no coordination, no economic impact, and no efficient use of scarce specialist talent.” 泓君 later linked the subsequent pullback in quantum stocks to this problem.
- Microsoft’s topological approach is a high-odds bet: if it works, it skips levels; if the evidence is insufficient, it remains at the starting line. Majorana zero modes encode information in a globally entangled structure and could theoretically resist local noise by design; one logical qubit might be equivalent to a dozen or more of Google’s physical qubits. But the episode contrasted Willow’s more than 100 qubits with Microsoft’s “0.5 qubit.” 尤亦庄 stressed that the paper demonstrated a potential readout method, but “we still haven’t seen very clear evidence” that the device contains genuinely topologically protected Majorana zero modes.
- Amazon’s Ocelot pushes error correction into the physical structure through “cat states,” and has reported lifetimes measured in seconds. It turns a single fragile quantum state into a redundant, multicellular-like structure, extending the lifetimes of its alive and dead eigenstates to the second scale—roughly 1000x longer than those of other qubits. Ocelot currently has 5 qubits, which by Amazon’s definition can be treated directly as noise-resilient logical qubits.
- The five mainstream approaches have no single overall champion; the competition is fundamentally a trade-off across scale, speed, quality, and cost. Neutral atoms can scale relatively easily to 1000 qubits and operate at room temperature; trapped ions have only a dozen, a few dozen, or roughly 32, and may be 100x slower, but are more precise; photons are cheap and travel at light speed but are lossy and difficult to make interact; superconducting systems are the most mature and have relatively low error rates, but require millikelvin cooling; topological systems offer the brightest “quality” upside, though their scale and speed remain unproven.
- AI and quantum computing are not substitutes; they are two-way infrastructure that lowers each other’s bottlenecks. AI can already optimize quantum-circuit compilation, learn surface-code error correction from real-device data, and build digital twins. In the other direction, quantum sampling could accelerate generative-model deployment, quantum neural networks have been theorized to offer an N-to-N² memory scale, and QAOA-type search has been summarized as reducing classical time T to √T. “Learn from nature and use big-data methods to drive the design of error-correcting codes” is a combination already taking shape.
- Before universal fault-tolerant computing, value may be realized in quantum chemistry, drug and materials screening, graph optimization, portfolio construction, and route planning. The ceiling for classical simulation of quantum systems is roughly 20 qubits on a laptop, 30 on a server, and 40–50 on a large server; 70, 80, and even 105 qubits are already entering territory beyond classical coverage. 尤亦庄 believes these quantum-simulation and optimization tasks “do not require a fully error-corrected quantum computer” and are already under development.
- Large-scale, meaningfully useful quantum computing is still estimated to be 10 to 15 years away, but the industry inflection point may arrive as a launch event rather than a linear progress bar. 尤亦庄 imagines that one day people may think 100 qubits “can’t do anything,” only for Google to announce the next day that it has built the system; deployment could look more like the sudden opening of ChatGPT. The ultimate pace will depend largely on when leading companies complete the engineering. As for teleportation, manipulating spacetime, and obtaining the universe’s “root access,” the episode explicitly places them in the distant category of “not forbidden by physical principles, but still science fiction from an engineering standpoint.”
Deep dive
1. GTC Exposed a Lack of Coordination, Not a Flourishing of Approaches
泓君 opened with an Nvidia historical analogy: even early 3D graphics cards, despite pixel dropouts and insufficient Z-buffer precision, solved a clear pain point—running Quake at high image quality and frame rates. 黄仁勋 had hoped quantum computing would find a similarly “very low-bar” breakthrough opportunity.
But this GTC put 12 prominent quantum companies and scientists from Microsoft and Amazon around the same table, revealing hardware and algorithms moving in different directions. Unlike last year’s historic conversation with 8 Transformer authors, which focused on a shared technical foundation, the discussion lacked a common center, while scarce specialist talent was difficult to deploy efficiently.
John, an investor in D-Wave and IonQ, offered the sharpest post-event verdict: “There is no unified path, no consensus, no coordination, no economic impact, and no efficient use of scarce specialist talent.” 泓君 later connected the broad decline in quantum-related stocks to this problem, calling it the industry’s biggest issue.
2. Microsoft Is Betting on Majorana Zero Modes to Build a “Natural Quantum Hard Drive”
尤亦庄 explained that Station Q has spent years pursuing topological quantum computing, encoding quantum information in a pair of Majorana zero modes. His core analogy was a coffee spoon tracing 2 vortices across a liquid surface: the vortices record not the state of any individual water molecule, but the pattern of the overall flow.
When 2 Majorana zero modes are separated far enough, they are unlikely to annihilate each other again. Local disturbances such as foam, ripples, and thermal fluctuations also struggle to alter the global entangled structure. The information is therefore stored as if in a background fluid, allowing the qubit to become a noise-resistant “quantum hard drive” from the moment it is created.
This differs from manually designed error-correction schemes. Superconducting physical qubits typically first minimize error rates, then use many physical qubits to encode a logical qubit; if the topological approach works, it obtains protected logical information directly at the physical layer—“winning from the starting line.”
The scale comparison explains the odds of Microsoft’s bet: the episode said Willow already had more than 100 qubits, while Microsoft had only “0.5 qubit.” But if Microsoft actually produces one logical qubit, it could be equivalent to a dozen or more of Google’s physical qubits. Under pessimistic assumptions, one logical qubit could require physical qubits in the millions; the range itself shows that the cost of error correction remains unsettled.
3. Microsoft’s Controversy Is Not the Theory, but Whether the Evidence Rules Out False Signals
尤亦庄 drew a clear line around the controversy: topological quantum computing and the Majorana approach are “feasible and very promising” in theory. The real problem is that experimental systems are complex, and the same signal can be produced by multiple mechanisms. Seeing a signal does not prove that zero modes exist, much less that they are topologically protected.
History has made the field especially cautious. Over the past decade-plus, multiple groups—not just Microsoft—have tried to realize Majorana zero modes. Some papers claimed to have found them, only for the signals to be identified later as false, with some papers even retracted. Independent replication and cross-validation therefore matter more than a single release.
On Microsoft’s paper, 尤亦庄 first marked the limits of his expertise: “I’m not a complete expert in this area.” His reading was that the paper described an experimental method for reading out information from a Majorana qubit, but its conclusion depended on the premise that the device had already realized a Majorana qubit. In the data, he said, “we still haven’t seen very clear evidence.”
The episode was recorded on March 12, when Microsoft was reportedly expected to release more data on March 17. 尤亦庄 was looking for more than additional data: he wanted replication by other experimental groups and a demonstration of topological logical operations by moving Majorana zero modes. A paper that has spent time in peer review may not reflect the laboratory’s latest progress.
4. Ocelot Trades for Second-Scale Stability by Creating a “Quantum Multicellular Organism”
Amazon’s Ocelot uses “cat states.” An ordinary qubit is like a single-celled organism that can be killed by one environmental disturbance; a cat state is like a multicellular organism whose life-or-death state is maintained collectively. Damage to a few components does not disable the whole, after which error correction can “repair the wound and treat the disease.”
Its key metric is that the cat state’s lifetimes in at least its alive and dead eigenstates reach the second scale—roughly 1000x longer than those of other qubits. The approach does not eliminate noise; it first places quantum information in a more redundant structure, then handles other types of errors.
Ocelot currently has 5 qubits. By Amazon’s definition, these cat-state qubits can be called logical qubits directly because they have built-in noise resistance. Both Amazon and Microsoft are trying to improve stability at the physical layer, but their mechanisms and scaling challenges are different.
5. Neutral Atoms Win on Scale and Flexibility; Trapped Ions Win on Precision
The neutral-atom approach uses lasers like tiny tweezers to arrange atoms, then moves the beams so selected atoms can approach, become entangled, and exchange information. Unlike circuits fixed onto a superconducting chip, the qubits can be moved and rearranged. Because all atoms of the same type are “born looking identical,” the approach also reduces differences caused by fabrication errors and the need for compensation.
This route can reach 1000 qubits relatively easily and can be controlled with lasers at room temperature. Atom Computing, Pasqal, and QuEra are representative companies. The bottleneck is not the supply of atoms, but laser-control precision and system noise; on “quality,” the episode judged superconducting systems to retain a slight lead for now.
Trapped-ion systems suspend charged atoms in the air with electric fields. Their stable oscillations resemble precision clocks, while lasers “conduct a perfect dance troupe” to execute computations. Current systems mostly have a dozen or a few dozen qubits; 泓君 mentioned roughly 32. Operations may be 100x slower than the first 2 approaches, but error rates are lower. They can also operate at room temperature and integrate with chip processes. IonQ and Quantinuum are representative companies.
Trapped-ion scaling is more likely to rely on multiple traps and multiple devices working together: one batch of ions performs calculations, moves to a waiting area, and new ions are loaded, rather than scaling a single trap indefinitely. The approach can also use deliberately engineered physical qubits to “build” topological states, then construct low-error logical qubits from high-error physical qubits. Microsoft, by contrast, hopes natural electrons will spontaneously form topological states.
6. Photons Are Poor at Collisions but Could Become the Interface for a Quantum Internet
Photons can encode 0 and 1 through polarization or path. They travel fast, are relatively insensitive to environmental interference, and can operate at room temperature and low cost. Their main advantage is also their main drawback: “The light from 2 flashlights simply crosses through.” Photons are difficult to make interact with one another, even though computation fundamentally consists of information carriers exchanging information through interaction.
The photonic approach therefore relies heavily on measurement-induced quantum entanglement. But measurement outcomes are probabilistic, and light is lost in fibers and devices, so “the bit may disappear halfway through the calculation.” Designing measurements, controlling large numbers of photons, and managing loss are the core problems facing groups such as Jiuzhang, Xanadu, and PsiQuantum.
尤亦庄 believes photons may occupy a unique role in quantum networks. A superconducting or neutral-atom computer in Beijing could first translate a quantum state into photons, send it by fiber to Shanghai, and convert it back at the local device. “Photonic quantum computing is particularly well suited to serve as the interface connecting different quantum computers over the internet,” rather than competing with every other platform for general-purpose computation.
7. Superconducting Systems Remain the Giants’ Main Track, While Subroutes and Modular Strategies Continue to Diverge
In a superconductor, charge moves collectively like a wave. Fabricating the superconductor into a circuit is like building channels that allow current waves to fluctuate; different modes of fluctuation encode 0 and 1, while external electric fields control and couple them. IBM, Google, Amazon, and Rigetti have all invested in this relatively mature route, which has attracted substantial capital.
Superconducting systems are themselves splitting into subroutes. Transmon relies more heavily on different charge levels in a capacitor and can be understood roughly as “more electric”; IBM and Google are representative. Fluxonium uses magnetic flux in a superconducting loop and is “more magnetic.” Rigetti explores both, reflecting the greater willingness of startups to pursue more aggressive, high-difficulty approaches.
Rigetti also emphasizes modularity: there is no need to etch 1000 qubits onto a single chip if multiple smaller chips can be connected. The focus shifts to solving the connection between chips and the transfer of quantum information. This reduces the manufacturing pressure on a single chip, but transfers the problem to interconnects, bandwidth, and cross-chip quantum-information transfer.
The broader map also includes topological computing in the wider sense—Majorana is only one route, with anyons another theoretical path—as well as silicon-based electron-spin qubits pursued by Intel and others. The latter can leverage existing semiconductor-process advantages, but the episode judged it to be at an earlier stage and not yet mainstream.
8. “Scale, Speed, Quality, and Cost” Better Reflect Real Competitiveness Than Qubit Count
尤亦庄 proposed 4 evaluation criteria: scale means the number of qubits, speed means operations per unit time, quality means low error rates and high precision, and cost means manufacturing and maintenance expense. Superconducting systems are relatively mature and currently have lower error rates, but require millikelvin cooling; neutral atoms scale easily and are relatively economical at room temperature; trapped ions are slow but precise; topological systems, if successful, would maximize “quality.”
“10 chips with 32 qubits each” is not equivalent to one 320-qubit chip. If a task can be divided across smaller chips and run in parallel, modularity works well. If an algorithm requires many qubits to remain coordinated, quantum information must repeatedly cross a bandwidth-constrained bus; communication costs can erase the gains from scaling. There is therefore no universal answer to connectivity outside the context of the task.
尤亦庄 offered a simple benchmark for classical simulation: an Apple laptop reaches roughly 20 qubits, a powerful computer or server around 30, and even the world’s largest servers generally no more than 40–50. A chip with 70, 80, or even 105 qubits has essentially moved beyond the computational reach of classical computers, but being “impossible to simulate” does not automatically mean having commercial value.
9. Willow Demonstrated Sampling Speed, but Not Automatically Practical Value
For the benchmark task promoted by Willow, a classical computer would take longer than the age of the universe, while the chip needs only a dozen or so seconds—or a similar amount of time. 尤亦庄 summarized the result as proving speed, but not guaranteeing that the output is 100% correct; on certain tasks, it has already demonstrated very high speed.
泓君’s counterpoint is worth preserving: this is a mathematical task designed to test quantum devices, not a concrete problem in daily life or scientific research. Quantum advantage first appearing in random sampling does not mean a general-purpose software stack like classical AND, OR, and NOT gates already exists, much less that users can “tell a quantum computer to do whatever they want.”
尤亦庄 further cautioned that today’s quantum computers struggle even to multiply 2 numbers, but they do not need to compete on tasks where classical computers are strongest. Quantum computing must find tasks native to its architecture rather than porting classical software over one item at a time.
Random sampling may look like “generating a pile of not-very-useful random numbers,” but it underpins generative models for images, language, and other domains. The scenario is that, once a model has finished training, a quantum computer could be deployed as a high-speed generator: early ChatGPT could crash under traffic, while future quantum sampling might answer the questions of “100M people” simultaneously in a very short time. This is a directional view, not an achieved capability.
10. AI Has Already Entered Quantum Compilation, Error Correction, and Digital Twins
The same high-level quantum algorithm can correspond to many gate sequences. AI can search for the circuit with the fewest operations and the best layout, then compile the algorithm onto a specific device. Shorter circuits mean less accumulated noise, so software optimization translates directly into effective hardware performance.
Human-designed error-correction schemes typically rely on idealized models and struggle to cover correlated error bursts caused by events such as cosmic rays hitting a chip. AI can observe a device while collecting data, learn how “this quantum computer in its current state” fails, and generate a targeted correction strategy.
DeepMind and Google Quantum AI have already validated this idea within the surface-code framework. Algorithms trained on ideal laboratory data perform reasonably well; training them on real experimental data makes them better still. 尤亦庄 summarized the principle as: “Learn from nature and use big-data methods to drive the design of error-correcting codes,” which may outperform designs based on human experience.
AI can also learn device behavior and form a digital twin that behaves like a quantum computer, allowing teams to test and fail on lower-cost simulators first. Natural-language programming would resemble a quantum version of Cursor or an AI agent, but 尤亦庄 stressed that the ambition is higher: it should not merely perform low-level steps, but understand frontier quantum phenomena that even top experts have not fully mastered.
11. Quantum Computing’s Promise for AI Centers on Memory, Generation, Search, and Expressiveness
泓君 compared the randomness, interconnectedness, and “a small disturbance affecting the whole system” of the human brain to quantum processes. 尤亦庄 offered only a cautious possibility: researchers including Professor Matthew Fisher of UCSB are studying quantum consciousness, and biochemical processes in the brain may be affected by quantum entanglement. “Our thinking may contain a quantum component—we just don’t know,” he said, without treating the analogy as a conclusion.
A more concrete direction is quantum neural networks. The episode cited theoretical results from 郜勋, 扈鸿业, and others suggesting that N qubits can carry an information scale corresponding to N² classical bits, potentially expanding the contextual memory of language models. Papers and mathematical proofs exist, but “there is no company” that has completed the implementation.
Turning this mathematical advantage into a quantum large language model will require both hardware and software to mature. 尤亦庄 estimated 10 to 15 years. Generative models could benefit more directly from faster quantum sampling; IBM’s use of a quantum chip to generate handwritten digits was presented as a research example, not proof that a general-purpose generation platform has matured.
On the training side, quantum search methods such as QAOA can use superposition to search neural-network parameters in parallel. The episode summarized the theoretical speedup as reducing classical time T to √T. IBM is also studying quantum circuits as kernel or feature functions: quantum entanglement supplies correlations unavailable to classical systems and may extract hidden relationships in classical data more effectively. 尤亦庄 said quantum algorithms have greater expressive power in classification tasks.
12. Quantum Chemistry and Combinatorial Optimization May Arrive Before Universal Fault-Tolerant Computing
The underlying behavior of molecules, drugs, and materials is governed by quantum mechanics. Rather than struggling to simulate these systems on classical computers, quantum devices can directly serve as quantum simulators. The relevant algorithms can search for a molecule’s ground-state configuration—the most stable structure—then rapidly screen candidate drugs and materials, reducing expensive experimental trial and error.
QuEra has demonstrated combinatorial optimization in graph theory by processing atom arrangements, showing higher efficiency than classical computing on the task described in the episode. Potential extensions include portfolio construction, quantum finance, and route planning for autonomous driving, though these still depend on how real-world problems map onto the hardware and on device quality.
尤亦庄 emphasized that research itself is an application. Google and IBM are already working with universities to use quantum chips beyond the reach of complete classical simulation to study quantum phenomena. Quantum chemistry, quantum simulation, and some optimization tasks also “do not require a fully error-corrected quantum computer,” so they may generate value before a general-purpose commercial system arrives.
13. Teleportation Must First Bridge a Scale Gap of 20 to 30 Zeros
尤亦庄 explained quantum teleportation through a future post office. Post offices A and B first establish a large number of entangled pairs. A jointly measures the quantum state of the package and a local qubit, then sends the classical measurement results to B. B uses those results to operate on its local entangled qubit and recover the package’s corresponding quantum state.
This does not copy the original object to B. The classical measurement results still have to reach B, and the original package is destroyed during A’s measurement. The process therefore looks like “moving an object instantly from one place to another,” but the original does not survive.
An ordinary package contains roughly 10^23 atoms. Describing its internal state precisely could require 10^20 to 10^30 or more qubits and entanglement resources; current devices have only hundreds or thousands, a gap of “20 to 30 zeros.” The episode explicitly said that teleporting packages, let alone people, remains science fiction.
A more distant extrapolation is that some theoretical physicists believe gravity may be fundamentally related to large-scale quantum entanglement. If humans could control entanglement precisely, they might alter gravity and the structure of spacetime, linking teleportation to wormholes. 尤亦庄 repeatedly qualified this: “Physical principles do not prohibit these things from happening.”
14. Commercial Use Could Arrive Suddenly, but Large-Scale Utility Is Still Estimated at 10 to 15 Years
尤亦庄 believes “application” should not be limited to the consumer market. Quantum devices already serve academic research and have reached scales that classical computers cannot simulate. Cross-industry and everyday use are difficult to forecast with a 10%, 20%, or 100% progress bar because, once a viable route is found, the remaining engineering scale-up could arrive in a concentrated burst.
His scenario is that one night people still think 100 qubits “can’t do anything,” and the next day Google announces that it has built the system. The transition would look more like the change before and after ChatGPT opened to the public. The timing of commercialization will depend heavily on a handful of leading companies, while startups and academia will play an important role in lowering knowledge barriers and helping other industries understand how to use the technology.
尤亦庄 estimated that “large-scale, meaningfully useful” quantum computing is 10 to 15 years away, while excluding current quantum simulation and optimization from the waiting list because they can develop on devices that are not yet fully fault-tolerant. The forecast remains an estimate, not a fixed countdown.
Faced with the split between academia and industry, 尤亦庄 said, “As I’ve gotten older, I feel that both are pretty good”: the ivory tower satisfies curiosity and intellectual challenge, while industry can truly serve society; industry, in turn, cannot do without a long-term supply of basic science. The ultimate vision is to understand “the universe itself as a quantum computer” and obtain “root access” to interact with its underlying laws, but the episode did not confuse that vision with a commercial roadmap.