Summary

Willow’s milestone is not that “quantum computing is already useful,” but that Google has physically demonstrated that quantum error correction can improve as systems scale. It completed a benchmark in under 5 minutes that classical computers would reportedly need about 10^25 years to finish, but this was not a practical task in medicine or AI; the real breakthrough is that as more qubits are added, the error rate no longer worsens in tandem and can instead decline. Roger sees it as quantum computing’s “Transformer moment”: the principles and algorithms have existed for years, but there is now a “live demo” showing that the system can scale.
Jensen Huang’s roughly 20-year timeline is better understood as defining when a large-scale fault-tolerant quantum computer will emerge, not when quantum computing begins creating value. Roger believes NVIDIA needs to stop Wall Street from pricing in quantum computing’s erosion of the GPU market over the next decade; but “don’t listen to what he says—look at what he does”: NVIDIA is already involved in partnerships and demos, while the guests’ team uses CUDA to connect GPUs and QPUs. Jared thinks the industry would have accepted 15–20 years without Willow, but the demo should pull forecasts toward within 15 years; Roger still sees a low probability of a complete system arriving within 15 years.
Over the next 10–15 years, QPUs and GPUs are more likely to scale together than quantum computing is to replace GPUs. For first-principles protein simulation and similar workloads, quantum computing could reduce a cost equivalent to 1 million GPUs to roughly the cost of 10,000 GPUs, eventually taking share from supercomputing workloads; in the near to medium term, GPUs will still be needed for quantum-chip design, control, runtime optimization and hybrid computing. The guests’ team is using CUDA as “glue” between GPUs and its proprietary QPU, while exploring quantum-enhanced AI that can train more generalizable models with fewer parameters and less data.
Willow is forcing the entire industry to redraw its road map, but bets cannot be ranked by qubit count alone. IBM shifted toward error correction and modularity after releasing a chip with about 1,000 qubits; Amazon has publicly backed Cat qubit; Microsoft has “most likely abandoned” its topological-qubit route in favor of partnerships with Quantinuum and Atom Computing; D-Wave’s quantum annealing cannot run general fault-tolerant computing; IonQ encountered scientific challenges scaling from about 11 entangled qubits to 20–30; and Rigetti’s momentum has faded after being founded too early. Roger’s screening criterion is direct: “show me actual demonstration.”
Quantum security has moved from a distant assumption to real capital expenditure for banks, central banks and communications companies. The estimate of roughly 3,000–4,000 qubits needed for Shor’s algorithm assumes a fully fault-tolerant, fully error-corrected machine; on such a machine, a public key could be used to derive a private key, making Bitcoin’s elliptic-curve signatures more directly exposed than banking systems whose keys are not public. The guests estimate such a machine could arrive in about 10 years, but say they “would not be extremely surprised” by a breakthrough in 5 years or sooner. U.S. institutions, NIST, the Monetary Authority of Singapore, HSBC and JPMorgan have already advanced post-quantum cryptography or QKD, driven by one question: “What if?”
Near-term commercialization is not about waiting for the complete system, but first selling data centers and high-value institutions a “partially useful” enhanced server. The guests say their company’s largest revenue comes from supplying quantum computers to data centers and that it was already profitable on a gross basis in the previous year; once the manufacturing method is known and trial-and-error is excluded, the bill of materials for one machine is about $1M. Serving finance, research and national infrastructure first, then reaching consumers indirectly through new materials, drugs, AI and financial products, is a more realistic path than direct consumer adoption.
Capital will arrive first from strategic investors and sovereign funds, with traditional VC potentially following after rate cuts and Willow’s validation. Roger summarizes the three sensitive sectors this way: “advanced semiconductors are a very legendary past, AI is a very vibrant present, and quantum is essentially an inevitable future.” During the rate-hike cycle, governments and strategic capital focused more on that future; he had not yet seen a clear wave of traditional VC entry, but expects rate cuts combined with Willow to make attention “ramp up.”

Deep dive

1. Quantum Computing Is Not a Faster CPU, but an Exponential Accelerator for Specific Problems

Roger and Jared Ren studied under Oskar Painter at Caltech from 2014 onward, then continued their research at Berkeley and Caltech respectively. Painter’s group later moved wholesale into AWS’s quantum-hardware division, while the two chose to commercialize a decade of experience building quantum chips.
Roger uses classical bits as an analogy: ordinary computers are built from bits that can take only 0 or 1, while a qubit can exist in a superposition of 0 and 1. Multiple qubits can also become entangled and “change simultaneously,” creating a computational space unlike classical, one-by-one enumeration.
Three classical bits have 8 possible states that must be processed individually; three qubits, in the guests’ plain-language explanation, can cover 8 states simultaneously. When the scale increases from 3 to 4, the classical state space rises to 16, while the quantum side can still operate on the whole space at once, so the advantage expands exponentially with scale.
泓君Jane asked whether harder computations naturally offer greater quantum advantages. Roger immediately narrowed the boundary: quantum computing is not a replacement for classical computers, just as GPUs did not replace CPUs; it may deliver exponential speedups only on specific problems such as matrix operations, factorization, quantum simulation and combinatorial optimization.

2. AI, Cryptography, Chemistry and Optimization Form the Four Clearest Application Chains

Roger’s underlying explanation is simple: “The world itself is quantum.” Simulating the quantum world on a classical computer is inherently difficult, making quantum chemistry, oil and gas, chemicals and drug discovery a set of applications especially well matched to quantum hardware.
AI connects through matrix operations. Roger compares the transition from CPU to GPU with quantum algorithms such as HHL: GPUs use large numbers of cores to process repetitive matrix operations in parallel, while quantum computing may further accelerate matrix inversion and compress the training and inference pipeline.
Jared’s order-of-magnitude claim is aggressive but worth preserving in his own terms: a “100B” model on the GPU side roughly faces the complexity of 100B multiplied by 100B, while the quantum side might reduce that to an operation count of “log 100B.” The larger the model, the wider the gap in energy and resource consumption could become.
The other two chains are more direct: Shor’s algorithm maps to public-key cryptography and security, while combinatorial optimization maps to logistics and other multi-variable decisions. The guests never claim that quantum computing will accelerate every workload; they repeatedly stress “certain specific problems.”

3. Jensen Huang’s 20-Year Call Defines the “Complete” Quantum Computer

泓君 cited Jensen Huang’s CES range: 15 years for “very useful” quantum computing would be too early, 30 years too late, and roughly 20 years reasonable. The comments were followed by declines in several quantum-computing stocks.
Roger’s conflict-of-interest analysis is that Wall Street will price in the growth and monopoly position of quantum computing over the next decade. If NVIDIA acknowledges that quantum computing could threaten GPUs within 10 years, its own valuation takes the hit. “Letting NVIDIA’s stock fall instead of quantum stocks is a very reasonable decision.”
But Roger also emphasizes, “Don’t listen to what he says—look at what he does”: NVIDIA is already a “big fish” in the quantum ecosystem and is actively pushing partnerships and demonstrations. CUDA launched in 2007, cuDNN in 2014, Transformer in 2016 and GPT-3.5 in 2022; even a full industry transition from the starting point took no more than 15 years.
The definition on which both sides largely converge is that a very useful quantum computer is equivalent to a large-scale fault-tolerant quantum computer, much as AGI is used in the AI conversation. It could ultimately develop new materials and more capable AI and become central to computing centers, rather than remain an experimental device that can perform narrow tasks today.

4. QPUs and GPUs Will First Create an Incremental Market, Before Anyone Asks Who Erodes Whom

NVIDIA defines itself as “a supercomputing infrastructure company,” spanning AI, weather forecasting and chemical simulation. Quantum companies are targeting the same high-complexity problems, so the long-term overlap is real; the question is when each arrives.
Roger uses first-principles protein simulation to illustrate the potential substitution. If a center would otherwise be willing to buy computing equivalent to 1 million future GPUs, a quantum solution might achieve the target at a cost equivalent to about 10,000 GPUs. That share could indeed be diverted, but only once the quantum machine can handle the full workload without relying on GPUs.
Over the next 10–15 years, quantum computing will instead need GPUs for chip design, control, runtime optimization and hybrid computing. The two technologies will “co-exist and grow together,” rather than fight over a fixed installed base. Compute demand itself will continue rising, with QPUs pushing outward the boundary of what can be computed.
The two companies’ architecture uses CUDA as “glue,” directly linking GPUs and their proprietary QPU so the two sides can exchange data in real time. On NVIDIA’s website, such quantum companies are called quantum backend providers, while NVIDIA calls itself a GPU backend provider. Another path is quantum-enhanced AI: using fewer parameters and less data to build more generalizable models. SSI’s hiring of quantum-machine-learning talent is viewed as an early signal.

5. Willow Proved Experimental Scalability, but Not Value on Practical Tasks

In December 2024, Google said Willow completed a standard mathematical test in under 5 minutes that would take a classical computer about 10^25 years. 泓君 drew a sharp line: this is a benchmark for measuring quantum progress, not a purpose-built computation for medicine, AI or commerce.
Roger believes the scale and performance of the Sycamore era were insufficient to prove that error-corrected computing could scale. The algorithms and principles had been broadly understood for 10 or 20 years; what Google, IBM and others actually had to solve over the past decade was: “Can we really build a chip ideal enough?”
Willow may not have the world’s highest single-qubit fidelity, but it used hardware and engineering optimization to show that as the system expands, the gains from error correction can reduce the error rate while reliability and computational power rise together. “The more the better” now has evidence at the physical-device level.
Roger does not expect Google to sell Willow as general-purpose hardware. It is closer to an internal proof of concept, designed to convince management to keep funding Google’s AI quantum team. His analogy: first prove that a machine-learning model can scale, then people ask, “How about we scale it actually?”

6. The Difficulty of Error Correction Is That Qubits Cannot Be Treated as Independent Failed Transistors

Jared uses a 1.2-volt GPU as an example: classical logic can classify roughly 1 volt, 1.2 volts or even 1.5 volts as logical 1, absorbing a range of physical variation at the lowest layer so users barely notice that classical computers also operate amid noise.
A superconducting qubit carries only about the energy of one photon, making it highly vulnerable to environmental and internal errors. It can undergo not only 0-to-1 and 1-to-0 flips, but also phase shifts. Entanglement makes the problem harder: qubits cannot simply be pulled out and corrected one by one, so error correction must target an interdependent group.
Roger explains that Willow uses the mainstream surface code. Through the surface code and optimized hardware, it demonstrates that expanding the system along this route can actually reduce the error rate and increase overall computational power.
Full error correction is the target for a large-scale fault-tolerant quantum computer. The estimate of roughly 3,000–4,000 qubits assumes full fault tolerance and full error correction; it refers to the scale needed to run Shor’s algorithm later in the discussion, not to today’s highly noisy physical qubits.

7. Willow Pulls the Timeline from 15–20 Years Toward Within 15, but Gives No Exact Date

Jared says that without Willow, the industry would have been more comfortable with 15–20 years; after the “live demo,” forecasts should converge toward “within fifteen years.” Jared is more optimistic than Jensen, but Roger still does not assign a high probability to a complete system arriving within 15 years and says the industry consensus may still be about 15 years away.
Asked what fundamentally changes before and after having the chip, Jared says Willow does not directly accelerate other research. Its significance is that it reduces uncertainty: Google’s management has been persuaded by evidence of scalability and can now expand the chips and investment with greater confidence.
Roger also challenges the narrative that every frontier technology must automatically wait another 20 or 30 years. He compares fission’s rapid path from small experiments to products with fusion’s long-term improvement, arguing that when demand is strong and the prerequisite science is sound, the first engineered deployment need not take as long as the historical examples suggest.
This is not an unconditional forecast. Investment, market demand and commercial logic will all affect the pace. He describes controlled fusion as being in a second stage—“already producing reactions, but still pursuing a more elegant form”—while quantum computing is still competing to build the first complete general-purpose system.

8. Google’s Error-Correction Route Is Forcing IBM, Amazon and Microsoft to Reconsider Their Positions

Roger says IBM historically focused more on stacking up qubit counts and commercializing early, hoping to find useful applications without error correction. After releasing a chip with about 1,000 qubits, IBM adjusted its road map in 2024 to put quantum error correction and modularity more explicitly at the center.
His assessment is openly opinionated: Google’s team is smaller, but its public road map has cleared away uncertainty. The paper had already gone through peer review before formal publication, so competitors could see the direction early; the major companies’ road-map changes in recent months were not accidental.
Amazon has publicly backed a newer superconducting route known as Cat qubit. Oskar Painter’s former Caltech research group later moved wholesale into AWS’s quantum-hardware division. Roger says Google has clearly had an impact, but does not claim that Amazon has changed its entire route.
Microsoft placed an earlier bet on topological qubits that would be fault-tolerant from the outset, but has yet to complete a proof of concept. Roger believes it has “most likely abandoned” that route. Microsoft is now filling the gap with software and partnerships, running logical qubits on ion and atom platforms with Quantinuum, Atom Computing and others.

9. D-Wave, IonQ and Rigetti Expose the Risks of Specialization, Scaling and Timing

D-Wave chose quantum annealing because, when it was founded, programmable digital quantum computers were widely considered nearly impossible to build. Annealing is simpler and may commercialize earlier, but it cannot be programmed generally, run general quantum algorithms or perform surface-code error correction.
泓君 asked whether this was the “wrong choice.” Roger narrowed his answer: “It’s not that I completely reject it.” D-Wave may still find a use case; the problem is that the digital, error-corrected route once thought harder could mature earlier and have a much larger impact.
Ion traps were long viewed as more promising than superconducting qubits, with excellent controllability and fidelity at small scale. But scaling from about 11 entangled qubits in 2017 to roughly 20–30 today has run into fundamental scientific challenges involving confinement and control. The guests put Quantinuum at up to about 32 qubits and observe a performance crossover as both routes scale: ion traps are increasingly approaching superconductors, while superconductors are increasingly approaching ion traps.
Rigetti’s founder participated in early work on high-fidelity superconducting qubits and IBM’s architecture, but the company launched in 2013 “way too ahead of the curve.” Roger still expresses respect, but says its technology and commercial performance are now falling behind. It was founded too early, before the Transformer moment, and may have burned through its momentum before the key inflection point.

10. Startups’ Opportunity Is Not to Invent New Physics, but to Rewrite the System Architecture

Asked whether the giants’ routes are therefore more solid, Roger does not fully agree. Amazon is not necessarily ahead, and IBM remained on its original route for more than a decade. Big companies have the money and confidence to keep going, but that can also reduce the incentive to experiment with new architectures.
From the outset, the two guests’ company bet on multi-chip modular quantum computing and obtained related patents. IBM’s later public work on long-range couplers and modularity, in Roger’s view, shows that the architecture has moved from a startup idea into a major-company road map.
Roger rejects a form of “leapfrogging” that bypasses a proven platform to reinvent an unvalidated qubit. The version he supports is system-level innovation: combining proven superconducting devices through a more agile, flexible and efficient architecture. “Show me actual demonstration” remains the final standard.
Google is the exception among the giants. Its UC Santa Barbara lineage gave it a tunable superconducting architecture more flexible than IBM’s, allowing it to run an error-correction experiment that IBM had considered difficult. The difference between large companies and startups is not only how much capital they have, but whether they can change their mindset in time.

11. After 2019, Quantum Computing Shifted from Academic Exploration to Industrial Engineering

Roger’s view is unequivocal: “After 2019, academia simply cannot compete with industry.” Most researchers from his Caltech and Berkeley cohorts, as well as later generations, went to AWS, Google, IBM and other companies because scaling quantum systems requires industry’s capital, equipment, teams and delivery objectives.
Academia’s role is not to compete with companies on early commercialization. When there is no ROI, it should “rapidly trial, rapidly develop and explore.” Superconducting and ion-trap platforms initially depended on universities to prove feasibility; once the questions became engineering problems, industry took over.
He sees Martinis’s group at UC Santa Barbara pushing superconducting-qubit fidelity above roughly 99%, the error-correction threshold, in 2013 as the turning point. Google concluded that quantum computing was no longer merely basic science and began systematic investment in hardware engineering.
This also explains why the two guests left academia to start a company: the goal had shifted from proving that a phenomenon exists to manufacturing deliverable equipment, scaling it and building a product. A paper was no longer the most direct objective function.

12. A $1M Bill of Materials Is Not the Bottleneck; Yield and Manufacturing Scale Are

Jared says that once the manufacturing method is known and repeated trial-and-error is excluded, the bill of materials for one quantum computer is about $1M—“cheaper than people imagine.” Willow has roughly 100 qubits according to the program’s figures, but the count itself is not a performance conclusion.
Roger explains the scaling problem through chip yield: if 1,000 qubits are designed, perhaps only about 300 can ultimately be connected and operate within the target range. A large nominal number that cannot form a stable system is merely “pure show.”
Ion-trap scaling is constrained by physics such as confinement, while superconducting systems face more microfabrication and design challenges. The guests believe superconductors can at least draw on 40–50 years of accumulated semiconductor-manufacturing processes, making the engineering path clearer.
Beginning in the second half of 2024, traditional equipment makers such as Applied Materials and wafer fabs began participating in quantum-chip fabrication. Roger sees this as an industrialization signal: once downstream companies show that larger scale brings higher returns, upstream suppliers can conclude, “We can take over now.”

13. Sovereign and Strategic Capital Will Validate Quantum Financing First; Rate Cuts May Bring VC In

Roger compresses the three policy-sensitive sectors into one line: “Advanced semiconductors are a very legendary past, AI is a very vibrant present, and quantum is essentially an inevitable future.” That positioning determines the order in which capital arrives.
During the high-rate cycle, traditional VC had not visibly increased its participation; growth was concentrated in strategic VC and sovereign capital. He cites John Martinis’s new company Qolab, which raised about $16M from investors including the Japan Development Bank.
His view is conditional: a rate-cut cycle could bring VC back to long-duration hard tech, while Willow reduces uncertainty around the route. Together, they could make investment “ramp up.” The scale would still depend on how long the next rate-cut cycle lasts; this is not an already realized funding boom.

14. Shor’s Algorithm Turns Bitcoin’s Public Key into the Most Obvious Quantum Attack Surface

泓君 separates Bitcoin’s risk into its mining mechanism and elliptic-curve signatures. Jared confirms that Bitcoin uses two cryptographic systems and explains that the direct exposure is the signature: a wallet’s public key is accessible on-chain, while Shor’s algorithm can solve the discrete-logarithm problem in polynomial time and derive the private key.
Bank keys, by contrast, are not public information. Unlike Bitcoin, whose public keys are exposed, banks are not directly exposed in the same way. Jared stresses that any system with a public key and a dependence on the same class of mathematical problems falls within Shor’s attack surface—not only Bitcoin.
“4,000 qubits” is easy to misunderstand. Roger clarifies that this means roughly 3,000–4,000 qubits under full fault tolerance and full error correction—the scale of the very useful quantum computer Jensen described—not a few thousand highly noisy physical qubits.
Roger reasonably estimates that a machine capable of decryption is still about 10 years away, but refuses to treat that as a law of physics. In 2021 and early 2022, people might still have believed that AI capable of passing the Turing test was 8–10 years away; a few months of progress rewrote the forecast. If the breakthrough comes in 5 years or sooner, he “would not be extremely surprised.”

15. Banks Are Buying “What If?” Insurance with Post-Quantum Cryptography and QKD

Roger says U.S. federal agencies have been required to migrate to quantum-resistant encryption, after which NIST finalized three post-quantum encryption standards. The Monetary Authority of Singapore has advised financial institutions to adopt post-quantum encryption and secure keys; HSBC has tested secure communications in the U.K.; and JPMorgan and others have public projects underway. QKD is another form of quantum communications.
Post-quantum cryptography mainly replaces the encryption mechanism and can be deployed through software upgrades; high-throughput environments may additionally reconfigure ASICs. QKD is slower to deploy because it protects communications at the physical layer and requires fiber networks and communications infrastructure.
A detail from one bank partner is representative: banks are pressuring carriers such as AT&T to deploy QKD networks to secure Wall Street communications. The potential exposure is a “billion-dollar loss,” and financial institutions will not take the risk simply because the timeline is uncertain.
Roger therefore does not believe quantum decryption will inevitably cause the financial or cryptocurrency system to collapse. Crypto can also update its cryptographic mechanisms; the key is to migrate in advance. Citing The Three-Body Problem, he says, “Arrogance is the biggest problem,” and, “I destroy you; it has nothing to do with you.” The defenses can be upgraded; the real danger is assuming the threat cannot arrive early.

16. Commercialization Starts with Data-Center Servers, with Value Reaching Individuals Through Institutions

The guests define the near-term product as a quantum-enhanced computing platform. It is not a fully fault-tolerant system capable of running every algorithm, but a way to use quantum hardware over the next 5–10 years to find the “first payoff.”
Potential end users include large banks, central banks and infrastructure projects; the company is essentially a server supplier. Roger says equipment sales—and its largest revenue stream—currently come from providing quantum computers to data centers. The company was already profitable on a gross basis in the previous year, though it disclosed neither revenue nor net profit.
For ordinary people, quantum computing will initially resemble early computers and GPUs that spent years operating in the background: it will first serve drugs, materials, AI and financial products, then reach individuals indirectly through those products rather than through every household owning a QPU.
The guests cite the classic mistaken prediction that the world would need only 5 computers to illustrate the demand flywheel: more users uncover more applications, rising demand drives production, and production lowers costs. Quantum computers depend little on rare metals; they are expensive today largely because scale has not yet arrived. Once the complete system appears, a mature supply chain will itself reshape the cost curve.

Solly Solomou, Founder & CEO @ LADbible Group: From Penniless Shop Offices to Public Company | E1259 Bee AI: The Wearable Ambient Agent