💻 Quantum

IBM Spent $10 Billion on Quantum Computing. Then It Bought a Lab That Was Laying People Off. The Nature Paper Explains Why.

IBM acquired HRL Laboratories, the 78-year-old research lab founded by Howard Hughes, days before a landmark Nature paper demonstrated 18 silicon spin qubits operating at 0.2% error rates. Three years ago, the state of the art was 2 qubits at 4% error. We calculated what it means that neither IBM nor Google is willing to bet on a single quantum technology anymore.

Marcus Chen

July 30, 2026 · ☕ 10 min read

Nine times more qubits, twenty times lower error rates, three years. On Tuesday, Nature published a paper from HRL Laboratories describing the most advanced silicon spin qubit processor ever demonstrated: 18 exchange-only qubits controlled by cryogenic electronics at 4 kelvins, connected via a superconducting ribbon cable, operating at error rates around 0.2%. One peer reviewer called it "a significant milestone in the technological maturity of semiconductor spin qubits." In 2023, the best anyone had managed was 2 qubits running at roughly 4% error.

Timing makes the paper hit harder. Six days before it went public, IBM announced it had signed a definitive agreement to acquire HRL Laboratories, a private R&D institution jointly owned by Boeing and General Motors. IBM did not disclose the acquisition price, and HRL, the former research arm of Howard Hughes' Hughes Aircraft Company, is the lab where Theodore Maiman demonstrated the first working laser in 1960. Earlier this year, it was conducting mass layoffs after losing U.S. government contracts.

From layoffs to the most consequential quantum acquisition of the decade in a matter of months. How that happened reveals something important about the entire quantum computing industry: after spending more than $10 billion, the world's leading quantum engineering teams still cannot tell you which technology will actually work. Nobody can. And IBM just said so with its checkbook.

The Two-Track Hedge

IBM has staked its quantum future on superconducting circuits. It operates the world's largest fleet of quantum computers over the cloud. It booked $1 billion in quantum business between Q1 2017 and Q4 2024. In May, the Trump administration awarded IBM $1 billion under the CHIPS Act to build Anderon, America's first pure-play quantum chip foundry, in Albany, New York. IBM matched with another $1 billion of its own. In June, IBM committed $10 billion over five years to advance its quantum roadmap, an amount larger than the entire GDP of some Pacific island nations and roughly equal to what the Apollo program cost in today's dollars when adjusted for its first five years alone, covering everything from R&D and capital expenditures to manufacturing scaling, ecosystem partnerships, and acquisitions that collectively represent the single largest corporate bet ever placed on a technology that has not yet demonstrated commercial superiority over existing computers. Its fault-tolerant machine, Starling, is targeted for 2029. A more powerful system called Blue Jay is planned for 2033.

And then IBM bought HRL and added a completely different quantum technology to its portfolio. A hedge. A confession.

"I and the team strongly believe that the future is going to be spins, or superconducting, or possibly a combination of them," Jay Gambetta, IBM's director of research, told Reuters. "They have a very strong spin qubit team, the strongest in the world. I would not pursue a second path that was not built on a foundation that could be integrated together."

IBM is not alone. Earlier this year, Google, which has also spent billions on superconducting quantum circuits, added neutral atoms as a second technology track. Both companies, having spent more money than anyone else on superconducting quantum computing, have independently concluded that superconducting quantum computing alone might not be enough.

What the Hedge Tells You

We can quantify this. IBM's total disclosed quantum expenditure now exceeds $12 billion: $10 billion in the five-year commitment, $1 billion in Anderon co-investment, the undisclosed but material cost of acquiring HRL, plus seven years of prior R&D. Quantum revenue over that same period: $1 billion. That is a 12-to-1 investment-to-revenue ratio, a figure that dwarfs any comparable technology bet. IBM is spending $12 for every $1 it earns from quantum computing. For context, Nvidia's ratio of R&D plus capex to revenue is roughly 0.4 to 1. Tesla's is about 0.35 to 1, and even SpaceX, in its most capital-intensive Starship development years, has never approached 12 to 1.

A company operating at a 12-to-1 investment-to-revenue ratio does not hedge because it is cautious. It hedges because the cost of being wrong is existential. Total loss. Losing a $12 billion bet on superconducting circuits to a competitor that backed the right alternative technology would not just be a setback. It would be a permanent structural disadvantage in a market that the Boston Consulting Group projects will reach $90 billion to $170 billion by 2040.

Gambetta confirmed this. He told Reuters that HRL's spin qubit technology will become critical after IBM's Blue Jay superconducting system ships in 2033. Spin qubits are not a supplement to IBM's existing program. They are IBM's insurance policy for the 2030s, hedging against the possibility that superconducting circuits hit a scaling wall that seven years of additional engineering cannot solve. Put differently: the director of research at the company that has invested more money in superconducting quantum computing than any other entity on the planet is explicitly planning for a future in which that investment may not be sufficient, and he just acquired the world's strongest team in the alternative technology to make sure IBM survives that scenario.

Why Spin Qubits Matter Now

Spin qubits encode information in the angular momentum of individual electrons trapped in wells carved into semiconductor wafers. The concept dates to the late 1990s, when Daniel Loss and David DiVincenzo, then at IBM, realized that electron spins could be controlled with electrical pulses the same way transistors are controlled in classical chips. Cheap. Familiar. Scalable. Spin qubits could theoretically be manufactured on the same 300-millimeter CMOS production lines that churn out billions of conventional silicon chips every year.

Execution was the problem, because for two decades, spin qubits remained stuck in a laboratory cul-de-sac, unable to compete. While superconducting systems scaled past 100 qubits and neutral-atom machines reached into the thousands, the best spin-qubit devices topped out at a handful of qubits with error rates that made useful computation impossible. Spin qubits were the quantum computing equivalent of a promising technology that could never get out of the lab.

HRL's paper changes the arithmetic, because it provides the first peer-reviewed evidence that spin qubits have reached competitive quality with far more established modalities. Here is the improvement trajectory:

The error rate improvement is more significant than the qubit count. A 20-fold reduction in error rate in three years, from 4% to 0.2%, puts spin qubits into the same quality band as superconducting circuits. Google's Willow chip, using superconducting transmon qubits, operates individual gates at about 0.1% to 0.3% error. HRL's spin qubits are now competitive on quality even though they trail dramatically on quantity. And RIKEN's preprint, if it survives peer review, suggests the floor may be far lower: below 0.01%, an error rate that superconducting circuits have not consistently achieved. And yet.

The Manufacturing Advantage Nobody Has Priced

There is a reason IBM's Gambetta specifically mentioned that HRL's team will start making chips at IBM's New York facility. Both superconducting and spin qubits can be fabricated on conventional semiconductor equipment. But spin qubits are physically much smaller than superconducting circuits, which means more qubits per wafer, which means lower cost per qubit at scale.

Consider what that means for the economics of fabrication. Anderon, IBM's new quantum foundry, will produce 300-millimeter quantum wafers, the same wafer size used by TSMC, Samsung, and Intel for their most advanced classical chips. A single 300-millimeter wafer can hold roughly 80 billion transistors in a modern classical process. Superconducting qubits require macroscopic circuit elements (Josephson junctions, resonators, coupling buses) that occupy orders of magnitude more area per computational unit than a transistor. Spin qubits, being formed from individual electron wells, are closer in scale to classical transistors.

At scale, this gap is enormous, and it explains why Gambetta specifically mentioned that HRL's chips would be manufactured in New York rather than remaining at HRL's Malibu facility. At IBM's current superconducting architecture, a 1,000-qubit chip requires a dedicated processor die roughly the size of a classical CPU. A spin-qubit architecture at similar error rates could theoretically pack 1,000 qubits into a fraction of that area, because each qubit is defined by a nanometer-scale potential well rather than a micrometer-scale superconducting loop. Cost follows density. That ratio determines the cost of the hardware that IBM will manufacture at Anderon. If quantum computing follows the same economics that drove classical semiconductors, the technology that puts more qubits per wafer at competitive error rates wins. HRL's paper is the first evidence that spin qubits can compete on quality, a fact that only matters because manufacturing physics have always suggested they can compete on density, and a technology that wins on both quality and density in a wafer-based production environment is one that the semiconductor industry, with its trillion-dollar infrastructure already optimized for exactly that kind of fabrication challenge, knows how to scale.

The Scoreboard Nobody Wants to Show You

Every major quantum computing company is now pursuing at least one technology track, and several are pursuing two. Here is where they stand:

Collective spending across these companies and others exceeds $22 billion. Nobody has achieved fault tolerance, and nobody has demonstrated unambiguous commercial advantage over classical computers for a real-world problem. Not once. Not yet. And the two biggest spenders just added second technology tracks because they are not confident their primary bets will scale.

This is not a criticism. It is an honest accounting of where the industry stands after more than two decades of research and over $22 billion in cumulative investment across all major players, including governments, corporations, and startups that have each bet on at least one of a half-dozen different physical implementations of the same theoretical promise. BCG's projection of a $90 billion to $170 billion market by 2040 requires that at least one of these approaches crosses the fault-tolerance threshold within the next several years. The two-track bets by IBM and Google suggest that the companies closest to that threshold believe the probability of any single approach succeeding is materially less than 100%.

Strongest Counterargument

Spin qubits are 18 qubits while superconducting systems are past 100 and neutral-atom systems are into the thousands. Quantity matters here. It matters enormously. The fact that spin qubits just achieved comparable error rates to superconducting circuits does not mean they can catch up on the metric that matters most for practical quantum computing: the total number of high-quality qubits operating simultaneously.

Error correction requires massive qubit overhead, which is the reason raw qubit counts dominate every quantum company's roadmap slides even though error rates determine whether those qubits can actually do anything useful once assembled into a fault-tolerant system. Current estimates suggest that producing a single logical (error-corrected) qubit requires somewhere between 1,000 and 10,000 physical qubits, depending on the error rate and the code used. HRL's 18-qubit system, even at 0.2% error, would need to scale by roughly 100 to 500 times just to produce one or two logical qubits. That gap is real. Brutally real. Superconducting systems, despite higher per-qubit costs, are already closer to that threshold in absolute numbers. IBM's roadmap calls for 100,000 superconducting qubits by the early 2030s. HRL's? Unpublished. That silence speaks. The manufacturing density argument is compelling in theory, but it requires HRL to demonstrate that spin qubits scale in practice the way classical transistors did, and classical transistor scaling took decades of engineering to achieve, required billions in tooling investment, and depended on materials breakthroughs that were far from guaranteed when the first integrated circuits shipped in the 1960s. IBM may be buying an insurance policy that never pays out.

Limitations

IBM did not disclose the acquisition price for HRL, making it impossible to calculate a precise cost-per-qubit for the spin technology. Our estimate of IBM's total quantum expenditure exceeding $12 billion aggregates publicly disclosed commitments and excludes undisclosed R&D embedded in broader IBM Research budgets, which would push the figure higher. Our $22 billion collective spending estimate for the industry is a lower bound based on public funding rounds, government awards, and disclosed corporate investment, not audited figures. Spin qubit scaling projections are theoretical; no group has demonstrated a silicon spin qubit processor with more than 18 qubits. Error rate comparisons across modalities should be treated with caution: different qubit types are measured against different benchmarks, and a 0.2% error rate for a spin qubit gate is not directly equivalent to a 0.2% error rate for a superconducting gate without accounting for gate times, connectivity, and cross-talk.

The Bottom Line

Howard Hughes' old research lab went from mass layoffs to the most important quantum computing acquisition of the decade in a matter of months, because it published a paper proving that silicon can make qubits as good as the superconducting circuits that IBM has spent $12 billion developing. That is the single most significant data point in quantum computing this year, and not because 18 qubits will change anything tomorrow. It matters because the number reveals that nobody, not IBM, not Google, not anyone spending billions, knows which quantum technology will actually scale.

If you work in semiconductor manufacturing, the Anderon foundry's 300-millimeter wafer capability for both superconducting and spin qubits means quantum chip production is about to enter the same supply chain you already operate in. If you invest in quantum computing stocks, the two-track hedges by IBM and Google should recalibrate your risk model: the companies closest to the technology are allocating capital as though each approach has a meaningful probability of failure. If you follow quantum computing from a distance, the number to remember is 20. That is how many times spin qubit error rates improved in three years. If that trajectory holds for three more years, the question of which quantum technology wins may be answered by a silicon chip fabricated on the same production line that makes your phone's processor.