Why Did Goldman Sachs Shut Down Its Quantum Computing Team?

Because the arithmetic came back at roughly 8 million logical qubits. An internal study concluded that a genuinely useful portfolio optimisation algorithm would need machines on that scale and would still take millions of years to run. For comparison, IBM’s fault-tolerant target for 2029 is 200 logical qubits — a factor of forty thousand short. JPMorgan looked at the same evidence and kept more than 50 researchers, betting that algorithmic decomposition closes the gap from the other end.

In 2026 two of the most sophisticated quantitative institutions in the world looked at the same evidence about quantum computing and reached opposite conclusions.

Goldman Sachs dismantled most of its quantum research team. An internal study had concluded that a genuinely useful portfolio optimisation algorithm would need on the order of 8 million logical qubits and would take millions of years to run. JPMorgan Chase kept going, with a group of more than 50 physicists, mathematicians and computer scientists working on everything from cryptography to machine learning.

Neither is being foolish. The disagreement is about timing, and the gap between what the hardware can now do and what a bank would need it to do is the most useful thing to understand about this field.

The Number That Explains the Split

Put IBM’s 2029 fault-tolerance target next to Goldman’s requirement and the shape of the problem is obvious. Starling is planned to deliver 200 logical qubits capable of more than 100 million operations per job. Goldman’s study pointed at eight million.

That is a factor of forty thousand, and it is not a gap you close by waiting a couple of product cycles. It is the difference between a machine that can demonstrate something and a machine that can replace a server rack doing the same job today.

A logical qubit, incidentally, is not a physical one. It is an error-corrected abstraction built out of many physical qubits — which is why IBM’s Kookaburra prototype needs roughly 4,158 physical qubits to work toward logical storage at all. When a vendor quotes a qubit count, the first question is which kind.

What Genuinely Got Built

The engineering progress in 2025 and 2026 is real, and it is worth separating from the commercial claims made on top of it.

Google’s Willow chip demonstrated below-threshold error correction: the point at which adding more physical qubits makes the logical error rate go down rather than up. That had been the field’s central obstacle for two decades, and clearing it moved quantum error correction from a physics question to an engineering one.

Microsoft and Quantinuum’s H2 system reached 12 logical qubits at a logical error rate around 2 in 1,000 — better than the underlying physical error rate, which is the whole point of error correction.

IBM has published a dated roadmap and has so far hit it: Loon in 2025 for high-rate qLDPC codes, Nighthawk at 120 qubits with a next iteration due at the end of 2026, Kookaburra in 2026 as its first logical-qubit storage prototype, Cockatoo in 2027 to entangle separate processors, and Starling in 2029 as the fault-tolerant system. IBM says it is on track to demonstrate verified quantum advantage by the end of 2026.

None of these is a product launch. They are proof points, and the companies making them mostly say so.

Why JPMorgan Stayed

JPMorgan’s bet is that the algorithms will meet the hardware partway. With Amazon, its researchers built a decomposition pipeline that breaks large portfolio optimisation problems into smaller pieces, cutting problem sizes by up to 80%.

That is the counter-argument to Goldman’s eight million: if you can decompose the problem far enough, you do not need the machine that solves it whole. Whether that holds all the way down to hardware that will actually exist is precisely what is in dispute, and it will not be settled by argument.

No bank is running quantum hardware in live operations. Everything in this area is still research, including the parts with press releases attached.

The Money Is Not Waiting for the Physics

Private capital has continued regardless. PsiQuantum raised $1 billion in September 2025 at a $7 billion valuation, led by BlackRock with Temasek and Baillie Gifford, then $1.5 billion in May 2026 at $10.5 billion.

Public markets have been far less settled. IonQ’s market capitalisation was about $21 billion in May 2026, roughly $16.7 billion by July and $17.6 billion in August — some 58% below its high.

Market-size estimates disagree so widely that they are close to useless without their definitions attached: one house puts the 2026 market at $1.9 billion growing to $8.0 billion by 2033; another puts 2025 at $3.62 billion and 2026 at $5.09 billion. When forecasts differ by more than a factor of two for the current year, they are measuring different things.

The One Part a Business Should Actually Act On

Almost nothing above is actionable for a small or medium business. This is, and it is not speculative.

NIST finalised the first post-quantum cryptography standards in August 2024: FIPS 203 (ML-KEM, key encapsulation), FIPS 204 (ML-DSA, digital signatures) and FIPS 205 (SLH-DSA, hash-based signatures). The NSA’s CNSA 2.0 framework sets the timetable for national security systems — quantum-safe algorithms for new systems by January 2027, application migration by 2030, full infrastructure migration by 2035.

The reason the deadlines run ahead of the hardware is harvest now, decrypt later: an adversary records encrypted traffic today and decrypts it whenever a capable machine arrives. The attack is passive, undetectable, and it does not require the quantum computer to exist yet. Anything you transmit today that must stay confidential into the 2030s — medical records, legal files, long-term contracts, source code — is exposed to it now.

Practically, for most businesses this is a supplier question rather than a project. Hybrid key exchange combining X25519 with ML-KEM-768 is already running in production browsers and major CDNs, so much of the migration arrives through software you already buy. The useful action is to ask your vendors — hosting, email, VPN, payments — what their post-quantum timeline is, and to know which of your own data still needs to be secret in ten years.