In February 2025, Amazon Web Services announced its first quantum computing chip, "Ocelot," built with Caltech and described in a paper published in Nature. The headline claim was a big one: up to a 90% cut in the cost of quantum error correction, the single biggest obstacle standing between today's experimental quantum processors and machines that can do useful work. This post explains what Ocelot actually is, what has happened to the quantum computing race since the announcement, and where things realistically stand today.
What Ocelot Actually Is
Ocelot is not a general-purpose quantum computer you can rent on AWS. It is a small research chip with nine qubits total: five "data" qubits that hold quantum information, and four "ancilla" qubits whose job is to detect when an error has occurred. That is a lab prototype, not a product — AWS has been explicit that Ocelot is the first step in a multi-year hardware roadmap, not a shipping system.
Cat Qubits: The Actual Innovation
The interesting part isn't the qubit count — it's the qubit design. Ocelot uses "cat qubits," named for Schrödinger's cat, which encode information in a superconducting microwave resonator as a superposition of two opposite-phase states. The physical trick is that a cat qubit's resistance to one specific kind of error — a bit-flip, where a 0 flips to a 1 — improves exponentially as you pump more photons into the resonator. AWS reported bit-flip times approaching one second alongside phase-flip times of about 20 microseconds, a lopsided error profile that turns out to be very useful, because it means engineers only need to build correction circuitry for one error type instead of two.
Why That Lopsidedness Matters
Standard superconducting qubits (the kind Google and IBM use) are symmetric: they're equally likely to suffer bit-flips and phase-flips, so correcting both requires a two-dimensional error-correcting code — typically the "surface code" — which burns through enormous numbers of physical qubits to protect a single reliable "logical" qubit. Estimates for surface-code fault tolerance commonly run to hundreds or even over a thousand physical qubits per logical qubit. Because a cat qubit suppresses bit-flips on its own, Amazon only needs a simpler one-dimensional repetition code to mop up the remaining phase-flips. That's the actual source of the "90% cheaper" claim: fewer physical qubits and fewer control wires needed per unit of reliable quantum computation, not a discount on the chip itself.
Physical Layout
Ocelot's hardware consists of two stacked silicon microchips, each about 1 cm², electrically bonded together. Oskar Painter, AWS's director of quantum hardware, has estimated that chips built on this architecture could eventually cost as little as one-fifth of comparable error-corrected approaches, and that this could pull AWS's timeline to a practical quantum computer forward by as much as five years. Those are projections, not measured results — Ocelot has not yet been scaled up to test them.
How Ocelot Compares to Its Rivals
Ocelot didn't launch into a vacuum — it landed in the middle of a very public three-way sprint between Amazon, Google, and Microsoft, each betting on a fundamentally different way to build a qubit. Here's how the approaches actually compare, based on what each company has published and demonstrated rather than what it has promised:
| Chip | Company | Qubit type | Qubit count | Headline claim | Independent verification |
|---|---|---|---|---|---|
| Ocelot | Amazon AWS | Superconducting cat qubits | 9 (5 data + 4 ancilla) | ~90% lower error-correction overhead | Peer-reviewed in Nature; architecture uncontested, cost claim not yet demonstrated at scale |
| Willow | Superconducting transmon qubits | 105 | First "below threshold" result — errors fall as the chip scales up | Peer-reviewed and widely replicated in method; benchmark task has no known practical use | |
| Majorana 1 | Microsoft | Topological qubits (indium arsenide/aluminum nanowires) | 8 | Roadmap to 1 million qubits on a single chip | Disputed — Nature's own reviewers said the paper's data doesn't establish the claimed particle signature |
| Condor / Kookaburra | IBM | Superconducting transmon qubits | 1,121 (Condor, 2023); modular scaling via Kookaburra from 2026 | Fault-tolerant system ("Starling") targeted for ~2029 | Peer-reviewed incrementally; roadmap dates have slipped before |
| Tempo | IonQ | Trapped ions | 64 algorithmic qubits (Q1 2026) | 2 million physical / 80,000 logical qubits by 2030 | Commercial system, benchmarks published by IonQ |
Google's Willow: A Different Kind of Milestone
Google's Willow chip, announced in December 2024, made a different and arguably more load-bearing claim than either Amazon or Microsoft: it demonstrated "below threshold" error correction, meaning that as researchers tested larger patches of physical qubits (3×3, then 5×5, then 7×7), the logical error rate went down exponentially rather than up. That is the specific behavior a real fault-tolerant quantum computer needs, and it's the reason Willow's result has held up well under scrutiny. Willow also ran a benchmark sampling task in under five minutes that Google says would take the Frontier supercomputer roughly 10 septillion (10^24) years — an eye-catching number, but one that applies to a contrived benchmark with no known practical application, not to a real-world problem like drug discovery or logistics. In March 2026, Google's quantum team added a second hardware track entirely, based on neutral atoms held in laser tweezers, suggesting even Google isn't betting everything on one qubit design.
Microsoft's Majorana 1: The Claim the Physics Community Rejected
This is where the story has moved most since the original announcement. Microsoft's Majorana 1, unveiled February 19, 2025 — a week before Ocelot — claimed eight "topological" qubits built from indium arsenide and aluminum nanowires, with CEO Satya Nadella describing a clear path to a million qubits on a single chip within years, not decades. The claim rests on detecting a theoretical particle state called a Majorana zero mode. Independent physicists pushed back hard and quickly: Henry Legg of the University of St. Andrews said publicly that "it doesn't look like Majoranas," and called the claim potentially damaging to public trust in quantum research. Nature's own editorial team, after bringing in additional reviewers, concluded the paper's data does not establish the presence of the particle Microsoft says it found. As of 2026, that dispute is still unresolved in the scientific literature, even as Microsoft has continued to promote successor chips. The practical takeaway for readers: treat Majorana 1's "path to a million qubits" claim as unverified marketing until independent labs replicate the underlying physics.
What's Happened Since the Original Announcement
None of the three companies has shipped a bigger successor chip yet. AWS has said only that further Ocelot-generation chips are coming, each expected to add qubits and lower error rates, following what Painter calls a "flywheel" of iterative hardware improvement — but no second-generation chip had launched as of late 2025. The cat-qubit approach itself has kept advancing outside Amazon too: the French startup Alice & Bob, a direct competitor in cat-qubit hardware, reported bit-flip stability exceeding one hour on a chip called Galvanic Cat in September 2025, and introduced a new "elevator codes" technique in January 2026 to further cut error-correction overhead — both signs that Amazon's core bet on cat qubits is shared, and being pushed forward, by others in the field.
Elsewhere, IBM and IonQ have kept executing on steadier, less headline-grabbing roadmaps. IBM is targeting a modular processor called Kookaburra for 2026 and a fault-tolerant system called Starling around 2029, while IonQ's Tempo system reached 64 "algorithmic qubits" in early 2026 using trapped ions, with a stated roadmap of 2 million physical and 80,000 logical qubits by 2030. Neither claim has generated the kind of controversy Microsoft's did, partly because neither is asking anyone to accept a disputed physics result on faith.
So When Will Any of This Actually Be Useful?
The original post repeated Oskar Painter's estimate of 10 to 20 years to practical quantum computing and noted that a commercially viable machine would need roughly a million qubits. Both figures still hold up reasonably well, but they need context. Most current industry roadmaps and independent analysts converge on 2027–2030 for the first demonstration of quantum advantage on a problem that's actually commercially relevant — most likely in chemistry simulation or optimization, not cryptography. The million-qubit, fully fault-tolerant machine that could do arbitrary useful computation is a longer story: a naive Moore's-law-style extrapolation from today's qubit counts puts that milestone as far out as 2044, while more optimistic doubling-every-year projections put it around a decade away. For the specific scenario people worry about most — a quantum computer breaking RSA-2048 encryption — credible estimates cluster around 2033 to 2040 or later, which is also why the encryption industry is already migrating to post-quantum cryptographic standards now rather than waiting.
With Ocelot still sitting at nine qubits and no confirmed successor chip yet public, Amazon remains, by its own account, several scaling generations away from anything resembling a useful quantum computer. What's changed since 2025 isn't that Amazon has caught up to Google or Microsoft — it's that the field has gotten more legible: Google's error-correction result has held up under scrutiny, Microsoft's headline claim has not, and Amazon's cost-reduction bet is now being tested and extended by competitors outside the company. That's a more honest way to read the "quantum computing race" than any single chip announcement, including the one this post started with.
-EditorZ
Post a Comment