IBM Spins Off the First Pure-Play Quantum Chip Foundry

IBM is spinning off its quantum chip manufacturing operation into a standalone foundry called Anderon: the first of its kind dedicated purely to quantum processors. Backed by a proposed $1 billion CHIPS Act award from the U.S. Department of Commerce and matched by $1 billion from IBM, Anderon will be headquartered in Albany, New York, operating as a 300-millimeter quantum wafer foundry. A separate business whose entire purpose is fabricating quantum chips, including for customers who have nothing to do with IBM's own quantum computing stack.

The semiconductor industry has seen this pattern before with classical chips. IDMs (integrated device manufacturers) that both design and fab their own silicon eventually face a choice between optimizing for internal needs and opening the fab to outside customers. When they open up, foundry economics start to apply and the whole industry changes. IBM is attempting that transition in quantum. The surface similarity to the classical story is real, but it hides some differences that determine whether this actually works.

What quantum chip fabrication actually involves

Superconducting qubits (the architecture IBM uses) are fabricated on silicon or sapphire substrates using processes that look superficially similar to classical semiconductor manufacturing but differ in ways that matter enormously. The Josephson junctions at the heart of each qubit are thin-film structures, typically aluminum on aluminum oxide, where the tunnel barrier thickness needs to be controlled to sub-nanometer precision. The operating frequency of a qubit is directly tied to its capacitance and junction inductance, both set during fabrication. You can't trim a qubit after the fact the way you can trim a resistor.

The fabrication requirement that most separates quantum from classical is cleanliness: specifically, materials purity and surface defects. Two-level systems (TLS), microscopic defects at material interfaces, are one of the dominant sources of decoherence in superconducting qubits. A contaminant or surface imperfection that would be irrelevant in a CMOS process can destroy qubit coherence. Substrate surface preparation, dielectric deposition, and metal patterning all have tighter process tolerances than equivalent classical nodes. The failure modes are different in kind, not just degree.

IBM's quantum fabrication work, and the capabilities being transferred into Anderon, is centered at its Albany NanoTech facility: a 300mm wafer fabrication site located at SUNY Polytechnic Institute's Colleges of Nanoscale Science and Engineering in Albany, NY. The spinoff takes those manufacturing capabilities (people, equipment, process knowledge) into a separate legal entity that can contract with external customers while IBM's quantum computing business continues on a separate track.

Why a pure-play foundry model makes sense right now

The quantum hardware space has proliferated fast. IonQ, Quantinuum, Rigetti, Google, Microsoft, Amazon: each pursuing different qubit modalities and different hardware roadmaps. Most smaller players don't have in-house fab capability at the level IBM has built. Rigetti runs its own Fab-1 facility in Fremont, California, which is part of why it has always been capital-intensive. Startups pursuing photonic qubits or neutral atoms often have no fab at all, relying on external fabs or academic facilities that weren't designed for this work.

A credible quantum foundry addresses a real gap. Companies working on quantum hardware need access to process nodes that classical fabs don't optimize for, tighter tolerances than classical fabs will run, and process engineers who understand why the qubit frequency drifted between wafers. That specialized service doesn't currently exist at scale outside a handful of internal R&D operations.

The TSMC analogy is tempting, and IBM has used it explicitly. It's useful but should be held loosely; the exact points where it breaks down are where the real risk sits.

The PDK problem: why external customers can't just send a design file

In classical semiconductors, a foundry publishes a process design kit: layer stackups, device models, design rules, SPICE parameters, everything a chip designer needs to tape out without visiting the fab. A well-specified PDK means a fabless company can design a chip, send GDSII to a foundry, and get back working silicon. The process is the product; the foundry's value is executing a known specification reliably at scale.

Quantum doesn't have this. Each qubit design is highly coupled to the specific process parameters of the fab it was developed in. Transmon qubit geometry, junction area, dielectric stack: all of it interacts. A design optimized for IBM's internal process won't necessarily hit target frequencies if the junction oxidation recipe changes by 10%. Process control in quantum fab is improving, but a customer can't yet hand over a design and reliably get back qubits hitting spec.

Building a foundry business requires building something like a PDK first: a well-characterized process specification that external customers can design against. That's a substantial body of engineering work that precedes any external revenue. What Anderon's timeline looks like for that, and whether the spinoff implies it's further along than has been publicly disclosed, remains an open question. The depth of whatever process documentation IBM eventually publishes will say more than any press release about how ready this actually is.

The cryogenic screening bottleneck

Classical foundry yield is the percentage of dies on a wafer that pass electrical test at probe. Quantum yield is harder to define and harder to measure. A qubit that passes DC characterization at room temperature may fail coherence testing at millikelvin. Wafer-level qubit screening at cryogenic temperatures is not a standard production process anywhere; it requires dilution refrigerators for screening, which are slow, expensive, and in short supply.

A quantum foundry that wants to quote yield to customers needs a qualification pipeline that handles this. The classical assumption (that electrical test at room temperature is a reliable proxy for device performance) doesn't carry over. This is the constraint most likely to derail the foundry's throughput, and IBM hasn't addressed it directly in public statements. The fabrication process itself is hard; the screening pipeline is arguably harder to scale.

Who the customers actually are

The most obvious customer segment is quantum hardware startups that need superconducting qubit fabrication but can't justify building their own fab. Fab construction at the level IBM operates runs into the hundreds of millions of dollars in capital equipment, plus the process expertise to run it. Most startups currently rely on small-scale academic fabs or internal processes running in leased cleanroom space.

A second segment is defense and government. DARPA has been funding quantum hardware development (ONISQ, the Quantum Benchmarking Initiative) and defense contractors building quantum sensing systems need access to high-quality superconducting device fabrication. Quantum sensors based on superconducting resonators and SQUID arrays have near-term defense applications that don't require fault-tolerant qubits. That market may be more immediately viable than the quantum computing customer segment, and the IP sensitivity is lower, which matters for the trust problem covered below.

There's also a longer-term segment: companies in adjacent fields that need superconducting device fabrication for non-qubit applications. Superconducting nanowire single-photon detectors, kinetic inductance detectors, and certain microwave components all use fabrication processes adjacent to qubit manufacturing. Some of that market exists today in astronomy and quantum networking research, and classical fabs don't serve it well. It's a smaller dollar opportunity but it's real and near-term.

The IP separation problem IBM hasn't solved yet

IBM's spinoff puts it in a structurally awkward position relative to its quantum computing competitors. Google and Microsoft both fabricate their own qubits internally. Google's Willow processor was fabricated in Google's specialized facility in Santa Barbara, and Microsoft's topological qubit work (the Majorana 1 chip) relies on a proprietary materials stack of indium arsenide and aluminum developed and fabricated in-house. Neither is a foundry customer candidate.

Smaller hardware companies (Rigetti specifically, and various European and Asian quantum hardware startups) could theoretically become foundry customers. The problem is that Rigetti is also a direct competitor to IBM's quantum computing cloud offering. Foundry relationships in classical semiconductors navigate this tension with strict IP separation between foundry and design businesses. TSMC spent years building that reputation. IBM is starting from zero on that front.

That separation is hard to make convincing when the foundry's process knowledge was built entirely in service of IBM's own qubit roadmap. The question a potential customer has to answer: does Anderon have sufficient independence that my process IP is safe from the people running IBM's competing quantum computing business? IBM's structure will need to answer that explicitly, not just assert it. If Rigetti or any other superconducting qubit company eventually signs on as a customer, that's the clearest possible signal the answer is yes.

What to watch

• Whether Anderon publishes something resembling a public process design kit for external customers, and how characterized the process parameters are when they do. The depth of that document will say more than any press release.
• Who the first external foundry customers are. Defense contracts and government programs seem more likely near-term than commercial quantum computing competitors: the economics work better and the IP sensitivity is lower.
• How the foundry entity handles cryogenic screening at scale. That bottleneck either gets solved or becomes the rate-limiting factor for throughput, full stop.
• Whether Rigetti, or any other superconducting qubit company, becomes a customer. That would be the clearest signal that the IP separation is credible.
• How the foundry prices access. Classical foundry pricing is capacity-based with well-understood economics. Quantum foundry pricing is genuinely novel and will reveal a lot about who IBM thinks its primary customers are.

Where this lands

It's unclear how fast the PDK and process standardization problem gets solved. IBM's internal process control for its own qubits has improved substantially; the jump from 127-qubit Eagle to 433-qubit Osprey to 1121-qubit Condor involved real yield engineering, not just scaling. But internal process control optimized for one design is different from a characterized foundry process that external customers can design against confidently. Those are different engineering problems, and IBM hasn't had to solve the second one before.

The TSMC analogy is useful for understanding the business model IBM is trying to build. It's less useful for predicting the timeline, because TSMC's foundry model emerged in an industry that already had CAD tools, standard cells, and process design kits: the ecosystem that makes fabless design possible. Quantum is missing most of that. IBM is trying to build the foundry while the design ecosystem is still being invented, and those two things have to co-evolve. That's a harder problem than the announcement implies.

If IBM pulls it off, Anderon accelerates the entire hardware ecosystem by giving startups access to a world-class superconducting fab without the capital burden of building their own. If the process standardization work isn't as far along as the announcement implies, the foundry spends its early years serving internal IBM workloads under a different legal entity name. Those are very different outcomes, and the announcement doesn't tell you which one you're looking at.

If you're working in quantum hardware (at a startup, a national lab, or a government program) and have a view on whether the PDK or the cryogenic screening problem is closer to being solved than the public record suggests, that context would be genuinely valuable. Those two things are what keep coming up when trying to figure out whether this foundry is real near-term or a longer bet than IBM is letting on.