FAULT-TOLERANT QUANTUM COMPUTING AND DRUG DISCOVERY: RESOURCE ESTIMATES FOR 2026

The promise of quantum computing in pharmaceutical research is specific and well-defined: quantum systems can represent and evolve quantum mechanical wavefunctions natively, whereas classical computers must approximate them — and those approximations become exponentially more expensive as molecular systems grow in size and electron correlation complexity. A quantum computer running the quantum phase estimation algorithm can, in principle, determine the ground-state energy of a drug molecule's active site to chemical accuracy (approximately 1 kcal/mol) in polynomial time, where the best classical methods require exponential resources for the same precision.

The gap between that promise and operational hardware is enormous — and precisely quantifiable. The field of quantum resource estimation has matured significantly since 2018, with detailed studies from Google, Microsoft, IBM, and academic groups establishing concrete qubit counts, gate depths, and T-gate budgets for specific pharmaceutical targets. This article synthesises those estimates, explains the engineering challenges behind the numbers, and situates them against the actual trajectory of hardware development in 2026 — covering superconducting systems, trapped ion platforms, and photonic approaches, including PsiQuantum's million-qubit photonic roadmap.

What Is Fault-Tolerant Quantum Computing and Why Does It Matter for Drug Discovery?

All current quantum computers are noisy intermediate-scale quantum (NISQ) devices. Physical qubits — whether superconducting transmon circuits, trapped ytterbium ions, or photonic modes — interact with their environment in ways that cause decoherence (loss of quantum information) and gate errors (wrong operations applied with some probability). On IBM's best hardware in 2024, single-qubit gate error rates are approximately 0.03% and two-qubit (CNOT) gate error rates hover around 0.3%. These numbers sound small, but a quantum algorithm of pharmaceutical relevance requires circuits with millions to billions of gate operations — meaning errors accumulate catastrophically.

Fault-tolerant quantum computing solves this by encoding each logical qubit across many physical qubits using a quantum error-correcting code. The code continuously measures error syndromes — detectable signatures of errors — without collapsing the encoded quantum information, and classical decoding algorithms correct identified errors in real time. Below a critical physical error rate threshold (approximately 1% per gate for the surface code), increasing the code distance — adding more physical qubits per logical qubit — reduces logical error rates exponentially. This is the qualitative transition that separates NISQ from fault-tolerant computation.

For drug discovery specifically, fault tolerance matters because the algorithms with genuine quantum advantage — quantum phase estimation for ground-state energy, quantum simulation of open quantum systems for protein-drug binding — require long, deep circuits that NISQ devices cannot execute with useful fidelity. Variational quantum eigensolvers (VQE), which are NISQ-compatible, have been demonstrated on small molecules such as H₂ and LiH, but scaling them to pharmaceutical-relevant molecules (50-200 heavy atoms) faces a barren plateau problem: gradients vanish exponentially with system size, making classical optimisation intractable. The drug discovery quantum advantage window is firmly in the fault-tolerant era.

The Physical-to-Logical Qubit Overhead: Why One Million Qubits Is the Minimum

The most frequently cited resource estimate for pharmaceutical quantum computation is from Babbush et al. (2018, Nature Chemistry), refined by Lee et al. (2021, PRX Quantum) and Kivlichan et al. (2020, PRX Quantum). For simulating the electronic structure of the FeMoco cluster — the active site of nitrogenase, with 54 atoms and 113 electrons — to chemical accuracy, Lee et al. estimated approximately 4,000 logical qubits and 10&sup9; T-gate operations. For a larger but more pharmaceutically general drug molecule with ~100 heavy atoms, estimates range from 1,000 to 4,000 logical qubits.

The physical-to-logical qubit ratio depends on the target logical error rate per operation and the hardware physical error rate. For surface codes at a code distance of d = 17 (reasonable for physical error rates near 0.1%), each logical qubit requires approximately 2d² ≈ 578 physical qubits in the data layer, plus ancilla qubits for syndrome measurement — totalling roughly 1,000 physical qubits per logical qubit. Encoding 4,000 logical qubits therefore demands ~4 million physical qubits. Encoding 1,000 logical qubits at the same code distance requires ~1 million physical qubits.

Improving physical error rates reduces the required code distance and therefore the physical qubit overhead. If physical two-qubit gate error rates fell to 0.01% (a factor of 30 improvement over IBM Heron's 2024 performance), code distances of d = 7 would suffice for many applications, reducing the overhead to approximately 100 physical qubits per logical qubit — bringing the total physical qubit requirement for the FeMoco simulation to ~400,000. Microsoft's topological qubit programme targets physical error rates of 10⁻⁶, which would reduce overhead to tens of physical qubits per logical qubit. These are not yet demonstrated at scale.

Surface Codes and Error Correction Thresholds: The Engineering Requirements

The surface code, introduced by Kitaev (1997) and adapted for superconducting hardware by Fowler, Martinis, and colleagues (2012, Physical Review A), dominates current fault-tolerant architecture designs because it requires only nearest-neighbour interactions on a 2D qubit grid. This matches the planar lithography of superconducting chip fabrication. Stabiliser measurements (which detect errors without reading the logical qubit directly) can be implemented using only 4-body Pauli measurements on adjacent qubits — achievable with existing two-qubit gate primitives.

The surface code's fault-tolerance threshold is approximately 1% per gate operation — higher than most competing codes, which is why it remains the hardware community's preferred choice despite its relatively poor encoding rate (1 logical qubit per d² physical qubits). Below threshold, logical error rates fall as p_L ∝ (p/p_th)^⌈(d+1)/2⌉, where p is the physical error rate and p_th is the threshold. At p = 0.3% (IBM Heron's 2024 two-qubit error rate) and p_th = 1%, increasing code distance from d = 5 to d = 17 reduces logical error rates by roughly five orders of magnitude — from ~10⁻⁴ to ~10⁻⁹ per logical operation.

Alternative error-correcting codes are attracting growing interest. Low-density parity-check (LDPC) codes — including the bivariate bicycle code and the gross code — achieve better encoding efficiency (more logical qubits per physical qubit) at the cost of requiring higher-connectivity qubit graphs. Google's Willow processor (105 qubits, released late 2024) demonstrated exponential error suppression with increasing code distance in surface codes — the first experimental confirmation that the basic scaling behaviour of quantum error correction operates as theory predicts. This is a milestone result, even though Willow is many orders of magnitude short of pharmaceutical relevance.

IBM Heron and the Below-Threshold Milestone in 2024

IBM's Heron processor, released in December 2023 and featured prominently in IBM's 2024 quantum roadmap, represents the current performance frontier for superconducting gate-based quantum computing. Its two-qubit (ECR gate) error rates of approximately 0.2–0.3% place it comfortably below the surface code's 1% fault-tolerance threshold — an important distinction that was not consistently achieved by any superconducting system before 2023. Single-qubit gate error rates on Heron are approximately 0.02–0.05%, and median T1 coherence times exceed 300 microseconds on the best qubits.

Heron uses a fixed-frequency transmon architecture with tunable couplers between qubits, allowing two-qubit gates to be performed without tuning the qubit frequencies — which was a major source of errors in earlier flux-tunable designs. The processor has 133 qubits in a heavy-hex connectivity graph (each qubit connects to at most 3 neighbours), which is optimised for surface code implementation but imposes routing overhead for algorithms requiring arbitrary qubit connectivity.

Critically, Heron's below-threshold performance does not mean IBM has demonstrated fault-tolerant computation — it means the hardware is operating in the regime where adding more physical qubits and increasing code distance would actually improve logical qubit fidelity, rather than making it worse. The 2025–2026 IBM roadmap targets 1,000+ physical qubits on a single chip with Heron-level error rates (the “Flamingo” modular architecture), with quantum interconnects linking multiple chips. IBM's stated goal is a “quantum-centric supercomputer” with over 100,000 physical qubits by 2033 — still 10x short of what pharmaceutical simulation would require without complementary advances in code efficiency.

T-Gates and Magic State Distillation: The Hidden Algorithmic Overhead

Quantum error correction protects Clifford gates (H, CNOT, S) natively — they can be implemented transversally (qubit by qubit across the code block) without spreading errors. However, universal quantum computation also requires non-Clifford gates, specifically the T-gate (a π/8 rotation). T-gates cannot be implemented transversally in the surface code without a technique called magic state distillation: preparing many noisy copies of a “magic state” (the +1 eigenstate of the T-gate) and distilling them into fewer, high-fidelity copies using purification circuits.

Magic state distillation consumes enormous physical qubit resources — typically 100–1,000 physical qubits per T-gate operation, in dedicated “distillation factories” that run in parallel with the main computation. For pharmaceutical algorithms with T-gate counts of 10⁹ to 10¹², the distillation overhead can dominate total physical qubit requirements, often exceeding the data qubit footprint by a factor of 10. This is why raw logical qubit counts from resource estimates must be multiplied significantly to get total physical qubit requirements.

Reducing T-gate counts is therefore a primary focus of quantum algorithm optimisation for pharmaceutical applications. Techniques include Trotterization order selection (higher-order product formulas reduce Trotter error per step and therefore total gate count), tensor hypercontraction (Lee et al. 2021 — reduces T-gate count by ~50x versus earlier algorithms for FeMoco simulation), and quantum signal processing / qubitization (Babbush et al. 2019). These algorithmic advances are as important as hardware improvements in closing the gap between current capabilities and pharmaceutical utility. The interplay between fault tolerant quantum algorithms, machine learning-assisted circuit compilation, and resource estimates for logical qubits and T-gates is now a distinct research subfield, with groups at Google Quantum AI, Microsoft Research, and the University of Sydney publishing major results annually.

Photonic Quantum Computing and PsiQuantum's Million-Qubit Roadmap

PsiQuantum, founded in 2016 by Jeremy O'Brien, Terry Rudolph, Mark Thompson, and Pete Shadbolt — all from the University of Bristol's photonics group — is pursuing a fundamentally different physical platform: photonic quantum computing using silicon photonic chips manufactured in GlobalFoundries' 300mm semiconductor fabrication lines. Rather than superconducting circuits cooled to near absolute zero, PsiQuantum's architecture uses single photons as qubits, with two-qubit interactions mediated by measurement-induced nonlinearity (the KLM scheme, after Knill, Laflamme, and Milburn).

The photonic approach has a critical advantage for scale: silicon photonic components — waveguides, beamsplitters, phase shifters, single-photon detectors — can be fabricated using existing CMOS semiconductor processes, enabling millions of components per chip with mature yield engineering. PsiQuantum's thesis is that reaching one million physical qubits is an engineering problem solvable with semiconductor manufacturing capacity, whereas superconducting systems face fundamental limits in qubit coherence, interconnect bandwidth, and dilution refrigerator scalability.

The core challenge for photonic quantum computing is two-qubit gate error rates. Linear optical two-qubit gates are probabilistic — they succeed only a fraction of the time — requiring photon multiplexing and adaptive switching to achieve effective deterministic operation. Achieving two-qubit gate error rates below 0.1% in a photonic system at scale has not yet been demonstrated publicly. PsiQuantum's investors (including Microsoft, BlackRock, and the Australian government's $940M AUD commitment in 2023) have sustained the company through its pre-demonstration phase. The photonic quantum computing two-qubit gate error rate question — specifically, whether PsiQuantum can reach below-threshold performance — remains the central open question for the platform. As of mid-2026, no below-threshold demonstration on a photonic system comparable to IBM Heron's superconducting benchmark has been published in peer-reviewed literature. The IBM Heron logical qubit coherence time (indirectly, through quantum memory experiments) currently represents the closest proxy for below-threshold operation in an accessible, documented system.

Realistic Timelines: What the 2030–2035 Consensus Actually Means

The 2030–2035 window cited as the mainstream consensus for fault-tolerant pharmaceutical quantum computing is not a single study's conclusion — it is a rough convergence of resource estimation benchmarks, hardware roadmaps, and expert survey results. The 2019 National Academies report “Quantum Computing: Progress and Prospects” concluded that “it will likely take many decades” to build systems capable of practically relevant quantum simulation, while noting that hardware was advancing faster than the committee's baseline expectations. The McKinsey Global Institute (2021) and Boston Consulting Group (2021) independently placed the first “quantum advantage” in drug discovery at 2030 in their central scenarios.

Three conditions must be met approximately simultaneously for the window to open. First, physical qubit counts must reach the millions with Heron-level or better error rates — requiring either continued superconducting scaling or a photonic breakthrough. Second, qubit connectivity and routing must improve to reduce T-gate overhead from circuit transpilation. Third, classical co-processing — the decoding of error syndromes in real time — must keep pace with the syndrome generation rate, which for a million-qubit surface code at 1 MHz clock speed means decoding on the order of 10¹² syndrome bits per second. This classical decoding bottleneck is itself a major research area, with groups developing FPGA-based and ASIC-based decoders (Google's Sparse Blossom decoder, published 2023; Amazon's FPGA decoder) specifically for this purpose.

The most cited early pharmaceutical target is not a full drug molecule simulation but a narrower proof-of-concept: accurate simulation of the FeMoco nitrogenase active site, relevant to developing nitrogen-fixation catalysts for agriculture. At ~54 atoms and 113 electrons, it is smaller than a drug molecule but far beyond classical exact methods. A fault-tolerant quantum simulation of FeMoco is estimated to require approximately 4,000 logical qubits and 10⁹ T-gates — representing a factor of ~1,000 reduction in physical qubit requirements compared to the largest pharmaceutical targets. If hardware scaling continues as projected, FeMoco simulation may be feasible in the early 2030s, serving as both a scientific milestone and an engineering validation of fault-tolerant pharmaceutical quantum computing. See also our post on protein folding with quantum computing for the biological simulation problem that shares many of the same algorithmic requirements, and our overview of the quantum drug discovery pipeline for how fault-tolerant simulation fits into the broader pharmaceutical development workflow.

Competing Architectures: Trapped Ions, Neutral Atoms, and the Near-Term Landscape

Superconducting qubits and photonics are not the only competing platforms in the fault-tolerant race. Trapped ion systems — using individual ytterbium or barium ions confined by electromagnetic traps — offer native all-to-all connectivity and some of the highest demonstrated gate fidelities of any platform. IonQ's Forte processor (2023) achieved two-qubit gate error rates of approximately 0.3–0.5% with 32 algorithmic qubits. Quantinuum's H2 processor reported two-qubit gate fidelities exceeding 99.9% (0.1% error rate) for small circuits in 2024 — the highest demonstrated in any platform at that date. The challenge for trapped ions is clock speed (two-qubit gates take ~100 microseconds, versus ~100 nanoseconds for superconducting gates) and scaling to thousands of ions in a single trap, which requires either photonic interconnects between ion trap modules or novel trap architectures. Quantinuum's QCCD (quantum charge-coupled device) architecture shuttles ions between trap zones to implement gates, but this adds latency.

Neutral atom arrays — using optical tweezer arrays to hold individual rubidium or caesium atoms — have emerged as a surprise contender. Harvard's group (Bluvstein et al., 2023, Nature) demonstrated 48 logical qubits and executed 228 logical gate operations with transversal gate implementation, representing the largest demonstration of fault-tolerant logical qubit operation at the time. Atom Computing and QuEra (a Harvard spin-out) are commercialising neutral atom platforms. The appeal is rapid array reconfiguration — atoms can be moved between positions using tweezer arrays, enabling dynamic connectivity — but two-qubit gate fidelities (using Rydberg interactions) are currently 0.5–1%, right at the surface code threshold rather than comfortably below it. For a comparative analysis of quantum computing architectures and their relative merits for pharmaceutical applications, see our post on quantum annealing vs gate-based quantum computing.