Frontier Watch · checked October 4, 2026
What’s real at the edge.
Quantum, photonic and robotic computing — the milestones that happened, why they matter, and the catch nobody puts in the headline.
Latest: Infleqtion — 30 entangled logical qubits on a commercial neutral-atom machine (September 24, 2026).
Quantum computing
When does a quantum computer do something useful that a normal one can’t?
State of play
2026 is the year of the logical qubit: several teams now show encoded qubits beating raw hardware — but mostly with error detection and discarded runs, not full error correction. Fault-tolerant machines remain roadmap items for 2029 and later.
InfleqtionResult
30 entangled logical qubits on a commercial neutral-atom machine
Thirty logical qubits encoded in 80 atoms on the Sqale system, about a thousand physical operations in total.
Why it matters — More logical qubits on hardware a customer can buy, rather than a lab one-off.
The catch: A distance-2 code: it detects errors but cannot correct them, and runs with a detected error were thrown away. No logical error rate published yet.
IQM · CSC (Finland)Roadmap
Blueprint for Europe’s first publicly owned fault-tolerant system
LUMI-IQ: 150 physical qubits in 2027, growing to 9 fault-tolerant logical qubits by 2029 using surface and colour codes.
Why it matters — A concrete, publicly funded timeline with the error-correction scheme spelled out — useful for judging everyone else’s dates.
The catch: A plan, not a result.
NVIDIAShipping
Ising: open AI models for calibrating and decoding quantum hardware
Open-source models that cut processor calibration from days to hours and decode errors up to 2.5× faster than the standard open decoder.
Why it matters — Error correction needs fast classical decoding; this is where today’s AI hardware meets tomorrow’s quantum hardware.
The catch: Speed and accuracy figures are NVIDIA’s own.
QuantinuumResult
Logical qubits beat the physical ones — at scale
On the Helios trapped-ion machine: up to 94 logical qubits with error-detecting codes and 48 with error-correcting codes, at roughly one logical error per ten thousand operations.
Why it matters — “Beyond break-even” is the line where encoding starts helping instead of hurting. Crossing it with dozens of qubits is new.
The catch: Partially fault-tolerant, and relies on post-selection (discarding flagged runs).
IBMShipping
Nighthawk: a 120-qubit square-lattice processor
120 qubits with more connections between neighbours than IBM’s earlier heavy-hex chips.
Why it matters — Better connectivity means deeper circuits before noise wins — a step on IBM’s road to error correction.
AWSResult
Ocelot: a cat-qubit chip built for cheaper error correction
A prototype chip using “cat qubits” that suppress one kind of error by design.
Why it matters — AWS estimates it could cut the cost of error correction by up to 90%.
The catch: Prototype; the 90% figure is AWS’s estimate.
MicrosoftResult
Majorana 1: the first chip built on topological qubits
A processor designed around a new kind of qubit that Microsoft says could scale to a million on one chip.
Why it matters — If topological qubits work as claimed, they would need far less error correction than any other approach.
The catch: Independent physicists have questioned whether the device shows topological behaviour. Treat as unproven.
Photonic computing
Can light replace electrons where it matters — moving data, and eventually computing?
State of play
Light is winning the wiring first: co-packaged optics went into mass production in 2026 inside AI data centres. Computing with light is still in the lab — real progress on switching energy and chip-to-space interfaces, but no general photonic processor yet.
NVIDIA · BroadcomShipping
Co-packaged optics enters mass production
NVIDIA began shipping Spectrum-X switches with optics built into the package; Broadcom’s 51.2 Tb/s Bailly switch is in volume manufacturing.
Why it matters — Putting optics next to the chip cuts network power by up to 70% versus plug-in transceivers — the first large-scale win for silicon photonics in computing.
The catch: Power figures are vendor-reported; yields and packaging capacity are the bottlenecks.
Monash University and partnersResult
A room-temperature chip that computes with light’s “valley” property
An atoms-thin chip that generates, routes and reads light-encoded information, processing two images at once (Nature Photonics).
Why it matters — A new degree of freedom to carry information, at room temperature — no cryogenics.
The catch: Lab demonstration.
University of PennsylvaniaResult
All-optical switching at about 4 femtojoules
Hybrid light–matter particles (exciton-polaritons) switched light with light, using about four quadrillionths of a joule (Physical Review Letters).
Why it matters — Switching without converting back to electronics is the missing piece for optical AI hardware; the energy is what makes it interesting.
The catch: Lab demonstration.
QuiX QuantumResult
Errors fall as a photonic quantum system grows
Photon distillation filtered out bad photons so that error rates dropped as the system scaled — a first for light-based quantum computers.
Why it matters — Photonic quantum machines run at room temperature; their weakness has been noise. This addresses it.
The catch: Reported in an arXiv preprint (uploaded January 2026); not yet peer reviewed at time of coverage.
MITResult
“Ski-jump” chips that beam light off the chip into open space
Curled micro-structures that project thousands of individually controlled laser beams — dense enough for full-colour images half the size of a grain of salt (Nature).
Why it matters — Getting light off a chip efficiently matters for AR displays, optical links and controlling millions of qubits.
Humanoid robotics
Do humanoids leave the demo stage and do paid work?
State of play
Production is real now — thousands of units a year from several makers, and the first home robot sold to consumers. What’s still unproven is how much useful, unsupervised work each robot does per day.
AGIBOTResult
15,000th humanoid off the line
Production went from 5,000 to 10,000 units in three months; the company shipped 5,168 robots in 2025.
Why it matters — The clearest sign yet that humanoids are a manufacturing business, not a research project.
The catch: Company-reported figures.
1XShipping
A factory for home humanoids opens in California
The Hayward plant builds the NEO home robot at 10,000 units a year, aiming for 100,000+ by the end of 2027. Early access is US$20,000 or US$499 a month.
Why it matters — The first humanoid sold to households, at a price people actually paid — the 2025 run sold out in five days.
The catch: Capacity targets are the company’s own.
Figure AIResult
350+ Figure 03 robots delivered, one built every hour
Figure reached a production rate of about one humanoid per hour at its BotQ facility.
Why it matters — A hard rate number from a US maker, months after the robot was introduced.
The catch: Figures as reported; sourced via an encyclopedia summary rather than a primary release.
Boston Dynamics · HyundaiShipping
Production Atlas, with the whole 2026 run already committed
The all-electric Atlas (56 degrees of freedom, 50 kg lift) went into production, with fleets going to Hyundai and Google DeepMind.
Why it matters — A carmaker that plans to deploy tens of thousands of robots is now a customer of its own humanoid.