In October 2025, the New England Journal of Medicine published results from the PRIMAvera trial—an international study across 17 sites in five European countries involving 38 patients with advanced dry age-related macular degeneration, all legally blind in their central visual field. The patients received a photovoltaic retinal implant about half the thickness of a human hair, placed beneath the retina, paired with augmented-reality glasses that project near-infrared light onto the chip. At 12 months, participants improved by an average of five lines on an eye chart. Some could read large print. Some could recognize objects. Some could cook, play cards, navigate rooms they hadn’t been able to see clearly in years.

The lead developer, Daniel Palanker at Stanford’s Byers Eye Institute, described the PRIMA system as “the only way to restore sight in AMD patients” who have already lost their photoreceptors. By June 2025, Science Corporation—the bioelectronics startup founded by Neuralink co-founder Max Hodak, which acquired Pixium Vision’s assets in April 2024—submitted a CE mark application for European approval. The restored vision is in shades of gray, not color. It’s crude by any standard of normal sight. And for people who couldn’t see the faces of their grandchildren or read a word on a page, it’s transformative.

Age-related macular degeneration affects roughly 200 million people globally. It’s one of the most common causes of blindness in people over 50, and until PRIMA, there was no treatment that could restore vision once the photoreceptors were gone. Drugs can slow progression. Nothing reversed the damage. The PRIMA trial is the first clinical evidence that an electronic implant can meaningfully restore central vision in this population—and it published in the NEJM, which is about as close to a stamp of legitimacy as medical science gets.

How a retinal implant actually works

The retina is essentially a biological sensor array at the back of the eye. Photoreceptor cells—rods and cones—detect light and convert it into electrical signals that travel through the optic nerve to the visual cortex of the brain. In diseases like macular degeneration and retinitis pigmentosa, the photoreceptors die, but the rest of the neural circuitry—the bipolar cells, ganglion cells, and the optic nerve itself—often remains largely intact. A retinal implant replaces the dead photoreceptors with an electronic substitute.

The PRIMA system works through a three-component chain. A camera mounted on augmented-reality glasses captures the visual scene. A pocket-sized processor converts the image into patterns of near-infrared light. The glasses project those infrared patterns onto the photovoltaic chip implanted beneath the retina. The chip’s pixels convert the infrared light into electrical current, which stimulates the surviving bipolar cells, which relay the signal through the remaining visual pathway to the brain. The photovoltaic pixels replace the dead photoreceptors. The brain—remarkably—learns to interpret the artificial signal and merges the prosthetic vision with whatever natural peripheral vision remains, creating what patients describe as a single continuous image.

The elegance of the photovoltaic approach is that the implant is entirely wireless. No battery. No external power source threading through the eye. The infrared light from the glasses simultaneously carries the visual information and powers the chip. Earlier devices, like Second Sight’s Argus II, required wired connections between external hardware and the implant, which created durability and surgical complications that ultimately contributed to the device’s commercial failure.

The Argus II cautionary tale

Any discussion of bionic eyes in 2026 has to reckon with the Argus II, because it’s the field’s most instructive failure. Second Sight Medical Products received FDA approval for the Argus II in 2013—making it the first retinal prosthesis approved in the United States—and implanted the device in roughly 350 patients with retinitis pigmentosa worldwide. The device had 60 electrodes (compared to PRIMA’s much higher pixel density) and restored crude vision: flashes of light, edges, shapes, movement. Not reading. Not face recognition. Basic spatial orientation.

Then Second Sight went bankrupt in 2020 and ceased operations in 2022. Patients were left with implants in their eyes and no company to maintain them, update the software, or replace failing components. The external hardware—glasses and processor—became orphaned technology with no manufacturer support. The IEEE Spectrum description was blunt: the Argus II was “a pure bridge to nowhere.” Lloyd Diamond, Pixium Vision’s outgoing CEO, said it directly when Science Corporation acquired the PRIMA assets: “It’s very important to us to avoid another debacle like Argus II.”

The Argus II failure wasn’t primarily a technology failure. The device worked, within its limitations. It was a business model failure—a medical device company that couldn’t sustain operations long enough to support implanted patients for the lifetime of the implant. This is the shadow that hangs over every bionic eye company in 2026: the device has to work, and the company has to survive. Patients are making a decades-long commitment to hardware inside their body; the manufacturer needs to make the same commitment to them.

Where the field stands across all approaches

Retinal implants are the most clinically advanced category but not the only one. The landscape in 2026 includes three distinct approaches to electronic vision restoration, each targeting different anatomical locations and different patient populations.

Subretinal implants, like PRIMA, sit beneath the retina and stimulate bipolar cells. They’re the furthest along clinically and are best suited for conditions where photoreceptors are damaged but the rest of the visual pathway is intact—primarily AMD and potentially retinitis pigmentosa, though results for RP with PRIMA are expected to be more limited because the retinal damage is more widespread.

Suprachoroidal implants, developed by the Bionics Institute in Australia, sit in the space between the retina and the outer wall of the eye. Their second-generation device demonstrated substantial improvements in functional vision, navigation, and quality of life over 2.7 years in a phase I/II trial, with 97 percent of electrodes remaining functional and no serious adverse events. The device has received FDA breakthrough device designation, and larger multi-center trials are planned. The suprachoroidal position is surgically less invasive than subretinal placement, which could matter for broader adoption.

Cortical visual prostheses bypass the eye entirely. An electrode array implanted directly on the visual cortex at the back of the brain stimulates the neurons that process vision, creating artificial sight even in people with severely damaged or missing eyes. This approach can theoretically help patients whose optic nerves are damaged—a population that no retinal implant can serve. Neuralink’s Blindsight system, which uses 36 flexible threads with roughly 3,000 electrodes sewn into the visual cortex by a surgical robot, has shown success in monkey tests. Neuralink received preliminary regulatory approval for early human studies in the U.S., Canada, the U.K., and the EU, and reported plans to implant the first human volunteers by 2026. The Orion Visual Cortical Prosthesis, originally developed by Second Sight (before its collapse) and now continued by Cortigent, has published five-year data from its early feasibility study.

A fourth approach—the Science Eye, being developed by Max Hodak’s Science Corporation—combines a retinal implant with optogenetic therapy: a genetically engineered virus delivers a gene that makes specific retinal cells light-sensitive at a particular wavelength, and a tiny implanted display with resolution sharper than an iPhone screen provides precise control over those sensitized cells. This hybrid biological-electronic approach is the most ambitious and the least clinically proven.

The resolution problem

The fundamental limitation of every bionic eye in 2026 is resolution. The human retina has roughly 120 million rods and 6 million cones. The PRIMA implant has 378 pixels per square millimeter across a 2-by-2-millimeter chip. The Argus II had 60 electrodes. Even the most optimistic next-generation devices are operating with electrode counts measured in thousands, not millions. The gap between what the implant provides and what a healthy retina delivers is roughly four to five orders of magnitude.

This is why the restored vision is grayscale, crude, and limited to central-field perception. It’s enough to read large print, recognize shapes, navigate rooms, and regain a measure of independence. It is not—and won’t be for the foreseeable future—enough to drive a car, recognize a face across a room, or see the world the way a sighted person does. The next-generation PRIMA device under development uses smaller pixels for higher density, and software improvements including electronic zoom and image stabilization are being tested. But the trajectory is incremental improvement within a fundamentally low-resolution paradigm, not a leap to natural vision.

Daniel Palanker draws the comparison to cochlear implants—devices that restore hearing by stimulating the auditory nerve. Early cochlear implants provided crude sound perception. Today, after decades of refinement, they enable many deaf patients to understand speech and enjoy music. Retinal implants may follow a similar arc: the first generation establishes the principle, each subsequent generation improves the resolution, and the technology becomes standard clinical practice over a timeline measured in decades.

The honest prognosis

PRIMA works. The NEJM data shows it. Patients are seeing things they couldn’t see before the implant, and the benefits are durable over years. The Australian suprachoroidal device works. The cortical approaches show promise. The field in 2026 is further along than it has ever been, with more clinical data, more companies, more approaches, and better-funded organizations than at any point in the history of vision restoration.

The caveats are significant. The PRIMAvera trial was not placebo-controlled—an anonymous retinal-degeneration researcher told Nature that the intensive training and motivation from receiving an exciting new device might have inflated the results. The resolution remains orders of magnitude below natural vision. The commercial viability question—whether any company can sell enough devices at a price patients can afford while sustaining operations for the multi-decade life of the implant—is unanswered. The Argus II proved that a technically successful implant and a commercially sustainable business are not the same thing.

But for 200 million people with macular degeneration, and for the broader population with retinitis pigmentosa, glaucoma, and other causes of irreversible blindness, the PRIMA trial represents something that didn’t exist before October 2025: evidence, published in the world’s most prestigious medical journal, that an electronic implant can restore meaningful central vision in humans. That’s not a cure. It’s not normal sight. It’s the beginning of a technology that may, over decades of refinement, make blindness from photoreceptor loss a treatable condition rather than a permanent one.

We cover retinal implants alongside brain-computer interfaces, neural stimulation for depression, and the full landscape of neuroprosthetic technology across our Neuroprosthetics course—including why the most important medical device of the 2020s is a chip the size of a pencil eraser that lets people read again.

Google’s Willow quantum chip, unveiled in late 2024, completed a benchmark calculation in roughly five minutes that would take a classical supercomputer an estimated 10 to the 25th years to perform. That number is so large it’s functionally meaningless—it exceeds the age of the universe by roughly 15 orders of magnitude. IBM has promised quantum advantage by the end of 2026. Microsoft debuted the world’s first topological qubit processor, Majorana 1, in February 2025. The global quantum computing market hit somewhere between $1.8 billion and $3.5 billion in 2025, depending on which analyst you trust, and is projected to reach $5.3 billion by 2029. Investment is pouring in. Milestones are being announced quarterly. The headlines suggest a revolution.

The practical reality in 2026: quantum computers are not commercially useful at scale. Most real-world applications remain experimental—research, simulations, and controlled pilots, not everyday business operations. Quantum computers are expected to outperform classical computers in specific, commercially meaningful tasks sometime after 2030, not before. The technology is real. The progress is genuine. The gap between what exists and what the headlines imply is enormous.

What a quantum computer actually is (in one paragraph)

A classical computer processes information as bits—ones and zeros. A quantum computer uses qubits, which exploit quantum mechanical properties called superposition and entanglement to exist in multiple states simultaneously and to correlate with each other in ways that classical bits cannot. This allows quantum computers to explore many possible solutions to a problem at once rather than checking them sequentially. For certain classes of problems—molecular simulation, optimization, cryptography, materials science—this parallelism offers exponential speedups over classical approaches. For most computing tasks—running spreadsheets, streaming video, training large language models, browsing the internet—quantum computers offer no advantage whatsoever and are in fact dramatically worse than your laptop.

Where things actually stand

The field in 2026 sits in what’s called the NISQ era—Noisy Intermediate-Scale Quantum computing. Modern quantum processors operate with dozens to a few hundred physical qubits, and those qubits are fragile. They’re sensitive to temperature (most superconducting quantum computers operate near absolute zero, about 15 millikelvins), electromagnetic interference, vibration, and essentially any interaction with their environment. These interactions cause errors—qubits lose their quantum state in a process called decoherence—and current error rates are high enough that computations longer than a few hundred operations become unreliable.

The fundamental challenge is that the qubits we can build are good enough to demonstrate quantum effects but not good enough to solve real problems. A useful quantum computer needs to run circuits with millions or billions of operations. Current machines can reliably execute circuits with roughly 5,000 operations before errors overwhelm the result. IBM’s Nighthawk processor, delivered in late 2025, achieves this 5,000-gate threshold. IBM expects to push this to 7,500 gates by late 2026 and 10,000 by 2027. These are genuine improvements. They’re also roughly five to six orders of magnitude below what’s needed for the applications that justify the investment.

The error correction problem

The path from “interesting but impractical” to “commercially useful” runs through quantum error correction—using multiple physical qubits to encode a single “logical” qubit that’s protected against errors. The math works. The engineering is brutal.

Google’s Willow chip achieved a critical milestone by demonstrating what’s called “below threshold” error correction—as they added more qubits, errors decreased exponentially rather than increasing. This is the first time that scaling up a quantum system made it more reliable rather than less, and it’s the foundational requirement for building large, error-corrected machines. But the milestone came with caveats: the demonstration was limited to quantum memory preservation rather than actual gate operations, logical error rates are still orders of magnitude higher than needed for practical algorithms, and vastly larger qubit arrays will be required for real-world applications.

IBM’s roadmap targets a fault-tolerant quantum computer—Quantum Starling—by 2029, featuring roughly 200 logical qubits encoded across approximately 10,000 physical qubits, capable of executing circuits with 100 million gates. That’s the machine that could actually do something useful. IBM has been hitting its interim roadmap milestones consistently, which matters because roadmap credibility is scarce in quantum computing. Their 2025 Loon processor demonstrated all the key hardware components needed for fault-tolerant operation, and they achieved real-time error decoding in under 480 nanoseconds—a ten-times speedup over previous approaches, completed a year ahead of schedule.

Microsoft took a fundamentally different approach with Majorana 1, pursuing topological qubits—a theoretical construct where quantum information is stored in the topological properties of exotic particles, making it inherently more resistant to errors. If it works at scale, topological qubits could leapfrog the error correction overhead that burdens other approaches. The emphasis on “if” is doing heavy lifting in that sentence.

The qubit zoo

One telling detail about where the field stands: there’s no consensus on what a qubit should even be made of. In classical computing, the transistor won decades ago. In quantum computing, at least five competing technologies are under active development with billions of dollars behind each.

Superconducting qubits (IBM, Google) lead in raw qubit counts and gate speeds but require extreme cooling and are sensitive to noise. Trapped ions (IonQ, Quantinuum) achieve higher fidelity and longer coherence times but are slower. Neutral atoms (Atom Computing, QuEra, Pasqal) offer scalability advantages—you can put 100,000 atoms in a single vacuum chamber—and are the basis for some of the earliest error-corrected machines. Photonic approaches (PsiQuantum, Xanadu) use photons and can operate at room temperature but face different engineering challenges. And Microsoft’s topological qubits remain largely unproven at scale.

An IEEE Spectrum analysis from January 2026 put it directly: we won’t build a powerful, functional quantum machine capable of solving large-scale problems in science and industry in 2026. Scientists have been working toward that goal since at least the 1980s, and it has proved difficult.

What quantum computers can actually do today

Molecular simulation: quantum computers can model the behavior of molecules and chemical reactions with a fidelity that classical computers struggle to match, because molecules are themselves quantum systems. This is the most natural application—using a quantum system to simulate a quantum system—and it’s where the earliest commercial value is likely to emerge. Drug discovery, catalyst design, and materials science are the target verticals. IBM is working with partners including Cleveland Clinic and Boeing on these applications.

Optimization: certain classes of optimization problems—logistics routing, portfolio optimization, scheduling—map well onto quantum architectures. D-Wave’s quantum annealing systems have found niche traction here, though debate continues about whether they offer genuine advantage over classical optimization algorithms.

Cryptography research: quantum computers can theoretically break the public-key encryption that secures most internet traffic, using Shor’s algorithm. No existing quantum computer is remotely close to doing this—it would require millions of error-corrected qubits—but the threat is taken seriously enough that NIST finalized post-quantum cryptography standards in 2024, and “quantum-safe” migration is underway at governments and financial institutions worldwide.

What quantum computers cannot do in 2026: run AI models better than GPUs, replace cloud computing, speed up your database queries, make your phone faster, or accomplish any general-purpose computing task more efficiently than a classical machine. The commercially meaningful applications are narrow, specialized, and mostly still in the pilot or research phase.

The honest timeline

IBM says quantum advantage by end of 2026, fault-tolerant quantum computing by 2029. Google says below-threshold error correction is achieved, with practical applications to follow. Microsoft says topological qubits will change the game. The market says $5.3 billion by 2029, possibly $20 billion by 2030.

The pattern is familiar if you’ve followed fusion, solid-state batteries, or autonomous vehicles: genuine technical progress, consistent milestone achievement, and a commercial timeline that keeps resolving into “a few more years.” Quantum computing is not vaporware. The physics works. The engineering is advancing. The gap between where we are and where we need to be is measured in orders of magnitude, and orders of magnitude don’t close on schedule.

The most honest framing: quantum computing in 2026 is where classical computing was in the early 1950s—room-sized machines operated by specialists, solving problems of academic interest, with a transformative future that’s visible in theory and invisible in daily life. The difference is that the 1950s computer scientists didn’t have venture capital, quarterly earnings calls, or a global media ecosystem incentivized to describe every milestone as a breakthrough.

We cover quantum computing alongside fusion energy, solid-state batteries, and 21 other civilization-scale technology challenges across our Moonshot 2169 course—including why the most overpromised technology of the 2020s is also, quietly, the one making the most consistent progress.