🧠 Neuro
14 Humans Have Brain Chips. Samsung Is Manufacturing the First That Writes Back.
Neuralink's 4th-generation chip, codenamed O1, switched from TSMC to Samsung's 4nm process. It is the first bidirectional brain-computer interface chip to enter foundry production. TSMC was too busy making GPUs.
June 16, 2026
At least fourteen people on Earth have Neuralink brain implants. Seven received the N1 chip through the PRIME study at Barrow Neurological Institute. Seven more enrolled in GB-PRIME across University College London Hospitals and Newcastle Hospitals, with a Canadian arm, CAN-PRIME, now enrolling six additional participants in Toronto. Every one of them has a read-only device: 1,024 electrodes threaded into motor cortex by a surgical robot, recording what the brain does but unable to tell the brain what to do. That is about to change.
According to reporting based on a Korea Economic Daily investigation, Samsung Foundry has begun manufacturing test chips for Neuralink's fourth-generation implant, internally codenamed O1, using Samsung's 4-nanometer process. Test chip production started last month, and if validation succeeds, first-batch shipment is projected for early 2027 with mass production following in late 2027. Neither Samsung nor Neuralink has confirmed the contract.
TSMC manufactured the third-generation chip and is no longer making the fourth. As WCCFTech reported, TSMC has been replaced by its Korean rival for a chip that will sit inside a human skull, bumped from the world's most advanced foundry because NVIDIA, AMD, Apple, and Qualcomm consume every available wafer-start on TSMC's cutting-edge nodes.
Why Samsung? Because TSMC Is Full.
Start with capacity math. Microsoft, Amazon, Alphabet, Meta, and Oracle are collectively spending over $300 billion on AI capital expenditure in 2026, and most of that bottlenecks through TSMC's advanced nodes. A single NVIDIA H100 die consumes roughly 814 square millimeters of silicon; on a standard 300mm wafer, that yields approximately 87 gross dies (fewer after edge losses and defect rejection), and at $25,000 per unit, one wafer of H100s is worth north of $2 million in end-product revenue.
Neuralink occupies a different position. Its N1 custom ASIC is estimated at 25 to 50 square millimeters, meaning one wafer could produce over 1,400 brain chips, but Neuralink's installed base is fewer than two dozen humans, and even projecting aggressively to 1,000 patients per year, entire annual demand fits on a single wafer run, a production volume so small it would not register as a line item in TSMC's quarterly report.
When every spare nanometer of foundry capacity is being fought over by hyperscalers building out the infrastructure to train and serve large language models, a company making chips for paralyzed patients gets deprioritized. Not because those chips are less important to the humans who need them, but because the wafers generate less revenue for the foundry.
Samsung's 4nm: Right Node for a Skull
Samsung's 4nm process recently crossed the 80% yield threshold, a milestone reported by industry trackers as marking "mature process" status, and maturity is exactly what brain implants demand. Semiconductor manufacturing's relentless push toward smaller nodes optimizes for transistor density and performance per watt, metrics that matter when you are training a frontier language model or rendering frames at 240 fps but that become secondary when the chip is sealed inside a titanium case, implanted through a craniotomy, and expected to function without service for years.
For that application, you want high yield so fewer defective chips means fewer defective implants, a well-characterized thermal envelope because the brain tolerates only fractions of a degree Celsius above baseline, and thousands of cumulative production hours establishing reliability. Samsung's 4nm line, serving Groq, IBM, and Ambarella across thousands of wafer runs, has that history. Samsung's 2nm gate-all-around process, the node you would choose if chasing transistor density, does not. When field failure means a second brain surgery rather than a replacement graphics card, process maturity outranks process shrink.
Gen 4: Bidirectional
Every Neuralink chip implanted in a human so far is unidirectional: it reads electrical signals from 1,024 electrodes, compresses roughly 200 megabits per second of raw data to 1-2 Mbps, and transmits wirelessly via Bluetooth, allowing patients to move cursors, type text, and play video games using neural activity alone. Noland Arbaugh, Neuralink's first patient, has controlled his computer with his thoughts for over two years.
O1 adds the other direction, and nobody has one yet. According to reporting from the Korea Economic Daily, the fourth generation can input data to the brain to activate physical functions, with the cited example being vision restoration through direct stimulation of visual cortex neurons, a capability that maps to Neuralink's Blindsight project, which received FDA Breakthrough Device Designation in September 2024 for a visual prosthesis bypassing the optic nerve entirely. Musk described early Blindsight resolution as "like Atari graphics."
Regulatory implications are substantial: a read-only BCI is a sophisticated sensor, but a write-capable BCI is a neurostimulation device carrying risks that recording does not, including seizure induction, tissue heating, and electrode degradation over thousands of stimulation cycles. No Neuralink chip with write capability has ever been implanted in a human, yet Samsung is manufacturing the hardware on an industrial foundry line before any clinical data exists to support bidirectional stimulation in a living brain.
A Consolidating Pattern
Samsung's Neuralink contract is not isolated. Tesla signed a $16.5 billion deal with Samsung for its AI6 and AI6.5 self-driving chips on Samsung's 2nm process. Samsung's foundry division, which accumulated operating losses exceeding 2 trillion won (about $1.36 billion) over prior years, narrowed its deficit to roughly 1 trillion won by late 2025 and may reach profitability by Q3 2026, driven largely by 4nm utilization from AI and memory clients.
If Tesla alone represents $16.5 billion in commitments against Samsung's annual foundry revenue of roughly $14-16 billion, Musk-affiliated entities could account for a double-digit percentage of Samsung's manufacturing business within two years. That creates leverage and a risk profile the semiconductor industry has rarely seen concentrated in one person's portfolio spanning electric vehicles, autonomous driving, humanoid robots, and brain-computer interfaces.
Limitations
This report relies on unnamed industry sources in the Korea Economic Daily. Neither Samsung nor Neuralink has confirmed the O1 contract. N1's precise die size is not public; our wafer-economics calculations estimate die area from the implant's external dimensions and comparable ASICs. Bidirectional characterization comes from secondary reporting, not Neuralink engineering documentation. Patient counts may exceed fourteen; one BCI research tracker reports 21+ patients enrolled across all PRIME variants as of May 2026. Samsung foundry profitability timelines conflict: DIGITIMES projects Q3 2026 while WCCFTech cites Samsung's own guidance of 2028.
Against This Story
Samsung may not have been chosen for reliability at all, and the strongest case against this narrative is rooted in simple supply dynamics: TSMC's 4nm lines are fully booked by NVIDIA, AMD, Apple, and Qualcomm, while Samsung's lines have spare capacity precisely because Samsung lost customers in prior years due to inferior yields that only recently crossed 80%. Neuralink may have switched because Samsung was the only major foundry with open slots at a competitive node, and a "mature node is ideal for implants" narrative could be post-hoc rationalization of a supply-constrained decision. If TSMC had capacity, Neuralink might never have left.
Bottom Line
AI infrastructure demand has consumed so much semiconductor capacity that brain implant chips are being rerouted to a different foundry, not as a market failure but as a market signal revealing how completely GPU volume economics dwarf medical neurotechnology. Fewer than two dozen humans have brain chips while hundreds of millions of GPUs have shipped, and silicon flows to wherever revenue concentrates.
Samsung's O1 contract is small in dollars but significant in what it represents: the first bidirectional brain chip, manufactured on a deliberately mature process node, by a foundry rebuilding on overflow from the AI boom, for a company whose entire patient base fits in one room. If O1 works and Blindsight delivers even "Atari graphics" vision to someone who has none, that room will matter more than every data center on Earth. But the semiconductor industry does not price things that way, and until BCI volumes register on foundry order books, the chips that restore sight will be manufactured on whatever capacity the chips that train language models leave behind.
What You Can Do
If you work in semiconductor manufacturing, recognize that BCI's foundry needs are tiny in volume but extreme in reliability. Dedicated medical-grade semiconductor capacity will eventually replace overflow allocation, and early partnerships with BCI companies could lock in long-term relationships before the market scales.
If you follow BCI technology, watch for Neuralink's Blindsight trial enrollment. O1's projected ship date of early 2027 suggests human bidirectional trials could begin in 2027 or 2028, though FDA stimulation pathways mean timelines will be measured in years, not months.
If you are an investor, note that Samsung's foundry recovery is increasingly a Musk-concentration story. Tesla AI6 and Neuralink O1 represent major manufacturing commitments from a single corporate ecosystem: bullish for utilization rates, but introducing customer diversification risk Samsung has not historically carried at this scale.