The AI Race May Not Be About AI at All

On 15 January 2026, OpenAI announced that it was joining the seed round of Merge Labs, a research lab whose stated mission is “bridging biological and artificial intelligence to maximize human ability, agency, and experience” (OpenAI, 2026). The post opened with a claim about computing history—“Progress in interfaces enables progress in computing”—and then placed brain-computer interfaces inside that lineage. Merge Labs, OpenAI wrote, is pursuing interfaces that “interface with the brain at much higher bandwidth by combining biology, devices, and AI.” The more specific proposal comes from Merge Labs’ own announcement: molecular interfaces and ultrasound are research directions intended to avoid implants into brain tissue. Those details describe the lab’s ambitions, not a demonstrated device.

Read alone, that is one company’s option on a speculative technology. Read alongside other documents from the months on either side of it, it becomes a signal about the direction of competition among the largest AI organisations. One is government policy. One is an international ethics instrument. One is a financing round for a company that designs medicines rather than chatbots. Together they support a thesis worth taking seriously: the model race that has dominated attention since 2022 may be the opening stage of a longer contest to connect artificial intelligence to living systems—to neurons, to cells, to the machinery of biology.

This series keeps one discipline throughout. It separates what has been demonstrated, what is plausible engineering, and what remains scenario. That discipline matters here because the language of the field flatters everyone involved. In this article, the demonstrated facts are modest and checkable: capital moved, policies were published, an international ethics standard was adopted. The plausible engineering is a set of interfaces that could be built, with timelines that nobody can honestly promise. The scenario is a world in which direct neural links to computation are an ordinary consumer product. That world has not arrived.

The interface OpenAI paid for

The details of the Merge Labs investment are revealing because of what they avoid. The company is not competing on the size or benchmark scores of a language model. Its premise is that the bottleneck in human–AI interaction is the interface itself—the keyboard, the screen, the slow conversion of intent into text. Merge Labs was founded by researchers Mikhail Shapiro, Tyson Aflalo, and Sumner Norman, with entrepreneurs Alex Blania and Sandro Herbig, and with Sam Altman participating in a personal capacity. Its announced approach combines neuroscience, bioengineering, and device work, and OpenAI said it would collaborate on “scientific foundation models and other frontier tools.”

Reporting on the round placed it at roughly $250 million at an $850 million valuation, with OpenAI writing the largest single check (Deutscher, 2026). Those figures come from journalism rather than from OpenAI’s own announcement, and the valuation should be treated as reported rather than confirmed. What is not in dispute is the strategic direction. A leading AI lab invested in a possible channel into the nervous system while continuing its model business.

The technical caveat belongs in the same paragraph as the enthusiasm. No non-invasive brain-computer interface today reads or writes neural activity at the bandwidth the company describes. Merging molecules, ultrasound, and machine learning into a safe, high-bandwidth, non-surgical interface is a research program, not a product line. The investment buys a position in a race whose outcome is unknown.

A seven-ministry plan, not a delivered capability

If the Merge Labs check shows where private capital is pointing, China’s policy machinery shows how a state can commit to the same direction at industrial scale. On 23 July 2025, seven ministries and agencies—industry and information technology, the national development commission, education, health, state-owned assets, the academy of sciences, and the drug regulator—jointly issued an “implementation opinion” on promoting the brain-computer interface industry (工信部联科〔2025〕164号) (七部门, 2025).

The document is explicit about being a plan rather than an achievement. It sets two staging posts. By 2027, key technologies should break through, electrodes, chips, and complete devices should reach internationally advanced performance, and BCI products should find early use in industrial manufacturing, healthcare, and consumer goods, anchored by two or three industrial cluster zones. By 2030, it foresees two or three globally influential leading enterprises and an industrial ecosystem whose “comprehensive strength” enters the world’s front rank.

The body of the document carries the specifics. It calls for implantable electrodes on and beneath the dura and inside the cortex, for non-implantable electrodes and helmet, headset, eyeglass, and earbud form factors, for high-channel-count acquisition and ultra-low-power processing chips, and for decoding software that uses AI to raise accuracy and responsiveness. It orders improvements to established devices such as deep-brain stimulators and cochlear implants, and it directs effort toward brain-intent recognition for device control. One clause addresses governance directly, calling for a data-governance framework to regulate the collection, storage, and use of user information and to “prevent brain-privacy leaks.”

This is a state deciding that the boundary between nervous systems and machines is an industrial sector to be built. The plan’s own framing—“enabling the collaborative interaction of biological intelligence and machine intelligence”—is the same thesis Merge Labs advances in the language of venture capital. The instrument differs. Where a startup raises a seed round, a government directs seven agencies and a schedule measured in years. And the plan, for all its ambition, is a statement of intent: the products it describes are targets, not reports.

The ethics arrived before the products

The third document suggests the convergence is taken seriously beyond the industries that stand to profit from it. On 11 November 2025, UNESCO’s General Conference adopted the Recommendation on the Ethics of Neurotechnology at its 43rd session in Samarkand (UNESCO, 2025). It is the first global standard-setting instrument covering the field.

The Recommendation draws its boundary unusually widely. It covers neural data, and it also covers data that can be used to infer mental states—so that much of the consumer technology that touches cognition without recording a brain signal still falls within its scope. Its preamble names the interests at stake: autonomy, privacy, mental and physical integrity, personal identity, freedom of thought, and the risk of discrimination. It distinguishes sharply between medical and non-medical use, and between treatment and enhancement. The instrument calls for consent and transparency, inclusive and affordable access, and particular caution where children, workplace monitoring, or products that influence behaviour are concerned. It warns against using neurotechnology to enhance productivity at the expense of a worker’s mental health, and against designs that exploit vulnerability to produce compulsive use. Drafting was steered by an expert group co-chaired by the bioethicist Hervé Chneiweiss and the legal scholar Nita Farahany.

Two facts about the Recommendation are easy to miss. First, UNESCO had reported a 700 percent rise in investment in neurotechnology companies between 2014 and 2021—the commercial acceleration preceded the governance. Second, a Recommendation is not a law. It formulates principles and invites member states to act. It shapes national legislation; it does not enforce it. The document is evidence that the convergence is real enough to require rulemaking. It is not evidence that the rules will hold.

The biology the AI money is entering

The fourth document shifts the subject from the brain to the cell, and it is where the phrase “the AI race” starts to strain. On 12 May 2026, Isomorphic Labs, the drug-design company spun out of Google DeepMind, announced a $2.1 billion Series B round led by Thrive Capital, with participation from Alphabet, GV, and new investors including MGX, Temasek, CapitalG, and the UK Sovereign AI Fund (Isomorphic Labs, 2026). The company said the capital would scale its “AI drug design engine,” known as IsoDDE, and push a pipeline of therapeutic programs toward the clinic.

The company’s own claims are company claims. Isomorphic reports that its engine has hit internal milestones and identified viable candidates with unusual speed; those results are proprietary and have not been independently replicated in the way a peer-reviewed paper would allow. What is independently checkable is the direction of the money and the shape of the ambition. A company founded in 2021 to commercialise protein-structure prediction has become one of the best-capitalised private efforts to design medicines from first principles, with partnerships spanning Novartis, Eli Lilly, and Johnson & Johnson. The funding announcement is not a report of clinical efficacy. Whether the candidates become safe, useful medicines requires patient trials; financing and internal design milestones cannot answer that question.

Two further moves complete the picture. In September 2026, Novo Nordisk announced a collaboration with Anthropic to apply Claude models, including a science-focused product, to drug-discovery workflows—the latest in a string of partnerships through which a large pharmaceutical company is wiring frontier AI into its research organisation (Novo Nordisk, 2026). Days later, Reuters reported that Anthropic had quietly built a wet lab in the Bay Area, staffed with biologists, to move beyond purely computational work and toward physical experiments, with an ambition that its models could eventually direct laboratory automation (Dastin and Erman, 2026). An AI company now owns more of the loop: reading sequence data, forming hypotheses, and running the bench.

The case against the thesis

A serious counterargument deserves a hearing: that the race is still about models, and biology is simply a lucrative customer.

On this reading, OpenAI’s Merge Labs check is a hedge and a talent play, while the company’s core business remains general-purpose models and biology is one application domain among many. China’s BCI plan is industrial policy of a familiar kind—a state picking a promising sector and subsidising it—and plenty of such plans have produced documents rather than industries. Isomorphic’s capital is drug-discovery capital, which has flowed into computational biology for decades without producing a wave of approved AI-designed medicines. And brain-computer interfaces have been “five years away” since the 1970s; the research literature contains far more demonstrations in patients with paralysis than products on shelves.

This objection identifies the weakest link in the convergence thesis: the leap from investment to capability. The investment is real; the capability is not yet demonstrated at the scale the story implies. Anyone who wants to believe the strongest version of the thesis has to explain why this decade differs from the last several—why foundation models change the feasibility of interfaces that earlier generations of signal processing could not deliver. That explanation exists and is plausible: large models are unusually good at extracting structure from noisy, high-dimensional signals, which is exactly what neural decoding and biological sequence design demand. It remains an argument about plausibility, not a record of achievement.

The honest position is that both readings fit the evidence so far. The convergence is visible in the allocation of capital and the publication of policy. Whether it becomes a durable reorientation of the field, or a well-funded detour, depends on results that have not yet been reported.

Two ways the same technology can cut

The reason this contest matters more than a rivalry over benchmark scores is that its subject is the boundary of a person. A neural interface is not a neutral pipe. It reads a signal that may reveal intention, emotion, or attention, and it can write a signal that may alter them. The same device that lets someone move a cursor with a thought can let an employer infer whether that thought was focused.

That duality is why the UNESCO Recommendation spends so much of its text on consent, on children, and on the workplace. Consent is a strong safeguard when a person can refuse without cost. It weakens when the alternative is unemployment, a worse grade, or exclusion from a service. It weakens further when the intervention itself changes the preferences that consent would express—a person who has been made more attentive, or more content, may not be the same person who agreed to the change. The instrument’s warning against “implicit and explicit coercion” is an attempt to name a problem that law alone cannot solve.

The engineering version of the same worry is objective selection. A system is only as good as the target it is told to maximise. A controller optimised for productivity may learn to suppress the signals of fatigue. One optimised for engagement may find that mild anxiety keeps people returning. One optimised for compliance may find it easier to reduce dissent than to address its causes. Capability does not choose the objective. Someone chooses, and the choice is where the moral weight sits.

What the evidence supports

It is worth restating the ledger, because the two directions of this story—the promise and the risk—are usually argued at different levels of rigour.

On the demonstrated side: private capital is moving into neural interfaces and AI-driven biology; a major state has committed to building a BCI industry on a defined schedule; an international ethics standard has been adopted; and AI companies are hiring biologists, building wet labs, and signing research agreements with pharmaceutical firms.

On the plausible-engineering side: high-bandwidth, non-invasive neural interfaces; foundation models that usefully design biological molecules and genomes; and laboratory loops in which models propose experiments that automated systems run. Each is a defensible research program. None is a finished capability.

On the speculative side: consumer neural interfaces that are safe, cheap, and widely adopted; AI-designed medicines that reliably reach patients; and the wholesale compression of drug development timelines that the largest financing rounds imply. These are scenarios, and their timelines are, at present, guesses.

Keeping those three columns separate is not pedantry. It is the difference between forecasting and wishful thinking, and it is the only way to read a field in which the press release often outruns the paper.

How to read the race

Three things are worth tracking, and they are more specific than the headlines. The first is whether the wet-lab move spreads. A single AI company building a biology lab is an anecdote; several doing it, with published results, is a change in the structure of the field. The second is whether the governance keeps pace. The Chinese plan and the UNESCO Recommendation both exist; the test is whether they shape actual products, or whether they arrive after the market has already set its terms. The third is whether the biology delivers clinical results rather than financing announcements. A drug that reaches patients, designed substantially by a model, would settle more arguments than any valuation.

The pattern to resist is the one that treats investment direction as technical inevitability. Money reveals intent. It does not reveal feasibility, and it says nothing about whether the resulting systems will enlarge human agency or narrow it. The same neural interface that restores a voice to someone who has lost one can be repurposed to monitor a worker’s attention. The same biological model that designs a medicine can design something less welcome. The convergence described here is a genuine shift in where the most capable organisations are pointing their resources. What they build with those resources remains, for now, an open question—and the answer will be decided less by the models than by the objectives they are given.