Unitree says cumulative production of its bipedal humanoid robots has reached approximately 18, 000. Large enough to take the factory claim seriously. Not a reason to hand one your house keys.
Three announcements keep getting sold with the wrong verb. Unitree says the robots have been built. Microsoft says a diagnostic setup scored a stack of hard cases. Locus Biosciences says a federal agency will fund an early trial of an AI-designed phage cocktail. Built, scored, and will fund are different products. If you are buying AI for business, the rule is the same for all three: ask what the figure excludes, who was compared, and whether the verb has already happened. Only the version that survives those questions belongs in an operating plan.
A production total is not a housekeeper
The 18, 000 figure comes from a Unitree update dated August 12, 2026, as reported by the trade outlet Gasgoo. It counts bionic bipedal humanoids only. Unitree said the total leaves out other humanoid-like machines and anything on a wheeled chassis. A production number that excludes wheeled and non-bipedal platforms is an inventory claim. It does not show a category ready for a home.
Hold the other figures apart, or someone will add them up. Approximately 18, 000 is that cumulative bipedal production claim. Gasgoo also cites a market firm it identifies only as SAG, which put Unitree shipments at 5, 900 units in the first half of 2026. One secondary estimate. The account does not give the firm’s full name or its definition of a humanoid, so it is not a census. Useful only as a warning: a half-year shipment count is not a lifetime build total.
The third number is a plan. Gasgoo reported that Unitree is pursuing an IPO on Shanghai’s STAR Market, the exchange’s board for science and technology companies, and that it plans to raise 6.1 billion yuan for model research, robot bodies, new products, and a factory. The planned plant would take two years to build. Designed annual capacity is 190, 000 units: 75, 000 humanoids and 115, 000 quadrupeds. Designed capacity is not this year’s output. Do not add it to the 18, 000 already claimed.
The cited customers are the tell. Gasgoo’s account points to research, tourism, and industrial pilots, not a confirmed home deployment. The company’s first humanoid, the H1, dates to 2023. Tourism demos and factory pilots can be real revenue. They are not a service you can leave alone with the dog.
If you are pricing AI automation, separate a machine you can count from a task you can trust. I would not let one of these into my home. That is a judgment, not a finding that the hardware is unsafe. The machines are not fake. The public record answers a manufacturing question and stays quiet on failure rate, cost per useful hour, and how much of any demo is still a person in a headset. Until those show up in writing, humanoids are an industrial bet wearing a consumer costume. Count the supply. Do not pretend you have bought a service.
A hard-case benchmark is not a clinic
The second verb is scored. Microsoft’s write-up, The Path to Medical Superintelligence, is a company account, not a peer-reviewed paper, and not an independent replication. The materials are dated September 2025. Read it as a vendor’s best case, then look at what was actually tested.
MAI-DxO is what Microsoft calls a model-agnostic orchestrator: a coordinator that can sit on different underlying models and manage a sequence of questions and tests, rather than one model issuing a diagnosis on its own. Paired with OpenAI’s o3, Microsoft says it reached the correct diagnosis in 85.5 percent of 304 recent New England Journal of Medicine cases in the company’s Sequential Diagnosis Benchmark. Twenty-one practicing physicians in the United States and the United Kingdom, with 5 to 20 years of experience, posted a 20 percent mean “across completed cases.” Microsoft’s summary of that gap is “more than four times higher.”
The multiple is less useful than the fine print. “Across completed cases” is unexplained. Nothing in the write-up shows that the physicians and the system had the same information, the same time, or the same set of finished cases. A 20 percent mean on this exercise is not a verdict on ordinary doctors. “More than four times” is not a staffing ratio.
The cases are NEJM Case Records, which Microsoft itself calls among the most diagnostically complex in medicine. Not Monday morning in primary care. A high score on the hardest published puzzles can still be the wrong tool for a cough, a refill, and a worried parent.
The test design is more interesting than the multiple. Microsoft argues that USMLE-style benchmarks, the multiple-choice exams used in physician licensing, are near saturated and flatter real skill, because they are mostly one-shot. Its sequential benchmark lets the system ask questions and order tests, then scores the result against the published diagnosis. Closer to an AI agent that can ask, test, and revise than to a chatbot guessing a letter.
Microsoft also says the orchestrated setup reached a diagnosis at lower virtual testing cost than the physicians or any single model it tried. Virtual testing cost means a modeled price of the tests ordered, not an invoice a hospital paid. No dollar figure came with the claim. Cheaper, without a price, is a direction. It is not a budget line. The company also says the orchestrator improved every foundation model it tried. That is a claim about a method. It is not a result you can drop into a budget.
Microsoft has a commercial reason to care that has nothing to do with emptying a clinic. It says Bing and Copilot already see over 50 million health-related sessions a day, and that a dedicated consumer-health effort began at the end of 2024. Distribution of that size will tempt health operators and software buyers to treat a benchmark as a staffing plan. It is not one.
Use it as a second reader on difficult cases, with a physician still accountable for the chart, and a vendor who can show the case mix, the incomplete cases, and the cost in money. “Medical superintelligence” is Microsoft’s frame, not a finding. Leave the superintelligence framing on the slide. Buy the workflow you can audit.
The phage headline ran ahead of the trial
The third verb is will fund, and the scare line got there first. “AI-designed virus” has been doing laps. The documented story is narrower. It is not finished.
On January 7, 2026, Locus Biosciences in Morrisville, North Carolina, said the National Institute of Allergy and Infectious Diseases, part of the National Institutes of Health, had awarded it $3.3 million. Up to $28 million is available if all milestones are met. The company’s release identifies the award as contract 75N93025C00035. The money is to support a proof-of-concept Phase 1b trial of LBP-PA01, a bacteriophage cocktail Locus says it designed with an AI-driven discovery engine. A Phase 1b is an early human study, meant to test whether the idea can be studied safely, not a verdict on widespread use. The target is hospital-acquired pneumonia and ventilator-associated pneumonia caused by antibiotic-resistant Pseudomonas aeruginosa.
A bacteriophage is a virus. It infects bacteria, not people. That difference in host is the whole point, and it is why phage therapy has been studied for years as one possible answer to drug-resistant infections. A headline that stops at “virus” sells the fright and skips the mechanism.
Locus describes the engine only at a high level: lab automation and machine learning that measure how phages interact with bacteria and help propose candidate mixes. A company design claim. Not a protocol, a dose, an enrollment figure, or a patient outcome. The release is written in the future tense. The award will fund and will support a trial. It does not report that a person has already been injected, and it does not report that the therapy worked.
The medical stakes, as Locus lays them out, do not need the science fiction. The company says hospital-acquired pneumonia is about 15 to 20 percent of hospital-onset infections, and that P. aeruginosa accounts for up to 24 percent of ICU respiratory infections in some studies. Locus also points to the CDC’s listing of antibiotic-resistant P. aeruginosa as a serious threat. Hospitals already pay for that in longer stays and failed antibiotics.
An AI-designed phage that eventually cleared early trials would matter to those budgets. A press release about funding the first study does not. AI automation in discovery has reached the point where a federal contract will pay for an early human study of a named candidate. That is past the slide deck. Not efficacy. Not a product. Not a reason to brief a board as if the injection already happened and the bacteria already lost.
Questions worth asking before the next budget meeting
- What should you ask a vendor on Monday morning?
Ask what the figure excludes, who was compared, and whether the verb is built, scored, or will fund. Only the version that survives those questions belongs in the operating plan.
- Would you let a humanoid like Unitree’s into your home?
Not on the evidence that would decide it. The cited work is research, tourism, and industrial pilots, and the public figures do not establish reliability, cost, or unsupervised household work. I would not, and that is a judgment about missing proof, not evidence the machine is unsafe.
- Does approximately 18, 000 robots mean housekeepers are close?
No. That figure is Unitree’s announced cumulative bipedal production, with wheeled and other non-bipedal machines excluded. It is a different metric from a half-year shipment estimate or a factory’s designed future capacity.
- Is Microsoft’s medical AI twice as good as doctors?
No. On this company benchmark, with methods not shown to match, Microsoft says MAI-DxO with OpenAI’s o3 reached the correct diagnosis in 85.5 percent of 304 hard NEJM cases, against a 20 percent mean for 21 experienced physicians across completed cases. Microsoft calls that gap more than four times higher, not twice, and it is not a measure of ordinary care.
- Has someone already been injected with an AI-designed virus to fight resistant bacteria?
The January 2026 Locus release does not say that. It announces NIH funding, $3.3 million now and up to $28 million if milestones are met, to support a future Phase 1b trial of the phage cocktail LBP-PA01.
Built, scored, and will fund are three different kinds of progress. A number earns a line in the operating plan only after you know what it leaves out.