Verily Trains All of Us Multimodal Model Ten Times Faster on NeMo

What changed

Verily and NVIDIA described a precision health AI deal on 4 November 2025. The work includes a model trained on NIH All of Us data. The model joins health records with genomics via polygenic risk scores. Verily posted the summary on its site and in a GitHub thread tied to NVIDIA NeMo Automodel.

Verily says researchers used NVIDIA NeMo Automodel on H100 GPUs. They reported roughly ten times faster training versus prior work on the same goal. The claim is about lab speed, not about auto clinical approval.

Why it matters

Most clinical AI tests still lean on health record fields alone. Diagnosis codes and vitals miss much of genetic risk. Adding genomics at large model scale changes what risk tools can test. Teams can explore new ideas without new wet lab work for every trial.

For biotech and health R and D leads, the hold up is often compute and MLOps on regulated data. It is not always model slides. A named stack on Verily’s Pre platform Workbench shows where many researchers may try similar pipelines next.

Hospital groups should note that public news is not a new consent form. Each use case still needs IRB review and data use checks. Privacy leads must map which fields leave the controlled enclave.

Who is affected

Precision medicine teams feel this first. So do hospital innovation groups and pharma stats units that use All of Us or similar biobanks. NVIDIA sellers and cloud planners should expect more genomics jobs on H100 clusters.

IRB and privacy officers must still review each use case. Public news is not patient consent for new models. Legal teams should read data notes in the primary post before they copy the pipeline.

Clinicians need plain words on what polygenic risk adds beyond diagnosis codes. They also need clarity on where human review stays required. Do not ship scores to patients without a clinical governance path.

What to do next

If you join All of Us research, read Verily’s tech post for data handling notes. Do that before you copy the pipeline. Compare NeMo Automodel versions against your stack on a small cohort.

Brief clinical leads on what polygenic risk adds beyond codes. State where human review stays required. Log model version, training date, and cohort filters in your lab notebook.

Start with read only evals on de identified slices if full data access waits on paperwork. That still shows whether the stack fits your GPU budget.

What to watch

Watch whether Verily publishes peer reviewed metrics beyond the November demo blog. Watch NVIDIA NeMo release notes for Automodel changes that touch genomic tensor paths.

Track NIH All of Us policy on use of derived models in commercial tools. Also watch partner posts for repeat studies from outside labs.

Ethics boards may ask how explainability works when genomics and record fields mix. Prepare answers before vendor sales calls reach your C suite.

Procurement should compare GPU rent to long term cluster buys if this pipeline becomes standard. Ten times faster training still costs real money at scale.

Patient advocates may ask how models use diverse cohorts in All of Us. Prepare plain summaries of who is in the train set and who is not. Gaps in the data can skew risk scores for some groups.

IT teams should verify that NeMo Automodel versions match the discussion thread before they open firewall ports to new endpoints. Version drift is a common source of failed repro runs in shared lab VMs.

Clinical trial designers should ask whether genomics features change inclusion rules for future studies. A faster train loop does not by itself prove better patient outcomes.

Store consent language that covers genomics use if your org mirrors this pipeline. Regulators may ask for proof that patients knew how models would use their data.

Archive the GitHub thread URL in your quality system so auditors can find the primary claim later.

Image: GitHub

Photo: National Cancer Institute / Pexels

Sources

• Primary — GitHub, Verily and NVIDIA NeMo Automodel discussion (4 November 2025). Describes the All of Us multimodal model and training speed claim.
• Secondary — Verily, Precision health AI collaboration perspectives post (4 November 2025). Public announcement of the collaboration scope.