NVIDIA and researchers built an AI that predicts blood sugar four years out

Ever looked at a cookie and wondered what it would do to your body an hour later? It turns out the same AI architecture behind ChatGPT can now tell you the answer, and even forecast your metabolic trajectory four years into the future.

NVIDIA, the Weizmann Institute of Science, and Israeli startup Pheno.AI unveiled GluFormer, a generative AI model that predicts an individual's future glucose levels and critical health metrics using past continuous glucose monitoring (CGM) records.

Reading blood sugar like a text prompt

Transformers run on sequence prediction: given a sequence of words, guess the next word. NVIDIA and its research partners realized that continuous health telemetry is basically the exact same thing.

CGM patches collect biometric readings every 15 minutes, generating an uninterrupted time series of how a body handles fuel. According to Gal Chechik, senior director of AI research at NVIDIA, continuous glucose monitoring can be treated as sequences of diagnostic tests tracing biological processes over time. The model simply learns the underlying grammar of human metabolism.

GluFormer was trained on 14 days of CGM data from over 10,000 non-diabetic participants, captured every 15 minutes as part of Pheno.AI's Human Phenotype Project. Training and inference ran on an NVIDIA Tensor Core GPU cluster.

Once dietary intake data is plugged into the model, GluFormer can forecast individual glucose spikes to specific foods, paving a direct path toward truly personalized nutrition.

Predicting far beyond glucose

What surprised researchers most was how well the model generalized. Despite training strictly on healthy participants, the team validated GluFormer across 15 separate datasets, successfully predicting health outcomes for people with prediabetes, type 1 and type 2 diabetes, gestational diabetes, and obesity.

Even wilder, the model looks past blood sugar to forecast broader clinical markers. Lead author Guy Lutsker noted that GluFormer successfully estimated visceral adipose tissue (fat wrapping around organs like the liver and pancreas), systolic blood pressure, and the apnea-hypopnea index for sleep apnea, a condition closely tied to type 2 diabetes.

Harvard Health Publishing and NYU Langone Health have both emphasized how CGM devices speed up diagnoses for prediabetes. GluFormer essentially supercharges that stream of data, transforming passive monitoring into an early warning radar.

Why it matters

Diabetes already touches roughly 10% of adults worldwide, a number projected to double to 1.3 billion by 2050, with global economic costs hitting $2.5 trillion by 2030. For individuals, tools like GluFormer could turn vague dietary advice into actionable precision care before prediabetes escalates into irreversible organ damage.

For the wearable and healthtech industry, this is a massive signal. Wearables have spent a decade acting as glorified digital logbooks, tracking what already happened. GluFormer shows that generative architectures can turn basic biometric sensors into predictive diagnostic tools, forcing wearable makers to rethink their software pipelines.

In our view, NVIDIA and its partners made a remarkably smart move by treating continuous glucose telemetry as a pre-training foundation rather than trying to parse messy medical charts. However, relying on AI to forecast health four years out carries real clinical risks. Long-range metabolic predictions are statistical correlations, not deterministic fates, and unexpected lifestyle shifts can throw them off. Still, turning LLM mechanics onto human biology looks like one of the most practical gen AI applications we have seen yet.