ARK • Disrupt

On September 4, Tesla’s new Robotaxi app ranked in the top five on the US iOS App Store and was the #1 travel app for consecutive days.¹ Previously accessible only by private invitation, the app now allows riders to sign up for Tesla’s pilot robotaxi services in Austin and San Francisco.

According to app store data, Tesla’s downloads on the first day surpassed the peak monthly onboarding rates of Uber and Waymo by ~40% and ~700%, respectively, as shown by the purple square at 80,000 below. Given that the Robotaxi service is available in just two cities, this early traction suggests strong rider interest in Tesla’s service, even before it scales its fleet or removes safety drivers.

Note: We infer Tesla downloads from app store rankings Source: ARK Investment Management LLC, 2025, based on data from SensorTower and iOS app store rankings as of September 5, 2025. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

Tesla’s Robotaxi debut also marked the 10th largest ride-hail onboarding event in US history—behind the major social and holiday events, such as New Year’s Eve and St. Patrick’s Day, that fueled Uber’s early adoption, as shown below. Hit on an ordinary Thursday in September without significant marketing support, Tesla’s app performance highlights consumer enthusiasm for Robotaxi and underscores its competitive positioning relative to Waymo, as its Robotaxi onboardings exceeded Waymo’s record day by 6.5x.

Source: ARK Investment Management LLC, 2025, based on data from SensorTower and iOS app store rankings as of September 5, 2025. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

While Tesla’s Robotaxi program still operates with safety drivers, Elon Musk has suggested that the service could be driverless by the end of 2025.² If Tesla can scale, Robotaxi is likely to deliver superior unit economics compared to both human-driven ride-hail and other autonomous fleets. Early traction indicates Tesla could expand its pilot program into new geographies more easily than expected, accelerating the path toward commercial-scale deployment once its systems are validated.

Last week, the Tennessee Valley Authority (TVA) announced an agreement to deploy up to 6 gigawatts of NuScale Small Modular Reactor (SMR) capacity—the largest SMR deployment program in US history—across its seven-state service region.³ The move comes amid growing nuclear momentum as hyperscalers seek clean, reliable power to meet AI-driven demand for computing power.

To fulfill President Trump’s Executive Order and quadruple US nuclear capacity by 2050, the industry must overcome a critical bottleneck: domestic enrichment.⁴ Once the west’s sole provider, the US now effectively has no commercial enrichment capacity and is decades behind Europe, Russia, and China, as shown below.

Note: China’s industrial-scale centrifuge program was initially reliant on Russian technology but has significantly accelerated domestic capacity over the past few decades. Source: ARK Investment Management LLC, 2025, based on data from Tenex as of June 2025.⁵ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

For more details, read ARK’s latest research detailing the reasons enrichment capacity is essential to energy independence, innovation, and national security in the US.

Powered by Epic and Microsoft, Cosmos Medical Event Transformer (CoMET)⁶ applies real-world health data to predictive models. Based on 118 million patients, 115 billion medical events, and 151 billion tokens, CoMET predicts the next medical event in a patient’s life: it converts sparse, disorganized medical records into multiple plausible scenarios—simulated timelines that can be queried for risks, lab values, or utilization.⁷ Focusing on individual instead of group records is the foundation for personalized medicine: “N-of-one medicine.”

Scale matters. Trained on three sizes (62M, 119M, and 1B parameters), CoMET’s larger models outperformed the smaller ones consistently. The largest, CoMET-L, not only outperformed task-specific baselines across diabetes, hypertension, and acute crises such as sickle-cell events, but also improved 30-day readmission forecasts (AUCROC 0.770 vs. 0.717) and reduced error in predicting office, emergency room, and hospital visits. That said, not all results were positive: hyperlipidemia outcomes lagged, reinforcing the idea that the power of prediction depends on data density and cohort size.

The scaling law in biology and medicine is revealing. While GPT-like models optimize at ~10 tokens per parameter, CoMET required 1,000 tokens per parameter. Medical data are more voluminous yet sparser and noisier than text. Because predictive events are rare and signal density per token is low, achieving comparable performance generally requires much larger cohorts and longer longitudinal coverage. Indeed, powered by 118 million patients, CoMET represents a fraction of the world's 8 billion humans, each with unique biological journeys and diseases.

In other words, the scaling of healthcare models works but is nascent. The next step will be fusing data associated with longitudinal medical events with that from deeper biological layers—genomes, protein structures, single-cell systems. Only then will these models shift from prediction to prevention, making N-of-one medicine a reality.

Last week, Ionis released positive topline data from its Phase 3 studies of olezarsen in people with severe hypertriglyceridemia (sHTG). Importantly, the data indicate significant reductions in triglycerides and acute pancreatitis events: up to a 72% placebo-adjusted reduction in triglycerides and an 85% reduction in acute pancreatitis events.⁸

Olezarsen is an RNA-targeted therapy that reduces expression of apoC-III, a key regulator of triglyceride metabolism.⁹ Triglycerides are one type of lipid—the fat and fat-like molecules in the blood that, at elevated levels, increase risk of pancreatitis and cardiovascular disease. Already approved by the FDA for familial chylomicronemia syndrome, the US market opportunity for olezarsen could expand 1,000-fold, from ~3,000 patients to more than 3 million.¹⁰

Currently, sHTG management combines a low-fat diet with chronic therapy, including prescription omega-3 fatty acids and fibrates. If approved, olezarsen would mark a paradigm shift in care for sHTG with a targeted therapy that addresses the biology of disease instead of symptoms.

According to our research, the next wave of disruption in lipid management could come from one-time, in-vivo gene-editing. In June, Eli Lilly announced plans to acquire Verve Therapeutics for $1.3 billion, gaining access to a Phase 1 in-vivo gene-editing candidate that has demonstrated LDL cholesterol reduction of up to 69%.¹¹ Last week, Editas announced a preclinical in-vivo gene-editing candidate designed to reduce LDL cholesterol.¹² Among the most advanced are two CRISPR Therapeutics‘ in-vivo programs in Phase 1 trials.¹³ One targeting ANGPTL3 demonstrated that a single dose can lower not only triglycerides up to 82% but also LDL cholesterol by 86%.¹⁴ Additional data later this year could help demonstrate that gene-editing might transform cardiovascular medicine by shifting care from chronic management to durable, one-time interventions.

OpenAI's recent foray¹⁵ into science highlights that autonomous research and development (R&D) could bridge the gap between bits and atoms, or computation and physical experiments. While OpenAI’s entry validates the massive potential of computation, its path into the physical lab is an open question. True autonomous science would close the loop.

Various players are tackling the challenge with distinctive strategies:

• Pure "Bits" Players: Companies like Japan’s Sakana AI are building in-silico "AI scientists"—agents that hypothesize, write, and iteratively rewrite their own code, improving with each cycle. While they are creating powerful, self-improving digital brains, these brains still need bodies to interact with the physical world.
• Integrated "Bits-to-Atoms" Stacks: Core to the current revolution, computational biotech companies like Recursion and Absci are building models and then wiring them into high-throughput robotic labs. Their innovation lies in a seamless feedback loop: an AI designs a compound (bit) and a robot synthesizes and tests it in the lab (atom), the data from which trains the next generation AI model. The proprietary engine gets smarter with each cycle, driving efficiency by reducing systematically the necessary experiments and time to discovery.
• The Unified Full-Stack System: Lila Sciences is building a fully integrated platform, developing both the AI and the autonomous labs to unify the learnings from different domain-specific labs into a central, overarching AI, and creating a powerful discovery engine that learns and improves across all science.

The endgame is not creating the smartest AI or the fastest robot. Instead, those who master the bits-to-atoms-to-bits feedback loop will capture the most value, turning scientific discovery into a rapid, iterative, and continuously learning system.

This movement is converging on the concept of the "self-driving lab," analogous to self-driving cars aiming for full autonomy. This shift could alter the unit economics of science, lowering the cost per experiment and cost per insight dramatically, and accelerating the pace of innovation.