ARK • Disrupt

Last week, the first public footage of a Tesla driverless Model Y testing in Austin surfaced online.¹ Concurrently, Elon Musk confirmed that Tesla is targeting, albeit tentatively, June 22 for the commercial launch of its robotaxi service.² Its robotaxi platform will mark a fundamental shift in Tesla’s business model, from one-time hardware sales to recurring revenue with software-like margins. Our research suggests that Tesla’s robotaxi platform could account for ~90% of its enterprise value in 2029.³

Its vertically integrated manufacturing and vision-only approach to autonomy could result in a major cost advantage relative to competitors, allowing Tesla to scale more quickly and efficiently. Our research suggests that, at scale, without depending on other automakers, LiDAR, and other unnecessary hardware, Tesla’s cost per mile could be ~30–40% lower than Waymo’s, as shown below.

“Early launch” assumes ten cars per remote operator with utilization rates roughly equivalent to ride-hail today. “At scale” assumes 100 cars per remote operator with an improved autonomous operation utilization rate. Source: ARK Investment Management LLC, 2025. This ARK analysis draws on a range of external data sources as of December 31, 2024, which may be provided upon request. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results. Forecasts are inherently limited and cannot be relied upon.

That cost advantage is compelling, especially since the rideshare and taxi price comparison app Obi recently found that consumers in San Francisco are willing to pay more for Waymo than for Uber or Lyft.⁴ As Tesla launches in the coming weeks and Waymo continues to expand its reach, ARK will be monitoring how quickly the autonomous ride-hailing market is scaling.

Last week, in collaboration with Mach33, ARK released its open-source SpaceX valuation model, projecting a base-case of ~$2.5 trillion in enterprise value by 2030. Check out our blog,⁵ download the model on GitHub,⁶ explore the assumptions, and share your thoughts!

The forecast presented is subject to revision by ARK Invest (“ARK”) and provided solely as a guide to current expectations. Forecasts regarding broad markets and individual issuers are not, and are not intended to be, representative of any ARK-managed investment product or the characteristics of any ARK portfolio. FORECASTS ARE HYPOTHETICAL AND HIGHLY SPECULATIVE, AND PRESENT MANY RISKS AND LIMITATIONS. While ARK believes that there is a sound basis for the forecasts presented, they are provided for illustrative purposes only and no representations are made as to their accuracy. The recipient is urged to use extreme caution when considering the forecasts, as they are inherently subjective and reflect ARK’s inherent bias toward positive expected results. Any positive results should be viewed as a measure of the relative risk of such companies, with higher forecasts generally reflecting greater risk. There is no guarantee that any results will align with the forecasts, and they might not be predictive. Some or all results may be substantially lower than projected results. Please refer to the blog and/or the GitHub model for more important information regarding the forecast methodology.

Last week, OpenAI launched o3 pro,⁷ extending the frontier of large language model (LLM) performance and validating ARK’s long standing thesis that the quality of foundation models will compound at an exponential rate. On the GPQA benchmark—an exacting test of PhD level scientific reasoning—o3 pro scored 84%, surpassing the 81% achieved by o3 and reaching the level of the most advanced systems—Anthropic, Google, and xAI—as shown below on an absolute log error scale. In our view, the speed at which OpenAI iterated from o1 to o3 pro continues to demonstrate the power of scale in AI training: larger datasets, better optimizers, and increasingly massive compute clusters are reinforcing one another in a virtuous cycle of capability gains. xAI’s massive GPU cluster in Memphis, TN, could be the reason that Grok 3 Beta reached o3-pro’s performance level four months ago.

Source: ARK Investment Management LLC, 2025. This ARK analysis draws on a range of external data sources as of June 10, 2025, which may be provided upon request. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.

OpenAI also slashed the price of the original o3 model by 80%,⁸ bringing its cost in line with that of GPT 4.1.⁹ This move could unleash a new wave of experimentation, as enterprises embed high reasoning agents more deeply into production workflows. The competition is likely to respond. Previously priced nearer to o3,¹⁰ Claude 4 Opus now faces significant price pressure, forcing Anthropic to decide whether to cut margins or cede share at this early stage of the most profound technology revolution in history.¹¹

The magnitude of this price decline is breathtaking. Despite offering better performance, o3-pro is 87% less expensive than o1 at its launch just six months ago. At an annual rate, the cost decline is 98%, surpassing the 92% drop that ARK documented in Big Ideas 2024. If that trend were to persist, inference costs could become negligible in the next few years, causing an explosion in AI agents that could accelerate the shift of global knowledge work to software. Indeed, the “AI arms race” could intensify as model builders like OpenAI and xAI deploy increasingly larger compute clusters that widen the gap between leaders and laggards.

Single-cell RNA sequencing has shifted from theoretical to a practical reality, igniting fierce competition to achieve higher sensitivity and throughput at lower costs. According to a recent benchmarking study,¹² 10x Genomics leads in cell recovery and resolution quality, while Fluent Biosciences is the lowest-cost provider, at ~$0.005 per cell, half the rate of its competitors. Scale Biosciences’ solution detects a broad array of features but suffers from sequencing efficiencies, and Parse Biosciences, though slightly lower than 10x Genomics in sensitivity, excels in fixed-cell T-Cell Receptor (TCR)—crucial for clinical applications and time-sensitive samples.¹³

Importantly, prices have dropped to $0.01 per cell or below, breaking below a critical economic barrier and opening up single-cell sequencing to a new zone of biological discovery. Lower costs and improved data quality at the scale of a billion-cells are enabling researchers to explore the mechanisms of disease and complex biological systems at a level of detail previously impossible.¹⁴

Parse, for example, recently partnered with the Wellcome Sanger Institute and Helmholtz Munich to create a cancer plasticity atlas analyzing more than a billion cells, using 3D organoids to map treatment responses.¹⁵ Immunai is generating a 50 million-cell dataset from the Parker Institute for Cancer Immunotherapy (PICI) pan-cancer cohort and powering an Annotated Multiomic Immune Cell Atlas (AMICA) foundation model based on multi-million-cells.¹⁶

Scaling single-cell datasets is essential to building biological intelligence. For perspective, today's biology foundation models are equivalent to ChatGPT-2, influenced by both the noise and the complexity of biological data. Techniques like Single-Cell Transcriptomics Analysis and Multimodal Profiling through Imaging (STAMP), which can profile 1 billion cells at a record-breaking $0.000075 per cell, are at the cutting edge of affordability and scalability.¹⁷

To reach GPT-4-level performance, the datasets supporting biology models will have to approach GPT-4-level datasets. Based on our research, billion-cell atlases are achievable now. Scale will be essential for the precision and generalization necessary to understand biology. Aligning with the 35-40 trillion cells in one human being, trillion-cell datasets will be the next frontier, moving the science toward ChatGPT-4-level understanding.

This week, Caris Life Sciences marks another important moment for healthcare innovators and investors with its initial public offering (IPO).¹⁸ Along with competitor Tempus AI—which went public last year¹⁹—Caris is leading a seismic shift from traditional binary diagnostics to a new age of continuous, multi-dimensional genomic data and testing.

Unlike conventional tests that provide discrete positive or negative results, Caris and Tempus both deliver rich molecular insights through extensive genomic profiling. Caris has a vast database of more than 530,000 whole-exome /whole-transcriptome (WES/WTS) sequences,²⁰ while Tempus holds hundreds of petabytes of sequences and other patient profiling data.²¹ These large-scale datasets offer detailed, actionable, and personalized insights into cancer biology, significantly improving targeted treatment decisions.

Importantly, the market potential for these companies is not zero-sum. As the ability increases to detect cancer as well as neurological and cardiovascular diseases earlier, multiple innovators with AI expertise and large datasets should grow the market, especially with continuous and longitudinal health profiling.

The value proposition lies in combining deep genomic data with orthogonal datasets that link genotype to phenotype, enhancing predictive and prognostic capabilities. By scaling this approach, Caris and Tempus seem poised to unlock unprecedented clinical and economic value, reshaping medicine—from static diagnostics to dynamic, continuously monitored, personalized healthcare.

Last week, Elon Musk’s xAI named Polymarket X’s official prediction-market partner. Living inside Polymarket’s interface, Grok already is providing conversational summaries (“Ask Grok”) on every market page. In addition, X plans to add Polymarket probabilities to posts and Grok answers, giving 600 million monthly users real-time odds on the world’s most high-profile debates.²²

Because traders risk capital—not just reputation—market-priced probabilities have a track record of beating polls and pundits. The last US presidential election is a good case in point.

In another example, the odds of “Israel military action against Iran before July” surged two days before the Israeli Defense Force’s (IDF’s) surprise attack on June 13, as shown below. Polymarket flagged the heightened risk well ahead of mainstream media coverage.²³ Apparently, the transparency and continuous pricing in decentralized markets helps filter noise and highlight signals that matter for policymakers, investors, and the general public.

Source: ARK Investment Management LLC, 2025, based on data from Polymarket, as of June 13, 2025. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.

Compared to the infrastructure associated with past prediction markets, smart contract blockchains now allow them to scale inexpensively. Thanks to Polygon’s proof-of-stake network, a Polymarket trade settles in seconds for less than $0.01 in gas fees, allowing thousands of traders to supply liquidity without intermediaries. In 2024, Polymarket generated network fees of $102 thousand against $9 billion in betting volume—illustrating costs orders of magnitude lower than regulated betting exchanges and highlighting the cost advantages of public blockchains.²⁴

AI agents should increase the breadth and utility of prediction markets significantly. Grok-powered assistants could watch live data streams, spin up micro-markets (“Will Apple ship 3nm M3 chips by Q4?”) in seconds, and auto-quote two-sided liquidity, collapsing the friction that has limited prediction markets to marquee events historically.

As autonomous agents arbitrage inconsistencies across thousands of Polymarket contracts, spreads should tighten, volumes should rise, surfacing the “price of truth” for everything from clinical-trial readouts to climate metrics. As a result, xAI’s tooling could transform prediction markets from a curiosity to an always-on oracle layer.²⁵ 

The convergence of smart contracts and AI could create a probabilistic data layer for the global economy. Among the investable implications are the following:

1. Social platforms that surface market odds could increase user engagement and trust. X, for example, could become a more reliable source of truth.
2. Token-secured oracle networks could capture more value as the demand for verifiable probabilities grows.
3. AI-driven liquidity could blur the lines between trading bots and research desks, accelerating capital formation in previously illiquid niches.