ARK • Disrupt

Last week, Tesla held an event in Austin introducing the two-seat, purpose-built Cybercab to its Robotaxi fleet, and then the following evening commercialized its first rides.¹ Tesla disclosed that, in just six weeks, the Robotaxi fleet increased its unsupervised driving miles 2.6x, from the 380,000 reported during Tesla’s second-quarter earnings call to one million miles.²

On its site, Tesla posted a “Robotaxi interest form” aimed at commercial fleet buyers, asking respondents to flag interest in purchasing a Cybercab fleet.³ As ARK outlined in Big Ideas 2026, fleet owners are an important component of the robotaxi value chain, as shown below, with third parties owning and operating the majority of fleets.⁴

Note: Numbers are rounded. ARK continuously adjusts its forecasts based on latest available information and technological and industry developments. Updates this year include giving the fleet owner category a higher multiple and adjusting the adoption curve for autonomous driving adoption based on updated forecasts from market leaders. “EBIT”: Earnings Before Interest and Tax. “Autonomous Technology Providers”: refers to companies that develop and provide autonomous driving software, such as Waymo. Source: ARK Investment Management LLC, 2026, based on data from AAA 2025, Capital IQ 2025, and New York City Taxi & Limousine Commission 2014.⁵ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Forecasts are inherently limited and cannot be relied upon.

Cybercab is the first vehicle built with Tesla’s “Unboxed” process, which assembles vehicle modules in parallel rather than moving a chassis along a conventional assembly line. The process cuts the size of the line by 50% while increasing output substantially.⁶ The Unboxed method also simplifies the vehicle and requires no rare earth metals, which Elon Musk explained is “extremely difficult to achieve” while maintaining the range.⁷ Cybercab is the first Tesla with fully electronic brake-by-wire and no hydraulic brake fluid: it trades lines and plumbing for an actuator at each caliper.⁸

The Cybercab is central to scaling Tesla’s robotaxi service at competitive prices and, in ARK’s view, to expanding autonomous ride-hail into a multi-trillion-dollar opportunity. Initial Austin fares suggest that, on a 15-minute ride, Cybercab prices ~40% below Model Y Robotaxi rides in the Tesla app and ~50% below Uber, albeit with longer wait times.⁹ At scale, we project that robotaxis could be priced as low as $0.25 per mile—approximately one-tenth the cost of human-driven ride-hail, as shown below, and roughly one-third the cost of personal car ownership.¹⁰

Note: Values are rounded. Source: ARK Investment Management LLC, 2026, based on data from Yipit 2025, AAA 2025, and GetTransfer 2025.¹¹ In addition to those sources, certain information presented may be the result of ARK’s internal analyses, which draw on various additional sources of information. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Forecasts are inherently limited and cannot be relied upon.

Last week, OpenAI released GPT-6 Astra,¹² its highest-end model for complex agentic work, with capability ambitions and prices above the GPT-5.6 family, in direct competition with premium models like Anthropic’s Fable 5.1. OpenAI is positioning Astra for increasingly complex tasks involving tools, enterprise applications, and advanced reasoning.

At first glance, based on broad composite benchmarks, its performance jump looks modest. Astra scored 61.2 on Artificial Analysis’s Intelligence Index, essentially matching GPT-5.6 Sol’s 60.9 and trailing Fable 5’s 62.1. That said, the broad-based benchmark misses Astra's differentiated focus on computer and browser use, speed, efficiency, and sustained execution. Notably, Astra scored 99.9% on ARC-AGI-3, compared to Sol’s 7.8% and Opus 5’s 30.2%.¹³ ARC-AGI-3 places agents in unfamiliar interactive environments without instructions and tests their ability to explore, infer goals, and adapt efficiently. With a focus on long, high-context workflows and practical outcomes, Astra used OpenAI’s Responses Application programming interface (API) harness.

API pricing for Astra is $10 per million input tokens and $50 per million output tokens, 2.5x the price of Sol but only 65% the number of output tokens compared to Opus 5 on Agents’ Last Exam and a cost per task reduction of 63% compared to Fable 5.1 on Terminal-Bench 4.0.¹⁴ In other words, OpenAI’s revenue per token could be higher per unit of constrained compute compared to its prime competitors. Despite higher pricing per unit, fewer tokens and iterations could lower the time and total cost of a completed job.

Some AI experts speculate that OpenAI used a new technique called “recursive depth”¹⁵ to make those strides: instead of adding new layers to the transformer model, Astra may process tokens repeatedly in the same layers, reducing the overall need for memory and output tokens. Initially, Astra will roll out to a limited number of organizations and then expand to ChatGPT Plus, Pro, Business, and Enterprise users, as the API, Microsoft Azure, and Amazon Web Services (AWS) Bedrock also offer access.

Early third-party results are encouraging,¹⁶ with Perplexity and Databricks indicating that Astra is advancing the frontier on their in-house benchmarks. The broader rollout should reveal whether those high-water marks will translate into dependable productivity in the real world.

Last week, Elon Musk confirmed that SpaceX—not Elon, as previously reported—acquired APR Energy,¹⁷ a natural gas turbine manufacturer, to speed turbine deployment by bringing in-house the casting of blades and vanes, both of which have constrained power deployment.¹⁸ Elon expects the acquisition to pull timelines forward by 18 months,¹⁹ supporting SpaceX's plan to scale terrestrial compute from ~2 gigawatts (GW) by year-end to 5–10 GW—“closer to 10 GW”—by the end of next year.²⁰

Power has become a primary bottleneck in the deployment of AI compute, as most manufacturers of natural gas turbines are sold out through the end of the decade.²¹ The shortage has attracted new entrants, including Boom Supersonic, which repurposed its Symphony engine core—originally designed for supersonic flight—into Superpower, a 42 megawatt (MW) turbine for data centers. As of last December, its backlog was above $1.25 billion.²²

While turbines are addressing the near-term bottlenecks, Elon expects that SpaceX and Tesla will reach ~200 GW of solar manufacturing capacity by 2029. For context, that capacity expansion is ~9x the incremental GWs added by all players in the US in 2025, though only two-thirds of the incremental solar capacity that China added in 2025, as shown below.

Most of that capacity will support SpaceX's AI satellite ambitions, which will begin launching on Starship in the fourth quarter of next year.²³

Note: For this analysis, ARK assumes SpaceX and Tesla’s stated target is for grid capability. Source: ARK Investment Management LLC, 2026. The chart reflects ARK’s internal modelling, based on a range of external sources. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Forecasts are inherently limited and cannot be relied upon.

In ARK’s view, energy production is the backbone of every economy, as it translates directly into economic activity. Despite recent concerns over AI's energy use, ARK's research suggests that top economies have grown steadily more efficient, requiring less energy per dollar of Gross Domestic Product (GDP), as shown below. With a multi-trillion-dollar market at stake, that efficiency drive should intensify as AI scales.

Note: “kWh”: Kilowatt-hour, unit of energy, representing the use or generation of 1 kilowatt of power for 1 hour. “GW”: Gigawatt, unit of power equal to one billion watts. Source: ARK Investment Management LLC, 2026, based on data from Ritchie et al. 2025, International Energy Agency 2025, and Our World in Data 2025 as of December 22, 2025.²⁴ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.