This week's newsletter covers Starbase Louisiana, OpenAI Jalapeño results, and Figure AI's Index app.
1. SpaceX Highlights Starbase Ambitions
Last week, SpaceX and Governor Jeff Landry announced Starbase Louisiana, a $100 billion commitment to build one of the largest infrastructure projects in history on roughly 125,000 acres at Pecan Island in Vermilion Parish.[1] The greenfield investment includes five complexes, each with two Starship towers, with ten pads at the outset, and eventually more than a dozen towers supporting ~30 flights per day. The land will include not only a launch site, but also onsite propellant production, power generation, deep-water shipping, vehicle processing, employee housing, and likely an airport. Construction will begin in 2027, with first launches no earlier than 2029.
Two geographic features were the deciding factors. Launch corridors facing south over the Gulf will give SpaceX efficient access to the polar orbits for its space-based compute constellation, and natural gas in Louisiana will power the methane-fueled rockets. Gwynne Shotwell was clear about the motivation: the company's existing infrastructure—two pads at Starbase in Texas and three soon-to-be in Florida—cannot support the ambitious plans for Starship’s flight cadence.[2]
For perspective on the scope of the project, Brazil’s 14 gigawatt (GW) Itaipu hydropower plant, the closest completed contemporary analog, cost ~$90 billion in today’s dollars, and California’s much delayed and uncompleted high-speed rail project is projected to cost ~$125 billion, as shown below.
Source: ARK Investment Management LLC, 2026, FHWA, ESA/NASA, California High-Speed Rail, Louisiana Economic Development, Itaipu Binacional, China Three Gorges, the Hong Kong Government, Global Infrastructure Hub, and Massachusetts, with historical costs converted to approximate July 2026 dollars using BLS CPI-U data as of August 31, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.
What Explains The Size Of This SpaceX Investment?
SpaceX needs the capacity because the payloads-in-waiting are worth potentially trillions of dollars a year, starting with billions for its Starlink communications constellation and, we believe, trillions for its Starmind constellation. Based on ARK’s projection of SpaceX’s monetization rate per communications satellite, a single reusable rocket filled with Starlink satellites could generate ~$4 billion in lifetime net cashflow relative to the $1 billion in combined launch, satellite manufacture, ground station installation, and customer acquisition costs. Importantly, the towers at Pecan Island should be able to catch Starship. Indeed, if Starship launches its tenth fully reusable commercial rocket successfully in 2027, as we anticipate, the post-tax IRR (Internal Rate of Return) would approach 100% at an annual rate, as shown below.
Source: ARK Investment Management LLC, 2026, based on data from SpaceX company filings and statements and ARK modeling as of August 31, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.
With that kind of return, the constraint on SpaceX will not be capital but the physical capacity to deploy it. The company should maintain similarly healthy, albeit moderately diminishing, returns even as it scales through hundreds of Starlink-carrying Starship flights.
Though the Starlink opportunity will eventually saturate, the galaxy is the limit for SpaceX’s AI opportunity. By its 100th AI satellite launch ARK’s research suggests all-in costs to manufacture and launch its satellites will already run roughly half that of terrestrial datacenter developers. At that time ARK’s research anticipates that SpaceX will still be spending substantially on research and development (R&D) and will mostly monetize its orbital constellation by renting out capacity as an infrastructure-as-a-service provider while it seeks to catch up to the performance frontier currently occupied by Anthropic and OpenAI. Even with those constraints, its early Starmind launches should be able to yield IRRs in the high 20s as can be seen below.
Source: ARK Investment Management LLC, 2026, on data from SpaceX company filings and statements and ARK modeling as of August 31, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.
As the buildout expands, datacenter economics on the ground are likely to get worse as developers cope with local opposition and have to find exponentially increasing amounts of power. Meanwhile, SpaceX should become increasingly expert at manufacturing its satellites and packing more satellites into each launch; its economics should get better. Its 1000th launch could enjoy upfront costs at less than 40% those of the terrestrial benchmark. The volume of compute that SpaceX will command simultaneously suggests that it should be able to catch up with the performance frontier, pull back on research and development use of its constellation, and deliver higher-monetizing AI software to end-customers. By its 1000th launch, the prospective IRRs of Starmind could double those of Starlink at its peak, even as SpaceX invests much larger dollar volumes into the AI constellation.
Source: ARK Investment Management LLC, 2026, on data from SpaceX company filings and statements and ARK modeling as of August 31, 2026. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.
What Will Be The Macroeconomic Impact Of This Investment?
Based on Louisiana's ~$344 billion in nominal gross domestic product (GDP),[3] a capital commitment of $100 billion, no matter how phased, will move the needle, especially because it will impact a parish of fewer than 60,000 people. Compared to the state’s GDP per capita of ~$56,000, its fact sheet on this project projects 3,000 direct new jobs and ~8,100 indirect jobs with salaries averaging $92,600 per year over ten years. Landry put the historical contrast in stark relief: for generations, industry has extracted oil, gas, and petrochemicals from Louisiana, taking their share of state GDP down from ~25% in 1999 to less than 20% today.[4] SpaceX is entering Louisiana not to extract, but to build.
Chronically underestimated in macro forecasts, disruptive innovation does more than displace the existing capital stock: it increases the expected return on new capital enough to incentivize physical infrastructure that otherwise would not be built. In 2015, no company would have considered investing $100 billion in an industrial complex on Pecan Island to serve the rocket launch business. Rocket reusability changed the expected return on capital on such a project, which summoned the capital. Now, the capital is buying propellant plants, power generation, and port infrastructure in a parish that had none.
More important than the initial investment will be the second-order impact. Infrastructure built to accommodate a technology on a steep cost-decline curve should generate a higher return on capital than the legacy stock it displaces, delivering productivity—cheaper access to orbit, always-on connectivity, cheaper compute per watt—that redeploys labor, energy, and land at the margin for more productive uses cases. That supply-side expansion is the reason the growth associated with disruptive technology is much larger than consensus models have incorporated.
2. OpenAI’s Jalapeño Could Accelerate The Shift Toward Custom AI Silicon
Last week, OpenAI published the first results for Jalapeño, its first custom inference chip developed with Broadcom. Relative to Kimi K2.5, the largest public model tested, OpenAI reported ~1.5 times higher peak performance per watt and 3.4 times lower latency, as shown below.[5] Across Kimi K2.5, DeepSeek R1, and GPT-OSS 120B, Jalapeño achieved the performance-latency frontier despite being OpenAI’s first attempt.
Note: Comparison systems differ by model. GPT-OSS 120B is compared with NVIDIA GB200, while DeepSeek R1 670B and Kimi K2.5 1T are compared with NVIDIA GB300. Metrics show OpenAI-reported peak mixed tokens per second per kW. Source: ARK Investment Management LLC, 2026, based on data from OpenAI 2026.[6] 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.
SemiAnalysis independently verified Jalapeño’s InferenceX results at OpenAI’s lab and found that it outperformed Blackwell on performance per watt across nearly all of the workloads tested.[7] Its output-token throughput per megawatt also exceeded NVIDIA’s latest public Rubin results in SemiAnalysis’s testing.[8] Rubin is the more appropriate comparison, given the timing of the two chips. That said, the results are early: Jalapeño remains on engineering silicon and SemiAnalysis has yet to test it on AgentX, its benchmark for longer, multi-turn agentic workloads.
Perhaps more interesting than the benchmark results is how quickly OpenAI got there. Using its own AI models during the design and optimization process, OpenAI took Jalapeño from initial Register-Transfer Level (RTL) to tapeout[9] in roughly nine months.[10] SemiAnalysis also noted that OpenAI used Codex to overcome one of the traditional disadvantages associated with custom silicon: developing kernels for Jalapeño and shortening the time necessary to build a mature software stack around new hardware.
As power becomes a larger constraint on AI infrastructure, squeezing more inference from each megawatt will become increasingly valuable. OpenAI and Broadcom plan to deploy 10 gigawatts of OpenAI-designed accelerators through 2029,[11] giving OpenAI the volume to spread chip-development costs across an enormous inference workload.
Jalapeño does not mean that OpenAI will stop buying NVIDIA graphics processing units (GPUs), particularly for training. Instead, its early performance suggests that frontier AI companies operating at sufficient scale can justify the cost of designing their own inference silicon, particularly if AI itself continues to compress chip-development timeline cycles.
3. Figure AI Unveils Index, The Largest And Most Diverse Robot Dataset In The World
Last week, Figure AI came out of stealth mode with Index, a consumer app that pays people to record everyday tasks on camera, sourcing real-world physical data to train humanoid robots.[12] Over four months in stealth, the app has surpassed 264,000 downloads across 108 countries, with users uploading more than 16 million videos and earning $15 million and creating the largest and most diverse robot training dataset in the world, according to Figure AI. The company has committed more than $1 billion to data and compute over the next 12 months. Notably, to outsource the tasks, users can book Creators—a human-powered on-ramp to robots-as-a-service.
The launch of Index underscores the importance of diverse real-world data, the current bottleneck for humanoid robot deployment at scale. Hardware is advancing rapidly, as demonstrated during last week’s World Humanoid Robot Games in Beijing, during which Chinese humanoids beat Usain Bolt's 100m world record.[13] Commercial deployments, however, remain few and far between. As a result, companies are beginning to collect data in-house: Tesla flagged its own data efforts on its latest earnings call,[14] and Unitree's CEO is allocating a large portion of proceeds from the company’s initial public offering (IPO) toward software development.[15]
ARK's research suggests that humanoid robots are ~200,000X more complex than autonomous vehicles. That complexity is likely to create a ~$26 trillion total addressable market, split roughly evenly between household and manufacturing applications, as shown below.
Note: “GDP”: Gross Domestic Product. GDP forecast based on analyses by ARK Investment Management LLC. We define take rate as the percentage of transactions value that the business retains. Source: ARK Investment Management LLC, 2026, based on data from International Federation of Robotics 2025, Knutsen et al. 2025, and 36Kr European Central Station 2025, as of December 18, 2025.[16] 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.
The development of humanoid robots is still in early innings. We look forward to monitoring the pace of scaling over the coming years.
[1] Louisiana Economic Development. 2026. “SpaceX Launches New Era of Commercial Spaceflight with $100 Billion Louisiana Campus.”
[2] Ibid. See also SpaceX. 2026. “SpaceX Reports Second Quarter 2026 Results.”
[3] Bureau of Economic Analysis. 2026. “GDP by State.”
[4] Louisiana Economic Development. 2026. “SpaceX Launches New Era of Commercial Spaceflight with $100 Billion Louisiana Campus.”
[5] OpenAI. 2026. “Jalapeño’s first results show industry-leading speed and efficiency in AI inference.”
[6] Ibid.
[7] Shan, B. et al. 2026. “OpenAI Jalapeño: Better Than Nvidia Blackwell.” SemiAnalysis.
[8] Ibid.
[9] The RTL to Tapeout process transforms a high-level Register-Transfer Level (RTL) hardware description into a final, manufacturable layout file delivered to a foundry. See ChipExpert. 2025. “From RTL to Tapeout : A Complete VLSI Flow Explained.”
[10] OpenAI. 2026. “OpenAI and Broadcom unveil LLM-optimized inference chip.”
[11] Broadcom. 2025. “OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators.”
[12] FigureAI. 2026. “Introducing Index: Building The World’s Largest and Most Diverse Physical Dataset.”
[13] Zhuang, Y. “2026. A Chinese Robot Beat Usain Bolt’s 100-Meter Record. Should We Be Impressed?” The New York Times.
[14] Yahoo!Finance.2026. “Tesla, Inc. (TSLA) Q2 FY2026 earnings call transcript.”
[15] Wang, Y. 2026. “Unitree IPO Turns 36-Year-Old Founder Into China’s First Humanoid Robot Billionaire. Forbes.
[16] International Federation of Robotics. 2025. “National Robot Density.” Knutsen, R. et al. 2025. “Quadruped State of The Market - Unitree, Boston Dynamics, ANYbotics, DEEP Robotics, and The Rising Application Ecosystem.” SemiAnalysis. 36Kr European Central Station 2025 2025. “Unitree Launches Listing Guidance: Favored by Capital, but Mass Production Yet to Come.”
