The Evolving AI Landscape: A Conversation between SpaceX CFO Bret Johnsen and BlackRock's Tony Kim
In a Nutshell
SpaceX has transformed into an AI infrastructure leader, achieving $100B+ AI run rate in one year through vertical integration across compute, hosting, and software. The company plans orbital compute deployment by late 2027 using modified Starlink satellites, targeting 100 gigawatts in space by decade's end. Key advantages include sub-one-year ROI on infrastructure investments, five-month contract flexibility, and the ability to leverage existing aerospace engineering talent for terrestrial and orbital data centers.
These notes were generated by AI and may contain inaccuracies.
Tony Kim describes AI development as accelerating dramatically. The period from 2023 to 2026 represents the fourth year of AI advancement, with one quarter now containing a year's worth of activity. Kim notes that a year ago, development moved at a rate of six months equaling a year's worth of progress, and he anticipates that by year-end, a month could represent a year's worth of activity. He characterizes this as being at the foothills of exponential change and potentially approaching the singularity.
Bret Johnsen spent 10 years in semiconductors before joining SpaceX 15 years ago. Initially hesitant due to lack of aerospace experience, he was convinced by the argument that SpaceX was fundamentally a technology company operating in aerospace. Johnsen has been SpaceX's only CFO and describes the transition from semiconductor product cycles to aerospace as a cultural fit despite initial concerns about aerospace seeming "old and stodgy."
The SpaceX mission control facility has managed 700 launches. The building houses the Dragon capsule that first flew in 2010. Over 15 years, the facility has expanded from a single building to 40 buildings feeding into the main assembly area. Johnsen notes that SpaceX has transitioned from being solely a space company to operating as both a space company and an AI company, with AI representing the majority of business activity.
SpaceX's AI business grew from zero to over $100 billion annualized run rate within one year. AI now constitutes more than half of the company's total annualized run rate, which is projected to exceed $100 billion by December. The company completed its largest IPO in history while simultaneously building out AI infrastructure.
SpaceX's organizational DNA involves embracing audacious challenges. The first successful rocket booster landing was celebrated like a rock concert because it was previously considered impossible. Johnsen states that AI represents a challenge the company knows it can accomplish, unlike the booster landing which had unknown feasibility.
The January acquisition of XAI provided significant momentum. Key advantages include Elon's leadership of both companies creating immediate mission alignment, X providing content differentiation, and the ability to apply SpaceX leadership and operational DNA to the AI business. The acquisition positioned SpaceX at the forefront of the AI product cycle.
SpaceX claims leadership in AI infrastructure regarding cost and speed. The company is building Colossus data centers and has demonstrated the ability to rapidly scale compute capacity. Johnsen emphasizes that cost will be a determining factor across the AI industry.
SpaceX plans to exit this year (2026) with slightly over 2 gigawatts of compute capacity. The target for next year (2027) is 5 to 10 gigawatts. The company is transitioning from hundreds of megawatts to over 2 gigawatts in a single year. Johnsen identifies regulatory constraints and power availability as potential limiters for future expansion.
SpaceX's core differentiation is vertical integration across all operations. For AI infrastructure, the company handles engineering, acts as general contractor for data center construction, manages power systems, cooling systems, and networking internally. This approach eliminates margin stacking typical in traditional contractor relationships and enables faster execution.
SpaceX applies aerospace engineering talent to data center projects. Engineers experienced with pressurized fuels, gases, cooling systems, and high-power applications on launch sites are reassigned to terrestrial compute projects. Elon Musk compared this to "the Yankees playing a little league team" in terms of technical capability advantage.
SpaceX achieves less than one-year payback on cutting-edge compute investments through a dual monetization approach. The company hosts third-party workloads while building capacity for its own training and inference needs. This strategy derisks infrastructure investments by generating revenue during the buildout phase.
SpaceX offers three-month contracts with three-month termination options, providing six-month commitments. This flexibility contrasts with traditional multi-year hyperscaler agreements. The shorter terms are driven by SpaceX's expectation of needing significant compute capacity for its own models in the near future.
SpaceX hosts compute for tier-one companies including Anthropic and Google. The combination of software products (Grok, X, Cursor) and hosting services creates multiple revenue streams from infrastructure investments.
Johnsen identifies approximately five companies competing at the frontier level of AI development: Anthropic and OpenAI at the absolute leading edge, with three others (including Grok) close behind. He states that recursive self-improving AI will require frontier positioning to maintain competitive parity.
Grok is demonstrating faster progress than competitors on a month-by-month basis. The coding solution has improved dramatically in recent months. Johnsen notes that Grok is not yet at the bleeding edge with Anthropic and OpenAI but is closing the gap rapidly.
Grokbot is described as growing faster than any product SpaceX has previously launched. User feedback includes statements that the product is "transformational." Enterprise rollout is planned for the coming months, with integration into Cursor expected.
Regarding AI safety concerns raised over the weekend, Johnsen references Elon's position that the industry should follow self-regulation models used by other industries. He describes this approach as collaborative and beneficial for all parties, without offending stakeholders domestically or internationally.
SpaceX CFO Bret Johnsen discussed the $28 trillion TAM presented at the time of the IPO, noting that AI accounted for $25-26 trillion of that figure. This represents approximately a quarter of global GDP. Tony Kim emphasized that this TAM covers only digital AI, with physical AI representing an even larger opportunity in future years.
The conversation distinguished between digital AI (the current focus) and physical AI, which both speakers believe will be significantly larger than digital AI over the coming years. The space portion of the business, while representing hundreds of billions in annual TAM opportunity, is relatively small compared to the AI opportunity.
The AI business consists of four pillars:
- Enterprise - the largest segment where major players are competing for market share
- Consumer - expected to be significant, with differentiation through X's 500-600 million monthly active users providing best-in-class real-world content
- Infrastructure - where SpaceX already has a differentiated solution before orbital compute
- Content - leveraging X's content advantage over competitors like YouTube and WhatsApp
Starship represents the catalyst for all three SpaceX businesses. The fully reusable rocket is expected to achieve 10x improvement in cost per kilogram to orbit within months. This will enable putting boots back on the moon, establish lunar bases, and unlock capabilities beyond low Earth orbit.
Starship launch 14 will deploy production version 3 broadband satellites. One Starship launch with V3 satellites will carry 20 times the capability of Falcon 9 launches. Each V3 satellite provides approximately 1 terabyte of bandwidth, with 50-60 satellites per flight delivering 50 terabytes of downlink bandwidth per launch.
The AI era will demand unprecedented connectivity that human brains don't require. Humanoid robots, autonomous vehicles, and agent-driven systems everywhere will consume vastly more data than current internet usage. Space-based connectivity in high bandwidth will be essential, and Starship enables next-generation direct-to-cell capabilities.
Starship enables orbital compute by using largely the same satellite bus configuration but with modified payloads featuring more solar arrays and deployable radiators. This creates an entire constellation for orbital compute capabilities.
Flight 15 may attempt first and second stage recovery. The goal is reliable first and second stage recovery this year, with meaningful reflights beginning next year. As soon as a couple of reflown second stages are achieved, orbital compute reaches cost parity with terrestrial compute.
Orbital compute benefits from technology-driven cost curves rather than terrestrial inflation. While terrestrial costs increase due to inflation, labor, regulatory burden, power, and cooling, orbital compute costs decrease through Moore's law improvements, chip cost reductions, and increasing flight frequency. Over time, launch costs become nearly negligible, approaching aircraft-like operations.
Orbital compute provides clean, free energy from space-based solar power without community impact or environmental harm. AI industry leaders generally agree space-based compute is the correct long-term solution, though most estimate it a decade away while SpaceX plans first orbital compute flights next year.
Terrestrial gigawatt compute currently costs approximately $50 billion. Space-based alternatives offer advantages in both cost and dramatically reduced time-to-market. A single Starship launch can deliver 7+ megawatts (150 kilowatts per satellite × 50+ satellites), with the potential to launch multiple rockets daily.
SpaceX targets 100 gigawatts per year in space by end of decade, representing 20% improvement to the entire US grid annually. Elon Musk has discussed a terawatt in space during the 2030s. The key advantage is availability and speed rather than just cost, as terrestrial power infrastructure cannot scale at required rates.
The hosting business demonstrates staying power to invest at the bleeding edge of compute. This enables hosting others' compute needs while building internal capabilities, creating a virtuous cycle for the industry. New compute platforms can be deployed without building additional data centers.
The Terapam initiative targets a million wafers per month production capacity, representing a terawatt (1,000 gigawatts) of compute potential. This addresses supply chain constraints from TSMC and memory manufacturers that cannot scale at required rates. Current industry growth of 30% annually cannot support the triple-digit growth rates needed.
Terapam development needs to proceed in parallel with orbital compute scaling. Memory may be an early focus area. Development versions from Tesla could demonstrate capability next year, with volume production targeted for 2028-2029 to support the orbital AI hockey stick inflection.
The primary bottlenecks identified are chip supply for next year and power infrastructure. Utilities typically plan for 1% annual growth, while AI demand requires 10-20x that capacity. Supply chain constraints extend to turbine blades and components with 18-24 month lead times. SpaceX is addressing this by developing internal investment casting capabilities.
Revenue durability comes from diversified monetization of the same compute across multiple paths. Short-term contracts are offset by multiple monetization channels. Software improvements enable monetization above hosting rates, providing an option on software productivity gains. The range of $30-50 per watt next year reflects this diversified approach rather than single-point pricing.
When analyzing AI as a stack of layers, each layer presents different cost structures for token generation along with varying margins and durability characteristics. Infrastructure layers often feature contractual duration through 10-year or 5-year contracts that provide greater durability, while higher application layers may offer more opportunity to capture value but typically lack the same contractual longevity.
Different ROIs and durabilities apply across each layer of the AI stack. SpaceX's integrated approach encompasses all layers of the stack, with a payback period of less than one year. This rapid return profile makes such investments highly attractive, as a structure where $10 invested returns $10 over five to six years represents a compelling investment opportunity.
A viewer question asked how soon data centers will appear in space. SpaceX plans to fly the first data center equivalent satellites by the end of next year, with 2028 representing the first meaningful amount of compute capability in space. The terminology of "data center in space" creates misconceptions about massive installations, when the reality involves racks in space as part of a satellite constellation.
The space compute satellites utilize largely the same satellite bus as existing broadband V3 satellites and direct-to-cell satellites, with modifications including additional solar panels, radiators, and compute chips instead of communications chips. This leverages the existing satellite platform that has already flown 11,000 satellites to space, with thousands more planned. The core technical capabilities of launch and satellite-to-ground connectivity are already established, resulting in minimal technical risk for deployment beginning end of next year.
The preferred terminology describes racks in space connected into data center capability specifically for inference workloads rather than traditional data center concepts.
Each satellite configuration includes a full 72 NVIDIA chips with solar panels, matching the standard terrestrial rack configuration. Each launch carries approximately five megawatts of compute capacity.
Current AI trends will continue at an accelerated pace over the next five to ten years. Physical AI represents an expanding frontier, encompassing SpaceX's data center initiatives both in space and on Earth. The terabits of connectivity being deployed in space will support not only human use but also a proliferation of machines requiring tokens and connectivity, potentially creating a parallel economy of machines.
BlackRock offers exposure to AI and tech innovation through the iShares AI and Tech Innovation Active ETF (ticker: BAI), which is actively managed and provides exposure to both US and non-US based AI and AI-related companies. The iShares NASDAQ 100 ETF (ticker: IQQ) provides exposure to the largest 100 non-financial companies listed on NASDAQ. All investments carry potential for loss, requiring thorough research before purchasing ETFs or equities.
Keep iShares by BlackRock in your library
Save the videos and channels worth coming back to, and find them again in one place.
