#461: Meta Unleashed Exciting New Tools At Llamacon, & More

During its inaugural LlamaCon developer conference last week, Meta released¹ a standalone Meta AI app, which includes a new “Discover” tab that should facilitate the use of AI by everyday users struggling to keep up with the rapid pace of change. By curating a feed that shows others interacting with the model, Meta is applying a “social media-meets-prompt engineering strategy” that not only will encourage collective learning but also will demystify AI through real-world examples.

Meta also announced its new Llama API, a turnkey platform for inference and end-to-end fine-tuning with open-weight models enabling customization. By empowering developers to download their tuned Llama models and to deploy them on-premises or across alternative cloud infrastructures, Meta is breaking from the “walled garden” approach and doubling down on the ethos of open-source innovation. The new flexibility could accelerate adoption across enterprises that demand data sovereignty, minimum latency guarantees, or specialized hardware configurations, broadening the total addressable market for large language model services.

To provide fast inferencing for the Llama API, Meta also has forged partnerships with Groq and Cerebras, who are engaging for the first time with a hyperscaler after serving mainly smaller partners or hosting their own services. By integrating these custom silicon architectures into its new API, Meta not only is unlocking ultra–low-latency inference for users but also is showcasing its willingness to knit together best-in-class external capabilities. For investors, the partnership framework underscores the strategic value of Meta’s flexible, open approach, which is assembling the building blocks of specialized hardware and software to meet the diverse performance needs of tomorrow’s AI workloads.

Tempus AI is transforming the drug discovery process by combining its vast genomic data with advanced AI and biological modeling, creating a unique business model to serve pharmaceutical and biotech companies. The company collects and refines patient data through its genomics business and collaborates with payers to build a rich data repository. That real-world data (RWD) then is licensed to pharma companies, providing them with valuable insights for drug development.

Based on the same data, Tempus is creating foundation AI models that other pharma companies can customize using their own internal data. Tempus’ Loop platform² allows pharmaceutical companies to validate their insights using patient-matched organoids—lab-grown tumor models that replicate human cancer biology better than traditional animal models. The new process is scalable, cost-effective, and aligned with real-world patient data. Pharma companies can map their insights from the foundation models to the organoids, rapidly identifying and validating drug targets for preclinical testing.

Through the Loop platform, not only is Tempus enabling pharma companies to customize and validate their insights, but it also is helping them advance drug discovery from hit identification to lead optimization and preclinical validation. Tempus monetizes the process by charging for access to the foundation models, customization capabilities, and target validation services, creating an ongoing revenue stream while refining its own platform and expanding its data capabilities.

This new approach—integrating AI, biological models, and data—offers a seamless, end-to-end solution that could reshape the traditional Contract Research Organizations (CROs) to which pharma/biotech companies outsource research services. Pharma companies are engaging directly with the platform—gaining access to more relevant data-driven insights, accelerating drug discovery, reducing reliance on traditional CRO services, and positioning Tempus as a key player in reimagining drug discovery.

ARK’s Autonomous Technology and Robotics team, in collaboration with Mach33,³ is finalizing our open-source SpaceX valuation model, with plans to publish in the coming weeks. In the meantime, please find the model’s logic in a graphical representation below.

Source: ARK Investment Management LLC, 2025. This ARK analysis draws on a range of external data sources as of May 2, 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. Forecasts are inherently limited and cannot be relied upon.

Source: ARK Investment Management LLC, 2025. This ARK analysis draws on a range of external data sources as of May 2, 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. Forecasts are inherently limited and cannot be relied upon.

Analytically, the model works like a flywheel. Beginning with cash, SpaceX builds rockets and satellites, generates orbital bandwidth, acquires Starlink customers, and reinvests the resulting cash. As the cycle continues, funds flow gradually toward Mars development until the Starlink constellation is complete. At that point, we assume that SpaceX goes “all-in” on Mars. Each Mars-bound rocket carries a mix of Optimus humanoid robots and materials. As Optimus costs decline and productivity improves over time, SpaceX builds its Mars “book value.”

ARK applies Wright’s Law⁴ to model various cost declines throughout this process. Interestingly, satellite performance (Gbps/kg) is the most sensitive input, directly impacting both revenue generation and capital expenditure and driving the efficiency and cost of bandwidth deployment.

From Starlink V1 to Starlink V2 Mini Optimized, satellite performance has followed a Wright’s Law learning curve. Nonetheless, its trajectory is the subject of many online debates, perhaps because of information from SpaceX itself. SpaceX has published Federal Communications Commission (FCC) filings, for example, for a next-generation satellite with a mass of ~2,000kg and, in separate materials, for future satellites with 1Tbps bandwidth. If those two specifications were to apply to the same satellite, the leap in performance would be massive, as shown by the red and yellow points in the chart below. Our research suggests, however, that the two sets of materials probably refer to different satellites. If so, the 1Tbps satellite would follow Wright’s Law, as shown by the green dot in the chart below.

Source: ARK Investment Management LLC, 2025, based on data from Starlink 2024 and U.S. Federal Communications Commission 2023.⁵ 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.

Last week, Adaptive Biotechnologies reported⁶ another strong quarter, including a record number of tests that are reinforcing the company’s leadership in the minimal residual disease (MRD) testing space—particularly for liquid cancers. Recently endorsed by the U.S. Food and Drug Administration (FDA),⁷ Adaptive’s clonoSEQ test is now the primary clinical endpoint in multiple myeloma trials.

ClonoSEQ offers a critical quality control (QC) tool that gives researchers a clearer picture of why patients in trials respond—or fail to respond—to specific treatments. Unlike traditional endpoints that rely on a few markers, clonoSEQ gives a much more comprehensive view, enhancing trial accuracy and predictive power.

As MRD testing becomes integrated into clinical practice, clonoSEQ will enable clinicians to monitor patients’ real-time responses to therapies. By using the same technology in the clinic as in trials, doctors can compare how patients react to different treatments, understand treatment efficacy more accurately, and adjust therapies.

Enhanced clinical trial success, real-time patient monitoring, and tailored treatments are creating a robust data moat for companies like Adaptive. As liquid biopsy technology proliferates, beginning with liquid cancers and eventually solid tumors, the data generated should continue to refine treatment strategies and offer unprecedented insights into disease progression. Adaptive appears well-positioned as a leader in both research and patient care.