#447: Blue Origin’s New Glenn Reaches Orbit As SpaceX’s Starship Advances Toward Mars, & More
- 1. Blue Origin’s New Glenn Reaches Orbit As SpaceX’s Starship Advances Toward Mars
- 2. Are AI Models Learning In Real Time?
- 3. Are AI Foundation Models That Incorporate Biological Engineering Outpacing Evolution?
- 4. In Biotech, Which Is Closer To The Mark: The Dour Public Market Or The Optimistic Private Market?
1. Blue Origin’s New Glenn Reaches Orbit As SpaceX’s Starship Advances Toward Mars

Last week, a decade in the making, Blue Origin's New Glenn rocket reached orbit successfully1 but failed, post-separation, to land its booster.2 The same day, during SpaceX’s Flight Test 7, Mechazilla chopsticks caught the Super Heavy Booster for a second time, but its next-gen Starship exploded post-separation, likely the result of a fuel leak.3
These events marked a historic day in the US quest to accelerate human spaceflight and included valuable learnings for future flights. ARK’s research suggests that Starship will enable SpaceX to continue down the Wright’s Law cost curve for satellite bandwidth capacity,4 expanding Starlink’s constellation from ~7,000 today5 to its long-term target of ~42,000 satellites6 and transporting humans to Mars as early as 2028, according to Elon.7
Sources: SpaceX 2025 (left) and Johnson 2025 (right).8 For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.
Last week illustrated a significant difference between Blue Origin and SpaceX as they race to space. Blue Origin worked methodically for a decade before debuting New Glenn while SpaceX has been “mov[ing] fast and break[ing] things.” Indeed, improved versions of Starship and Super Heavy Booster are already prepared for next month’s planned Flight Test 8.9 That said, Blue Origin seems to be evolving its strategy, as its next flight is scheduled for spring.10 Now facing investigations, however,11 the next phases of New Glenn and Starship are hostage to the Federal Aviation Administration’s (FAA’s) timeline.12
Look for more thoughts on this space in ARK’s Big Ideas 2025 coming soon.
2. Are AI Models Learning In Real Time?

In their recent paper “Titans: Learning to Memorize at Test Time,”13 Google researchers introduced a potential breakthrough in AI, specifically an architecture that enables models to store new information in long-term memory, enhancing their ability to adapt and evolve to meet the needs of individual users. While still early, the new architecture could be a fundamental advancement in AI.
Thus far, three major architectural innovations have accelerated AI: transformers, diffusion, and test-time compute. Introduced by Google researchers in 2017, the transformer architecture triggered the large language model revolution by allowing AI models to connect and weigh disparate but inter-related concepts. In 2019, the diffusion architecture allowed models to refine structured noise into probabilistic predictions, spurring advances in image and video generation that have been critical for robotics path-planning and prediction. Test-time compute allows models to solve exceedingly complex problems with internal error correction and reasoning. OpenAI’s o1 and o1 Pro have harnessed this architecture, which is likely to be the foundation for many agent-type systems.
Google’s new Titans architecture is likely to join this list of architectural innovations, allowing AI models to curate and refine their own long-term memory and save new information relevant not only for enterprise software agents solving problems but also for factory robots performing tasks. Instead of approaching each set of instructions as if introduced for the first time, the Titans-enabled AI system should become more expert and capable the longer it deploys in a particular context.
Indeed, Titan’s advancements could have dramatic implications for the stickiness of AI business models. The architecture would raise switching costs for businesses seeking to move from one AI provider to another: if an AI system has evolved to understand internal processes or local environments, switching to another system would require costly and potentially lengthy retraining. As with test-time compute, the Titans architecture also would shift compute needs for AI from training to inference. Instead of spending $ billions training an AI system upfront and hoping to re-coup the investment, AI model providers could shift the costs into operating expenses associated with provisioning the model to customers. With test-time compute, the customer pays for the time the model needs to think and solve a particular problem. With Titans, the customer would pay to imbed and retain local context in the AI system.
The Titans breakthrough suggests that AI research, engineering, and development still are in early innings. Researchers are exploring a variety of techniques to advance the efficiency of the transformer and diffusion architectures. Titans opens a new area for exploration and cost performance improvements, suggesting that research in test-time compute has a long way to go. As blazingly fast as AI has been moving, Titans could kick it into even higher gear.
3. Are AI Foundation Models That Incorporate Biological Engineering Outpacing Evolution?

Evolution is an engineering process that has shaped life on earth through billions of years of adaptation and optimization that has produced the extraordinary diversity and functionality of biological systems. Foundation models now seem to be demonstrating that they can harness the principles of evolution and surpass them with engineering solutions.
AI-driven models, for example, have tackled the century-old challenge of designing effective snake antivenom proteins. Snakebites claim over 100,000 lives annually. Traditional antivenoms derived from animal plasma are costly, have limited efficacy, and must be administered clinically.
Leveraging the power of machine learning, researchers are using computational protein design to create mini-binder proteins that neutralize lethal snake venom toxins. Remarkably, what once required months or years, the computational protein design process took only seconds,14 and yielded designs so precise that only a few dozen candidate proteins—as opposed to tens of thousands traditionally—were screened to identify highly effective binders. In animal and laboratory studies, the mini-binders showed a strong attachment to venom components that target neurotransmitters and cause tissue damage.
Published in Nature,15 the research highlights how machine learning has revolutionized protein design and is enabling the rapid development of transformative therapies. The resulting antivenom proteins promise safer, more accessible, and more cost-effective treatments that overcome significant limitations of more traditional approaches.
In another example of AI systems trained on evolutionary datasets, the ESM3 foundation model has simulated 500 million years of evolutionary progress. Now, within moments,16 the model can design entirely novel functional proteins that far surpass the performance of natural proteins, effectively "beating" evolution by exploring and optimizing design spaces inaccessible to natural processes.
Together, these advancements illustrate how AI-driven models are redefining biological engineering, solving long-standing challenges with unmatched speed and effectiveness, and opening new frontiers in therapeutic development.
4. In Biotech, Which Is Closer To The Mark: The Dour Public Market Or The Optimistic Private Market?

Biotech investors are witnessing a growing divergence between private and public markets, both in valuations and performance. Private investors are betting on long-term innovation, while public markets are focusing on short-term profitability. Will regulatory and economic shifts close that gap?
In 2024, the private market in biotech bounced back with 96 megarounds (nine-figure financing) that approached the 106 funded in 2021. JNJ acquired Intra-Cellular Therapies for $14.6 billion,17 for example, suggesting an urgency to replenish pipelines as $400 billion in revenue, or 30% of industry revenue, succumbs to patent cliff risks during the next nine years.18 While the initial public offering (IPO) market has languished in biotech, investors in the private market are concentrating on clinically validated late-stage assets and backing transformative science.
Meanwhile, the public markets are rewarding companies like Natera—a diagnostics leader with strong cash flow and consistent outperformance—and avoiding not only the broader diagnostics space but also innovation-heavy sectors like gene editing. In other words, public markets are favoring booked revenue over long-term potential. In our view, this misalignment will recalibrate as innovation-driven growth surprises on the high side of expectations.
The Trump Administration’s push for innovation-friendly policies—like repealing the Inflation Reduction Act’s (IRA’s) “pill penalty” and expediting drug approvals—could shift the public market’s focus. Coupled with efficiency reforms, those policy changes should align public markets with the confidence in private markets that high-impact innovation will be rewarded handsomely.
Last week, companies at JP Morgan’s (JPM’s) Healthcare Conference suggested that that alignment is beginning to take shape. Showcased by Nvidia’s partnerships with IQVIA and Illumina, for example, artificial intelligence (AI) is beginning to transform drug discovery. Highlighted by Bluerock’s Parkinson’s trial, advances in cell and gene therapies are pointing to the promise of durable, personalized medicine. Meanwhile, biotech innovation in China, which now leads a third of global licensing deals, epitomizes its globalization.
Private markets seem to be forecasting that public markets will recognize that innovation drives long-term value. While diagnostics, AI, and cell therapies already are proving their potential, regulatory reform in the new administration should accelerate recalibration in public markets.
-
1
Jonathan’s Space Pages. 2025. “Starlink Launch Statistics.”
-
2
Blue Origin. 2025. “Blue Origin’s New Glenn Reaches Orbit.”
-
3
Musk, E. 2025. “Preliminary indication is that we had an oxygen/fuel leak...” X.
-
4
Korus, S. 2025. “Starlink Is Riding Down Wright’s Law’s Cost Curve.” ARK Disrupt Newsletter. ARK Investment Management LLC.
-
5
Jonathan’s Space Pages. 2025. “Starlink Launch Statistics.”
-
6
Korus, S. 2023. “SpaceX Makes Progress With Second Starship Launch.” ARK Disrupt Newsletter. ARK Investment Management LLC.
-
7
ARK Investment Management LLC. 2025. “Elon's Historic Week: Cybercab And Starship | The Brainstorm EP 65.” YouTube. See also Musk, E. 2024. “The first Starships to Mars will launch in 2 years…” X. Nawfal, M. 2025. “ELON: FIRST MARS MISSION IN 2 YEARS...” X.
-
8
SpaceX. 2025. “Watch Starship’s seventh test flight…” X.
Johnson, J. 2025. “@cnnbrk @CNN @cnnio @pardon_Me_22 appears to be a meteor shower….” X. -
9
Musk, E. 2025. “Improved version of the ship & booster…” X.
-
10
Blue Origin. 2025. “Blue Origin’s New Glenn Reaches Orbit.”
-
11
U.S. Federal Aviation Administration. 2025. “FAA General Statements.”
-
12
Musk, E. 2025. “The booster flight was a success…” X.
-
13
Behrouz, A. et al. 2024. “Titans: Learning to Memorize at Test Time.” arXiv.
-
14
Callaway, E. 2025. “AI-designed proteins tackle century-old problem — making snake antivenoms.” Nature.
-
15
Vázquez Torres, S et al. “De novo designed proteins neutralize lethal snake venom toxins.” Nature.
-
16
Hayes, T. et al. 2025. “Simulating 500 million years of evolution with a language model.” Science.
-
17
Satija, B and Singh, P. 2025. “J&J doubles down on neurological drugs with $14.6 billion Intra-Cellular deal.” Reuters.
-
18
Eckhardt, J. 2025. “5 Insights From The 2025 JP Morgan Healthcare Conference.” Forbes.