Constraint Capital #6 - June 2026
Constraint Capital Issue #6 breaks down Second Space Age technologies through Space Exploration Technologies Corp. public registration filing.
In his May 20th, 2026 interview on Invest Like The Best, Gavin Baker /X/ outlines several profound transformations and strategic maneuvers taking place within the AI infrastructure and data center landscape. Constraint Capital #6 tackles these maneuvers and the technology bottlenecks behind the SpaceX’s identified $28.5 trillion TAM /SEC/.
The three most important technology shifts and product development strategies Gavin highlights are:
The Disaggregation of Prefill and Inference and its Impact on Hardware Lifespans
One of the most critical structural shifts in how AI models run is the architectural decoupling of the “prefill” phase from the “decode” phase during inference [49:46].
The Shift: Baker explains that prefill—where the model processes and understands the prompt or input context—is fundamentally a memory capacity-bound problem [51:17]. Conversely, decode—the subsequent generation of output tokens—is a memory bandwidth-constrained process [51:25].
The Competitive Strategy: This canvas allows chip designers to make highly targeted, optimized trade-offs [51:32]. Startups (such as Groq or Cerebras) can build hyper-specialized chips optimized specifically for one part of this loop.
The Financial Impact: Crucially, Baker points out that companies can now stick specialized custom chips (like Cerebras wafer-scale hardware or Groq LPUs) in front of existing, older hardware architectures like Nvidia Hoppers or Amperes [49:46]. This approach uses the legacy GPUs for the prefill stage and custom silicon for decode, extending the useful economic lifespan of data center GPUs from 1–2 years to potentially 10–15 years [49:28].
Ongoing question for investors: How long before a GPU depreciates? /CNBC/
The Transition to Custom ASICs and the “Terafab” Talent Strategy
As standard GPU architectures face physical limits and intense market constraints, product strategy is shifting toward hyper-specialized ASICs (Application-Specific Integrated Circuits) and alternative foundry setups [43:30].
The Shift: Major players are leaning away from general GPUs to develop custom internal accelerators (e.g., Amazon’s Trainium 3) [43:30, 43:46]. To win, startups can’t just try to make a “better GPU”; they have to choose unique trade-offs within the “iron triangle” of hardware design (optimizing tightly for speed, memory, or cost) and do something fundamentally difficult to replicate [44:14, 47:58].
The Competitive Strategy: Baker highlights a major domestic foundry initiative (a joint venture involving SpaceX/Tesla, dubbed the “Terafab,” partnering with Intel) to establish bleeding-edge semiconductor manufacturing capacity in the U.S. [27:48].
The Financial Impact: Rather than operating like traditional corporate entities, Baker states that hardware leadership requires an aggressive, human-centric recruiting approach to compete with Taiwan Semiconductor (TSMC) [28:45, 29:59]. This includes recruiting elite global engineering talent and building highly accommodating cultural enclaves specifically tailored to entice the absolute best hardware minds to move and work domestically [30:15, 30:28].
All content published on this newsletter is based on public information and independent research. Opinions are authors own and have been sanitized through AI engines which can make mistakes. This newsletter is not financial advice, and readers should always do their own research before investing in any security.
Which leads to the last of the major maneuvers and the focus on this article…
The Move Toward “Orbital Compute” (Racks in Space)