EP688 | 🥽

The Release of Locked-Up Shares Increases Liquidity, but the Initial Impact Does Not Necessarily Mean Long-Term Bad News

The host first discussed the gradual release of SpaceX shares ahead of a potential listing. The market had initially expected shares entering circulation to immediately push prices lower, but prices instead appeared to bottom and rise after the lock-up expired. The heavier selling pressure seemed to have emerged before the release. Since additional shares will continue to be released in stages, the short-term increase in supply could certainly cause volatility, but the first wave of impact is usually the most pronounced.

His core view is that increased liquidity is not necessarily a bad thing. When ownership becomes more widespread and shares are no longer controlled by a small group, price discovery moves closer to market consensus. Even for a company as large as SpaceX, the experience of small and mid-cap stocks can offer a rough analogy: the more buyers there are, the less influence any single seller has over the price. A lock-up expiration should not be understood only as selling pressure; it also means more market participants and greater trading depth.

Cerebras Splits Inference into Long-Context Processing and Low-Latency Decoding

The host revisited Cerebras’ chip architecture. AI inference can roughly be divided into two stages: first processing the prompt data in a long context, then generating tokens one by one. The former emphasizes overall throughput, while the latter places greater importance on memory bandwidth and latency. Cerebras uses large chips paired with SRAM in an effort to gain an advantage in per-token decoding.

The architecture announced by Cerebras and AMD$AMD assigns different tasks to the two companies: AMD’s Helios handles prompts and long-context processing, while Cerebras’ WSE handles low-latency decoding. This division of labor does not force every task onto a single chip; instead, hardware is allocated according to the needs of each inference stage. The host believes Cerebras’ recent increase in production capacity and improved supply-chain visibility are worth watching, but the supply of unlocked shares remains an investment factor that must be considered at the same time.

Another example is AMD’s approach of acquiring technology that writes model logic directly into chips, conceptually similar to an FPGA. This can reduce data movement and decision-making latency, making it particularly attractive for trading applications where differences of a few milliseconds matter greatly. By comparison, Cerebras’ SRAM approach is more of a compromise between efficiency and practicality.

800VDC Is Engineering-Driven, but a White Paper Is Not a Guarantee of Near-Term Revenue

The market has recently circulated NVIDIA$NVDA’s 800VDC white paper. The host believes white papers typically combine plans that can be implemented in the near term with more distant ideal solutions. They therefore need to be separated when reading them: which developments could happen in the next 1 to 2 years, and which are ambitious long-shot plans requiring new equipment, additional budgets, and higher yields?

Existing data centers cannot realistically be demolished and rebuilt simply to achieve a theoretically optimal voltage configuration. Actual adoption will be constrained by existing power centers, equipment upgrades, environmental reviews, costs, and supply chains. Even if a solution is technically more elegant, it may still lose commercially to an alternative that is more cumbersome but cheaper and more reliable if the former is too expensive or unstable.

The host’s cited view is that if 800VDC reaches 5% penetration over the next year, that would already qualify as meaningful progress.

Power Rack Comes First in 2026–2027, with Power Center Later

The host divides the power architecture into several stages. In the short term—namely 2026–2027—Power Rack will likely remain mainstream. Beneficiaries include AC-800V PSUs, 800V-to-54V conversion equipment, capacitors, connectors, and BBUs.