Six Five Pod Ep. 314: TerraFab, Memory's Comeback, and the Custom Silicon Debate

Patrick Moorhead and Daniel Newman cover Tesla and SpaceX's $16.8 billion TerraFab chip factory, Samsung's zHBM debut at FMS 2026, a week of frontier model launches, and a simulated debate for "The Flip" over whether custom silicon will capture the majority of AI workload dollars by 2028. The episode closes with earnings from SpaceX, Palantir, AMD, Astera Labs, IonQ, and Lattice Semiconductor.

The handpicked topics for this week are:

1. TerraFab Breaks Ground in Grimes County, Texas: Moorhead and Newman dig into Tesla and SpaceX's approved $16.8 billion TerraFab chip factory near Texas A&M, where Intel is confirmed as a digital foundry partner despite minimal SEC disclosure so far. Moorhead raises the open question of who handles the analog side of production that Intel doesn't cover, and Newman points to Musk's track record of hitting ambitious goals years behind his original stated timelines. Both hosts frame the investment as a genuine step toward expanding domestic chip manufacturing. (The Decode)

2. Samsung's zHBM Headlines a Blockbuster Memory Show at FMS 2026: The hosts cover the Future of Memory and Storage show in Santa Clara, where Samsung's BVNAND and zHBM concepts headlined alongside new announcements from Kioxia, SanDisk, and SK Group's high-bandwidth flash spec through OCP. Pat traces the industry's swing from oversupply and negative margins to today's memory shortage, driven by AI demand that has grown 5-10x faster than expected. Daniel adds that stacking memory directly on the accelerator, as zHBM proposes still needs to solve for the physical constraints of heat before it becomes production-ready.

(The Decode)

3. Three Frontier Model Launches Land in a Single Week: Alibaba shipped Qwen 3.8 Max, DeepSeek released V4 Flash at 14 cents per million tokens, and Meta debuted its first coding agent, prompting a discussion on how fast the frontier is moving. Newman argues every company will need a new benchmark built around operational workloads and real-world token efficiency. He points out that Chinese model usage and frontier lab revenue are both climbing simultaneously, and makes the case that healthy frontier labs are essential to funding the open source models built in their wake.

(The Decode)

4. Anthropic and AMD Both Make Moves on Custom Silicon: Anthropic confirmed it's developing custom AI chips, reportedly with Samsung and reportedly Broadcom. AMD announced its acquisition of Taalas, a company that etches model weights directly into silicon for major inference speed gains. Patrick traces his own multi-year prediction that heterogeneous computing would become standard, pointing to Google's TPU program as proof it can work at scale. Daniel frames both moves as evidence that AI companies are racing to control total cost of ownership (TCO) as memory and compute costs keep climbing. (The Decode)

5. The Flip: Will Custom Silicon Capture the Majority of AI Workload Dollars by 2028? Patrick argues every frontier lab is now pursuing vertical integration, citing Anthropic's new chip program, AMD's acquisition of Taalas, and NVIDIA's $20 billion license of Grok's inference technology as evidence the shift toward specialized silicon has become a step function. Daniel counters that every lab building custom silicon is simultaneously signing record contracts with NVIDIA and AMD, arguing that compute itself functions as the real moat regardless of chip architecture. (The Flip)

6. SpaceX Posts 92% Growth in Its First Public Earnings Report: SpaceX's debut public earnings showed 92% growth, narrowing losses, and rapidly expanding AI compute revenue layered on top of its launch and satellite businesses. Daniel flags new AI deals with Google and Anthropic as driving much of that growth, while Pat highlights that the company's stated $100 billion run rate target depends heavily on December performance and questions how soon TerraFab related capex will hit the income statement. (Bulls and Bears)

7. Palantir Delivers I