Jan 19, 2025 · 24-Hour AI Briefing: Tesla Signals a Dojo Restart, Micron Moves for DRAM Capacity, and TSMC’s 3nm Is Sold Out Through 2027
Three headlines, one underlying reality: AI competition is dropping down the stack into compute, memory, and leading-edge process capacity constraints. The winners may be those who can secure predictable supply and convert capex into reliable delivery and stable cost curves.
1. Musk says the AI5 chip design is in good shape, and Tesla will restart work on Dojo
Commentary:
In Aug 2025, Musk publicly said Dojo had hit an “evolutionary dead end,” effectively dissolving the effort. Five months later, the Dojo 3 restart signal suggests AI5 provides a new technical foundation and architectural path forward.
Unlike earlier generations tied to D1 and complex wafer-scale packaging, Dojo 3 is described more like a cluster approach: integrating large numbers of AI5/AI6 chips onto boards, shedding legacy constraints.
The most realistic success path isn’t replacing GPUs wholesale, but winning inside Tesla’s most stable and highly customized training loops—an internal “training factory” acceleration layer that complements GPUs and cloud rather than competes head-on.
This restart also reads like preparation for the next phase: when autonomy and robotics scale data loops, compute certainty and cost curves become strategic variables. Do you think Dojo can actually restart and deliver?
2. Micron reportedly plans to acquire a PSMC wafer fab facility in Taiwan for $1.8B
Commentary:
The DRAM market is splitting structurally: HBM is oversubscribed due to AI server growth (Micron’s 2026 HBM capacity is reportedly sold out), while traditional DDR4/DDR5 supply tightens as Samsung and SK Hynix shift more capacity upmarket—raising the risk of “traditional DRAM scarcity” in consumer electronics and general servers.
Building a new fab is slow and uncertain. Buying a mature facility is about immediate usable capacity and infrastructure, reducing supply volatility risk over the next few years.
Strategically, this looks less like a short-term patch and more like a positioning move around “memory sovereignty” in the AI era—whoever can convert capacity into deliverable products fastest gains leverage in a mispriced supply-demand cycle. The question is whether $1.8B buys sustainable advantage or just a temporary window.
3. TSMC’s 3nm capacity is reportedly sold out through 2027, pushing Apple, NVIDIA, and others to consider Samsung and Intel as second sources
Commentary:
A 3nm sell-out through 2027 underscores TSMC’s dominance—and exposes structural fragility in the global supply chain as AI demand surges.
The lock-up is driven by a resonance of AI chips and flagship mobile SoCs: NVIDIA Blackwell / next-gen Rubin, AMD MI300, Google TPU v7, Tesla AI5, Amazon Trainium3, and Apple’s A19/A20 and related chips all leaning into 3nm and beyond.
For Apple, leading nodes protect power/perf and product cadence. For NVIDIA and HPC buyers, density and efficiency define system competitiveness. If a second-source era truly arrives, Samsung and Intel’s opportunity depends on three words: yield, delivery stability, and credible ecosystem/packaging integration. Can Samsung or Intel capitalize on this window?
Closing:
Dojo reboot signals, DRAM capacity M&A, and 3nm sold out to 2027 all point to the same thing: supply-side certainty is becoming the real battleground. As compute, memory, and leading-edge nodes get locked years in advance, the competitive edge shifts to predictable capacity, repeatable delivery, and controllable cost curves. Who do you think has the best hand going into the next phase—vertical integrators, foundry leaders, or memory players buying capacity fast?
Further reading (top AI events in the last 72 hours):