The Second Silicon Spine: The Week Microsoft Called Maia 300 Its 'Primary' Inference Chip and Four Countries Rebuilt Around It
Black Tuesday was not an AI chip correction — it was a fork, from training GPU+HBM toward an inference stack of ASIC+LPDDR+HBF+edge that four countries formalized in a single week
On Black Tuesday — October 13 — the KOSPI shed 5.46%, the TAIEX dropped 1.95%, and the Nikkei lost 1,253 points, with SK Hynix down 9.8% and Samsung Electronics down 7%. Layer the previous day's Open Compute Project (OCP) summit announcements on top, and the character of the selloff becomes clear. Microsoft called the Azure Maia 300 its "primary inference system." AMD unveiled EPYC Verano (192 cores, integrated LPDDR5X) and the Helion rack (32K cores) for AI agents. SK Hynix announced MARS, which accelerates time-to-first-token (TTFT) by up to 5x, and HBF, which expands memory capacity by 14x. In the same week, four countries moved in the same direction — away from the two-year axiom that "AI silicon = training GPU + HBM" and toward a second silicon spine built for inference.
The US signal has three layers. First, Microsoft announced Maia 300 for US-and-Japan datacenter deployment starting in 2027 and explicitly called it the "primary inference system." Not supplementary, not alternative — primary. Second, AMD's Verano integrates LPDDR5X directly inside the server CPU, and the Helion rack unveiled with Agentic targets rack-scale inference infrastructure in a CPU market AMD now sizes at ~270 trillion KRW. Third, as Axios put it, "every major Nvidia customer is now a competitor." Combined with Broadcom's warning on AI-demand durability and CXMT's assault on Micron's server-memory margins from China, the Nvidia price umbrella started visibly compressing. The implication: training capex is still booked through 2027, but inference capex now has at least four credible non-Nvidia stacks competing for it.
Korea answered the fork most clearly on the memory side. SK Hynix's MARS, unveiled at OCP in the US, accelerates TTFT by up to 5x — an optimization aimed explicitly at inference latency, not training throughput. HBF increases capacity 14x, pushing model and cache layers closer to the compute die. Samsung formally announced its "post-HBM memory strategy" emphasizing speed, scalability, and durability, and Korean trade press flagged that "LPDDR will be the next star after HBM" — the same week AMD's Verano integrated LPDDR5X into a server CPU. Yet foreign investors sold record-earnings names; the IMF upgraded Korea's 2026 growth from 2.6% to 3.3% citing semiconductors; September semi exports hit $43.1B (+216% YoY), after $38.2B in August and $32.7B in July. Exports surge while P/E compresses. The market is asking a specific question: can an inference cycle absorb the GPU wave as it crests?
Taiwan is answering from the ATP (assembly-test-packaging) side. Foxconn (2317) showcased vertical integration of Nvidia's Vera Rubin rack at OCP 2026 through Ingrasys and Foxconn Industrial Internet; ASUS launched the XA P2N-E2 and Vera-based AI servers on the same stage. More consequential was Nvidia's prepayment agreement with KYEC (京元電, 2449-TW), which locks US AI-chip testing capacity in exchange for upfront cash — a hard signal that near-term training demand is not collapsing. But TAIEX still fell 1.95%, led by TSMC (2330), and Barclays flagged DDR4 spot up 24% in three months (vs. DDR5 +20%) as Micron's exit from legacy memory tightens the entire stack. TSMC's October 15 earnings call will test whether 2027 capex rises toward $80B — the empirical bar for whether training capex extends through the fork, or whether inference stacks (Maia, Verano, Trainium) start absorbing the marginal dollar.
Japan contributed two signals on the inference side: edge and hyperscale. Asmec and the AIE-KT78/68 launched Nvidia Jetson Thor-based edge AI terminals for robots, enabling cloud-free inference at the device — a direct Japanese expression of what Korean trade press called "LPDDR as the next star after HBM," since robot inference pushes LPDDR volume. In parallel, NTT entered final-stage negotiations for a $6.3B datacenter for Anthropic in Japan, representing geographic diversification of North American inference demand. Tokyo Electron Devices launched Zenity, an AI-agent security platform, and etch-chamber-liner CAGR forecasts were upgraded. The Nikkei breaking its 70,000-yen support (-1,253 points) pulled semi stocks down, but underneath the headline a vertical expansion from training → cloud inference → edge inference is clearly in motion — and Japan is positioned on the equipment and edge sides of all three layers.
Connect the four signals and the fork becomes legible. The training capex cycle is booked through 2027 — Blackwell → Vera Rubin → Rubin Ultra — but Axios's "every major customer is now a competitor" is now showing up in P/E compression, not just headline risk. The inference capex cycle, by contrast, has barely begun: Maia 300 (US and Japan, 2027), Verano (CPU + LPDDR5X), Helion (32K-core rack), MARS (5x TTFT), HBF (14x capacity), Jetson Thor edge, and the $6.3B Anthropic datacenter in Japan all landed in a single week. Memory hierarchy is being rewritten in parallel: HBM4e stays on the training rack; LPDDR5X, HBF, and even DDR5 extend into inference racks and edge devices. Black Tuesday's selloff prices "the training cycle may fatigue," not "the inference cycle won't come." Which side of the fork an investor stands on determines the direction of the next P/E re-rating.
Three positioning notes stand out. (1) Memory vendors (SK Hynix, Samsung, Micron) hold HBM capacity but will be judged on how fast they can shift lines to LPDDR and HBF. SEMI projects global DRAM capacity rising from 2.08M to 2.94M wpm by 2030 (+41%, 7.1% CAGR); what fraction gets built for inference layouts, not training layouts, is the swing variable for 2027-28 margins. (2) The ATP chain (TSMC, Foxconn, KYEC, Pengyuan/朋億) can serve a rack where Vera Rubin and Maia 300 coexist — a memory stack mixing DDR4, DDR5, LPDDR5X, and HBF paradoxically yields more test and assembly revenue, not less. Pengyuan's Q3 revenue hitting NT$4.58B (+124% YoY, a single-quarter record) is the first visible footprint. (3) Edge (Jetson Thor, Zenity) and robot-memory inventory cycles rotate asynchronously from the hyperscale cycle — the October 2026 drawdown may be the first time that asynchrony is actually exposed in prices, which is bullish for diversified exposure and bearish for pure-training beta.
The reading is simple. October 13 was not an "AI chip correction" — it was an AI chip fork. One axis is training GPU + HBM fatigue starting to price in; the other is inference ASIC + LPDDR + HBF + edge being formalized across four countries in the same week. SK Hynix's MARS being announced at US OCP, Microsoft's Maia flowing into Japanese datacenters, Foxconn showcasing both Vera Rubin and Vera-based ASUS servers from the same stage — these three events compounding in one week is not coincidence. Investors should stop asking "how much more can HBM re-price?" and start asking "who captures which layer of the inference-rack memory hierarchy?"
Key Sources: - Microsoft unveils Azure Maia 300, says chip will be its 'primary inference system' (DataCenterDynamics, 2026-10-13) - SK Hynix MARS Enables 5x Faster AI Inference; HBF Capacity Rises 14x (TrendForce, 2026-10-13) - Foxconn Deepens AI Rack Vertical Integration, Showcases Vera Rubin Platform at OCP 2026 (cnyes, 2026-10-13) - Nvidia Prepays KYEC to Lock In US AI Chip Testing Capacity (cnyes, 2026-10-13) - NTT to build $6.3bn data center for Anthropic in Japan (DataCenterDynamics, 2026-10-13) - Edge AI Computer AIE-KT78/68 Launches with NVIDIA Jetson Thor for Cloud-Free Robotics (PR TIMES, 2026-10-13) - plus 310 more
If this analysis was helpful · ☕ Support Us · ✈️ Telegram