AI Infrastructure & Industry Weekly Analysis — 2026-10-16: OCP standardization narrows the room for infrastructure differentiation, and in Korea small land parcels have entered the policy agenda

Briefings covered: 2026-10-13, 2026-10-16 news briefings

Layer key: ①Physical infrastructure ②Semiconductors and systems ③Cloud and compute supply ④Platforms and tooling ⑤Models ⑥Applications ⑦Demand and adoption ↔Policy and capital

This Week's Key Points

At OCP 2026, NVIDIA, AMD, Google and Meta signed the same facility specification (Open DC Spec v0.7), and power, cooling, rack and optical interfaces are converging on common standards (10/12). AMD Helion and Astera Labs Orion introduced "racks that take any chip," and Meta put Broadcom and NVIDIA silicon together in one switch chassis. In the same week, Palantir 'certified' Armada's containerized modules (10/1–2), and Flex made its integrated power, cooling and IT module business a candidate for a spin-off listing (10/12). Integrated modules assembled at the factory have become mainstream at large campuses too.

In Korea, small land parcels such as idle expressway sites entered the data center policy agenda for the first time. The government is reviewing about 20 abandoned expressway sections and idle rest-area sites as AI data center candidates (10/5). By contrast, the Gangneung 1GW plan has been blocked by its substation (completion of a 400MW substation at the end of 2028, 10/15). The Pohang AI data center closed a KRW 600 billion PF only after confirming its power supply and securing a pre-lease (10/12).

In capital markets, warnings about stranded AI infrastructure assets have started to be reflected in equipment stock prices. After Gartner warned of the risk of large-scale impairment of AI infrastructure assets (reported figures conflict: US$3.5 trillion by 2030 and US$4 trillion by 2032), infrastructure equipment stock Vertiv fell 6.09% in one day on 10/14. Investing in modular units generation by generation is drawing attention over building a large campus all at once in terms of capital efficiency.

1. Industry Implications

(1) Standards convergence narrows the room for infrastructure differentiation, and competition moves to semiconductors and software certification [①②]

Open DC Spec v0.7 targets 2027–28 systems and sets power architecture, cooling and layout as templates. It specifies rack heights up to 100 inches and loads up to 8,000lb. Add the compute-agnostic racks of AMD Helion (ORv3 specification, launched by Dell, HPE and Supermicro in November) and Astera Orion (1Q27), and the design principle of "standardizing interfaces on the assumption of generation turnover" becomes an industry default rather than a differentiator for a particular vendor.

In terms of industry structure, two things change. First, the room to differentiate with proprietary specifications in ① physical infrastructure and rack-level ② systems shrinks. The speed of proving standards compliance in documents (compliance matrices, certification) becomes more important than the design itself. Second, because hyperscalers standardized interfaces to avoid chip lock-in, competition moves up to ② semiconductors (which accelerator plugs into that rack) and ④ platforms (which software certifies that module). Meta's dual sourcing of Broadcom TH6C and NVIDIA Spectrum-6 in the same chassis is the same trend in networking ASICs.

(2) As 250kW per rack and 800 VDC become defaults, the unit of module design is changing [①②③]

If up to 250kW per rack under MGX 1.2 (per media reports; primary specification unverified) becomes the default, one container holds one or two GPU racks. NVIDIA's 800 VDC power rack is on the roadmap for 2H26, and 2MW per row for 2027. Flex AI Modules also ship an 800 VDC power rack and CDU as one bundle. The product unit of modular data centers is being redefined not as "rack count" but as "1–2 compute racks + 800 VDC power rack + CDU". At the other end is the AMD Helion enterprise rack (EPYC + up to 16 MI350P, air-cooling option). This is the gap in which air-cooled configurations survive at sites that cannot bring in coolant.

There is another signal in the ③ cloud and compute supply layer. The report that OpenAI returned to an NVIDIA configuration because of Cerebras capacity shortages (10/14) shows that the risk of alternative accelerators lies not in performance but in supply volume and delivery times. Supply diversification has increased as a design option, but actual procurement remains NVIDIA-centric.

(3) Grid recovery requirements and stranded-asset warnings intensify the capital-efficiency debate [① ↔Capital]

The OCP Grid Resilience Spec requires a return to normal load within two seconds after a grid disturbance. It is a demand that operators absorb the shocks large data centers impose on the grid, and grounds for battery energy storage systems (BESS) becoming a default specification rather than an option. Venture Global's frame contract for 16 power islands and transformers (10/9) is a signal that slots for large power equipment are being taken up even by non-power companies. Power equipment lead times remain the hardest bottleneck in layer ① this week as well.

On the capital side, Gartner's stranded-asset warning and Vertiv's sharp fall mean the market has started to price in the financial risk of "large campuses replaced wholesale with each GPU generation." A structure of "buying small and replacing module by module with each generation" gains persuasiveness on capital-efficiency logic. However, the Gartner original is unverified, so only the direction should be read, leaving out the figures.

(4) In Korea, GW-class plans are tied to substation schedules, while MW-class and smaller projects proceed on the existing distribution network [①③ ↔Policy]

Three cases point in the same direction. The review of about 20 abandoned expressway sections and idle rest-area sites (10/5) involves land with small parcels and limited grid connection, where large campuses cannot fit and only modular builds of a few MW or less are suitable. The Gangneung 1GW plan is tied to a 400MW substation (end of 2028) and has become a political issue (10/15). Pohang closed KRW 150 billion in equity and KRW 450 billion in PF (SC Capital investment) only after confirming 40MW of power intake and 32MW of IT and securing an anchor tenant on a pre-lease (10/12).

The implication is that the financial sector's screening criteria have been revealed. PF looks not at land but at confirmed power supply and an anchor tenant. The funding gap between mid-sized projects that meet these conditions and GW-class plans tied to substation schedules will widen for the time being. The idle-site plan, with scale, schedule and operators undecided, is at a typical early stage in which whoever presents a business structure first sets the criteria.

(5) Software certification is becoming a condition of sale for modular data centers [③④⑥]

Palantir's designation of Armada as its 'first certified modular data center partner' (10/1–2) shows a purchasing pattern in which public and defense customers buy hardware and certified operating software as a bundle. In this structure, the competitiveness of ③ module suppliers is determined not by specifications but by which ④⑥ software companies' certification lists they appear on. In markets with large sovereign AI demand (including Korea), mutual certification with Korean LLM and platform companies will play the same role.

2. What to Watch

• Timing of the 800 VDC transition — Why now: NVIDIA power rack in 2H26, standard on Flex AI Modules. Check: NVIDIA partner shipment notices in the second half and disclosure of the Flex Axiom spin-off (1Q27). Once power racks actually ship, AC/54V configurations in 2026 designs will be classified as transitional specifications.
• Open DC Spec v0.7 → 1.0 schedule and expansion of signatories — Check: revisions to OCP's public specification and additional hyperscaler signatories. The more signatories, the larger the discount on non-standard designs.
• Maximum IT kW with air cooling and delivery times for AMD Helion's November OEM launches — Check: product specifications from Dell, HPE and Supermicro. If air-cooled enterprise racks actually meet delivery dates, sites constrained on coolant get an option for the first time. Whether capacity shortages recur, as in the Cerebras case, is the counter-evidence.
• Whether operators will be solicited for the 20 idle expressway sites, and the scale per site — Check: MOLIT and Korea Expressway Corporation announcements. If a solicitation comes out, the first public reference market for Korean modular data centers opens.
• Gangneung substation schedule and replication of Pohang-type PF conditions — Check: KEPCO grid plans and PF disclosures for the next regional AI data center. If "confirmed power supply + pre-lease" repeats, it hardens into the standard condition for Korean data center finance.
• Original Gartner stranded-asset figures — Check: release of the original report. Once the figures are confirmed, they become the reference point for the equipment-stock valuation debate.

3. Data Reliability Notes

• Citable: Joint signing of Open DC Spec v0.7 (NVIDIA, AMD, Google, Meta; OCP public specification 10/12); Meta IceCube dual sourcing; AMD Helion November OEM launch; Astera Orion 1Q27; review of about 20 idle expressway sites (Korea Economic Daily 10/5, operators undecided); Gangneung 400MW substation at end of 2028 (10/15); Pohang KRW 600 billion PF structure (10/12); Vertiv -6.09% on 10/14.
• Marked as company claim: Flex's '50% shorter deployment time'; MGX 1.2 at 250kW per rack (primary specification unverified).
• Unverified / Use with caution: Gartner impairment figures (conflicting across reports, not checked against the original); mention of Haengdamdo as a candidate site (single source); the specific scale of the Cerebras capacity shortage.

*This article is industry analysis based on public information and is not a recommendation to buy or sell any security.*