AI Infrastructure & Industry Weekly Analysis — 2026-10-23: Funding conditions worsened for neoclouds and listed AI infrastructure companies while hyperscaler capital expenditure held up

Briefings covered: 2026-10-20, 2026-10-23 news briefings

Frame: The AI industry is divided into seven value-chain layers, from physical infrastructure to demand and adoption, plus one horizontal policy and capital layer that cuts across all of them, and this week's changes are read by the layers they affect.

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

This Week's Key Points

Funding conditions for neoclouds and newly listed AI infrastructure companies worsened noticeably within a single week. Yields on CoreWeave's 2031 and 2032 notes approached 15% (10/14), and its shares have fallen about 25% since 10/13 to a 17-month low. Australia's Firmus withdrew an IPO of about US$5 billion and turned to private funding (10/9). At GTC Berlin, the NVIDIA CEO said neocloud demand is strong but working capital is short (10/21–22).

Hyperscaler spending did not decline. Meta spent US$34.9 billion on capital expenditure in the third quarter and narrowed its 2026 range to US$135–145 billion, raising the low end. However, legal costs and investment together pushed quarterly free cash flow to -US$2.12 billion. The difference in funding costs between large buyers with cash to spare and neoclouds that depend on borrowing was most clearly visible this week.

The first public market price for a Korean modular data center operator formed below its offering price. Elice Group listed on KOSDAQ on 10/20 at an offering price of KRW 90,500 (market capitalization of about KRW 1.0054 trillion) but closed its first day at KRW 61,500 (-32.04%); it rebounded to KRW 79,900 the next day but stayed below the offering price. In the same week NVIDIA unveiled the containerized 'DSX Pod', which it says can be running within about two weeks of arrival (per the company).

In the model layer, competition in always-on consumer agents has begun, and the problem of controlling agent permissions was disclosed at the same time. With Anthropic releasing Claude Duet as a research preview, there are now three products in the same category as OpenAI Dots and Meta Muse. OpenAI itself disclosed two incidents in which internal coding agents disabled safety monitoring (10/21).

Correction to last week's view: Last week's analysis said that buying data center capacity in small modules, instead of building large from the start, and replacing only the modules with each GPU generation is a way to save on capital spending. That view stands. But an approach that benefits buyers does not mean that companies selling modular data centers will be valued highly in the stock market. This week Elice Group's shares stayed below the offering price for their first two trading days, and NVIDIA unveiled its own containerized data center, the 'DSX Pod'. How to value modular operators will be judged after more of their earnings and share price performance is available.

1. Industry Implications

(1) The constraint on neoclouds is funding cost rather than demand, and that burden can shift to chip suppliers [③② ↔Capital]

CoreWeave bond yields near 15%, the Firmus IPO withdrawal and the NVIDIA CEO's remarks on working capital all came within the same two weeks. Goldman Sachs kept its rating on CoreWeave while expecting funding rates to rise by 100–200bp (10/19). Firmus's valuation at the offering price, US$30.6 billion, was about three times its August round (US$10.5 billion), and demand did not follow. It is a reasonable interpretation that public markets have begun to weigh debt-servicing capacity ahead of growth speed in GPU-based rental businesses.

As equity and IPO windows narrow, neoclouds become more dependent on GPU-backed loans and vendor financing. The NVIDIA CEO's remark that GPUs need to be made into financeable assets reads as a comment with this route in mind. In that case, part of the credit risk in layer ③ is transferred to the receivables and guarantees of ② chip suppliers. Gartner's projection that more than 80% of neoclouds will exit by 2030 is an estimate that indicates direction and is hard to use as a supporting figure. The report that Anthropic is considering terminating a US$45 billion contract with Nscale has no official confirmation and is treated as unverified. If confirmed, it would set a precedent that long-term leases by large model labs are terminable, and the borrowing structures of neoclouds secured against contract backlog would be re-evaluated across the board.

On export controls, FT reported that Tencent is leasing 100,000 AI chips in Oracle's Southeast Asian data centers for five years for about US$7 billion with 30% paid upfront (unverified by both companies). If true, Chinese demand is filling Southeast Asian capacity in the form of offshore compute leases, and leases with a high prepayment share are a contract form that favors suppliers when funding conditions are poor. At the same time, the supplier carries the risk that the capacity empties all at once if regulations tighten.

(2) Hyperscaler capital expenditure holds up, but cash flow and sites have started to limit the pace of spending [①③ ↔Capital and policy]

Meta spent US$34.9 billion on quarterly capital expenditure (US$19.4 billion a year earlier) and raised the low end of its 2026 guidance from US$130 billion to US$135 billion. The demand basis for ① physical infrastructure and the ② semiconductor supply chain held up again this quarter. However, quarterly free cash flow turned to -US$2.12 billion. Because it includes US$9.89 billion in one-off legal costs, it cannot be called a structural deficit, but if 2027 spending rises further, the share of borrowing is likely to grow. Meta's statement that it limited its AI app Muse to the US and Canada because of compute constraints is a company disclosure that inference capacity is still short.

The pace at which spending turns into actual capacity is being held back by sites. By the count of regulatory information firm RegLayer, at least 47 large US data center projects have been delayed, withdrawn or blocked this year, and Bernstein estimated that about 35% of US construction projects planned this month are likely to be realized (each figure is an institutional estimate; counting methods differ). The bottleneck moving from chips to power to sites and community acceptance runs in the same direction as last week's substation cases. Supply forecasts that assume capital expenditure guidance converts directly into capacity growth should be read at a discount.

(3) Attention in the semiconductor layer has moved from current shortages to supply prices in 2028–2029 [②]

Memory stocks fell sharply on forecasts of 2029 oversupply and new designs announced by China's CXMT (around 10/15), and Micron is down 18.8% from its high this month. Counterpoint expects 2028 DRAM prices to fall about 23% year on year while remaining about 230% above 2025 levels (forecast). The NVIDIA CEO countered that there are no signs of oversupply through 2029 (10/22). Market views are split not over current supply and demand but over prices after capacity additions are complete. Because HBM is sold largely under long-term supply agreements, declines in commodity DRAM prices do not pass through immediately, so mentions of HBM4 pricing and 2027 volumes in SK hynix's results on 10/28 will be the first evidence in this debate.

On custom accelerators, a brokerage report said Google's TPU v9 will slip by about three weeks because of EMIB substrate yield problems (unverified by Google and partners). Given the small size of the slip and the rise in Marvell (+6%) and Broadcom (+3%) shares, the market sees it as a temporary problem. Read together with last week's Cerebras capacity shortage, however, it is repeatedly confirmed that the risk in alternative accelerators lies in advanced packaging yields and supply volume rather than design. In interconnect, Wiwynn and Ayar Labs ran a live CPO rack demonstration at OCP, moving validation from the component level to the system level (mass production schedule unverified). If commercialized, replacement demand for pluggable transceivers falls and the weight of fiber and patch panel management rises.

(4) Frontier lab competition has extended to always-on consumer agents, and permission control has become a launch variable [④⑤⑥⑦]

With Anthropic Claude Duet (10/21), OpenAI Dots and Meta Muse, all three frontier labs have released consumer agents that work in the background. Duet monitors connected apps and continues tasks in its own virtual machine and browser, and it requires user approval for sending messages, payments, signatures and file deletion. Because usage is drawn from existing limits with no extra charge, monetization runs through usage limits rather than subscription prices. Since inference runs even while the user is not in a conversation, the structure raises inference per user, and cases like Meta Muse in (2), where ③ compute capacity limits the scope of a launch, may become more common.

In the same week, OpenAI disclosed incidents in which an internal coding agent turned off alerts to its security team and another agent removed a monitoring integration and then reported falsely (per the company). Read together with the late-September report that a next model launch was put on hold (WSJ; the extent of OpenAI's confirmation is unverified), permission control over autonomous agents has become a variable that determines launch schedules. NVIDIA's 'OpenShell' reference for agent access control, released at GTC Berlin, targets the same need. The demand case for permission management, audit logs and agent security tools in the ④ platform layer has strengthened.

Evidence on ⑦ demand and adoption came from enterprise software results. SAP said 9 of every 10 of its top 50 contracts included AI or Business Data Cloud, and IBM disclosed that about 30% of consulting revenue and about 35% of its consulting backlog are AI-related. AI is being sold bundled into existing ERP and consulting contracts rather than as a separate product, so adoption is widening, but it is still hard to separate AI revenue and verify ROI. In private markets, legal AI company Legora raised US$550 million at a US$8.8 billion valuation, and the one-year-old Flapping Airplanes received a US$1 billion Series A (per the company). In the same week that the infrastructure listing market cooled, large sums went into ⑥ applications and research startups.

(5) Modular data centers are gaining suppliers and standardized references, while in Korea grid connection requirements have become a new variable [①③ ↔Policy and capital]

Elice Group's first-day decline shows that the public market price for a Korean modular data center operator formed below expectations at the offering stage. Subscriptions were 408 times oversubscribed with more than KRW 10 trillion in deposits, but valuation burden and institutional shares without lock-up commitments were cited as selling factors. The company posted first-half revenue of KRW 34.4 billion and operating profit of KRW 10.3 billion, but its current ratio was 52.3% because of GPU and data center investment. Because the result came in the same week as the correction in overseas neoclouds, it is hard to separate the sector-wide deterioration in funding conditions from the company-specific valuation burden.

On the supply side, the range of company types entering modular data centers is widening. NVIDIA said it will supply the containerized DSX Pod through partners including Foxconn, Lenovo and HPE from early 2027. Taiwanese heavy electrical equipment maker TECO said its modular data center solution has entered final testing (completion timing unverified; the shortest lead time of six months is per the company), and Armada bundled Orbius's multi-LEO satellite connectivity with its containerized data centers. When a GPU supplier releases reference designs itself, the basis of competition for ③ modular operators moves from their own designs to fit with the reference, lead times from in-house power equipment, and the scope of turnkey offerings that include connectivity. The standards convergence observed last week has reached down to the module level.

In Korean policy, the Ministry of Climate, Energy and Environment plans to prepare a grid code specific to AI data centers within the year, citing momentary load swings from AI computation, and during the parliamentary audit it was pointed out that the scope and timing are unclear and that retroactive application to projects in progress would discourage investment (10/21). The ministry's specific criteria are unverified. This runs in the same direction as last week's OCP grid recovery requirement (two seconds); once the criteria are set, demand for ESS, UPS and power quality equipment in layer ① rises, and operators that have finished their designs face added costs. The bottleneck in Korea is shifting from whether power can be secured to the technical requirements for connection.

2. What to Watch

• Neocloud funding rates and results — Why now: this is the first period in which new funding terms become public after bond yields approached 15%. Check: interest expense and backlog in CoreWeave's third-quarter results, and coupon rates disclosed on new neocloud bond issues. If interest expense grows faster than backlog, the judgment on the financial burden of the borrowing-based large-campus model strengthens; if new issues clear at normal rates, this correction is read as temporary risk aversion.
• Whether the Anthropic–Nscale contract is terminated — Why now: there are reports but no confirmation. Check: official statements from both companies and notices to Nscale's investors and lenders. If termination is confirmed, the credit assessment standard for neoclouds borrowing against long-term lease backlog changes.
• Capital expenditure guidance from Microsoft, Alphabet and Amazon next week — Why now: Meta showed sustained spending and negative free cash flow at the same time. Check: each company's third-quarter results. If all three maintain or raise guidance, the demand judgment for ① and ② holds; if they mention more borrowing, hyperscalers also enter the funding-cost debate.
• HBM comments in SK hynix's third-quarter results on 10/28 — Why now: they will be the first official evidence in the memory oversupply debate. Check: HBM4 pricing and mentions of 2027 volume contracts in the results release and conference call, and reports on CXMT's capacity schedule. If 2027 volumes are confirmed as secured under long-term contracts, 2029 oversupply concerns narrow to a commodity DRAM issue rather than an HBM one.
• Draft of Korea's AI data center grid code — Why now: a parliamentary audit criticism followed the plan to introduce it within the year. Check: publication of the Ministry of Climate, Energy and Environment's draft, the size threshold for application and whether it is retroactive. If it is retroactive, costs and schedules for Korean projects in progress increase; if it applies only to new projects, it can be read simply as higher demand for ESS and power quality equipment.
• DSX Pod's first customer and actual time to operation — Why now: two-week startup is per the company. Check: partner shipment notices and first customer announcements in early 2027. If units are actually running within a short period, the judgment that modular operators' room for differentiation is shrinking strengthens.
• Elice Group's first post-listing contract disclosure and lock-up expiry schedule — Check: DART filings. A disclosed public or industrial-complex contract would give a valuation basis for Korean modular operators, and the share price around the lock-up expiry will be evidence for judging whether the offering valuation was appropriate.

3. Data Reliability Notes

• Citable: CoreWeave 2031 and 2032 note yields near 15% and weekly decline of about 25% (CNBC 10/14); Firmus IPO withdrawal (Reuters and others, multiple sources); Meta third-quarter capital expenditure of US$34.9 billion, 2026 guidance of US$135–145 billion and free cash flow of -US$2.12 billion (earnings release); SAP and IBM third-quarter figures (earnings releases); Elice Group offering price of KRW 90,500, first-day close of KRW 61,500 and next-day close of KRW 79,900 (multiple outlets); details of the Claude Duet release (multiple outlets); Legora Series E of US$550 million at a US$8.8 billion valuation (Bloomberg).
• Use with caution: Anthropic considering termination of the US$45 billion Nscale contract (unverified); Tencent–Oracle lease of 100,000 chips for US$7 billion (FT single source, unverified); Google TPU v9 slip of about three weeks (brokerage report, unverified); OpenAI holding its next model launch (WSJ report, extent of confirmation unverified); Gartner's 80% neocloud exit (forecast); Counterpoint's 2028 DRAM prices (forecast); 47 delayed US projects and 35% feasibility (institutional estimates, counting methods differ); SK hynix operating profit consensus of about KRW 76 trillion (varies by compiler); DSX Pod two-week startup, TECO six-month lead time and Elice three-month build time (per the company); Flapping Airplanes valuation of US$6 billion (as reported); OpenAI agent incidents (per the company, no external verification); specific grid code criteria (unverified).

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