AI Infrastructure & Industry Weekly Analysis — 2026-11-06: Cloud revenue and enterprise AI spending kept growing as US midterm votes went against data centers and AI chip stocks fell sharply

Briefings covered: 2026-11-06 news briefing

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; this week's changes are assessed by the layer they affect.

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

This Week's Key Points

In the November 3 US midterm elections, data center ballot measures held in 10 states all ended in favor of a ban or moratorium on development, except one in Harwood, North Dakota. Ohio accounted for 21 of them, more than half. By Fast Company's count, 28 of the 38 measures with results showed more than 70% support for a ban or moratorium, and anti-data-center advertising exceeded US$173 million. An analysis also found that 34 of 36 governor candidates who won or are leading support at least one data center regulation (The Information, secondhand citation). Counting is still under way, so tallies differ by outlet, and the final result of the Texas governor race is unverified.

On November 5, two days after the election, AI chip stocks fell together. AMD reported on November 3 that third-quarter data center revenue rose 97%, but its stock fell about 13% on November 5, while NVIDIA fell 6–8.5% (conflicting across outlets). Reports cited concern about tighter data center regulation and delays in China's approval of NVIDIA H200 imports as causes.

On November 2 OpenAI announced "Grid Protectors," under which its self-developed campuses will reduce power use when the local grid is under strain. The pilot site, a campus in Effingham County, Georgia, will provide up to 1,000 MW of flexible demand, and OpenAI will cover the full cost of infrastructure and power supply. On November 5 reports followed that OpenAI had released a "National Community Compact" covering community agreements and no use of non-disclosure agreements.

Demand indicators continued to point upward this week. AWS revenue rose 39% to US$46 billion, and Amazon kept its 2026 capital expenditure at about US$220 billion while saying 2027 would "increase meaningfully." Menlo Ventures put annualized US enterprise AI spending as of September 2026 at US$176.8 billion, 5.4 times the US$33 billion of a year earlier.

Correction to last week's view: Last week's analysis treated the power and permitting bottleneck in the US and Korea mainly as a matter of technical and regulatory requirements, such as the NERC large-load standards and grid codes. This week's ballot measures and governor races show that site acquisition in the US is also a matter of local politics and resident consent, so the schedule risk for new US capacity is now judged higher than last week. In addition, the enforcement decree of Korea's AI data center special act, which last week's analysis expected to be pre-announced by the end of the month, is unverified as of this week's briefing.

1. Industry Implications

(1) With data center sites and permitting now an election issue in the US, the risk for existing capacity and for new projects has diverged [① ↔Policy]

Most ballot measures with results went toward a ban or moratorium, and most winning or leading governor candidates support regulation. Most of the votes were local measures at the county or township level, so they do not directly stop construction across the US. But anti-data-center advertising of more than US$173 million indicates that data centers were treated as an election issue over electricity bills and local burdens. If state governments move to moratoriums or tighter permitting, the effect will be broader than that of the local votes.

Within the ① physical infrastructure layer, outcomes depend on position. Existing capacity held by operators that have already broken ground or signed grid interconnection agreements becomes scarcer as new supply is delayed. By contrast, developers who secure land first and then seek permits, and projects still in planning, face schedule delays and higher costs. This is a reasonable inference, and the point to check is what measures the new governors actually take after taking office. Additional votes are scheduled in Arkansas and North Dakota in December.

(2) Model labs have started to bear campus power costs and demand-response obligations directly [①⑤ ↔Policy]

OpenAI committed to covering the full infrastructure and power supply cost of its Georgia campus and to providing up to 1,000 MW of flexible demand. The size, speed and duration of curtailment will be agreed individually with the utility. Given that the announcement came just before the election and was followed by plans for community agreements, it largely reads as a response to the backlash over electricity bills. The total campus size of 3.2 GW is reported only by some outlets, so the share of campus load that 1,000 MW represents cannot be determined.

If this approach spreads, it has two effects. First, ESS, on-site generation and power control equipment that allow load to be reduced and restored become part of the baseline campus specification, raising build cost per MW. This is the same direction of cost increase as the NERC large computational load standards covered last week. Second, halting computation during grid stress lowers a campus's effective utilization. Training jobs are relatively easy to reschedule, but latency-sensitive inference services must shift to capacity in other regions, which favors operators with capacity in multiple regions. Whether the structure in which a ⑤ model-layer company absorbs ①-layer costs will extend to other hyperscalers has not yet been confirmed.

(3) Results in the semiconductor layer improved, but share prices first reflected the political risk to the pace of US data center construction [② ↔Capital]

AMD reported third-quarter revenue of US$13.4 billion, a gross margin of 54% and net income of US$2.4 billion, with data center revenue up 97%. CEO Lisa Su projected revenue growth of more than 80% in 2027, but this comes from a single source report. That the stock fell about 13% two days later despite strong results and guidance suggests investors priced in the completion schedule of the data centers that will house the accelerators as a risk, more than accelerator demand itself. Benzinga put the decline in US tech stock market capitalization on November 5 at US$1.7 trillion.

If the regulation in (1) translates into actual construction delays, the bottleneck for the ② layer shifts from chip supply to installation space. In that case shipment schedules would be adjusted to customers' completion schedules, and hyperscalers may allocate more capacity outside the US. The volume affected by China's delay in approving H200 imports is unverified, so its contribution to this week's decline cannot be separated out.

In Korea, AI-related stocks also fell on November 4–5, but the domestic semiconductor industry judged that AI-driven memory demand will hold (Korea Economic Daily, Kyunghyang Shinmun). As noted last week, the companies have said HBM volume is contracted through 2027, so if US construction is delayed, the first effect is more likely to fall on the timing of general-purpose server memory demand than on HBM. At a Federation of Korean Industries meeting on November 2, industry asked for procurement rules to be revised so that domestic NPUs can bid on public contracts, and for more opportunities for demonstration projects. After two domestic NPU companies secured mass-production funding last week, this request confirms that securing public demand is the remaining task.

(4) With Amazon's cloud growth also accelerating, all three hyperscalers now show grounds for recovering their investment [①②③ ↔Capital]

AWS revenue was US$46 billion (+39%) and operating income US$17.8 billion, and Amazon kept 2026 cash capital expenditure at about US$220 billion. This was the third hyperscaler checkpoint listed in last week's What to Watch, and the result points in the same direction as Microsoft (Azure +48%) and Alphabet (Google Cloud +103%). All three have signaled higher capital expenditure in 2027, so the demand outlook for ① physical infrastructure and ② semiconductors holds. CEO Andy Jassy's projection of more than 40% AWS growth in 2027 and 2028 is a company forecast.

Some figures are not settled. Third-quarter capital expenditure is reported as US$75.9 billion or US$77.3 billion depending on the basis, and the US$578 billion backlog is a report of an earnings-call remark that does not appear in the earnings materials. No figure was given for the size of the 2027 increase. What this week adds is the question of where the money is spent. The three companies' willingness to spend is confirmed, but if the regulation in (1) spreads, the same amount could move to other US states or overseas, changing which regions benefit at the ① layer and on what schedule.

(5) Growth in enterprise AI spending is coming from higher usage by existing adopters and is concentrated in the top three model labs [⑤⑥⑦]

In the Menlo Ventures survey, Anthropic, OpenAI and Google together accounted for 69% (US$122.7 billion) of US enterprise AI spending. The share of companies that have adopted AI rose only from 55% to 65%, but spending per company rose about fivefold. This means growth in the ⑦ demand and adoption layer comes from existing customers' usage rather than new customers, consistent with last week's growth in Copilot customers on usage-based billing. The survey covered about 400 US enterprise decision-makers, so the spending totals should be treated as estimates.

By category, coding grew 7.7 times to US$34 billion, monetizing faster than workplace chatbots (US$25 billion, 3.0 times). If spending concentrates in the top three ⑤ model labs, mid-tier model providers and general-purpose ⑥ application providers with weak differentiation are at a disadvantage. Last week's analysis noted that Anthropic's long-term compute obligations depend on future revenue growth; this survey indicates that enterprise spending, the source of that revenue, is concentrating in the top labs. Because the figures are survey-based, they need to be checked again against the quarterly revenue disclosures of model labs heading to listing.

2. What to Watch

• The new governors' first data center measures — Why now: most winning or leading candidates support regulation, and the Texas result is unverified. Check: state executive orders, moratoriums and state legislative bills after January inaugurations, and the final Texas count. If state-level moratoriums or tighter permitting appear, the schedule risk for new US capacity is judged higher; if measures stay local, the effect is confined to the counties concerned.
• Additional ballot measures in Arkansas and North Dakota in December — Check: county election results. If opposition prevails again, the opposition is treated as a sustained trend rather than a one-off election effect.
• Demand-response and cost-bearing commitments by other hyperscalers — Why now: OpenAI made the first commitment to bear full costs and provide flexible demand. Check: hyperscaler announcements and the Georgia Public Service Commission's conditions for rate approval. If similar commitments follow, higher build cost per MW and lower effective utilization are reflected as industry-wide costs.
• AMD MI450 shipment volumes and hyperscaler capacity allocation outside the US — Why now: chip results and share prices moved in opposite directions. Check: AMD shipment announcements and hyperscaler new region and campus announcements. If shipments proceed as planned and overseas allocation rises, the view that the bottleneck has moved from chips to sites is strengthened.
• Amazon's 2027 capital expenditure figure and backlog in its fourth-quarter results — Check: fourth-quarter earnings materials. If the size of the 2027 increase and the backlog appear in official materials, this week's use-with-caution figures become confirmed figures.
• Quarterly revenue of model labs heading to listing, and coding agent pricing — Why now: a survey found enterprise spending concentrating in the top three labs and in coding. Check: Anthropic's amended S-1 and post-listing quarterly disclosures, and coding agent pricing change notices. If disclosed revenue growth matches the survey trend, the top-lab concentration view holds; if coding agent prices fall sharply, the view on the pace of monetization is lowered.
• The pre-announced enforcement decree of Korea's AI data center special act and NPU procurement rules for public AI projects — Why now: the timing of the pre-announcement is unverified, and industry has asked for procurement rule changes. Check: the Ministry of Science and ICT's pre-announcement notice (provisions on domestic equipment and semiconductors) and Public Procurement Service and public AI project notices. If provisions on domestic semiconductors or revised procurement rules are included, the view of public demand for domestic NPUs is raised.

3. Data Reliability Notes

• Citable: that the ballot measures ended in favor of a ban or moratorium except the one in Harwood, and Ohio's 21 measures (Ballotpedia and CNBC tallies, multiple outlets); AMD third-quarter revenue of US$13.4 billion, 54% gross margin and US$2.4 billion net income (SEC 8-K); OpenAI Grid Protectors' up to 1,000 MW of flexible demand in Georgia and full cost coverage (company announcement, multiple outlets); AWS revenue of US$46 billion (+39%), operating income of US$17.8 billion, Amazon revenue of US$201.9 billion (+12%) and 2026 capital expenditure kept at about US$220 billion (earnings release, multiple outlets); Menlo Ventures' annualized US$176.8 billion, 5.4 times and 69% share for three companies (survey original).
• Use with caution: the number of ballot measures (counting under way, varies by outlet); 28 of 38 measures with more than 70% support and anti-data-center advertising above US$173 million (Fast Company tally); 34 of 36 governor candidates supporting regulation (secondhand citation of The Information's analysis); the Texas governor result (unverified); AMD data center revenue of US$8.5 billion (reported) and the 2027 outlook of more than 80% growth (single source); NVIDIA's 6–8.5% decline (conflicting); the US$1.7 trillion decline in tech stock market capitalization (Benzinga tally); the volume affected by the H200 import approval delay (unverified); the 3.2 GW Georgia campus size (some outlets); the National Community Compact (reported); Amazon third-quarter capital expenditure of US$75.9 billion or US$77.3 billion (conflicting); the US$578 billion backlog (report of an earnings-call remark, not in earnings materials); more than 40% AWS growth in 2027 and 2028 (company forecast); Menlo's category spending and roughly fivefold spending per company (estimates based on a survey of about 400 respondents); the domestic view that memory demand will hold (industry opinion as reported); whether the AI data center special act enforcement decree has been pre-announced (unverified).

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