# The State of the AI Market in Wartime: Chips, Power and the Defence AI Companies AI capacity is now bought by the gigawatt and used to plan a war. Power, memory and packaging set the pace, and the defence money has gone to Palantir, SpaceX, OpenAI and the autonomy makers. Canonical HTML: https://www.livetradingnews.com/the-state-of-the-ai-market-in-wartime-chips-power-and-the-defence-ai-companies Last modified: 2026-10-28 --- By Shayne Heffernan · 2026-10-28 Tags: $PLTR, $NVDA, $SPCX, $GOOGL, $AMD, $AVGO, $MSFT, $AMZN, $META, $MU, $AVAV, $ORCL, $TSM, $GEV, $LMT, $RTX, $NOC, $LHX, defence AI, Iran war Signed: ML-DSA-87, anchored on Armature L1. Nothing in this article is investment advice. The AI market in late 2026 is no longer a story about whether demand is real. It is a story about how fast physical capacity can be built to meet it, and this year it became a story about war. In the 38 days of Operation Epic Fury, the campaign against Iran that began on 28 February, US forces used Palantir's Maven Smart System for strike missions against [13,000 targets](https://breakingdefense.com/2026/05/insatiable-appetite-for-ai-maven-usage-surged-for-strikes-on-iran-pentagon-ai-chief-says/), according to the Pentagon's chief digital and AI officer. The same chips, memory and power now sell by the gigawatt: the four largest hyperscalers have called for about [$730 billion of capital spending](https://www.cnbc.com/2026/10/21/meta-earnings-q3-2026-live-updates.html) this year. The binding constraints have moved down the stack: from model capability to accelerators, from accelerators to high-bandwidth memory (HBM) and advanced packaging, and from packaging to electricity, transformers and grid interconnection. The war has put a price on the last of those. Brent crude settled at [$113.88 on 27 October 2026](https://www.nationthailand.com/news/world/40072812), and European gas touched [€90/MWh on 20 October](https://gmk.center/en/news/european-gas-prices-reached-e90-mwh-in-october/), about [three times its level before the war](https://www.france-epargne.fr/news/gaz-le-ttf-depasse-90-euros-le-mwh-une-premiere-depuis-decembre-2022). Three dynamics define the market. First, demand has broadened. Training frontier models still absorbs enormous compute, but inference, the everyday work of serving models to people and to software agents, has become the volume driver, and enterprises are moving from pilots to production with budgets attached. Second, pricing has split. Leading-edge accelerators and rack-scale systems command premium prices, while the cost of delivering a unit of inference keeps falling. Third, supply is concentrated. NVIDIA remains the clear leader, AMD is a credible second source, and the hyperscalers are scaling custom silicon with partners such as Broadcom, all of it funnelled through a small number of TSMC fabs and packaging lines. The share prices already show where the money in this war has gone. From the last close before it, on 27 February 2026, to 27 October 2026, Palantir, $PLTR (NASDAQ), rose [44.9%](https://query1.finance.yahoo.com/v8/finance/chart/PLTR?period1=1793059200&period2=1793145600&interval=1d) and NVIDIA, $NVDA (NASDAQ), [23.0%](https://query1.finance.yahoo.com/v8/finance/chart/NVDA?period1=1793059200&period2=1793145600&interval=1d), while Northrop Grumman, $NOC (NYSE), fell [23.9%](https://query1.finance.yahoo.com/v8/finance/chart/NOC?period1=1793059200&period2=1793145600&interval=1d) and L3Harris, $LHX (NYSE), [28.8%](https://query1.finance.yahoo.com/v8/finance/chart/LHX?period1=1793059200&period2=1793145600&interval=1d), per Yahoo Finance closes. The market is paying for the software and compute layers of warfare, and the platform makers have not shared in it. The next US decision on Iran is [most likely to come after the midterms](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump) on 3 November, US officials told CNN, which reported on 26 October that US forces could launch new operations [within 96 hours of an order](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump) from President Trump. US officials also told the Wall Street Journal that a [third carrier strike group is being sent](https://stripes.com/branches/navy/2026-10-02/makin-island-middle-east-iran-strikes-november-23032862.html) to the region, with arrivals expected by the end of November. For investors the read is the same in peace and in war. Returns will go, unevenly, to whoever controls the scarce inputs: power, memory, packaging and distribution. In defence they go to the companies inside the military's AI loop: Palantir; SpaceX, $SPCX (NASDAQ), which owns xAI; OpenAI; Alphabet, $GOOGL (NASDAQ), whose Gemini was the first model on the Pentagon's GenAI.mil; and the private autonomy makers Anduril and Shield AI. Anthropic, whose Claude was inside the loop when the war began, is [fighting a Pentagon supply-chain risk designation](https://www.cnbc.com/2026/09/25/pentagon-anthropic-ai-risk-appeals-court.html) in court as it prepares to list. *"The binding constraint on AI is no longer the model. It is the transformer on the substation, the memory stack on the package, and the permit on the desk." Shayne Heffernan* ## Demand is now quoted in gigawatts The clearest signal of demand is the unit in which it is quoted. OpenAI alone has announced 26 GW of accelerator capacity across three suppliers: a [10 GW letter of intent with NVIDIA](https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems) in September 2025, a [6 GW agreement with AMD](https://www.amd.com/en/newsroom/press-releases/2025-10-6-amd-and-openai-announce-strategic-partnership-to-d.html), $AMD (NASDAQ), that came with a warrant for up to 160 million AMD shares, and [10 GW of its own accelerators with Broadcom](https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration/), $AVGO (NASDAQ), to be complete by the end of 2029. In February NVIDIA replaced its milestone investment with a [$30 billion equity stake](https://www.techtimes.com/articles/317839/20260605/nvidia-openai-investment-shrinks-100b-30b-compute-lock-war-continues.htm) in OpenAI's $110 billion funding round. Comparable campus plans have been disclosed by every large buyer. Meta, $META (NASDAQ), is building its Hyperion campus in Louisiana to bring [2 GW online by 2030 and scale to 5 GW](https://techcrunch.com/2025/07/14/mark-zuckerberg-says-meta-is-building-a-5gw-ai-data-center/). Epoch AI, which tracks the sites from satellite imagery and filings, projects Fairwater, the Wisconsin campus of Microsoft, $MSFT (NASDAQ), at [about 2.3 GW of IT power by mid-2028](https://epoch.ai/data/ai-data-centers/directory/microsoft-fairwater-wisconsin). It puts Project Rainier, which Amazon, $AMZN (NASDAQ), built in Indiana for Anthropic, at [about 1 GW already operating](https://epoch.ai/data/ai-data-centers/directory/anthropic-amazon-new-carlisle). OpenAI's Stargate programme, built with Oracle, $ORCL (NYSE), and partners including CoreWeave, reached [nearly 7 GW of planned capacity](https://openai.com/index/five-new-stargate-sites/) in September 2025, xAI's Colossus 2 in Memphis runs at an estimated [946 MW](https://epoch.ai/data/ai-data-centers/directory/colossus-2), and Google's $15 billion hub in Visakhapatnam includes a [gigawatt-scale campus](https://blog.google/intl/en-in/company-news/our-first-ai-hub-in-india-powered-by-a-15-billion-investment/). A gigawatt-class campus implies hundreds of thousands of accelerators, dedicated generation or long-term power contracts, and several years of construction. Hyperscaler capital expenditure tells the same story at a larger scale. The four largest hyperscalers had called for roughly [$730 billion of capital spending in 2026](https://www.cnbc.com/2026/10/21/meta-earnings-q3-2026-live-updates.html) by 21 October, per CNBC, and analysts polled by FactSet expect about $1 trillion next year. Guidance has gone one way. Alphabet raised its 2026 range to [$195 to $205 billion](https://w.media/ai-infrastructure-demand-pushes-alphabets-2026-capex-guidance-to-us-205-billion/) in July, Amazon raised its plan to [about $220 billion](https://fortune.com/2026/07/30/andy-jassy-amazon-capex-demand-aws-pga-tour/) and said it still would not have enough capacity to meet 2026 demand, and Meta lifted the floor of its range to [$135 to $145 billion](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx) on 21 October. Alphabet and Microsoft report today and Amazon tomorrow. Beneath the hyperscalers, enterprise behaviour has changed. Open-ended experimentation is giving way to fewer, larger deployments tied to measurable outcomes, led by customer service, software development, document processing, claims handling and internal knowledge retrieval. In Madrona's August survey of 150 senior enterprise decision-makers, [74% plan to expand AI budgets](https://www.madrona.com/wp-content/uploads/2026/08/Madrona-Research-Harnessing-Enterprise-Value-ROI-of-AI.pdf) over the next 12 months and 45% already have a dedicated AI budget line, which is a sign of durability. The conversion has room to run: [83% still turn fewer than half of their pilots into production](https://www.madrona.com/wp-content/uploads/2026/08/Madrona-Research-Harnessing-Enterprise-Value-ROI-of-AI.pdf). The most important demand trend is inference. Training a frontier model is a large but episodic event. Serving that model to hundreds of millions of people, and increasingly to software agents that call models again and again to finish multistep tasks, is continuous and compounding. Reasoning models spend more compute at inference time to produce better answers, and agentic workflows multiply the queries per task. Google said in May that it processes [over 3.2 quadrillion tokens a month](https://blog.google/innovation-and-ai/sundar-pichai-io-2026/), about seven times the figure a year earlier. Inference demand can grow faster than user counts. Alongside its February funding round, OpenAI said it would use [3 GW of dedicated inference capacity](https://www.datacenterdynamics.com/en/news/openai-raises-110bn-with-50bn-from-amazon-and-30bn-from-both-nvidia-and-softbank/) and 2 GW of training capacity on NVIDIA's Vera Rubin systems. It is inference that will fill most of the gigawatts now under construction. Sovereign demand adds a further layer, and in the Gulf it now sits inside a war. Saudi Arabia's HUMAIN, owned by the Public Investment Fund, anchored the largest commitments at LEAP in September, including a [250 MW phase with AMD and Cisco from 2027](https://www.globaldatacenterhub.com/p/saudi-arabias-leap-2026-15-billion). The first 200 MW phase of Stargate UAE, inside a planned 5 GW campus, had a third-quarter target and [no published operating capacity by 12 September](https://thegulftape.com/articles/stargate-uae). The physical risk is no longer theoretical. Iranian drones [directly struck two AWS data centres in the UAE](https://www.cnbc.com/2026/03/02/amazon-says-drone-strikes-damaged-3-facilities-in-uae-and-bahrain.html) in March and damaged a site in Bahrain, and on 15 September AWS said it [could not restore resources hosted only in its Bahrain region](https://www.cnbc.com/2026/09/15/aws-cant-restore-service-to-bahrain-uae-6-months-after-iran-strikes.html) or in one UAE zone. *"Training is an event. Inference is a utility bill, and utility bills compound." Shayne Heffernan* ## Prices have split between hardware and tokens Pricing has split, and the split is essential to reading company results. At the top of the stack, leading-edge accelerators and complete rack-scale systems sell at premium prices. Demand for each new generation exceeds supply at launch and lead times stretch. NVIDIA's second quarter to 26 July produced revenue of [$96.2 billion, up 106%, with a 75.0% gross margin](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027), against 66.4% at TSMC, $TSM (NYSE), itself high for a manufacturer, and it guided $108 billion for the third quarter. At the bottom of the stack, the price of a unit of intelligence keeps falling. Epoch AI found the price of reaching a fixed level of model capability [fell by 9 to 900 times a year](https://epoch.ai/data-insights/llm-inference-price-trends) depending on the task, and by 40 times a year for GPT-4-level performance on PhD-level science questions. Each hardware generation delivers more inference throughput per dollar and per watt. Lower-precision formats such as FP8 and FP4 raise effective throughput. Batching, speculative decoding and better memory management raise utilisation. Smaller distilled models now match much larger models from a year or two earlier, and competition among model providers, strong open-weight models included, passes much of the saving to customers. The two trends do not conflict. Falling unit costs expand the addressable market, which raises total demand for premium hardware. As the price of a token falls, uses that were uneconomic become viable and aggregate spending rises. The risk sits with intermediaries whose pricing power depends on scarcity of capability rather than scarcity of capacity. Model providers without distribution or cost advantages face sustained margin pressure. The tightest pricing in the supply chain is in high-bandwidth memory and advanced packaging. HBM is dominated by three suppliers, with SK hynix holding about half the market, Samsung a third and Micron, $MU (NASDAQ), about 18%, per [StorageNewsletter](https://www.storagenewsletter.com/2026/10/02/micron-technology-fiscal-4q26-and-fy26-financial-results/). SK hynix reported an [80% operating margin](https://www.tradingview.com/news/prnewswire:10159c85e082e:0-sk-hynix-announces-3q26-financial-results/) for the third quarter yesterday, said HBM4 shipments had begun in earnest, and has long-term agreements with 14 customers covering more than half of its 2027 and 2028 capacity. It expects demand to exceed supply through 2030. Micron says most of its 2027 HBM is already [contracted at significantly higher prices](https://www.storagenewsletter.com/2026/10/02/micron-technology-fiscal-4q26-and-fy26-financial-results/), and Amazon cited [higher memory costs](https://fortune.com/2026/07/30/andy-jassy-amazon-capex-demand-aws-pga-tour/) when it raised its capital plan. Advanced packaging, particularly TSMC's CoWoS family (chip-on-wafer-on-substrate), which joins logic dies and HBM stacks on a silicon interposer, is the other bottleneck. Industry estimates cited by TrendForce put CoWoS capacity at about [90,000 wafers a month in mid-2026, rising to 120,000 by year-end](https://finance.biggo.com/news/5784aaf1-fcbc-4f76-8856-491b9e7175f6), and TSMC's chairman said on 15 October that CoWoS is still its customers' [first packaging choice](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf). ## The breakthroughs are in racks, memory and packaging The most consequential hardware shift is the move to rack-scale systems. Vendors now design the whole rack as one machine. NVIDIA's GB200 NVL72 links 72 Blackwell GPUs and 36 Grace CPUs in one NVLink domain that [acts as a single, massive GPU](https://www.nvidia.com/en-us/data-center/gb200-nvl72/), which matters most for serving large mixture-of-experts and reasoning models whose memory and communication demands exceed any single chip. NVIDIA began production shipments of its next system, Vera Rubin, in early August, per its [second-quarter investor presentation](https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/NVDA-F2Q27-Quarterly-Presentation-final-1.pdf). AMD's Helios rack pursues the same goal, and Reuters reported in July that it was [in full production](https://www.igorslab.de/en/amd-helios-in-production-mi455x-epyc-venice-openai-challenge-nvidia-ai-racks/) and due to ship within months, and that OpenAI was expected to begin using it later this year. AMD says OpenAI, Meta, Anthropic, Microsoft and Oracle are among the [buyers adopting Helios](https://www.storagereview.com/news/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin). Rack-scale design moves the engineering challenge to interconnect, power delivery and liquid cooling, and it raises switching costs for buyers. Memory architecture is now the central design problem. For inference, performance is often limited by memory capacity and bandwidth rather than by arithmetic. HBM4 doubles the interface of each memory stack to 2,048 bits from 1,024 on HBM3E, and on AMD's MI455X that lifts capacity by half to 432 GB and bandwidth 2.9 times to 23.3 TB/s, per [StorageReview](https://www.storagereview.com/news/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin), against 288 GB on today's leading parts. Around the HBM, designers are adding tiers: pooled memory over CXL (Compute Express Link), larger on-package caches, and software that keeps the key-value caches of long conversations in the cheapest tier that still meets latency targets. One research team showed a shared CXL memory pool for that cache that [cut time to first token by 89.6%](https://arxiv.org/abs/2511.20172v2) and raised throughput 7.35 times against network-based designs. At the process level, TSMC's 2 nm (N2) node started [volume production in the fourth quarter of 2025](https://www.tsmc.com/english/dedicatedFoundry/technology/logic/l_2nm), its first with gate-all-around (nanosheet) transistors, and N2P follows in the second half of 2026. Samsung introduced the [Exynos 2600](https://www.koreajoongangdaily.com/business/samsung-unveils-exynos-2600-mobile-chip-slated-to-power-galaxy-s26-series/11989445), which it calls the industry's first mobile processor on a 2 nm gate-all-around process, and Intel's 18A, which pairs RibbonFET transistors with PowerVia backside power, ran [about 25% above its output target](https://www.datacenterdynamics.com/en/news/intel-posts-fastest-yoy-growth-since-2011-with-q2-2026-revenue-totaling-161bn/) in the second quarter. TSMC's own backside-power node, A16, is now [slated for 2027](https://www.trendforce.com/news/2026/04/23/news-tsmc-unveils-latest-roadmap-a12-a13-set-for-2029-without-high-na-euv-a16-volume-production-delayed-to-2027/). For AI accelerators, packaging now matters as much as lithography. TSMC is taking its CoWoS interposers from 3.5 times the reticle size to 5 and 9 times, and aims to [mass-produce 9.5-reticle packages with 12 or more HBM stacks by 2027](https://www.trendforce.com/news/2025/05/02/news-wafer-level-packaging-showdown-tsmc-scales-up-cowos-reticle-size-as-intel-readies-foveros-s/). AMD's MI455X [hybrid-bonds its compute dies onto cache dies](https://www.storagereview.com/news/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin) and is the largest chip yet built on CoWoS-L. Co-packaged optics has left the laboratory: NVIDIA's Spectrum-6 switches launched in July with [co-packaged optics versions due to ship later in 2026](https://www.datacenterdynamics.com/en/news/nvidia-debuts-spectrum-6-switches-to-power-next-gen-ai-networking/). The industry's roadmap increasingly reads as a packaging roadmap. Software efficiency is the least visible and perhaps the most powerful breakthrough. NVIDIA says it has [more than doubled Blackwell's performance since launch through software alone](https://blogs.nvidia.com/blog/blackwell-inferencemax-benchmark-results/), and in the two months after OpenAI's gpt-oss-120b launched, TensorRT-LLM updates cut the cost per million tokens on a B200 fivefold, from $0.11 to $0.02, per [NVIDIA's technical blog](https://developer.nvidia.com/blog/nvidia-blackwell-leads-on-new-semianalysis-inferencemax-benchmarks/). Inference engines such as vLLM and SGLang compete on the [same daily benchmark](https://developer.nvidia.com/blog/nvidia-blackwell-leads-on-new-semianalysis-inferencemax-benchmarks/). Quantisation, sparsity, mixture-of-experts designs, distillation and speculative decoding deliver comparable quality with fewer operations. These gains do not reduce demand for hardware. They lower the cost per task and so widen adoption. They also make hardware advantage partly portable, which strengthens the case for second-source and custom silicon. ## Power is the binding constraint, and the war has widened the gas-price gap Power is now the primary constraint on AI deployment. Morgan Stanley put US data centre power demand for 2026 to 2028 at 97 GW in October, and the net shortfall at 32 GW, 34% of demand, after capacity under construction, spare grid capacity and on-site generation, per a [summary of the note](https://www.itiger.com/news/1130640716). The bank's figure moves with each vintage, from [about 45 GW in November 2025](https://mlq.ai/news/morgan-stanley-warns-of-looming-45-gigawatt-us-power-shortage-driven-by-ai-data-center-boom/) to [38 GW in August](https://cryptobriefing.com/morgan-stanley-ai-data-centers-38-gigawatt-power-gap/) and 32 GW now, and every vintage is tens of gigawatts. The Department of Energy and Lawrence Berkeley National Laboratory found data centres used 4.4% of US electricity in 2023 and project [6.7% to 12% by 2028](https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers). The market already prices it. PJM, the grid operator for 13 states from Illinois to Virginia, cleared its 2028/29 capacity auction in July at the $325/MW-day cap and [6,831 MW short of its reliability requirement](https://www.pjm.com/-/media/DotCom/about-pjm/newsroom/2026-releases/20260714-pjm-capacity-auction-procures-138318-mw-of-generation-resources.pdf), the first auctions in its history in which the whole region fell short. Rack densities show why. Data centres built around racks of under 10 kW have had to support [about 120 kW for NVIDIA's NVL72](https://www.datacenterdynamics.com/en/analysis/nvidia-gtc-jensen-huang-data-center-rack-density/). AMD's Helios rack draws [225 to 245 kW](https://www.storagereview.com/news/amd-mi455x-and-helios-432gb-hbm4-72-gpu-racks-and-a-real-answer-to-vera-rubin). NVIDIA put its Kyber rack for Rubin Ultra at [about 600 kW](https://www.datacenterdynamics.com/en/news/nvidias-rubin-ultra-nvl576-rack-expected-to-be-600kw-coming-second-half-of-2027/), and SemiAnalysis reported in July that Kyber has [slipped to 2028](https://www.sdxcentral.com/news/nvidia-pushes-kyber-release-to-2028-following-manufacturing-concerns-report/); NVIDIA did not comment. Schneider Electric puts the practical limit of air cooling at [25 to 30 kW a rack](https://blog.se.com/datacenter/2026/01/28/how-leverage-air-cooling-nvidia-ai-gpus/), so direct-to-chip liquid cooling is becoming standard for AI builds, and facilities must be redesigned around higher-voltage distribution, heavier floors and new heat-rejection systems. ![Exhibit 1, where AI's constraint binds: models, accelerators, memory and packaging, racks and power, with the grid, transformers, turbines and fuel under power](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-96e850d137d4457e.svg) **Exhibit 1, its links and their sources.** - Epoch AI found the price of reaching a fixed level of model capability fell by 9 to 900 times a year, depending on the task. ([Epoch AI, LLM inference price trends](https://epoch.ai/data-insights/llm-inference-price-trends), 12 March 2025, study) - NVIDIA's second quarter to 26 July produced revenue of $96.2 billion with a 75.0% gross margin. ([NVIDIA, second-quarter results](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027), 26 August 2026, company) - SK hynix has long-term agreements with 14 customers covering more than half of its 2027 and 2028 capacity, and expects demand to exceed supply through 2030. ([SK hynix, third-quarter results](https://www.tradingview.com/news/prnewswire:10159c85e082e:0-sk-hynix-announces-3q26-financial-results/), 27 October 2026, company) - Data centres have had to move from racks of under 10 kW to liquid-cooled racks of about 120 kW for NVIDIA's NVL72 systems. ([DatacenterDynamics](https://www.datacenterdynamics.com/en/analysis/nvidia-gtc-jensen-huang-data-center-rack-density/), 21 March 2025, reported) - Morgan Stanley puts US data centre power demand for 2026 to 2028 at 97 GW and the net shortfall at 32 GW, 34% of demand. ([Morgan Stanley, via Tiger Brokers](https://www.itiger.com/news/1130640716), 5 October 2026, reported) - The median US power project built in 2025 took 61 months from interconnection request to commercial operation, against 22 months in 2008. ([Lawrence Berkeley National Laboratory, Queued Up](https://eta-publications.lbl.gov/sites/default/files/2026-06/queued_up_2026_edition.pdf), June 2026, study) - Wood Mackenzie's 2025 survey put average lead times at 128 weeks for power transformers and 144 weeks for generator step-up transformers. ([POWER Magazine, on Wood Mackenzie](https://www.powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/), 2 January 2026, reported) - GE Vernova's gas turbine backlog and slot reservations rose to 116 GW in the second quarter, and it expects them to reach at least 125 GW by the end of 2026. ([GE Vernova, second-quarter results](https://www.gevernova.com/news/press-releases/ge-vernova-reports-second-quarter-2026-financial-results-raises-2026-financial), 22 July 2026, company) - European gas (TTF) reached €90/MWh on 20 October for the first time in almost four years, on uncertainty over Hormuz shipping. ([GMK Center](https://gmk.center/en/news/european-gas-prices-reached-e90-mwh-in-october/), 20 October 2026, reported) The grid is the slowest-moving part of the system. The median US power project built in 2025 took [61 months from interconnection request to operation](https://eta-publications.lbl.gov/sites/default/files/2026-06/queued_up_2026_edition.pdf), against 22 months in 2008, per Lawrence Berkeley National Laboratory. Power transformers averaged [128 weeks of lead time](https://www.powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis/) in Wood Mackenzie's 2025 survey, and generator step-up transformers 144 weeks. Gas turbine order books now run to 2030 and 2031: GE Vernova, $GEV (NYSE), is taking reservations for 2031 delivery and [Mitsubishi Heavy's new orders deliver between 2028 and 2030](https://oilprice.com/Energy/Energy-General/The-Gas-Turbine-Shortage-Just-Became-AIs-Biggest-Constraint.html). GE Vernova's backlog and slot reservations rose to [116 GW in the second quarter](https://www.gevernova.com/news/press-releases/ge-vernova-reports-second-quarter-2026-financial-results-raises-2026-financial), and it expects backlog and slot reservations to reach at least 125 GW by year-end. The industry is responding on several fronts. Microsoft has a [20-year agreement](https://neutronbytes.com/2025/06/06/meta-signs-20-year-power-deal-with-constellation/) to buy the power of the shuttered Three Mile Island unit that Constellation is restarting as the Crane Clean Energy Center, and in August a [fuel licence approval moved the restart closer](https://www.constellationenergy.com/news/2026/08/constellation-reports-second-quarter-2026-results.html) to its 2027 target. Meta signed for the 1,121 MW Clinton plant in Illinois, per [Neutron Bytes](https://neutronbytes.com/2025/06/06/meta-signs-20-year-power-deal-with-constellation/), and Amazon for [1,920 MW from Talen's Susquehanna plant](https://world-nuclear-news.org/articles/new-supply-agreement-expands-talen-amazon-partnership). Small modular reactors arrive later: Google's first, a 50 MW Kairos Power unit, is due at [full operation in 2030](https://www.datacenterdynamics.com/en/news/tva-signs-50mw-ppa-with-smr-developer-kairos-for-google-data-centers-in-tennessee-and-alabama/). Regulators are trading speed for flexibility. Texas now requires [curtailment protocols for new loads of 75 MW or more](https://www.bracewell.com/resources/texas-senate-bill-6-ushers-in-major-overhaul-of-large-load-interconnection-and-grid-access-rules/), and PJM asked federal regulators in August to let [loads of 50 MW or more connect before matching capacity exists](https://datacenterknowledge.com/energy-power-supply/pjm-s-new-deal-for-data-centers-bring-power-or-face-cuts), curtailed first in an emergency. A Duke University study found the US grid could absorb [76 GW of new load curtailed 0.25% of the time](https://www.powermag.com/duke-researchers-grid-flexibility-key-to-accommodate-load-growth/). Energised land has become a scarce asset in its own right: powered land in Northern Virginia sells for [at least $6 million an acre against about $4 million without power](https://www.techtarget.com/it-infrastructure/news/366651918/AI-data-center-land-buying-hits-6B-as-power-premiums-rise), and CBRE puts the region's data centre vacancy at [0.2%](https://www.cbre.com/press-releases/north-american-data-center-demand-continues-to-outpace-supply-despite-record-construction). The war has priced the constraint outside the United States. European gas reached [€90/MWh on 20 October](https://gmk.center/en/news/european-gas-prices-reached-e90-mwh-in-october/) for the first time in almost four years, on uncertainty over Hormuz shipping, against about [€32 before the war began](https://www.france-epargne.fr/news/gaz-le-ttf-depasse-90-euros-le-mwh-une-premiere-depuis-decembre-2022), and EU storage is 74% full against 83% a year earlier. US gas has stayed cheap: the Energy Information Administration puts Henry Hub at [$3.09/MMBtu for February to December 2026](https://www.eia.gov/outlooks/steo/archives/oct26.pdf), with ample inventories limiting upward pressure. US wholesale power is still forecast to average [11% more in 2026 than in 2025](https://www.eia.gov/outlooks/steo/archives/oct26.pdf), mainly on weather and with PJM up 41%, as of the same October outlook, and the average industrial price reached [9.64 cents/kWh in August](https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_3) against 8.98 a year earlier. The war has widened the gas-price gap between Europe and the United States, which favours US sites for AI, while the US grid's own limits stay where they were. *"A gigawatt of AI capacity is a power-plant decision before it is a chip decision." Shayne Heffernan* ## The chipmakers: NVIDIA leads as the field widens NVIDIA remains the defining company of the AI cycle. Its advantage rests on more than silicon: the CUDA software ecosystem, NVLink and networking, rack-scale system design and a supply chain secured years in advance give it a lead that competitors cannot close in one product cycle. Bloomberg data put its share of AI accelerator sales at [86% in 2025](https://finance.yahoo.com/markets/stocks/articles/nvidia-stock-investors-just-got-090800739.html), unchanged on 2024, and its data centre revenue reached [$89.0 billion in the quarter to 26 July](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027), up 117%. Researchers cited by I/O Fund expect its share to fall to [about 70% by the end of 2026](https://io-fund.com/ai-stocks/nvidia-fiscal-q2-answer-to-losing-market-share) as alternatives ramp. Any fall in share comes from a market growing fast enough to support other suppliers, and from the largest buyers diversifying to manage cost and supply risk. NVIDIA's own sales are still rising: it guided [$108 billion](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027) for the current quarter. AMD has established itself as a credible second source. Its Instinct accelerators compete on memory capacity, an advantage for inference, and its ROCm software has matured to the point where model developers run production work on it. AMD's data centre revenue rose [107% to $6.7 billion](https://ir.amd.com/news-events/press-releases/detail/1295/) in the second quarter. It says its Helios racks are being deployed by Anthropic, HUMAIN, Meta, Microsoft, OpenAI and Oracle, and it has agreed with Anthropic to deploy [up to 2 GW of MI450 GPUs](https://ir.amd.com/news-events/press-releases/detail/1295/). AMD reports third-quarter results on [3 November](https://ir.amd.com/news-events/press-releases/detail/1300/amd-to-report-fiscal-third-quarter-2026-financial-results). Its task is execution at scale: securing packaging and HBM supply, and keeping software parity as models and frameworks change. Custom silicon is the third pillar. Google is on its eighth generation of TPU, the [TPU 8t for training and 8i for inference](https://www.techradar.com/pro/google-cloud-unveils-eighth-generation-tpus-built-to-support-an-agentic-era) announced in April, and it now offers NVIDIA's Vera Rubin beside them. Amazon made [Trainium3 generally available](https://press.aboutamazon.com/2025/12/trainium3-ultraservers-now-available-enabling-customers-to-train-and-deploy-ai-models-faster-at-lower-cost) in December 2025 and names Anthropic among its Trainium customers, and it will supply OpenAI with [2 GW of Trainium capacity](https://siliconangle.com/2026/02/27/openai-raises-110b-730b-valuation-expands-aws-nvidia-partnerships/). Meta and Broadcom agreed a [multi-generation MTIA partnership through 2029](https://www.nasdaq.com/press-release/broadcom-announces-extended-partnership-meta-deploy-technology-support-multi) with an initial commitment above 1 GW, and Microsoft's [Maia 200 is in production](https://datacenterdynamics.com/en/news/microsoft-unveils-azure-maia-300-says-chip-will-be-its-primary-inference-system), with Maia 300 due in 2027. OpenAI's own accelerator with Broadcom is [targeted to start in the second half of 2026](https://openai.com/index/openai-and-broadcom-announce-strategic-collaboration/). Broadcom, which designs custom chips and the networks that join them, booked [$16.7 billion of AI semiconductor revenue](https://pulse2.com/broadcom-q3-fiscal-2026-revenue-jumps-86-to-29-6-billion-as-ai-semiconductor-revenue-reaches-16-7-billion/) in its quarter to 2 August, up 221%. Custom chips make most sense for buyers with very large, well-understood workloads, where giving up flexibility for a lower cost per operation is worthwhile. They are unlikely to displace merchant GPUs broadly, but they cap pricing power and give the hyperscalers bargaining power. Underneath all of them sits TSMC. Nearly every leading AI accelerator, merchant or custom, is made on its advanced nodes and packaged on its CoWoS lines. Its third-quarter revenue rose [41.2% to $46.72 billion](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/526b65af7305df2dde6b9ece6812b1fdf1e49f9b/3Q26%20EarningsRelease.pdf) with a 66.4% gross margin, it guides AI accelerator revenue to grow at a [mid- to high-50s percentage rate a year](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf) from 2024 to 2029, and it is spending [$60 to $64 billion](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf) this year. This concentration is the single most important structural risk in the sector. TSMC is building [13 leading-edge and advanced packaging fabs in Taiwan against five to six overseas](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf). Its second Arizona fab is due in volume production in the [third quarter of 2027](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf) and a 3 nm fab in Japan in [2028](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf), and the overseas fabs cost more, diluting gross margin by [close to 2% in 2026](https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf), as of its 15 October 2026 call. Packaging capacity, HBM allocation and wafer starts are effectively rationed among customers, and the allocation shapes competitive outcomes as much as product design does. *"NVIDIA's share can fall while its revenue rises. That is what a market growing faster than any single supplier looks like." Shayne Heffernan* ## AI went to war in 2026 In the Iran campaign AI ran at the centre of US targeting, at a scale the Pentagon has now put figures on. Defence officials say Maven was used to identify and strike [more than 2,000 distinct targets in the first 96 hours](https://www.defensenews.com/industry/techwatch/2026/10/21/marine-corps-makes-maven-available-service-wide/) of Operation Epic Fury, and that it now processes about 6 million pages of text a day. Over the 38 days of the campaign it supported strike missions against [13,000 targets](https://breakingdefense.com/2026/05/insatiable-appetite-for-ai-maven-usage-surged-for-strikes-on-iran-pentagon-ai-chief-says/). Classified use of Maven rose 89% in a month and peak daily AI use reached [about 20 billion tokens, up 4,425%](https://breakingdefense.com/2026/05/insatiable-appetite-for-ai-maven-usage-surged-for-strikes-on-iran-pentagon-ai-chief-says/), a Pentagon spokesperson told Breaking Defense. Central Command used the [LUCAS one-way attack drone](https://defensescoop.com/2026/05/20/pentagon-selects-shield-ai-to-plug-swarm-software-into-lucas-drone-company-says/) in combat for the first time, and Shield AI says the Pentagon chose its Hivemind software as the drone's [AI pilot](https://defensescoop.com/2026/05/20/pentagon-selects-shield-ai-to-plug-swarm-software-into-lucas-drone-company-says/), so one operator can command a swarm. Commercial models were inside the loop from the first weekend. Two sources told CBS News that the US military [used Anthropic's Claude in the attack on Iran](https://www.cbsnews.com/news/anthropic-claude-ai-iran-war-u-s/), and Bloomberg reported that Claude was [one of the main tools installed in Maven](https://techcrunch.com/2026/03/05/its-official-the-pentagon-has-labeled-anthropic-a-supply-chain-risk/) for operators in the Middle East. The President had ordered every federal agency, [the day before](https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban), to stop using Anthropic's technology, with a six-month phase-out. The Pentagon has since said it moved off Claude: its chief technology officer, Emil Michael, said in September that [all Maven work had moved months earlier](https://www.claimsjournal.com/news/national/2026/09/11/340073.htm), and the BBC reported in early October that the department had confirmed it had [stopped using Anthropic's products](https://www.androidheadlines.com/2026/10/pentagon-stops-using-removes-anthropic-claude-ai-military.html), and that staff had used Claude inside Maven until the week before. The same speed carried a stale record into a strike. Two Tomahawk missiles hit the Shajarah Tayyebeh elementary school in Minab on 28 February, and more than 150 people were killed, at least 123 of them children, according to Bloomberg's account of the Pentagon's investigation as summarised by [AI Weekly](https://aiweekly.co/alerts/pentagon-rewires-palantir-maven-after-iran-school-strike). The school was on the target list because outdated intelligence had it catalogued as an IRGC facility. Central Command has since revised how it refreshes and vets planned targets and taken [dozens of upgrades to Maven](https://aiweekly.co/alerts/pentagon-rewires-palantir-maven-after-iran-school-strike), including features that make it re-review the underlying intelligence for disqualifying factors. For investors the lesson is precise: AI in a kill chain is only as good as the record underneath it, and the systems that win the next contracts will be the ones whose decisions can be checked. The department is making AI permanent. Secretary Hegseth's January strategy memo commits it to becoming an ["AI-first" warfighting force](https://meritalk.com/articles/pentagon-unveils-ai-acceleration-strategy/). GenAI.mil, launched in December with Google's Gemini, now carries Gemini, OpenAI's ChatGPT Mil and xAI's Grok, and [more than 2 million people logged in](https://nextgov.com/artificial-intelligence/2026/09/genaimil-saw-more-2-million-users-one-week-top-dod-official-says/416183) during a single week in September, against about 80,000 Pentagon AI users before it existed. A 9 March memo from Deputy Secretary Steve Feinberg moved Maven to become a formal [programme of record](https://defensescoop.com/2026/04/03/palantir-maven-feinberg-directive/), which DefenseScoop reported was due by the end of the fiscal year in September. On 1 May the Pentagon cleared [eight companies to run AI on its classified networks](https://breakingdefense.com/2026/05/pentagon-clears-7-tech-firms-to-deploy-their-ai-on-its-classified-networks/): AWS, Google, Microsoft, NVIDIA, OpenAI, SpaceX, Reflection and Oracle. The fiscal 2027 request totals [$1.5 trillion](https://defensescoop.com/2026/04/21/dod-plans-largest-ever-investment-drones-anti-drone-weapons/), with $53.6 billion for autonomy, drones and contested logistics. Golden Dome, the missile shield, has a planned cost of [$185 billion](https://defensescoop.com/2026/08/12/guetlein-golden-dome-in-jeopardy-due-to-2027-funding-uncertainty/), with capabilities due in summer 2028, and the Congressional Budget Office puts a comparable shield at [$1.2 trillion over 20 years](https://defensescoop.com/2026/05/12/golden-dome-cbo-cost-estimate-missile-defense-architecture/). ## The companies inside the military's AI loop The defence money for AI runs through five layers: decision software, frontier models, space and networks, autonomy and drones, and classified compute. Exhibit 2 maps the companies in each, listed and private told apart. ![Exhibit 2, the companies inside the US military's AI loop by layer, listed and private told apart, from Palantir and SpaceX to Anthropic, OpenAI and Anduril](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-4615a9e5f57ea41f.svg) **Exhibit 2, its links and their sources.** - The fiscal 2027 request proposes more than $2 billion for the programme, against more than $240 million in fiscal 2026, including more than $1.5 billion to expand access to Palantir's Maven Smart System. ([DefenseScoop](https://defensescoop.com/2026/05/28/dod-fy27-budget-cjadc2-maven-smart-system-palantir/), 28 May 2026, reported) - The Pentagon's chief digital and AI officer said Operation Epic Fury used Maven to conduct strike missions against 13,000 targets in 38 days. ([Breaking Defense](https://breakingdefense.com/2026/05/insatiable-appetite-for-ai-maven-usage-surged-for-strikes-on-iran-pentagon-ai-chief-says/), 12 May 2026, reported) - The Chief Digital and AI Office awarded Scale AI a $500 million contract in May 2026, five times its deal of September 2025. ([OODAloop](https://oodaloop.com/briefs/technology/meta-backed-scale-ai-wins-500-million-defense-department-deal/), 7 May 2026, reported) - More than 2 million people logged in to GenAI.mil in a single week in September, against about 80,000 Pentagon AI users before it launched. ([Nextgov/FCW](https://nextgov.com/artificial-intelligence/2026/09/genaimil-saw-more-2-million-users-one-week-top-dod-official-says/416183), 23 September 2026, reported) - The department launched GenAI.mil on 9 December 2025 with Google's Gemini for Government. ([Decrypt](https://decrypt.co/351691/war-department-launches-googles-gemini-military-ai-push), 9 December 2025, reported) - OpenAI's classified agreement allows all lawful purposes but bars using its AI to independently direct autonomous weapons where law or policy requires human control, in a cloud-only deployment. ([OpenAI](https://openai.com/index/our-agreement-with-the-department-of-war/), 2 March 2026, company) - A divided D.C. Circuit panel upheld the second supply-chain risk designation of Anthropic 2-1 on 25 September and delayed its effect to allow a rehearing. ([CNBC](https://www.cnbc.com/2026/09/25/pentagon-anthropic-ai-risk-appeals-court.html), 25 September 2026, reported) - xAI's Grok for Government joined Gemini and ChatGPT on GenAI.mil at the end of August 2026. ([DefenseScoop](https://defensescoop.com/2026/08/31/grok-chatgpt-added-to-genai-mil/), 31 August 2026, reported) - Golden Dome's director said its planned cost rose from $175 billion to $185 billion, with capabilities due in summer 2028. ([DefenseScoop](https://defensescoop.com/2026/08/12/guetlein-golden-dome-in-jeopardy-due-to-2027-funding-uncertainty/), 12 August 2026, reported) - The Space Force awarded SpaceX a $2.29 billion fixed-price contract to build the Golden Dome data network backbone on Starshield. ([The Next Web](https://thenextweb.com/news/space-force-spacex-229bn-data-network-contract), 27 May 2026, reported) - Central Command said it used LUCAS in combat for the first time against Iran in Operation Epic Fury. ([DefenseScoop](https://defensescoop.com/2026/05/20/pentagon-selects-shield-ai-to-plug-swarm-software-into-lucas-drone-company-says/), 20 May 2026, reported) - Anduril raised $5 billion in May 2026 at a $61 billion valuation, double its previous valuation. ([CNBC](https://www.cnbc.com/2026/05/13/anduril-valuation-defense-tech-funding-boom.html), 13 May 2026, reported) - Shield AI said the Pentagon selected its Hivemind software as the AI pilot for the LUCAS one-way attack drone. ([DefenseScoop](https://defensescoop.com/2026/05/20/pentagon-selects-shield-ai-to-plug-swarm-software-into-lucas-drone-company-says/), 20 May 2026, reported) - On 1 May the Pentagon announced agreements for eight companies to deploy AI on its classified networks, with Anthropic absent. ([Breaking Defense](https://breakingdefense.com/2026/05/pentagon-clears-7-tech-firms-to-deploy-their-ai-on-its-classified-networks/), 1 May 2026, reported) - NVIDIA is among the eight companies cleared on 1 May to deploy AI on the Pentagon's classified networks. ([Breaking Defense](https://breakingdefense.com/2026/05/pentagon-clears-7-tech-firms-to-deploy-their-ai-on-its-classified-networks/), 1 May 2026, reported) - AWS, Microsoft and Oracle are among the eight companies cleared on 1 May to deploy AI on the Pentagon's Impact Level 6 and 7 networks. ([Breaking Defense](https://breakingdefense.com/2026/05/pentagon-clears-7-tech-firms-to-deploy-their-ai-on-its-classified-networks/), 1 May 2026, reported) - The fiscal 2027 request totals $1.5 trillion, including $53.6 billion for autonomy, drones and contested logistics. ([DefenseScoop](https://defensescoop.com/2026/04/21/dod-plans-largest-ever-investment-drones-anti-drone-weapons/), 21 April 2026, reported) - The Defense Innovation Unit awarded Scale AI the Thunderforge prototype in March 2025, to bring AI into theatre-level planning. ([Defense Innovation Unit](https://www.diu.mil/latest/dius-thunderforge-project-to-integrate-commercial-ai-powered-decision-making), 5 March 2025, official) - OpenAI added language on 2 March explicitly barring domestic surveillance of US persons. ([OpenAI](https://openai.com/index/our-agreement-with-the-department-of-war/), 2 March 2026, company) - Judge Rita Lin ruled one supply-chain risk designation of Anthropic unlawful on 27 August. ([TechCrunch](https://techcrunch.com/2026/08/28/anthropic-gets-its-first-court-win-over-the-pentagons-supply-chain-risk-label/), 28 August 2026, reported) - Pentagon CTO Emil Michael said in September that all Maven work had moved off Anthropic months earlier. ([Bloomberg, via Claims Journal](https://www.claimsjournal.com/news/national/2026/09/11/340073.htm), 10 September 2026, reported) - Anthropic filed its S-1 publicly on 26 October to list on Nasdaq under the ticker ANT. ([Fortune](https://fortune.com/2026/10/26/anthropics-s-1/), 26 October 2026, reported) - The Space Force awarded SpaceX a $4.16 billion agreement for satellites that track aircraft, drones and cruise missiles. ([DefenseScoop](https://defensescoop.com/2026/05/29/spacex-amti-contract-space-force), 29 May 2026, reported) - The Air Force awarded production contracts for the first Collaborative Combat Aircraft, Anduril's FQ-44 Fury among them, in June. ([Tectonic Defense](https://www.tectonicdefense.com/air-force-awards-production-contracts-for-cca/), 18 June 2026, reported) - Shield AI raised $2 billion in March 2026 at a $12.7 billion valuation. ([The Next Web](https://thenextweb.com/news/shield-ai-2-billion-hivemind-autonomous-defence), 26 March 2026, reported) Palantir is the listed company most directly wired into US targeting. Its second-quarter revenue rose [93% to $1.935 billion](https://letsdatascience.com/news/palantirs-q2-revenue-jumps-93-as-us-commercial-sales-rise-14-90b0812e), with US government revenue up 90% to $809 million and US commercial revenue up 149%, and it raised its 2026 guidance to [$8.150 to $8.158 billion](https://letsdatascience.com/news/palantirs-q2-revenue-jumps-93-as-us-commercial-sales-rise-14-90b0812e). The Army's enterprise agreement runs [up to $10 billion over 10 years](https://defensescoop.com/2025/07/31/army-palantir-software-enterprise-agreement-10-billion/), the Maven contract ceiling is [nearly $1.3 billion](https://defensescoop.com/2025/05/23/dod-palantir-maven-smart-system-contract-increase/), and the fiscal 2027 request asks for [more than $1.5 billion](https://defensescoop.com/2026/05/28/dod-fy27-budget-cjadc2-maven-smart-system-palantir/) to widen access to Maven. NATO bought its own version of [Maven in 2025](https://breakingdefense.com/2025/04/nato-picks-palantirs-maven-ai-for-military-planning-amid-trans-atlantic-tension/), and from 1 October the Marine Corps opened Maven to [all its commands](https://www.defensenews.com/industry/techwatch/2026/10/21/marine-corps-makes-maven-available-service-wide/). The Wall Street Journal reported that Palantir and Anduril are in the consortium building [Golden Dome's software](https://finance.yahoo.com/markets/stocks/articles/palantir-stock-pltr-rises-joins-071818571.html). Palantir reports third-quarter results on [2 November](https://finance.yahoo.com/markets/stocks/articles/palantir-announces-date-third-quarter-200500055.html), after the close. SpaceX is the listed company in the space layer, and it has owned [xAI since 2 February](https://x.ai/news/xai-joins-spacex). The Space Force awarded it [$2.29 billion](https://thenextweb.com/news/space-force-spacex-229bn-data-network-contract) to build Golden Dome's data network on its Starshield satellites and [$4.16 billion](https://defensescoop.com/2026/05/29/spacex-amti-contract-space-force) for satellites that track aircraft, drones and cruise missiles. It is one of 12 companies holding [space-based interceptor prototype contracts](https://fortune.com/2026/04/25/spacex-anduril-lockheed-raytheon-northrop-golden-dome-interceptor-contracts/) worth up to $3.2 billion in all, and in July it won [$1.6 billion for 18 military launches](https://airandspaceforces.com/space-force-to-launch-18-comms-targeting-missions-by-end-of-2027). The National Reconnaissance Office's proliferated constellation is [believed to be Starshield](https://spaceflightnow.com/2026/06/06/spacex-to-launch-2-starshield-satellites-during-saturday-night-starlink-mission/), though the office has never said so. About [20% of SpaceX's revenue](https://www.algemeiner.com/2026/05/26/pentagon-spars-spacex-starlink-price-hike-iran-war/) comes from the US government, according to an SEC filing cited by Reuters on 26 May 2026. xAI's Grok for Government sits on [GenAI.mil](https://defensescoop.com/2026/08/31/grok-chatgpt-added-to-genai-mil/), and in February xAI became the [second company cleared for classified settings](https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban). The Wall Street Journal reported in July that SpaceX was in talks to give the Pentagon [data-centre capacity worth billions of dollars](https://www.thestar.com.my/tech/tech-news/2026/07/18/musk039s-spacex-in-talks-to-supply-the-pentagon-with-computing-power-wsj-reports) to run AI models. SpaceX shares closed at [$148.58 on 27 October 2026](https://query1.finance.yahoo.com/v8/finance/chart/SPCX?period1=1793059200&period2=1793145600&interval=1d), 7.7% below their first close of $160.95 on 12 June 2026. Anthropic is the most contested company in the loop, and it is about to list. The Chief Digital and AI Office gave it a [$200 million prototype agreement](https://www.anthropic.com/news/anthropic-and-the-department-of-defense-to-advance-responsible-ai-in-defense-operations) in July 2025, and its [Claude Gov models](https://www.anthropic.com/news/claude-gov-models-for-u-s-national-security-customers) were built for national security customers. Anthropic says talks with the Pentagon broke down in February over two exceptions it asked for, [mass domestic surveillance of Americans and fully autonomous weapons](https://www.anthropic.com/news/statement-comments-secretary-war). The Pentagon said it did not intend either use and required AI companies to allow [all lawful purposes](https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban). On 27 February President Trump ordered every agency to [stop using Anthropic's technology](https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban), with a six-month phase-out, and Secretary Hegseth directed that it be designated a supply-chain risk. The courts have split on that. Judge Rita Lin ruled one designation [unlawful on 27 August](https://techcrunch.com/2026/08/28/anthropic-gets-its-first-court-win-over-the-pentagons-supply-chain-risk-label/), and a divided D.C. Circuit panel [upheld the other 2-1](https://www.cnbc.com/2026/09/25/pentagon-anthropic-ai-risk-appeals-court.html) on 25 September; Anthropic asked for a [rehearing on 13 October](https://www.unite.ai/anthropic-asks-d-c-circuit-to-rehear-pentagon-supply-chain-risk-case/). Commerce Secretary Howard Lutnick has said Anthropic is [back on side with the administration](https://www.claimsjournal.com/news/national/2026/09/11/340073.htm), Bloomberg reported in September, while Michael drew a distinction between the Pentagon's national security needs and those of the rest of the government. Anthropic was absent from the 1 May classified-network agreements and is not on [GenAI.mil](https://nextgov.com/artificial-intelligence/2026/09/genaimil-saw-more-2-million-users-one-week-top-dod-official-says/416183). Its S-1, filed publicly on [26 October](https://fortune.com/2026/10/26/anthropics-s-1/), seeks a Nasdaq listing under the ticker ANT. The filing shows first-half revenue of [$16.3 billion](https://fortune.com/2026/10/26/anthropics-s-1/), up 1,300%, and a net loss of $80.7 billion, of which $72.4 billion is an accounting charge on convertible notes. Government work is [under 1% of its revenue](https://thenextweb.com/news/anthropic-pentagon-rehearing-bid), The Next Web reported on 14 October 2026, and it has agreed to spend [up to $120.5 billion with SpaceXAI](https://finance.yahoo.com/technology/live/anthropic-ipo-live-updates-frontier-ai-lab-discloses-revenue-and-risks-in-highly-anticipated-s-1-filing-141702607.html), SpaceX's AI arm, for compute into 2029. OpenAI took the other route. Hours after the ban it said it had reached an agreement to [deploy on the Pentagon's classified networks](https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban), for all lawful purposes, cloud-only, with its own safety stack, and with a bar on using its AI to [independently direct autonomous weapons](https://openai.com/index/our-agreement-with-the-department-of-war/) where law or policy requires human control. On 2 March it added language [barring domestic surveillance of US persons](https://openai.com/index/our-agreement-with-the-department-of-war/), and it said publicly that [Anthropic should not be designated a supply-chain risk](https://openai.com/index/our-agreement-with-the-department-of-war/). ChatGPT Mil went live on GenAI.mil at the [end of August](https://defensescoop.com/2026/08/31/grok-chatgpt-added-to-genai-mil/). OpenAI was heading towards an [$852 billion valuation](https://abc7news.com/post/anthropic-races-public-offering-debut-confidential-sec-filing/19213604/) as of its March round, per AP, which reported in June that it had not filed to list, and Fortune reports that Anthropic will [list first](https://fortune.com/2026/10/26/anthropics-s-1/). Alphabet's Gemini was the first model on GenAI.mil, which launched on [9 December 2025](https://decrypt.co/351691/war-department-launches-googles-gemini-military-ai-push), and users built [more than 50,000 agents](https://nextgov.com/artificial-intelligence/2026/09/genaimil-saw-more-2-million-users-one-week-top-dod-official-says/416183) within two weeks of Gemini's Agent Designer arriving in March. Google's air-gapped cloud has been authorised for [Top Secret missions](https://cloud.google.com/blog/topics/public-sector/google-public-sector-achieves-top-secret-and-secret-cloud-authorization/) since 2024. Alphabet's shares rose [2.8%](https://query1.finance.yahoo.com/v8/finance/chart/GOOGL?period1=1793059200&period2=1793145600&interval=1d) from 27 February 2026 to 27 October 2026, and it reports third-quarter results today. Anduril is the private autonomy maker the Pentagon is building around. It raised [$5 billion at a $61 billion valuation](https://www.cnbc.com/2026/05/13/anduril-valuation-defense-tech-funding-boom.html) in May and says its revenue doubled to [$2.2 billion in 2025](https://techcrunch.com/2026/05/13/anduril-raises-5b-doubles-valuation-to-61b/). In March the Army folded its 120 Anduril contracts into one [enterprise agreement with a $20 billion ceiling](https://www.nextgov.com/defense/2026/03/army-anduril-enter-new-20b-enterprise-agreement/412153), centred on Anduril's Lattice software. The Air Force awarded production contracts for the first Collaborative Combat Aircraft, Anduril's FQ-44 Fury and General Atomics' FQ-42, [four months early](https://www.tectonicdefense.com/air-force-awards-production-contracts-for-cca/), and the first Ohio-built YFQ-44A rolled off Anduril's Arsenal-1 line on [27 July](https://www.militarytimes.com/industry/techwatch/2026/07/28/first-anduril-yfq-44a-rolls-off-production-line-for-us-cca-program/). Palmer Luckey has said he would [definitely take Anduril public](https://www.cnbc.com/2026/05/13/anduril-valuation-defense-tech-funding-boom.html). Shield AI raised [$2 billion at a $12.7 billion valuation](https://thenextweb.com/news/shield-ai-2-billion-hivemind-autonomous-defence) in March, and says the Pentagon has picked its Hivemind software to [pilot LUCAS](https://defensescoop.com/2026/05/20/pentagon-selects-shield-ai-to-plug-swarm-software-into-lucas-drone-company-says/), with a demonstration due this autumn. It is one of three candidates, with Anduril and RTX's Collins Aerospace, for the Air Force's [primary autonomy provider](https://www.airdatanews.com/us-air-force-moves-cca-combat-drones-into-production-phase/) for the first increment of combat drones, to be chosen by mid-2027. Scale AI, 49% owned by Meta, runs [Thunderforge](https://www.diu.mil/latest/dius-thunderforge-project-to-integrate-commercial-ai-powered-decision-making), the theatre-planning AI for Indo-Pacific and European Commands, and won a [$500 million contract](https://oodaloop.com/briefs/technology/meta-backed-scale-ai-wins-500-million-defense-department-deal/) in May. The computing power sits with the cloud and chip companies. AWS, Microsoft, NVIDIA and Oracle are among the [eight cleared for the classified networks](https://breakingdefense.com/2026/05/pentagon-clears-7-tech-firms-to-deploy-their-ai-on-its-classified-networks/), Microsoft's Azure OpenAI Service has run at [Top Secret since January 2025](https://defensescoop.com/2025/01/16/openais-gpt-4o-gets-green-light-for-top-secret-use-in-microsofts-azure-cloud/), and Amazon has said it will invest [up to $50 billion](https://www.thestar.com.my/tech/tech-news/2026/07/18/musk039s-spacex-in-talks-to-supply-the-pentagon-with-computing-power-wsj-reports) in AI and supercomputing capacity for US government customers. The platform primes are in the programmes and have not shared in the rerating. Lockheed Martin, $LMT (NYSE), RTX, $RTX (NYSE), and Northrop Grumman joined Golden Dome as [prime contractors](https://finance.yahoo.com/markets/stocks/articles/palantir-stock-pltr-rises-joins-071818571.html). L3Harris holds a [$500 million counter-drone ceiling](https://defensescoop.com/2026/09/29/pentagon-task-force-army-announce-billions-in-counter-drone-tech-awards/) and took part in [Shield AI's $240 million round](https://thenextweb.com/news/shield-ai-2-billion-hivemind-autonomous-defence) in March 2025. From 27 February 2026 to 27 October 2026 Lockheed fell [5.3%](https://query1.finance.yahoo.com/v8/finance/chart/LMT?period1=1793059200&period2=1793145600&interval=1d), RTX [4.0%](https://query1.finance.yahoo.com/v8/finance/chart/RTX?period1=1793059200&period2=1793145600&interval=1d), Northrop [23.9%](https://query1.finance.yahoo.com/v8/finance/chart/NOC?period1=1793059200&period2=1793145600&interval=1d) and L3Harris [28.8%](https://query1.finance.yahoo.com/v8/finance/chart/LHX?period1=1793059200&period2=1793145600&interval=1d), per Yahoo Finance closes. AeroVironment, $AVAV (NASDAQ), fell [49.8%](https://query1.finance.yahoo.com/v8/finance/chart/AVAV?period1=1793059200&period2=1793145600&interval=1d) over the same period; in March it [cut its guidance](https://www.itiger.com/news/1121127243) and took a $151.3 million impairment after a Space Force stop-work order. On this record the market has paid for software, models, compute and space. The platform makers now have to show the same AI inside their own products. ## The next US decision on Iran is most likely after 3 November The war has run in phases since 28 February: the opening campaign, in which Supreme Leader Ali Khamenei was [killed on the first day](https://www.khaleejtimes.com/world/mena/iran-supreme-leader-khamenei-killed-us-israel-strikes?amp=1); a ceasefire from 8 April and a memorandum [signed on 17 June](https://www.armscontrol.org/node/15394); a [collapse on 8 July](https://www.goodmorningamerica.com/news/story/us-iran-ceasefire-mou-broke-timeline-134622392); and US strikes into September, the last confirmed on Iranian tankers on [8 September](https://gcaptain.com/u-s-destroys-five-more-irgc-tankers-after-iran-targets-navy-warship/), though CNN's sources say [limited strikes on radars and launchers](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump) have continued since full combat stopped. Iran's Assembly of Experts named his son [Mojtaba Khamenei](https://en.prothomalo.com/international/middle-east/yym9wflmzj) Supreme Leader on 8 March. On 8 October Trump wrote that the US ["will not be attacking Iran at any time prior to the Midterm Elections"](https://www.cnn.com/2026/10/08/politics/trump-iran-attack-election), and the US blockade of Iranian ports stays [in full force](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump). The last fortnight has moved the other way. On 21 October the State Department [ordered non-emergency staff out of Saudi Arabia](https://www.al-monitor.com/originals/2026/10/us-preparing-issue-ordered-departure-some-personnel-gulf-embassies-say-us), citing Houthi attacks. In a memo dated 23 October, the Director of National Intelligence, Jay Clayton, told Secretary of State Marco Rubio that Iran had [substantially increased efforts to target US personnel](https://bdnews24.com/world/middle-east/3l9fhdpf8h). The same day Trump told a rally the US ["may do another little Epic Fury"](https://www.aljazeera.com/news/2026/10/25/centcom-reportedly-has-plans-for-iran-strikes-will-trump-approve-them). On 26 October CNN reported that US forces are prepared to launch new, potentially major operations [within 96 hours of an order](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), against a target list that includes [energy infrastructure, military sites and nuclear facilities](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump). Israel voted on 27 October and is deadlocked: with about two-thirds of votes counted, Netanyahu's bloc had [54 of the 61 seats it needs](https://www.israelnationalnews.com/news/435272). After each of Exhibits 3 to 6 comes my own graphic of the same ground, two after Exhibit 5, corrected to the sources. ![Exhibit 3, the escalation timeline from Epic Fury on 28 February to the reinforcements reported due by late November, each event with its source](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-ea002bf6b0463e63.svg) **Exhibit 3, its links and their sources.** - The Pentagon named the US military operation against Iran that began on 28 February 2026 Operation Epic Fury. ([AFP, via Malay Mail](https://www.malaymail.com/news/world/2026/02/28/pentagon-us-dubs-iran-assault-operation-epic-fury/210772), 28 February 2026, reported) - Supreme Leader Ali Khamenei was killed in the opening US-Israeli strikes on 28 February; Iranian state media confirmed his death on 1 March. ([Reuters, via Khaleej Times](https://www.khaleejtimes.com/world/mena/iran-supreme-leader-khamenei-killed-us-israel-strikes?amp=1), 1 March 2026, reported) - The US-Iran memorandum of understanding was announced on 14 June and signed on 17 June. ([Arms Control Association](https://www.armscontrol.org/node/15394), 1 July 2026, reported) - The ceasefire and the memorandum collapsed on 8 July, after Iran struck three vessels in Hormuz and the US hit back. ([ABC News](https://www.goodmorningamerica.com/news/story/us-iran-ceasefire-mou-broke-timeline-134622392), 9 July 2026, reported) - The last confirmed US strikes were on Iranian oil tankers: three on 5 September and five more on 8 September, after IRGC missile attacks on US warships. ([gCaptain, on CENTCOM's statement](https://gcaptain.com/u-s-destroys-five-more-irgc-tankers-after-iran-targets-navy-warship/), 8 September 2026, official) - Trump wrote on 8 October: "We will not be attacking Iran at any time prior to the Midterm Elections." ([CNN](https://www.cnn.com/2026/10/08/politics/trump-iran-attack-election), 8 October 2026, reported) - Officials told CNN a go-order could come any time before the end of the year, most likely after the US midterms. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) - The State Department ordered non-emergency US government employees and their families to leave Saudi Arabia on 21 October, citing Houthi attacks. ([Reuters, via Al-Monitor](https://www.al-monitor.com/originals/2026/10/us-preparing-issue-ordered-departure-some-personnel-gulf-embassies-say-us), 21 October 2026, reported) - Director of National Intelligence Jay Clayton wrote, in a memo dated 23 October, that Iran has substantially increased efforts to target US personnel. ([Reuters, via bdnews24](https://bdnews24.com/world/middle-east/3l9fhdpf8h), 24 October 2026, reported) - At a rally on 23 October Trump said the US "may do another little Epic Fury". ([Al Jazeera](https://www.aljazeera.com/news/2026/10/25/centcom-reportedly-has-plans-for-iran-strikes-will-trump-approve-them), 25 October 2026, reported) - Saudi Arabia's defence minister, Pakistan's army chief and Turkey's intelligence chief held talks in Tehran on 26 October. ([CNN](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl), 26 October 2026, reported) - Two US officials told CNN that US forces are prepared to launch new operations against Iran within 96 hours of an order from Trump. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) - With about 66.3% of votes counted, Netanyahu's bloc had 54 seats, short of the 61 needed for a majority. ([Israel National News](https://www.israelnationalnews.com/news/435272), 28 October 2026, reported) - The US midterm elections are on Tuesday 3 November. ([CNN](https://www.cnn.com/2026/10/08/politics/trump-iran-attack-election), 8 October 2026, reported) - Israel's final tally is expected by 30 October and official results are published on 4 November. ([Al Jazeera](https://www.aljazeera.com/news/2026/10/28/five-key-takeaways-from-israels-parliamentary-elections), 28 October 2026, reported) - The Theodore Roosevelt strike group brings more than 7,000 sailors and the Makin Island group about 2,000 Marines. ([Straight Arrow News, on the WSJ](https://san.com/cc/us-sends-another-aircraft-carrier-thousands-more-troops-to-middle-east/), 2 October 2026, reported) - The Wall Street Journal reported the Pentagon is sending a third carrier strike group and another Marine expeditionary unit, adding 9,000 to 10,000 sailors and Marines by the end of November. ([Stars and Stripes, on the WSJ](https://stripes.com/branches/navy/2026-10-02/makin-island-middle-east-iran-strikes-november-23032862.html), 2 October 2026, reported) - CENTCOM struck IRGC targets around the Strait of Hormuz on 1 September. ([CNBC](https://www.cnbc.com/2026/09/01/us-strikes-iran-after-new-hormuz-strait-shipping-attacks-centcom.html), 1 September 2026, reported) - CNN's sources say limited strikes on radars and launchers have been the pattern since Trump halted full combat operations. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) - Stars and Stripes reported Makin Island's destination unconfirmed and the size of any buildup unclear. ([Stars and Stripes](https://stripes.com/branches/navy/2026-10-02/makin-island-middle-east-iran-strikes-november-23032862.html), 2 October 2026, reported) ![Shayne Heffernan's graphic: the escalation timeline, October 2026, with its dates corrected to the sources above](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-f7c1900e14a170c3.png) The alliances around the war have hardened. Saudi Arabia, Pakistan and Turkey signed a joint defence pact in [Mecca on 7 August](https://www.aljazeera.com/news/2026/8/7/turkiye-saudi-arabi-pakistan-sign-joint-defence-agreement-whats-in-it) and named it the [Mecca Defence Alliance](https://www.aljazeera.com/news/2026/8/31/pakistan-saudi-arabia-turkiye-meet-whats-next-for-their-defence-pact) on 31 August, building on the Saudi-Pakistan agreement of [September 2025](https://www.aljazeera.com/news/2025/9/18/watershed-how-saudi-pakistan-defence-pact-reshapes-regions-geopolitics). It was invoked on [5 October](https://www.aljazeera.com/news/2026/10/5/mecca-alliance-agrees-to-activate-collective-deterrence-measures), when Turkey and Pakistan agreed to deploy forces to Saudi Arabia against the Houthis, and Pakistan's military has confirmed its troops are [in the kingdom](https://unb.com.bd/category/World/pakistan-confirms-troops-deployed-in-saudi-arabia-under-makkah-defense-pact/197332). The Saudi-led coalition is flying [100 fighter jets](https://www.aljazeera.com/news/2026/10/5/mecca-alliance-agrees-to-activate-collective-deterrence-measures) in support of the Yemeni government's Dawn of Yemen offensive, which has turned towards [Houthi-held Sanaa](https://www.cnn.com/2026/10/04/middleeast/yemen-offensive-houthis-intl). The State Department calls the Houthis ["Iranian-supported"](https://www.thestar.com.my/news/world/2026/10/22/washington-orders-us-embassy-staff-families-to-leave-saudi-arabia), and Iran denies [directing regional armed groups](https://thenationaldesk.com/news/nation-world/iranian-assassination-plots-targeting-american-officials-are-in-advanced-stages-reports-intelligence-memo-houthi-rebels-war-conflict-nuclear-weapon-gasoline-gas-prices-oil-fuel-economy-inflation-middle-east-marco-rubio-jay-clayton-president-donald-trump). The diplomatic track has not closed. Saudi Arabia's defence minister, Pakistan's army chief and Turkey's intelligence chief held talks in [Tehran on 26 October](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl), which Riyadh said support [mediation by Qatar and Pakistan](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl); Pakistan has been [a key intermediary between Washington and Tehran](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl) since the war began. Iran's president accuses Washington of [derailing the negotiations](https://www.aljazeera.com/news/2026/10/25/centcom-reportedly-has-plans-for-iran-strikes-will-trump-approve-them) and says efforts to reach an agreement should continue. Inside Washington, the Pentagon's leak hunt has [polygraphed about 50 Joint Staff members](https://www.yahoo.com/news/politics/articles/paranoid-pete-takes-pentagon-leak-153945906.html) over stories on the war. ![Exhibit 4, who stands where in the Iran war: the US and Israel, Iran and the Houthis, the Mecca Defence Alliance and the diplomatic track](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-6d2c2bc404db26a8.svg) **Exhibit 4, its links and their sources.** - The New York Times reported that the US and Israel have drawn up joint plans for an attack on Iran after the midterm elections. ([Al Jazeera, on the New York Times](https://www.aljazeera.com/news/2026/10/25/centcom-reportedly-has-plans-for-iran-strikes-will-trump-approve-them), 25 October 2026, reported) - The US naval blockade of Iranian ports is in force; CENTCOM's commander visited the Strait of Hormuz on 25 October. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) - About 50 Joint Staff members were polygraphed in a hunt for the source of stories on military matters, the New York Times reported. ([The Daily Beast, via Yahoo News](https://www.yahoo.com/news/politics/articles/paranoid-pete-takes-pentagon-leak-153945906.html), 5 September 2026, reported) - Israel appears to have delayed immediate strikes and is building up for potential joint operations with the US, likely after the midterms, CNN's sources say. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) - Iran's Assembly of Experts named Mojtaba Khamenei the third Supreme Leader on 8 March. ([AFP, via Prothom Alo](https://en.prothomalo.com/international/middle-east/yym9wflmzj), 9 March 2026, reported) - The State Department describes the Houthis as "Iranian-supported" and blames them for attacks on Saudi civilian airports and critical infrastructure. ([AFP, via The Star](https://www.thestar.com.my/news/world/2026/10/22/washington-orders-us-embassy-staff-families-to-leave-saudi-arabia), 21 October 2026, reported) - Saudi Arabia, Pakistan and Turkey signed the joint defence pact in Mecca on 7 August 2026. ([Al Jazeera](https://www.aljazeera.com/news/2026/8/7/turkiye-saudi-arabi-pakistan-sign-joint-defence-agreement-whats-in-it), 7 August 2026, reported) - The Saudi-led coalition confirmed it is providing air support to the Yemeni government's offensive with 100 of its fighter jets. ([Al Jazeera](https://www.aljazeera.com/news/2026/10/5/mecca-alliance-agrees-to-activate-collective-deterrence-measures), 5 October 2026, reported) - Yemen's government forces have begun what they call a strategic offensive operation against Houthi-held Sanaa. ([CNN](https://www.cnn.com/2026/10/04/middleeast/yemen-offensive-houthis-intl), 4 October 2026, reported) - Pakistan's military spokesman confirmed in October that Pakistani troops are deployed in Saudi Arabia under the joint defence agreement. ([UNB, on Arab News](https://unb.com.bd/category/World/pakistan-confirms-troops-deployed-in-saudi-arabia-under-makkah-defense-pact/197332), 8 October 2026, reported) - The alliance was invoked on 5 October, when Turkey and Pakistan agreed to deploy forces rapidly to Saudi Arabia to counter Houthi attacks. ([Al Jazeera](https://www.aljazeera.com/news/2026/10/5/mecca-alliance-agrees-to-activate-collective-deterrence-measures), 5 October 2026, reported) - Saudi Arabia's Foreign Ministry said the Tehran visit supports mediation by Qatar and Pakistan. ([CNN](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl), 26 October 2026, reported) - Saudi Arabia's defence minister, Pakistan's army chief and Turkey's intelligence chief held talks in Tehran on 26 October. ([CNN](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl), 26 October 2026, reported) - Iran has repeatedly denied directing attacks by regional armed groups. ([The National News Desk](https://thenationaldesk.com/news/nation-world/iranian-assassination-plots-targeting-american-officials-are-in-advanced-stages-reports-intelligence-memo-houthi-rebels-war-conflict-nuclear-weapon-gasoline-gas-prices-oil-fuel-economy-inflation-middle-east-marco-rubio-jay-clayton-president-donald-trump), 24 October 2026, reported) - Pakistan has been a key intermediary between Washington and Tehran since the war started in February. ([CNN](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl), 26 October 2026, reported) ![Shayne Heffernan's graphic: the actor knowledge graph of the Middle East conflict network, October 2026, corrected to the sources above](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-8ec9fbedf307f580.png) The force is already in place. USNI places two carrier strike groups, USS George Washington and USS George H.W. Bush, in the [Arabian Sea](https://news.usni.org/wp-json/wp/v2/posts?slug=usni-news-fleet-and-marine-tracker-oct-26-2026&_fields=date,link,title,content), with the Boxer amphibious ready group and the 11th Marine Expeditionary Unit beside them, and the Navy has kept two carriers there for [six months](https://news.usni.org/wp-json/wp/v2/posts?slug=october-marks-highest-number-of-strait-of-hormuz-attacks&_fields=date,link,title,content). More than 50,000 US personnel are in the region. US officials say the Theodore Roosevelt strike group and the Makin Island amphibious group, with the 13th Marine Expeditionary Unit of [more than 2,000 Marines](https://euronews.com/2026/10/02/us-sends-third-carrier-strike-group-to-middle-east-in-biggest-buildup-since-2003), will arrive by the end of November, adding [9,000 to 10,000 sailors and Marines](https://euronews.com/2026/10/02/us-sends-third-carrier-strike-group-to-middle-east-in-biggest-buildup-since-2003) between them, per AP. As of 26 October USNI placed the Roosevelt [in the Philippine Sea on exercise and the Makin Island in Yokosuka](https://news.usni.org/wp-json/wp/v2/posts?slug=usni-news-fleet-and-marine-tracker-oct-26-2026&_fields=date,link,title,content), Japan; when the groups sailed, the Navy confirmed only [routine operations](https://stripes.com/branches/navy/2026-10-02/makin-island-middle-east-iran-strikes-november-23032862.html). Roosevelt is expected to [relieve the George Washington](https://stripes.com/branches/navy/2026-10-02/makin-island-middle-east-iran-strikes-november-23032862.html), which returns the count to two carriers once the rotation ends, and Maritime Executive expects weeks of overlap with US naval and Marine forces at [nearly double strength](https://maritime-executive.com/article/us-sends-third-carrier-strike-group-to-the-gulf). Two US officials told CNN that [sufficient forces are already deployed](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump) to carry out the potential attacks. On land, US forces work from Al Udeid in Qatar, Al Dhafra in the UAE and [Prince Sultan Air Base](https://gulfnews.com/world/mena/iranian-attack-on-saudi-base-wounds-at-least-10-us-troops-and-damages-several-planes-1.500488785) in Saudi Arabia, where an Iranian missile wounded at least 10 US troops on 27 March, and they [left Al Asad](https://www.militarytimes.com/news/pentagon-congress/2026/01/20/us-completes-withdrawal-from-al-asad-airbase-iraq-says/) in Iraq in January. ![Exhibit 5, US forces around Iran: two carrier groups and an amphibious group in the Arabian Sea, the reinforcements reported due by late November and the land bases](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-95d5fb6b5ba4b0a4.svg) **Exhibit 5, its links and their sources.** - USNI reports the Navy has kept two carriers in the Arabian Sea for six months, and an amphibious group with a Marine expeditionary unit on station since early March. ([USNI News](https://news.usni.org/wp-json/wp/v2/posts?slug=october-marks-highest-number-of-strait-of-hormuz-attacks&_fields=date,link,title,content), 22 October 2026, reported) - USNI places two US carrier strike groups in the CENTCOM area, both in the Arabian Sea: USS George Washington and USS George H.W. Bush. ([USNI News Fleet and Marine Tracker](https://news.usni.org/wp-json/wp/v2/posts?slug=usni-news-fleet-and-marine-tracker-oct-26-2026&_fields=date,link,title,content), 26 October 2026, reported) - No US carrier is reported in the Persian Gulf: USNI places both carriers in the Arabian Sea. ([USNI News Fleet and Marine Tracker](https://news.usni.org/wp-json/wp/v2/posts?slug=usni-news-fleet-and-marine-tracker-oct-26-2026&_fields=date,link,title,content), 26 October 2026, reported) - One amphibious ready group is in the region, the Boxer group with the 11th Marine Expeditionary Unit, operating in the Arabian Sea. ([USNI News Fleet and Marine Tracker](https://news.usni.org/wp-json/wp/v2/posts?slug=usni-news-fleet-and-marine-tracker-oct-26-2026&_fields=date,link,title,content), 26 October 2026, reported) - In the Red Sea the US has two guided-missile destroyers and no carrier or amphibious group. ([USNI News Fleet and Marine Tracker](https://news.usni.org/wp-json/wp/v2/posts?slug=usni-news-fleet-and-marine-tracker-oct-26-2026&_fields=date,link,title,content), 26 October 2026, reported) - AP estimated the combined deployment at 9,000 to 10,000 more troops by the end of November. ([Euronews, on AP](https://euronews.com/2026/10/02/us-sends-third-carrier-strike-group-to-middle-east-in-biggest-buildup-since-2003), 2 October 2026, reported) - Roosevelt is expected to relieve the carrier George Washington, so the number of carriers may stay at two once the rotation ends. ([Stars and Stripes](https://stripes.com/branches/navy/2026-10-02/makin-island-middle-east-iran-strikes-november-23032862.html), 2 October 2026, reported) - The Makin Island group carries the 13th Marine Expeditionary Unit of more than 2,000 Marines. ([Euronews](https://euronews.com/2026/10/02/us-sends-third-carrier-strike-group-to-middle-east-in-biggest-buildup-since-2003), 2 October 2026, reported) - US forces left Ain al-Asad air base in Iraq in January 2026, and the Iraqi army has taken full control. ([Military Times](https://www.militarytimes.com/news/pentagon-congress/2026/01/20/us-completes-withdrawal-from-al-asad-airbase-iraq-says/), 17 January 2026, reported) - Al Udeid in Qatar and Al Dhafra in the UAE host US forces, and Iranian state media said the IRGC targeted both on 28 February. ([Bloomberg, via The Boston Globe](https://www.bostonglobe.com/2026/02/28/world/iran-targets-us-bases-in-gulf-in-retaliation-for-airstrikes), 28 February 2026, reported) - An Iranian missile attack on 27 March wounded at least 10 US service members at a military base in Saudi Arabia, Prince Sultan Air Base. ([AP, via Gulf News](https://gulfnews.com/world/mena/iranian-attack-on-saudi-base-wounds-at-least-10-us-troops-and-damages-several-planes-1.500488785), 27 March 2026, reported) - Incirlik, near Adana, is a Turkish base and a key NATO air base that has hosted US troops for decades. ([AFP, via Al-Monitor](https://www.al-monitor.com/originals/2026/03/too-risky-iran-hit-turkey-over-us-assets-analysts), 2 March 2026, reported) - The Iraqi Defence Ministry and Military Times reported the US withdrawal from Ain al-Asad in January. ([Military Times](https://www.militarytimes.com/news/pentagon-congress/2026/01/20/us-completes-withdrawal-from-al-asad-airbase-iraq-says/), 17 January 2026, reported) - The Theodore Roosevelt strike group brings more than 7,000 sailors. ([Straight Arrow News, on the WSJ](https://san.com/cc/us-sends-another-aircraft-carrier-thousands-more-troops-to-middle-east/), 2 October 2026, reported) - As of 26 October USNI placed the Roosevelt in the Philippine Sea on exercise and the Makin Island in Yokosuka, Japan. ([USNI News Fleet and Marine Tracker](https://news.usni.org/wp-json/wp/v2/posts?slug=usni-news-fleet-and-marine-tracker-oct-26-2026&_fields=date,link,title,content), 26 October 2026, reported) ![Shayne Heffernan's graphic: CENTCOM force posture before and after the reported late-November reinforcements, corrected to the sources above](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-3120e764c91c2f61.png) ![Shayne Heffernan's graphic: the regional operations map of Iran, Yemen and the Gulf, October 2026, in two panels, corrected to the sources above](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-2688ea26356b7569.png) ## Four scenarios, and the market read of each Exhibit 6 sets out my four scenarios for US action on Iran, by timing and scale. The reporting points to the fourth. Officials told CNN that a go-order could come any time before the end of the year, [most likely after the midterms](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), and that Trump has not made up his mind, though they see him leaning towards major action. They also caution that Trump has [appeared in the past to lean towards actions only to back off at the last moment](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump). The New York Times reported that the US and Israel have drawn up [joint plans](https://www.aljazeera.com/news/2026/10/25/centcom-reportedly-has-plans-for-iran-strikes-will-trump-approve-them) for an attack after the vote. ![Exhibit 6, four scenarios for US action on Iran by timing and scale, with the reporting on each and the oil market's risk premium](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-50f3a30e5d06648c.svg) **Exhibit 6, its links and their sources.** - Trump wrote on 8 October that the US would not attack Iran before the midterm elections on 3 November. ([CNN](https://www.cnn.com/2026/10/08/politics/trump-iran-attack-election), 8 October 2026, reported) - Two US officials told CNN that sufficient forces are already deployed in the region to carry out the potential attacks. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) - CNN reports Trump has appeared reluctant to resume the war in earnest before the midterms, mindful of the political downsides. ([CNN](https://www.cnn.com/2026/10/08/politics/trump-iran-attack-election), 8 October 2026, reported) - Officials told CNN a go-order could come any time before the end of the year, most likely after the midterms, and that Trump has not decided. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) - Saudi Arabia said the Tehran visit supports mediation by Qatar and Pakistan. ([CNN](https://www.cnn.com/2026/10/26/middleeast/iran-meeting-regional-powers-intl), 26 October 2026, reported) - The New York Times reported that the US and Israel have drawn up joint plans for an attack on Iran after the midterm elections. ([Al Jazeera, on the New York Times](https://www.aljazeera.com/news/2026/10/25/centcom-reportedly-has-plans-for-iran-strikes-will-trump-approve-them), 25 October 2026, reported) - SEB's Bjarne Schieldrop said crude carries a significant risk premium for a US or Israeli attack on Iran within weeks. ([The Nation Thailand, market report](https://www.nationthailand.com/news/world/40072812), 27 October 2026, reported) - US officials caution that Trump has appeared in the past to lean towards actions only to back off at the last moment. ([CNN](https://www.cnn.com/2026/10/26/politics/us-iran-strike-plans-trump), 26 October 2026, reported) ![Shayne Heffernan's graphic: the scenario matrix for US action on Iran, October 2026, corrected to the sources above](https://livetradingnews-media.nyc3.digitaloceanspaces.com/media/2026/10/28/cmpgg3-d577141f32023676.png) Brent settled at [$113.88 on 27 October 2026](https://www.nationthailand.com/news/world/40072812), down $1.93 on the regional diplomacy, after an October high of [$116.72 on 23 October 2026](https://ts-api.cnbc.com/harmony/app/charts/1Y.json?symbol=%40LCO.1). October has had [the most attacks on ships](https://news.usni.org/wp-json/wp/v2/posts?slug=october-marks-highest-number-of-strait-of-hormuz-attacks&_fields=date,link,title,content) since the war began, 27 by 22 October according to the International Maritime Organization, and Kpler put Hormuz flows at about [14 million barrels a day](https://finance.yahoo.com/energy/articles/more-oil-flowing-persian-gulf-183858775.html) in the week to 18 October, 82% of pre-war volumes. SEB's Bjarne Schieldrop says crude carries a [significant risk premium](https://www.nationthailand.com/news/world/40072812) for a US or Israeli attack within weeks. My read of each scenario for investors: - **Q1, a limited strike before 3 November.** Politically costly, and the reporting says Trump has been [reluctant to resume the war](https://www.cnn.com/2026/10/08/politics/trump-iran-attack-election) before the vote. A short spike in crude, inside the premium already paid. - **Q2, a mass bombing campaign before 3 November.** The highest-risk case and the largest upside for oil, because it puts Hormuz flows back at risk while they are still below pre-war levels. - **Q3, a deal after the midterms.** The blockade holds and the talks run through Qatar and Pakistan. The risk premium comes out of crude, and European gas follows it down. - **Q4, a major campaign after the midterms.** The path the reporting leans to, with the officials' caution attached. Crude holds its premium into December, Gulf sovereign AI campuses stay inside the theatre, and the defence AI layer keeps its demand. In all four, the software, model, compute and space companies sit on budget lines that are becoming permanent. Maven is on track to become a programme of record, GenAI.mil runs at the scale of the whole department, and Golden Dome and the fiscal 2027 request fund the rest. Their demand does not depend on which scenario arrives. The scenarios move the price of the inputs: oil, gas and the risk premium on anything built in the Gulf. ## Where AI is being adopted Defence is now the most visible buyer, and it is not the only one. Financial services firms are among the most advanced adopters. Banks, insurers and asset managers use AI in fraud detection, anti-money-laundering, credit decisions, research summaries, code modernisation and client service. The constraint is regulatory: model risk management, explainability and data governance slow deployment, and they also create a moat for firms that build compliant systems early. Healthcare adoption is accelerating in administrative and documentation work, where ambient clinical note-taking, prior-authorisation processing and coding assistance offer clear returns. Diagnostic imaging and drug discovery are longer-cycle opportunities with significant regulatory gates. Manufacturing is moving from predictive maintenance and quality inspection towards AI in design, process optimisation and supply chain planning. Digital twins of factories and products, combined with simulation, let engineers test changes before committing capital. The return is real and slower to arrive, because it depends on joining AI to operational technology that was never designed to share data. Retail and consumer businesses use AI in demand forecasting, pricing, personalisation, customer service and content. Agentic commerce, in which software agents search, compare and buy on behalf of consumers, could change how retailers reach customers and how margins are shared with platforms. Energy is both a customer and a constraint. Utilities and oil and gas companies use AI for grid management, asset inspection, reservoir modelling and trading. AI demand is also reshaping the sector's investment cycle: it supports new gas generation, extends the life of nuclear plants and drives investment in transmission, storage and grid equipment. For energy companies, AI is as much a source of new load as a productivity tool, and in Europe the war has made that load more expensive to serve. Across the stack, infrastructure is where most AI dollars flow today: accelerators, memory, networking, servers, data centre construction, cooling, power equipment and the cloud operators that rent capacity. It has captured most of the economic value of the cycle so far, and it remains the most capital-intensive and cyclical layer. Software and platforms sit above it, where value is shifting from raw model access to the platforms that join models to enterprise data, identity and workflows. Services are a growing and underappreciated layer: systems integrators and managed service providers are being hired to move enterprises from pilots to production. Physical AI is the emerging frontier, with robotics, autonomous vehicles, drones and industrial automation driven by models that perceive and act in the physical world. Defence drones such as LUCAS are a form of physical AI already in combat. ## Outlook through the late 2020s The base case through 2028 and 2029 is continued growth in AI infrastructure spending, with the rate of growth moderating from the pace of 2024 to 2026 as the base gets larger and power limits bite. Inference will become the dominant share of AI compute, and spending will broaden from a handful of hyperscalers to sovereign buyers, enterprises, specialised clouds and defence ministries. Accelerator generations will keep an annual cadence, rack power densities will keep rising, and liquid cooling will become standard for new builds. Pricing will keep diverging. Premium systems will hold value while supply is tight, and competition from AMD and custom silicon, with software portability, will gradually narrow margins at the top. Unit inference costs will keep falling and widen adoption. Memory and packaging capacity will expand and the current extreme tightness should ease over time, though each new generation will reset the cycle. The principal risks are clear. A financing shock or a slowdown in monetisation could force hyperscalers to pause capital spending and produce a sharp inventory cycle for suppliers. Power and permitting delays could strand capital in partly energised campuses: two finished data centres in Santa Clara may [sit empty for years](https://fortune.com/2025/11/10/nvidia-hometown-santa-clara-california-data-centers-empty-power-grid/) waiting for power. Disruption around Taiwan, export controls or trade policy could interrupt supply. Circular financing arrangements, in which suppliers invest in customers who then buy their products, deserve scrutiny because they can overstate end demand: NVIDIA committed [more than $40 billion](https://thenextweb.com/news/nvidia-40bn-ai-equity-investments-2026) to AI equity stakes in the first four months of 2026, $30 billion of it in OpenAI, and AMD's warrant for up to [160 million of its shares](https://www.amd.com/en/newsroom/press-releases/2025-10-6-amd-and-openai-announce-strategic-partnership-to-d.html) vests as OpenAI's purchases scale. The Gulf adds a fifth risk: a wider war raises the price of the input AI depends on most, and it keeps sovereign AI campuses in the Gulf inside the theatre, as the [March strikes on AWS](https://www.cnbc.com/2026/03/02/amazon-says-drone-strikes-damaged-3-facilities-in-uae-and-bahrain.html) showed. On the upside, a breakthrough in agentic reliability could accelerate enterprise adoption faster than current plans assume. For investors and operators, the practical implications are to favour scarce physical inputs and companies with durable distribution, to treat power access as a strategic asset, and to price in steady deflation at the application interface. The winners of the late 2020s will be those who turn capital into energised, productive compute fastest, and those who turn falling token costs into products customers pay for. In defence, the winners will be the companies whose AI a commander can trust and an auditor can check. ## Conclusion The AI market in October 2026 is defined by real demand meeting hard physical limits. Demand has broadened from training to inference and from experiments to production. Prices have split between premium hardware and deflating inference. Breakthroughs in rack-scale systems, memory, packaging and software keep raising capability per dollar. Power has become the decisive constraint, and the set of chip suppliers is widening while it remains concentrated on a single foundry. The war has shown both halves of the market at once. AI helped US forces strike [more than 2,000 targets in the first 96 hours](https://www.defensenews.com/industry/techwatch/2026/10/21/marine-corps-makes-maven-available-service-wide/) after 28 February 2026, and the power and data centres that AI depends on came under fire in the same theatre, from the [drone strikes on AWS](https://www.cnbc.com/2026/03/02/amazon-says-drone-strikes-damaged-3-facilities-in-uae-and-bahrain.html) to the gas price. None of this suggests the cycle is ending. It suggests that the next phase will be decided less by who has the best model and more by who can build, power and operate capacity, and deliver it to industries and governments in a form they can trust. That is a more industrial, more capital-intensive and more durable market than the one that began in 2023. It rewards discipline above hype. ## What to watch - Alphabet and Microsoft report today and Amazon tomorrow, and their capital spending guidance is the next read on demand, [per CNBC's calendar](https://www.cnbc.com/2026/10/21/meta-earnings-q3-2026-live-updates.html). - Israel's final tally is expected by 30 October and official results on [4 November](https://www.aljazeera.com/news/2026/10/28/five-key-takeaways-from-israels-parliamentary-elections). - Palantir reports on [2 November](https://finance.yahoo.com/markets/stocks/articles/palantir-announces-date-third-quarter-200500055.html), and AMD on [3 November](https://ir.amd.com/news-events/press-releases/detail/1300/amd-to-report-fiscal-third-quarter-2026-financial-results), the day of the [US midterms](https://www.cnn.com/2026/10/08/politics/trump-iran-attack-election). - The D.C. Circuit's answer to Anthropic's [rehearing petition](https://www.unite.ai/anthropic-asks-d-c-circuit-to-rehear-pentagon-supply-chain-risk-case/), and Anthropic's Nasdaq listing, which Fortune reports is expected in [Thanksgiving week](https://fortune.com/2026/10/26/anthropics-s-1/). - The Theodore Roosevelt and Makin Island groups, [reported due by late November](https://stripes.com/branches/navy/2026-10-02/makin-island-middle-east-iran-strikes-november-23032862.html), and whatever order follows the vote. Stocks mentioned: $PLTR (NASDAQ), $NVDA (NASDAQ), $SPCX (NASDAQ), $GOOGL (NASDAQ), $AMD (NASDAQ), $AVGO (NASDAQ), $MSFT (NASDAQ), $AMZN (NASDAQ), $META (NASDAQ), $MU (NASDAQ), $AVAV (NASDAQ), $ORCL (NYSE), $TSM (NYSE), $GEV (NYSE), $LMT (NYSE), $RTX (NYSE), $NOC (NYSE) and $LHX (NYSE). Shayne Heffernan, Ph.D., is the founder of the KnightsBridge Group ([shayneheffernan.com](https://www.shayneheffernan.com)). *Each figure links to its source, read on 28 October 2026. This article is analysis, not a solicitation to buy or sell any security.* --- This Markdown mirrors https://www.livetradingnews.com/the-state-of-the-ai-market-in-wartime-chips-power-and-the-defence-ai-companies. The HTML page is canonical. Site index for AI clients: https://www.livetradingnews.com/llms.txt