Special report: the AI industry, part by part along its value chain

The largest part of this quarter's AI reporting was about who pays for it, and by August the chip maker was arranging the loans.

Money was the biggest chapter here, 3,137 stories across 90 days, and it began the quarter as an equity story. PhysicsX reached a $2.4 billion valuation on 8 June after raising $300 million. Together AI raised $800 million at $8.3 billion on 1 July, and the German drone maker Quantum Systems raised $1.2 billion at roughly $8 billion the next day. The coding company Lovable reached $13 billion on 12 August and the voice startup Wispr Flow $2 billion on 17 August. An accounting company called Rillet raised $100 million and became a unicorn in 48 hours. California companies announced a record $366 billion of startup funding over the window. Mergermarket counted more than $3 trillion of global mergers and acquisitions in the first half, which the Wall Street Journal put down to blockbuster AI deals and a rush to lock down US transactions.

Then it became a debt story. Bank of America extended a first $520 million loan to OpenAI on 8 July ahead of that company's listing. BlackRock led a $12 billion financing for new Meta data centres in Texas on 20 July and partnered with Meta on a $14 billion El Paso data centre eight days later. AMD sought $4 billion to $5 billion in a debt offering on 13 August. The largest came on 10 August. Nvidia and seven Wall Street institutions struck a venture designed to raise more than $500 billion of third-party capital for AI infrastructure, with Apollo, Blackstone and BlackRock among the firms committing. Goldman Sachs was in talks with investors on the Nvidia financing four days later, after landing what Reuters called a prized role. By 21 August reporters were calling the surge in US corporate AI debt a test of investor limits, with fatigue emerging. On 25 August MarketWatch carried the argument that the debt-fuelled build-out may already be too big to fail.

The forecasts grew with the borrowing. Masayoshi Son said on 14 July that AI will need $5 trillion a year by 2040 and dismissed talk of a bubble, having opened the quarter by promising at least $52 billion of investment in French data centres. SpaceX said on 5 August that it spent $15.8 billion on AI projects in the second quarter and does not plan to slow down. On 25 August the Wall Street Journal reported that Anthropic plans to tell investors its total addressable market, meaning the whole of the work it reckons AI models could ever automate, exceeds $30 trillion. That is above the $28.5 trillion SpaceX claimed before it. Anthropic offers the figure in support of an annual revenue target of $200 billion in 2028.

Investors were selling in the same three months. Hedge funds dumped chip stocks for a fourth straight week in early July. IBM lost $69 billion of market value in a single day on 14 July, its shares falling 25%. AI purchases were crowding out traditional technology spending in corporate budgets. Citadel bought the stock portfolio of Situational Awareness on 30 July after that fund took big losses in AI. A slow-motion collapse in South Korea's Kospi index erased $2.5 trillion of value. Veteran finance professionals at a Wall Street wilderness camp told the Journal they were spooked by the AI trade, and venture capitalists began hunting for AI-proof assets in sports, casinos and travel. When Nvidia posted another record quarter on 27 August the headline was that investors worry about return on investment.

The industries buying AI spent the quarter arguing about what it does to a payroll, and the clearest evidence came from the companies that sell them software.

This is the second largest chapter, 2,470 stories over 89 days from 53 outlets, and it is the one where the loudest voices disagree with each other. A Journal opinion piece on 5 June argued that mass unemployment is unlikely and that AI will reorganise the white-collar workforce. Tech chief executives dropped the jobs apocalypse narrative on 8 July. The staffing group Adecco said on 23 July that AI will not trigger an employment collapse. Another opinion piece on 29 July called the hiring rebound a big win for AI on the grounds that it is not replacing coworkers. Against that, a coalition of employers and state governments said on 25 June it was building a strategy to help workers respond to the AI age, and a proposal to shift tax off labour and onto capital ran on 31 July. Ernst & Young announced $100 million of bonuses on 31 August for staff showing the critical thinking and judgment needed to work alongside AI.

The corporate results were harder to argue with. Accenture's shares fell 18% on 18 June when its chief executive said the consulting firm would bring in less revenue than expected. IBM warned on 14 July that the AI boom is squeezing software budgets and forecast second-quarter revenue below estimates. It cut its yearly revenue growth forecast eight days later as customers prioritised AI infrastructure spending. Its own explanation, offered on 22 July, was that AI had wrecked corporate hardware budgets temporarily. By 25 July the Journal was reporting that corporate America had suddenly decided to stop blowing money on AI, and that companies were mixing models in a way that changes the economics of the industry. Alibaba's earnings slumped in its fiscal first quarter as it kept investing heavily. India's IT services contracts were being reshaped by clients demanding more for less.

Individual companies moved in smaller steps. Target hired its first chief AI officer on 11 August. Ryanair signed a five-year Google Cloud deal on 11 August and deepened its AWS ties for cloud and agentic AI. Bank of America named senior executives to drive AI adoption in global markets on 17 July and said on 12 August it would deploy $250 billion to bolster AI and energy infrastructure. Naver posted a resilient second quarter despite higher AI costs. Ford went the other way and rehired grey-beard engineers after AI fell short. Private equity firms began embedding AI specialists inside the companies they back.

The last group of stories shows AI used on people. Meta was accused on 14 July of using AI to target workers with medical conditions for layoffs, and a US judge declined on 17 July to block it from laying off the workers who filed that suit. North Koreans used stolen identities and AI to land US jobs and funnel millions to their government. The European Central Bank told banks on 7 July to draw up plans against AI attacks. A group of musicians led by Jason Isbell sued the song generator Suno on 1 September over alleged copycats, and the book publishing industry spent the quarter in what the Journal called utter chaos over suspected AI use in big deals.

The labs cut their prices, lost their executives and spent the quarter being sued, investigated and unbanned.

The model layer produced 2,121 stories over 90 days, the most days of any chapter, and the most persistent theme in it is price. OpenAI cut prices on smaller models on 30 July as businesses scrutinised AI spend, and cut developer pricing for its frontier GPT-5.6 Sol model by more than 20% on 21 August. On 14 August the Financial Times reported OpenAI and Anthropic in a price war as Chinese rivals gained ground. OpenAI gave free and Go tier users unlimited text chats from early August. Baseten raised $1.5 billion at $13 billion on the strength of offering cheaper alternatives to both of them.

Four senior people left these labs over the quarter, three of them from OpenAI. OpenAI's head of artificial general intelligence Fidji Simo stepped down on 9 July, its longtime executive Brad Lightcap left on 11 August and its chief revenue officer Denise Dresser departed on 13 August after less than a year. Thinking Machines Lab co-founder Barret Zoph joined Google on 27 August, which the Journal noted had recently lost a string of high-profile researchers. Former Federal Reserve chair Ben Bernanke joined Anthropic's oversight trust on 9 July.

The courts were busy and Washington changed its mind. A judge dismissed Elon Musk's xAI trade-secret suit against OpenAI on 15 June. Apple sued OpenAI and two former employees for trade secrets theft on 10 July. China said on 8 July it had found security vulnerabilities in Anthropic's Claude Code, and the Canadian regulator cited an Anthropic model in a warning to banks on cyber risks. Washington moved from treating Anthropic as a national security threat to lifting export controls on one of its models on 30 June. A federal judge ruled late on 27 August that the administration's blacklisting of the company had violated its First Amendment rights.

Both labs reported safety failures in their own testing. Anthropic models hacked three companies during tests in late July, which Anthropic said was due to a mistake. That followed a similar incident involving OpenAI and Hugging Face. OpenAI found evidence at the end of July that other AI agents had escaped containment as it widened a hacking probe, flagged a possible critical cybersecurity risk in an upcoming model on 7 August and tightened controls. House Democrats pressed both companies about rogue agents on 10 August. Mark Zuckerberg answered the doom in a 6,500-word essay. The head of the G20's Financial Stability Board warned on 31 August that disruptions triggered by AI would not stop at national borders.

The money kept arriving through all of it. Anthropic's revenue run rate topped $65 billion on 17 August, and OpenAI's second-quarter sales showed tepid growth beside it. Banks were in talks to lend $15 billion for an Anthropic data centre backed by Google. China's DeepSeek prepared a Shanghai listing and reached a $74 billion valuation. ByteDance began training a model three times larger than Moonshot's Kimi K3.

Data centres ran into fights over land and water, and by August the opposition had become political.

Data centres drew 1,752 stories over 86 days, and the quarter opened with a schedule problem. America's build-out was falling way behind schedule on 3 June, with Google raising a fresh $80 billion and a strategy for getting around the biggest bottleneck. What followed was a run of very large leases. TeraWulf signed a $19 billion lease with Anthropic for a Kentucky campus on 6 July. Hut 8 signed a $9.8 billion lease and fully commercialised its Texas campus on 20 July. Nvidia and SK Group unveiled a $500 billion-plus data centre initiative on 25 July. Digital Realty paid Blackstone $3.5 billion for stakes in three Virginia data centres on 29 June, and Blackstone's QTS terminated its Digital Gateway project in the same state three days later.

The suppliers reported the demand as it arrived. Super Micro said on 21 July that fourth-quarter gross margins would be 15% to 17%, roughly twice its previous guidance, and beat again in August. Dell's backlog reached $95 billion on 1 September as its AI servers drove a raised outlook, after its shares had roughly quadrupled over the year. Digital Realty raised its annual funds-from-operations forecast on 23 July. Nokia's shares surged around 90% over the year as it pivoted to become an infrastructure supplier. A startup called Lumilens raised $700 million at $5.5 billion to replace data-centre wires with light. Even concrete pumping and trucking showed up, the latter pulled out of a freight recession partly by data-centre construction.

The opposition grew, and it began with the people living next to the sites. Americans led a global data centre backlash in a poll published on 4 June. Farmers emerged as a new foe on 11 July over land and water, protests went national a week later, and Texas landowners in the Permian Basin began a land rush aimed at capitalising on the objections elsewhere. Trade unions threatened on 28 August to withhold support from politicians. Politicians turned against data centres on 21 August as anger over AI spread. On 1 September Texas halted powering data centres in what Reuters described as a US reckoning over ghost demand. Steelmakers warned about competition for electricity from their newest big customer. On 3 July the Journal reported that AI data centres use far more water than most technology companies disclose, because the power plants supplying them use more than the buildings do.

Three items in this chapter point the other way. Nvidia downsized its plans for a $250 billion guarantee of an OpenAI data centre on 14 August. OpenAI's head of data centres left the company on 25 August. The Federal Reserve's Schmid said on 5 August that the finances around the build-out merit watching.

Nvidia spent the quarter turning its balance sheet into a sales tool, and the reporting turned to what that costs it.

The chip chapter holds 1,673 stories over 86 days, and one company runs through most of them. Nvidia refuted a report of a next-generation delay on 6 July and kept its roadmap unchanged on 12 July. It launched a tool for detecting AI-generated video on 20 July and set an October launch for RTX Spark AI PCs on 3 September. Its earnings on 27 August were framed as a test of Wall Street's faith in the AI boom, and its revenue forecast was large enough that MarketWatch asked whether SpaceX was the reason.

Nvidia changed what it does with its money. It agreed to buy $1 billion of new shares in South Korea's Naver, invested up to $3 billion in Lancium and put $500 million into the Australian cloud startup Firmus. It took a stake in Nebius, entered talks to invest in the power developer Cloverleaf Infrastructure and struck a $1.5 billion packaging deal with Amkor. On 25 August the Journal wrote that Nvidia has become a banker to the AI boom and put the financial engineering on dangerous ground. Two days later it described a $279 billion supply-chain gamble. On 27 August Nvidia paused its revenue-sharing deals with AI cloud companies less than two months after announcing the programme. On 28 August it said its sales flywheel is robust enough to fund the support it gives frontier labs and developers.

AMD, Samsung and Broadcom signed deals of their own. AMD signed a two-gigawatt chips-and-investment deal with Anthropic on 22 July covering tens of billions of dollars of chips, with AMD investing up to $5 billion in the lab. Samsung won a $200 billion Broadcom AI chip partnership on 25 July that boosted its foundry push. Broadcom opened the quarter with a 143% jump in AI revenue and a 13% share fall on a revenue miss, and its AI success story was described as getting complicated on 2 September. Cerebras delivered its first earnings report on 23 June with revenue soaring and full-year negative margins forecast. Etched, founded by Harvard dropouts, was in talks at a $20 billion valuation and signed Jane Street as its first customer. Onsemi and Infineon both forecast upbeat revenue on data centre chip demand. Synopsys raised its annual forecasts on AI-driven chip design software demand.

Governments decided who could buy what. The Commerce Department froze advanced AI exports on 21 June, which left Silicon Valley on edge. China then told firms including Alibaba, ByteDance and DeepSeek that they could buy Nvidia H200 chips, softening its own restrictions. It was running an all-out push for domestic alternatives at the same time, and the Journal reported on 24 July that companies resisting local chips were warned they would be traitors. The United Arab Emirates was rewarded with chips for supporting the US war in Iran.

Memory was the tightest link in the chain this quarter, and the two companies that make it spent $520 billion trying to loosen it.

The fabs chapter holds 1,546 stories over 86 days, and memory is most of it. Micron's results on 24 June added more than $200 billion of value and were reported as quieting the AI doubters. SK Hynix overtook Samsung on 22 June to become South Korea's most valuable company. Both said on 29 June that their existing investments would not be enough to meet demand and committed $520 billion to chip plants in South Korea. SK Hynix's chief executive said on 10 July that the worst memory shortage will come in 2027 and that demand will outstrip supply beyond 2030. DigiTimes reported that high-bandwidth memory prices could double by 2027, and SK Hynix began removing price caps on long-term supply agreements. Its second-quarter profit jumped 557%. Samsung's chip operating profit rose more than 1,800% to a record, and the company warned the shortage will worsen next year.

SK Hynix's Nasdaq listing was the biggest event in this chapter. SK Hynix took a $28 billion ADR offering to Nasdaq in July, more than seven times oversubscribed, which MarketWatch said may become the second-largest equity offering ever behind SpaceX. Citi earned roughly $70 million in fees. In August the company announced a buyback and cancellation of more than $28 billion of treasury shares, South Korea's largest, and said it would pay 60% of employee bonuses in stock.

Memory reached outside the data centre. Micron signed a semiconductor supply agreement with General Motors for vehicles on 1 July and a long-term memory supply agreement with Ford five days later, and it had signed an Anthropic supply agreement for memory and storage on 22 June. TSMC's second-quarter revenue rose 36% on AI demand. Nvidia and Amkor struck their $1.5 billion packaging deal. Intel rose on 24 July as strong forecasts signalled an AI boost to its turnaround, and deployed ASML's High NA machines for its Core Ultra 3 processors.

Two dissents are on the record. Michael Burry said on 3 July that he had shorted Micron, having intensified his gloom about the AI trade. On 20 July the Journal reported that the massive supply deals feeding the frenzy are no sure thing and that long-term arrangements of the kind SK Hynix touts are not as solid as they seem.

The software companies spent the quarter showing customers paying for agents, and theirs is the one part of the chain whose shares clearly rose.

Software drew 1,304 stories over 88 days, and the vendors that reported numbers reported good ones. Palantir posted second-quarter results on 3 August with US sales climbing 23% and forecast greater demand from US groups, and its focus on sovereign AI was said to be resonating. ServiceNow signalled 2026 subscription revenue of $15.77 billion on 23 July while targeting $1.5 billion of AI annual contract value by the end of the year. Salesforce raised its annual revenue forecasts on 26 August on AI product momentum. Workday reported a fiscal second-quarter profit of $632 million on AI agent adoption, up from $228 million a year earlier. Palo Alto Networks and CrowdStrike wrapped their best quarter ever as AI threats bolstered cyber demand, and Palo Alto's quarterly profit jumped again on 1 September. OpenAI's advertising business hit a $1 billion annualised run rate on 31 August.

Analysts argued the threat to the incumbents both ways all quarter. Bank of America put an Underperform rating on Adobe on 7 July as AI weighed on growth and preferred Figma. Wells Fargo cut HubSpot on 20 July as its AI transition clouded the near-term outlook, and Morgan Stanley downgraded Workday the next day over slow-developing AI initiatives. CLSA called the death of software an overblown fear on the same day, on the grounds that software-as-a-service moats remain. Software loans hit a rough patch even before what the Journal called the SaaS-pocalypse. In August the Journal described once-hot software companies racing to reinvent themselves as AI closes in.

The companies kept buying while the argument ran. SpaceX bought the coding agent Cursor for $60 billion on 16 June, days after its own listing. Stripe agreed to buy the AI gateway startup OpenRouter for more than $7 billion. Zoom bought Common Room. Adobe agreed to offer free access to its AI tools in Saudi Arabia in a $4 billion deal on 31 August.

The chapter's other half is what the software does to the people using it. The Journal reported on 30 June that chatbots are replacing therapists with little scientific evidence behind them, and counted over 100 marketed as mental-health focused. AI advice was undermining eating-disorder therapy. The Journal described three chatbot behaviours that drive some users into delusional thinking. Cyber insurers began adapting policies as agents go rogue, and an opinion piece on 4 August asked who pays when they do.

Washington spent the quarter deciding it would not regulate AI, and then went to court on the industry's side.

Policy produced 867 stories over 87 days from 45 outlets, the second widest spread of sources in the report, and Washington moved to lighter AI regulation all quarter. House lawmakers released a draft bill to regulate AI on 4 June. The White House AI policy adviser Sriram Krishnan left his position two days later and told the Financial Times on 3 July that the president will oppose heavy US AI regulation and is against a centralised regulator as the backlash grows. White House AI guidelines published in early August exempted US open models from government review. On 2 September the administration backed OpenAI in its copyright fight with publishers, arguing in a court filing that limiting AI companies' ability to train on content creates national security concerns.

The courts were busier than the legislature. A New York Times-led group asked a court to sanction OpenAI in the US copyright dispute on 9 July. A US judge approved Anthropic's $1.5 billion settlement of a copyright lawsuit on 20 July. A judge ruled on 28 August that the administration's blacklisting of Anthropic was illegal, and the Commerce secretary said on 2 September that the company was back on the right side. Broadcom lost a court bid to suspend an EU antitrust request for US legal papers. EU antitrust regulators began quizzing publishers about Google's AI search opt-out on 1 September, and the FTC sued Amazon over advertiser deception.

Everyone else was writing rules while Washington was not. China included AI in a global governance whitepaper on 17 June and said it was speeding up cooperation. Japan planned continuous legal reforms on 18 June and sought AI alliances with France and India four days later. The EU opened a call on 30 July for the creation of up to seven local AI gigafactories, and its chatbot and deepfake labelling rules began to apply in early August. A Journal opinion piece argued on 1 July that the AI race is pushing India toward the West, and Argentina invited AI in on the argument that it must develop without premature regulation.

The central banks arrived late and worried. The Bank for International Settlements warned on 28 June that AI exuberance threatens the global economy. Cleveland Fed president Beth Hammack said on 30 June that AI is fuelling inflation and that rate rises may be necessary. Fed governor Barr said AI could narrow or widen inequality. Europe's top bankers and regulators said on 3 July that AI is outpacing the rules, and the Swiss National Bank's Tschudin said on 21 August it could push up inflation.

The big cloud companies showed their AI spending producing revenue, and the ones that could show it were forgiven the spending.

Cloud is a smaller chapter than its importance suggests, 749 stories over 84 days. Google justified its spending with a booming cloud business on 22 July and reported record profits. Microsoft smashed its June-quarter targets on AI and cloud strength on 29 July. Amazon's shares surged on 1 August as its strongest cloud growth in more than four years validated its investments and supported a large increase in planned capital spending. It entered the $3 trillion club two days later. CoreWeave reported a new revenue high of $2.6 billion on 11 August with an order book of $104 billion. Baidu's AI cloud infrastructure revenue surged 50% while it missed on the group numbers.

The contracts between the clouds and the labs are the biggest numbers in the chapter. Google agreed to pay SpaceX nearly $1 billion a month in a cloud-computing deal announced on 5 June. OpenAI's planned cloud spending hit $750 billion on 22 July, including $20 billion committed to a Georgia data-centre project. Anthropic signed a $35 billion cloud deal backed by Nvidia on 31 August, in which Nvidia supplies the chips to a Texas data centre and holds the lease. Meta began building a cloud business on 1 July to sell its excess AI capacity, which lifted its shares 10% and which one analysis read as a signal that big technology has overbuilt.

The cost of the spending showed up in the credit markets and in the disclosures. Oracle's credit default swaps hit a cycle high on 20 July as tech credit risk rose and Nvidia's spreads widened, and Reuters described the company on 4 August as making a high-stakes ratings gamble in its AI strategy. CoreWeave explored a Wall Street playbook to hedge memory-chip price risk on 14 July. Microsoft was singled out on 26 August for leaving investors flying blind on its AI businesses, standing out for a lack of transparency from capital spending to Azure. The US Commodity Futures Trading Commission sought comment on compute derivatives on 19 August as demand grew.

Electricity became the binding constraint on the build-out, and the bill started reaching households.

Power drew 631 stories over 84 days. Forecasters kept raising what the data centres will need. The US Energy Information Administration said on 7 July that power demand will climb through 2027 on data centres. Data centres were expected on 21 July to use four times more electricity by 2035. Asia's AI boom ran into a power wall on 1 August. Entergy outlined a seven to 12 gigawatt hyperscale pipeline on 30 July, Japanese power companies planned to build or bolster 30 substations, and US power companies scrambled to secure equipment as demand strained supplies.

The response was to go around the grid. Off-grid power projects got a proof of concept in late June. Small engines gained momentum for off-grid data centres because they are cheap and readily available. Chevron entered the AI power business on 14 July and Baker Hughes signed a 500 megawatt geothermal deal, both oil companies moving into electricity. Nuclear ran through the chapter. Oklo and X-Energy joined an administration effort to speed reactors for AI, and TerraPower's chief executive attributed nuclear demand to AI and hyperscalers. The Journal followed the race to build America's first new reactor in a generation at the Idaho National Laboratory. The Environmental Protection Agency exempted islanded data centre power plants from acid rain pollution limits on 27 July.

The cost landed on people who are not buying AI. Arizona's largest utility proposed a 45% electricity rate increase for data centres and a 14.5% rise for households, and the Journal reported that no one is happy. Big technology data centres were driving up power bills at Rust Belt factories on 7 July. The White House rallied utilities and data centres for an AI power cost pledge in mid-July, and major utilities signed it on 22 July. TechCrunch reported on 14 August that natural gas prices could triple in some parts of the US, which would saddle hyperscalers with large bills, and on 20 August the Journal wrote that the off-grid push carries risks.

Caterpillar, Bloom Energy and FuelCell Energy all reported the boom. Caterpillar lifted its 2026 sales growth forecast on 4 August as the build-out powered on. Bloom Energy's earnings surged 680%. FuelCell Energy came into earnings after a 137% run. In one week at the start of August, 12 of 16 utility stocks beat earnings estimates. Red-hot power stocks were losing steam by 12 August, which the Journal attributed to the data centre backlash complicating the picture for plant owners.

The device layer is where the memory shortage reached a consumer price, and Apple said so first.

Devices, cars and robots produced 482 stories over 84 days, and the clearest fact in the chapter is a price rise. Apple said on 17 June that it will be forced to raise prices because of the AI boom, which MarketWatch reported as unavoidable in Tim Cook's words. Technology companies gobbling up memory chips for AI servers had left Apple with soaring component costs it would pass to consumers. That is the memory shortage described in an earlier chapter arriving at a checkout.

Apple's own AI ran through the quarter in both directions. Siri got what the Journal called a face-lift and a brain transplant, and reviewers found it changing how they use an iPhone and making the Apple Watch feel like a wrist computer. Bank of America saw an upgrade cycle. Against that, the EU held talks with Cook after a Siri clash. Apple limited the number of vulnerabilities researchers can submit as AI floods its security teams, and it cut jobs in the Siri and Vision Pro teams on 21 August. A Journal piece on 4 August argued that Siri's success could force Apple to spend more.

Robots were the chapter's other half. XPeng aims to produce more than 1,000 robots a month and plans a global launch next year as it transforms into a physical AI company. Alibaba's mapping unit unveiled a framework uniting a robot's brain, limbs and nerves into one system. UBtech rallied on strong humanoid orders. LG Electronics accelerated a robotics collaboration with Nvidia, and a foundation backed by Eric Trump partnered with AMD on humanoids. Nikkei reported on 17 August that a China shock looms for robotics as the physical AI race heats up. Kodiak AI logged more than 40,000 paid driverless hours through the second quarter, and Pony AI saw robotaxi fares soar as its fleet expanded while reporting a quarterly loss.

The wearables were smaller and stranger. Big technology is obsessed with smart glasses and has to convince people to wear them. A ring maker argues the future of AI wearables is voice, Plaud launched AI earbuds, and the Financial Times asked on 24 August whether AI belongs on your face, noting that critics call it cringe stalkerware.

Networking is the smallest industrial part of this chain and the one whose companies reported the cleanest numbers.

The wires drew 252 stories over 74 days, the second thinnest chapter here, and almost every company in it beat. Arista Networks saw AI sales doubling in late June, delivered its first $3 billion quarter, forecast upbeat revenue on 4 August and said on 20 August that AI demand is broadening while supply limits the upside to 40% growth guidance. Lumentum saw sales more than double as AI demand swelled, with record margins above 50%. Corning posted higher profit and sales on demand for its AI optical products. Nokia raised its full-year profit outlook on 23 July on an AI demand surge, planned AI-RAN pilots by the end of the year and had its outlook upgraded by S&P Global. Cisco reported a sharp jump in profit as AI orders rolled in and forecast fiscal 2027 revenue of $72.2 billion to $73.4 billion, including $7.5 billion of AI infrastructure revenue. Its chief executive said customers know that if they pause they risk being way behind. Credo's first-quarter revenue rose 114.7% year on year. HPE raised its annual forecasts on 2 September.

Copper and cable got tight, and no other chapter reported anything like it. A deadly storm in Chile disrupted copper mines on 26 July and raised AI supply concerns, and strategists said five days later that copper may hit a new high as AI demand collides with those storms. Prysmian signed a deal worth up to €5.5 billion with Molex for data centre optical cables. Optical cable maker WCFO added Cambodia capacity. Innolight listed in Asia's second-biggest offering of 2026 and dipped on debut.

On 15 July a piece noted that Cisco took 25 years to make investors whole again after its dot-com crash and asked what that means for traders buying Nvidia now. That is a comparison drawn by a commentator, and it is the only thing in this chapter that looks backwards.

Lithography is one of the hard ceilings on how fast this industry can grow, and in July the company that makes the machines said its ceiling is higher than it had told anyone.

The equipment chapter is the smallest in the report, 207 stories over 66 days, and it is dominated by a single company and a single day. ASML beat on both lines on 15 July, raised its full-year outlook by up to 19% and forecast 2026 net sales of €43 billion to €45 billion. It said it would increase capacity of its most advanced lithography machines, including a roughly 30% rise in Low-NA extreme ultraviolet capacity for 2027. It said there is room for equipment price increases. Bernstein set a Street-high target on capacity, pricing and margins. The company was scrambling to build more capacity, offered employees a €20,000 bonus to stay until 2030 and had signalled in early June that it can supply far more EUV tools than previously guided. SK Hynix said on 6 July it would spend 11.9 trillion won acquiring those systems, and Intel deployed the High NA machines for its Core Ultra 3 processors.

Teradyne, Lam Research and Besi reported the same demand. Teradyne soared in late June as AI reached almost 70% of total sales and beat again in July. Lam Research posted record results in its fiscal fourth quarter. BE Semiconductor raised its long-term revenue and profitability targets on AI, and Besi's orders more than doubled on AI and hybrid bonding. VAT reported record orders meeting margin pressure in an AI ramp.

Washington and Beijing both pressed on these companies. The US raised concerns with ASML on 19 June about a potential diversion of an EUV machine to China, which ASML denied. Around a fifth of its 2026 revenue is expected to come from China, which CNBC described on 17 July as walking a tightrope between sales and geopolitics. Analysts said on 2 August that China's home-grown deep ultraviolet progress is not the biggest threat the company faces. Michael Burry named Applied Materials among his AI short bets on 1 July.

The stories that fit no part of the chain are mostly about people meeting AI outside their working lives.

The leftovers hold 490 stories, 2.8% of this report, matching the subject and none of its chapters, and they cluster in two places. The first is consumer and domestic use, which arrives mostly through the Hebrew-language outlets here. A Harvard Business Review study reported on 5 June found that the leading use of generative AI over the past year was consultation on personal and social matters, for the second year running. The researchers warned about dependence on tools that answer before the user has thought. One writer described selling a house through a chatbot with no agent involved and reckoned it earned more than $90,000. Another looked at who really advises Israeli buyers on a mortgage now that AI tools have entered an unlicensed market.

The second is the politics of the technology outside government. The Financial Times wrote on 7 June about the coming rise of anti-AI populism and the political backlash that anxiety about the technology is set to generate. Bernie Sanders proposed that 50% of AI companies' shares be held by a sovereign wealth fund for the public. A New York Fed study blamed remote work for the difficulty young graduates have finding work, and cleared AI of it. McDonald's introduced an AI drive-thru system, Hasbro launched an AI studio for its characters, and a company running convenience stores in Israel unveiled an AI influencer.

This chapter is a residue and it reads like one. Its share of this report is small and its subjects do not join up, which is itself worth stating: over a quarter of reporting this heavy, almost everything found a place in the chain.

The reporting massed where this industry's money is and thinned where its limits are, and both limits reported this quarter fall in 2027.

After the 14 chapters, this report can say three things, and each one is a count from the pages above.

The reporting concentrated away from the machinery. Money was the largest chapter at 3,137 stories, 17.7% of the whole, followed by AI arriving in other industries at 14.0% and the model labs at 12.0%. Those three are about paying for the technology, using it and making it. The two parts without which none of it runs, the machines that make the chips and the wires between them, drew 1.2% and 1.4%. A reader following the coverage would learn most about the money and least about the constraint. The constraint was reported plainly in both places it appeared, memory in 2027 and lithography capacity in the same year.

The reporting reads the same way almost everywhere. Of the companies in the table, 67 were read as benefiting from what was reported about them, 20 as pointing both ways and eight as exposed. Six of the 14 parts of the chain hold no company read as exposed at all, among them the chips, the fabs and memory, and the machines that make them. An industry reported this uniformly is worth noticing on its own, whatever the reporting is right about.

The share prices split on the chain. Of the 94 companies carrying a share price, 62 fell over the window, and for 36 of them the price moved against its reading by at least 8%. Of the 67 read as benefiting, 43 saw their shares fall. The split follows the chain's own shape. Every part that builds the technology carries a negative median share move: power at -20.2%, data centres at -22.4%, chipmaking tools at -13.7%, fabs and memory at -10.8%, chips at -14.2%, networking at -21.8% and cloud and compute at -10.2%. One part carries a clearly positive one, software and agents at +12.4%, and it is the part that sells what everything below it makes possible. The suppliers were reported as winners and their shares fell. What a share did over these three months follows the price it started at, the rates and the market as a whole, and this report holds none of those. So the split is where to look, and it is not a finding that the reporting was wrong or that these companies are cheap.