13 posts from 7 reports.
The market has treated the only risk to AI capex as a shortage of demand or funding. Morgan Stanley points at something else entirely: politics. Public opposition to data centers is remarkably broad, with a recent Gallup poll finding seven in ten Americans against local construction, and it cuts across party lines (roughly two-thirds of Republicans, nearly 80% of independents, 83% of Democrats). That breadth is turning into policy. New York enacted the first true state-level pause in July, Texas ordered a review and paused new grid connections, and around 500 jurisdictions have now passed bans or moratoriums. Projects worth an estimated $156bn were cancelled or delayed in 2025, with another $130bn affected in the first quarter of 2026.
The positioning read is counterintuitive. Demand for compute already far exceeds available supply, so slower capacity additions would deepen and prolong the existing shortage. That raises the scarcity value of the hyperscalers' installed base and their pricing power, which strengthens the case for high-quality hyperscaler equities. The same slowdown hurts the AI adopters, because dearer and scarcer compute delays the productivity gains their valuations assume. There is even a credit angle: a smaller build-out means less debt issuance, and since much of this summer's spread widening came from absorbing heavy supply, lighter issuance could let spreads tighten.
This is not Morgan Stanley's base case, and whether the backlash becomes a sustained capex slowdown is still unclear. The tail worth watching is geopolitical: a prolonged US slowdown could let China narrow or erase the US lead in AI.
If the build-out slows, own the high-quality hyperscalers whose installed compute gets scarcer and more valuable; go lighter on the AI adopters that depend on cheap, plentiful compute.
Morgan Stanley rebuilt its framework for measuring how AI spending flows into GDP, and the result is striking. Broad AI-related investment added 0.8 percentage points to annualized real GDP growth in the first half of 2026 and 0.6 points on average since 2025. Even the narrower "AI-only" measure, which strips out spending that would have happened anyway, contributed 0.5 points in the first half. That is about a third of all US growth coming from one theme. In dollar terms, AI-only investment reached an annualized $450bn in the second quarter, split into $391bn of infrastructure and $59bn of software, against roughly $426bn of hyperscaler capex in 2025.
The composition tells you where the cycle stands. Most of the contribution is still the physical build-out rather than adoption, with computers and peripherals making up about 63% of AI infrastructure and semiconductor imports accelerating at a 57% annualized rate in the first half. Morgan Stanley expects infrastructure spending growth to slow in 2027 and 2028 as software and adoption take over, even as the six largest hyperscalers are forecast to spend more than $1.3trn in 2027, up about 60% from this year.
The investment point is the dependence itself. Because AI capex feeds straight into growth, and through the wealth effect into consumption and valuations, a genuine slowdown would hit the economy and the market at the same moment. Right now this is a single-cycle economy.
Respect how much the whole market leans on a single capex cycle; the largest macro risk is any pause in hyperscaler spending, so watch the build-out numbers as closely as earnings.
The consensus fear is that AI hits white-collar, high-income households hardest. Morgan Stanley agrees these college-educated, high-income, city-dwelling households are the most exposed to job displacement, then argues they are the biggest net beneficiaries in the base case. The reason is ownership of assets. The top 20% of earners, college-educated households, and those over 55 each account for roughly 80% of household equity exposure, so they capture the wealth effect from rising markets, on top of productivity-driven wage gains, new AI-related jobs, and disinflation.
The math is what makes the spending resilient. To offset a 1% fall in labor income, a top-20% household needs only about an 8% rise in equity wealth, and as little as 4% at a high propensity to consume, while the bottom 40% would need far more. Because the top 20% drive 38% of all consumer spending and even larger shares of the discretionary categories that matter for markets (45% of vehicles, 46% of entertainment, plus big-ticket furniture and luxury), their resilience keeps discretionary demand firm. In the base case AI lifts unemployment only 0.2 to 0.5 points at the peak, and only temporarily.
The whole thesis rests on asset prices holding up, since the wealth effect is doing the heavy lifting. In the bear case, rapid AI diffusion spikes unemployment by anywhere from 0.6 to 4.1 points, hitting these households first before demand weakness spreads, and the damage is worst when it comes alongside a falling stock market. Own the high-end consumer, but understand the trade is really a bet that equities do not fall.
Stay constructive on high-end and discretionary consumer names; the wealthy shopper is better insulated than the "AI kills white-collar jobs" story implies.
Goldman visited a dozen optical companies across Hong Kong, Taiwan and Suzhou, and came back more positive. Demand is driven by AI server-rack ramps, the spread of 800G and 1.6T modules, and specification upgrades. Goldman raised its 800G-and-above module demand forecasts by 39% and 36% for 2027 and 2028, to 144m and 171m units. Pricing looks healthy rather than collapsing, with 800G prices falling only single digits in 2027 and 1.6T holding above $700. Innolight is the global leader and Goldman's top pick, rated Buy with an A-share target of Rmb2,645, about 35.6 times 2027 earnings.
What separates this from the 2000 telecom bubble is that the demand is visibly real. Management says every module they produce is being deployed in data centers, return on investment runs one to two years, and customers are already planning beyond 2030. Because end-customers now co-develop products and plan capacity with the supply chain, the information gap that caused double-booking last cycle is much smaller. The genuine constraint is upstream: tight supply of DSPs, lasers and PCBs could cap shipments in 2027, which is a supply ceiling rather than a demand problem.
The risks are slower 800G-plus adoption, faster share normalization at the leaders, and geopolitics around component supply. But the supply-chain tone points to further upside to estimates Goldman has already raised.
Own the optical-module leaders, above all Innolight; Goldman just raised its 800G-plus demand forecasts and the supply chain says every module is being deployed.
Co-packaged optics is the next step in AI networking, and as a new technology it will take time to commercialize. Goldman's read is that the equipment suppliers, the makers of testing and coupling tools, are the earliest beneficiaries. Expensive final modules require heavy testing during production, and finding a failure early is far cheaper than finding it late, so demand for testing tools comes first. RoboTechnik, rated Buy with a target of Rmb796, already posted second-quarter revenue up 172% quarter on quarter, 141% above Goldman's estimate.
The order inflection is the interesting detail. RoboTechnik and FiconTEC say the tools they expect to ship in the coming year will exceed half of everything they shipped in the past 25 years. They aim to cut wafer-level testing time by 50 to 60% and quadruple die-level testing speed next year. Photonics IC production has to catch up to the electrical-IC ecosystem within one to two years, a process that took decades the first time, so the tool demand is heavily front-loaded. Goldman estimates CPO switch volumes of 10k, 92k and 131k units across 2026 to 2028.
Competition and slower optical end-demand are the risks. But buying the equipment is the lower-risk way to ride co-packaged optics before the technology itself is proven at scale.
For a first-order CPO play, own the photonics testing and coupling equipment leaders such as RoboTechnik, where orders are already inflecting hard.
Morgan Stanley's BluePaper argues that China is entering "Industry 5.0", the embedding of AI directly into its vast factory base, and that this anchors a decade-long capex supercycle. Cumulative industrial investment reaches Rmb340trn ($50trn) over 2026 to 2035, of which Rmb80trn ($12trn) is incremental versus a path without Industry 5.0. Industrial capex rises from 17.4% of GDP in 2025 to about 19.5% by 2035, becoming the main counterweight to a structural decline in property and old-style infrastructure spending as the population ages.
The case for China rests on assets it already has. It holds about 28% of global manufacturing value-add, the deepest supplier clusters, and 43% of the world's operational industrial robots, backed by coordinated policy including a state VC fund targeting nearly Rmb1trn for robotics and AI. Morgan Stanley expects China to lift its share of global manufacturing to 30% by 2035. The cycle starts slowly, with capex growth of 4 to 5% over the next two years while excess capacity and chip constraints bite, then accelerates to 6 to 7% from 2028 as capacity is absorbed and semiconductor localization advances. The stock list spans BYD, CATL, Midea, Mindray, Sany Heavy, Han's Laser and Zijin Mining.
The main risk is that supply again runs ahead of demand, keeping China stuck in deflation, alongside chip and software bottlenecks and trade friction. The offset is that this cycle is retrofit-and-replacement driven, which skews to higher value-add per unit and helps absorb the existing glut.
Own China's industrial, automation and materials leaders on Morgan Stanley's Industry 5.0 list; this is a structural capex supercycle rather than an ordinary cyclical upswing.
The AI data-center boom, combined with energy-security spending, is shifting the global electricity system from stagnation to expansion. Morgan Stanley sees grid capex growing at 12% a year in Europe and 8.5% in the US through 2027, and global gas-turbine demand rising from 80GW in 2025 to more than 100GW over the next decade. Supply is the bottleneck: lead times have stretched to two or three times normal and equipment prices rose 30 to 50% in 2025.
This is where Chinese suppliers come in. They hold only 2 to 11% of the global market in transformers, high-voltage switchgear and turbine blades, yet they carry more spare capacity and shorter lead times than Western peers. Morgan Stanley argues these share gains can go non-linear, as they did in solar panels and EV batteries, where China moved from a few percent to dominance faster than anyone modelled. Every additional point of global share adds roughly $824mn of revenue each in transformers and switchgear, and up to about 40% of EPS upside in 2027 for the names it covers. The preferred picks are Sieyuan, Yingliu, Sungrow and ZTT.
Broadening trade barriers are the obvious threat to this export channel. But persistent shortages abroad are exactly the catalyst that drives the non-linear share gains.
Own China's power and grid equipment exporters (Sieyuan, Yingliu, Sungrow, ZTT); the worldwide data-center power crunch is handing them share, pricing and margin.
Morgan Stanley's robotics model sees a $2.5trn (Rmb18trn) China robotics market over 2026 to 2035, growing at a 30% CAGR. Without Industry 5.0 the market would grow at 15%, in line with its history, so the AI-driven upgrade adds roughly $1.5trn of value. China is already the largest robotics market, with domestic suppliers holding more than half the market and the country owning 43% of the world's operational industrial robot stock.
The driver is demographics meeting cost. An aging workforce and rising wages push manufacturers to swap labor costs for capital, which lifts the return on automation. Robots spread beyond fixed industrial arms into collaborative robots, mobile units and early humanoids, while the machines themselves gain embedded AI through machine vision, sensors and edge chips. China's scale also feeds a data flywheel: its huge base of connected factories generates the production data to train better industrial-AI models, compounding the advantage. Preferred names include Horizon Robotics, Leaderdrive and Han's Laser.
The risk is an undisciplined build-out that recreates overcapacity in yet another frontier industry.
Own China's robotics and automation leaders (Horizon Robotics, Leaderdrive, Han's Laser); Industry 5.0 roughly doubles the market's growth rate.
Morgan Stanley initiates coverage of digital assets with a "convergence" view: these assets are moving inside the financial system rather than replacing it. Tokenized real-world assets now top $30bn, roughly six times their level at the start of 2025, with more than $17bn in cash-like rates products. BlackRock's tokenized liquidity fund alone is $2.75bn, and Franklin Templeton, JPMorgan, KKR and Apollo are all issuing. About a quarter of the companies in the Bloomberg World Financials Index, representing roughly 60% of its market cap, already have live, piloted or announced digital-asset initiatives.
The real usage is in institutional plumbing rather than retail speculation. JPMorgan's Kinexys platform processes more than $7bn a day, Broadridge's blockchain repo runs at $357bn a day, BNY estimates $200bn of tokenized deposits globally, and SWIFT has launched a blockchain with 17 banks. Morgan Stanley's key caveat is that adoption of the technology does not automatically mean the crypto tokens rise in value. Value can just as easily accrue to the incumbents that capture the flows, which is the argument for owning the rails rather than the coins.
The longer-term risks are fragmented standards, weak interoperability, and eventually quantum computing threats to cryptography. But the direction of travel is clear.
Play the on-chain shift through the incumbents building the rails (BlackRock, JPMorgan, Broadridge, the exchanges); they capture the flows whatever the tokens do.
Stablecoin supply exceeds $300bn, which sounds large until you note it is only about 0.25% of global M2. Gross transfers average roughly $7.5trn a month, but once you filter for genuine payments the figure drops to about $63bn a month. Most of the activity is crypto trading, wallet movements and inorganic transactions rather than payments to merchants.
The mechanism is what limits the business. Stablecoin velocity runs near 300 times a year, similar to a pure settlement system, so the network can process enormous volume without needing large outstanding balances. Issuer-paid yield is banned under the US GENIUS Act and the EU's MiCA regime, so idle stablecoins carry an opportunity cost, and users move spare cash into tokenized money-market funds instead. Morgan Stanley's payments analyst James Faucette used exactly this logic to downgrade Circle, the issuer of USDC, noting that even generous adoption assumptions do not translate into meaningful USDC balances. Stablecoins also have to compete with FedNow, the RTP network and card rewards.
The limiting factor is the float itself: balances never build up enough to matter, however large the transaction volumes look.
Fade the stablecoin-issuer hype; Morgan Stanley's own analyst downgraded Circle, because high velocity and a ban on yield keep balances from compounding.
Nokia has pulled back nearly 40% from its June highs, far more than other European AI beneficiaries, even as demand stays exceptionally strong. Its second-quarter order book hit a record €2.8bn, almost three times the first-quarter level, and Network Infrastructure grew 12% in the first half against a 6 to 8% medium-term target. Morgan Stanley's telecoms team says the stock is priced as if the AI networking story has broken, while its 2028 EBIT estimate of €3.7bn sits about 15% above the top end of company guidance.
The catalysts are lining up. Nokia's Pennsylvania test and packaging capacity is set to rise tenfold by the end of the third quarter, guidance has room to move higher, and possible inclusion in the Euro Stoxx 50 later this month would force index funds to buy. The read-through from Oracle, whose cloud infrastructure revenue jumped 121% to $7.4bn on AI data-center demand, is that spending on networking and optical has further to run.
A contrarian AI-networking buy: Nokia trades as if the story broke, yet orders and estimates say otherwise, with index inclusion as a near-term catalyst.
The software index is up 37% off its April lows but has given back most of its late-August gains, and the obvious AI winners have already had their move. The hunt now is for cheaper ways into the same structural themes: MongoDB and Snowflake in AI apps, Datadog and Dynatrace in observability, Cloudflare and Akamai in agents and edge inference, Palo Alto and Okta in security, Atlassian and GitLab in developer tools.
These are more than valuation trades. MongoDB is a core beneficiary as enterprises build AI applications, Akamai offers edge-inference exposure at about 15 times earnings, and Okta offers identity security at roughly 27 times free cash flow. AI is already lifting GitLab usage, with secure repositories up 60%, code pushes up 50% and CI/CD pipelines up 40%. Positioning is not crowded either, since the record SaaS buying in the second half of August was mostly short covering rather than fresh demand.
Rotate into second-line software names with real AI usage and un-crowded positioning: MongoDB, Datadog, Cloudflare, Okta and GitLab.
Morgan Stanley's European brands analyst reiterates an Overweight rating and lifts the price target to €395. The underappreciated part of the story is what he calls relational value: owner-exclusive experiences deepen loyalty and repeat buying, so about 84% of cars sold go to existing Ferrari owners, with demand still running ahead of supply.
That closed ecosystem is what insulates Ferrari from broader luxury softness. The earnings upgrades come from a richer mix and personalization rather than higher volumes, with Morgan Stanley raising FY27 and FY28 EPS by about 1% and 2%. It remains a scarcity brand with genuine pricing power at a time when the rest of luxury is struggling.
Own Ferrari as a defensive luxury compounder; Morgan Stanley raised its target to €395 on mix and personalization rather than volume.
