10 posts from 6 reports.
Wood spent the week at Jefferies' Korea conference, and the trade data was the story. Semiconductor exports soared 193% year on year to $133bn in the three months to August, a record 44% of Korea's total exports. The pull is AI: memory chips are the scarce input for AI servers, and demand is running far ahead of supply. Korea's nominal GDP grew 26.4% year on year in the second quarter, the fastest in nearly 47 years, almost entirely because of chips.
SK Hynix is the clearest example of the cycle. Its board approved W54.3tn ($38bn) to build new fabs, and management is confident enough to return huge amounts of cash at the same time. Hynix is buying back and cancelling W40tn ($29bn) of stock, the largest cancellation ever by a Korean company, and paying workers 10% of operating profit as bonuses. Foreign appetite is strong too: its US-listed shares trade about 40% above the local shares.
Wood keeps his exposure through Samsung Electronics, SK Hynix, Kioxia and TSMC, and has been adding to Hynix and Kioxia. The main risk is that this is a cycle, not a permanent boom, and memory is famously volatile. There was already a sharp shakeout in July, when leveraged Korean ETFs that had ballooned to $50bn collapsed. If AI capex slows, memory turns down fast.
Own the memory "picks and shovels": SK Hynix, Samsung Electronics, Kioxia and TSMC.
Europe is rearming at a pace not seen in decades. German defence spending is set to reach 109bn euros next year and 184bn euros, around 3.5% of GDP, by 2030. The EU has also committed a 90bn euro Ukraine support programme for 2026-27, most of it for defence procurement. This is sustained government demand rather than a one-off order, which is what makes it investable.
The politics behind it are getting harder, and that matters for how durable the spending proves. Chancellor Merz is deeply unpopular, his approval down to 13%, and the AfD, which wants to end the Ukraine war and reopen Russian energy, won 43.8% in the Saxony-Anhalt state vote. High energy costs feed the anger: German gas now costs about 9.8 times US gas, up from 2.8 times a year ago, which is hollowing out European industry at the same time.
The read is exposure to the European defence suppliers riding this build-out. The risk is exactly the politics above: if anti-war parties keep gaining, some of these budgets could face pushback well before 2030.
A multi-year tailwind for European defence contractors that supply the build-out.
Korea stands out for pairing a high yield with strong public finances. The ten-year government bond yields 4.46%, up from 2.56% in April 2025, while government debt is just 48% of GDP, very low by G7 standards. The Bank of Korea has been hiking, to 3.0%, to stay ahead of inflation running at 3.1%, so the real yield on offer is genuine.
The won adds to the appeal. It is still cheap on a real effective basis even after rising 16.6% against the dollar since its July low, and Korea runs a large current-account surplus of about 13% of GDP. Foreigners have been heavy sellers of Korean equities, so the currency is not being propped up by hot money that could reverse.
For a bond investor this is a way to earn a solid real yield in a fiscally sound country, with currency upside if the won keeps recovering. The risk is that Korea's economy is geared to the same AI-chip cycle, so a memory downturn would weigh on both growth and the won.
A rare combination of a real yield, sound public finances and a cheap currency.
JPMorgan's call is that Meta's AI push is finally turning into real products. Its new Muse AI agent reached #3 in the US app store on its second day, with early usage running at ten times the training cohorts, and its Muse Spark models are now competitive with Claude and GPT. Meta can put these in front of about 4 billion users, a distribution edge no startup can match. JPM raised its December 2027 target from $640 to $820, roughly 25% above the current $654.
The nearer-term engine is still advertising, where AI is improving targeting and recommendations. JPM sees revenue growing from $201bn in 2025 to $305bn in 2027, with earnings power to match. On top of that sit two new profit pools the market is barely pricing: a commission or subscription take on Muse, and paid access to Meta's models for outside developers.
At about 16 times 2027 earnings, Meta is not expensive for that growth, and the stock has actually lagged, down 1% this year against a market up 12%. The main risk is spending: JPM models $243bn of capex in 2027 and $284bn in 2028, which pushes free cash flow to roughly negative $65-70bn a year. If AI monetization lags that outlay, the shares stay under pressure.
Buy Meta as its AI models and agents start to land; JPM sees roughly 25% upside.
Oracle's cloud arm is scaling fast. Infrastructure revenue grew 120% year on year, the company delivered a record 850 megawatts of capacity and more than 300,000 GPUs in the quarter, and those chips are running near full at 98% utilization. Even older GPUs coming up for renewal were re-signed at a 20% premium, a useful sign that this capacity holds its value rather than becoming obsolete.
The important detail is how the backlog is funded. Of the $664bn in signed contracts, the new deals this quarter come with customer prepayments or "bring your own hardware" arrangements, so they need little extra capital from Oracle. That speaks directly to the biggest worry hanging over the stock, which was that Oracle would have to borrow enormous sums to build data centers for other people.
The shares have already fallen from about $328 to $153, and Deutsche Bank keeps a Buy with a $300 target, implying roughly 100% upside on its 2027-28 earnings estimates. The real risks are margin and cash: gross margin fell almost 8 points as lower-margin infrastructure grew, and free cash flow was negative $5.4bn on $28.5bn of capex. The payoff depends on that spending converting into profit over the next two years.
The financing fear looks overdone; DB rates Oracle Buy with a $300 target, about double the price.
BofA tries to put a real number on the AI capex wave and lands on about $5tn through 2030, rising to $7.7tn at the top end. The amount that needs outside money is smaller than the headline suggests. Including non-AI spending, the eight biggest builders, among them Microsoft, Amazon, Alphabet, Meta and Oracle, are expected to generate $6.6tn of operating cash flow against $6.9tn of capex, a gap of only about $0.3tn.
That average hides a sharp split, which is the real point. The strongest four companies can fund the build entirely from their own cash flow. The weaker players, including the "neoclouds" such as CoreWeave and Nebius, face a $0.7tn shortfall and must lean on debt, take-or-pay contracts and supplier support. BofA puts total external funding at about $1.2tn, roughly $300bn a year, which the market can absorb as long as it is paced out.
For credit investors the message is to lean toward the high-quality hyperscaler bonds that are comfortably covered by cash flow, and to demand more yield from the weaker data-center and neocloud names where the funding gap actually sits. The risk to watch is timing: if all this supply arrives at once, or spreads widen, orderly issuance can quickly turn disruptive.
Favor high-quality hyperscaler credit; demand more from the weaker "neocloud" and data-center borrowers.
A new corner of the bond market is being built to fund data centers directly. About $568bn of AI-related debt has been issued this year, and deals now include large asset-backed bonds tied to specific campuses, such as a $14bn secured bond for a Michigan data center. For investors this offers direct exposure to the buildout at higher yields than plain hyperscaler bonds.
The part to watch is circular financing. Chip suppliers like Nvidia and Broadcom are increasingly backstopping demand or partnering on the funding, which means the same companies are selling the hardware, supporting the buyer and booking the revenue. That can flatter demand and mask credit risk if the end-buyers turn out to be weaker than they look on paper.
The takeaway is that data-center bonds are a genuine and growing source of AI yield, but quality varies widely and the vendor-financing links deserve real scrutiny. Favor deals backed by long contracts with strong, named tenants over those that lean on their suppliers to prop up demand.
A fast-growing way to earn yield from AI, but treat the circular-financing links with care.
BofA's core argument is that today's higher rates are not a glitch but a return to normal. The rise is mostly in real rates and it is global, driven by two forces: large post-pandemic fiscal deficits and a surge in AI-related investment that lifts the economy's underlying equilibrium rate. The US deficit is running above 6% of GDP, and interest costs now exceed both defence and Medicare spending.
Washington is trying to hold long-term yields down anyway. Treasury Secretary Bessent's plan to buy long-dated bonds funded with short-term debt is an attempt to cap the long end of the curve. BofA argues this is not a genuine market failure, so it is unlikely to work, and it may backfire: if investors sense the government is fighting fundamentals, or is worried about its own financing, they demand a higher risk premium, not a lower one.
The read is to expect higher-for-longer real rates and to treat sovereign risk as something that shows up in currencies and inflation rather than outright default. That favors keeping bond duration short and holding gold and other real assets as a hedge against financial repression. The thesis only breaks if AI investment collapses or governments finally cut their deficits.
Stay short duration and favor gold and real assets over long developed-market bonds.
BofA flags Japan as an under-appreciated driver of global rates. As Japanese yields normalize after years of yield-curve control, the country's investors, who are among the largest holders of foreign bonds in the world, have more reason to bring their money home. That repatriation adds upward pressure on yields elsewhere, including US Treasuries.
Tokyo is also intervening to support the yen, which often means selling US Treasuries and spilling over into the American bond market. So Japan is pushing global yields up through two channels at once: rising domestic yields that pull capital home, and currency intervention that sells foreign bonds along the way.
For investors this argues for caution on long-dated Japanese government bonds, where normalizing policy means price downside, and for watching the yen as an early signal of global rate pressure. A sharp move home by Japanese capital would be felt well beyond Japan.
Watch Japanese bonds and the yen as a global rate signal; be cautious on long JGBs.
Adobe's AI numbers are genuinely good. Its AI-first products now run at over $650mn of annual recurring revenue, up more than 150% in a year, and its Firefly app revenue grew 40% in a single quarter. The company also crossed 1 billion monthly users, with creative free users up more than 70%. On the surface this looks like a business riding the AI wave rather than being run over by it.
Yet the stock has fallen from about $530 in 2023 to $249, because the market cannot decide whether generative AI is a threat or a gift for Adobe. The fear is that cheap AI image and video tools erode its creative-software moat. The bull case is that Firefly turns AI into a fresh revenue stream. Growth is also cooling at the edges: total bookings grew 8%, down from double digits over the prior nine quarters.
Deutsche Bank sits on the fence with a Hold and a $260 target, right around today's price, because one solid quarter will not settle the debate. The sensible approach is to treat Adobe as a show-me stock: wait for AI-first revenue and bookings to reaccelerate as proof the moat is holding, rather than buying on the AI headlines alone.
A show-me stock: wait for AI revenue to reaccelerate before buying; DB rates it Hold at $260.
