8 posts from 6 reports.
In June, Broadcom joined Blackstone and Apollo to lease its custom AI chips to users like Anthropic, funded by borrowed money. Broadcom guarantees most of that debt, so if the chips end up worth less than the loans, it covers the gap. The first deal put $35bn of chips into service, with a plan to scale to 20 gigawatts by 2028.
BofA estimates the guarantee could grow to about $370bn by 2029, with a realistic worst-case loss near $42bn. That is more than twice Amazon's entire bond debt, and Nvidia reportedly plans a similar $350bn program. Broadcom can afford it, with cash flow rising from $27bn toward a forecast $98bn by 2027, so BofA only trimmed its bond view to neutral. The pattern is the point: the AI buildout increasingly runs on debt the chip sellers guarantee. The break is customer concentration, since these custom chips have no resale market.
Watch the vendor-financing chains behind AI. They are a hidden risk across the trade, even for a company as strong as Broadcom.
For three years the AI story was about GPUs, while the ordinary CPU was a supporting act. That is changing as AI shifts to running fleets of automated agents, which each need a CPU to coordinate their work. BofA now sees the server CPU market reaching $210bn by 2030, up from about $35bn last year, growing 36% a year.
In the training era there was one CPU for every four GPUs. For agentic AI that moves toward one to one, and it is extra demand rather than a swap. BofA's top pick is AMD, whose next chip packs 256 cores against 192 for Intel and 88 for Nvidia. AMD's share of server CPU revenue could reach 31% by 2030 as Intel slides toward 22%. The read is that the AI trade is wider than the GPU names everyone owns. The risk is that this year's chip gains came from higher valuations, so earnings must catch up.
Own the CPU winners as agentic AI arrives. AMD is the top pick, with Nvidia the top pick across the sector.
Everyone watches the chips. The tighter squeeze is one layer below, in the materials that make up the circuit boards inside AI servers. Morgan Stanley expects demand for AI circuit boards to grow more than sixfold by 2030, with the whole market rising from about $58bn to $135bn. The tightest links are high-end glass fabric, a specialist copper foil and optical fiber, where shortages should last into 2028 and hand suppliers real pricing power.
The setup is attractive because these stocks have already pulled back. The AI materials group rose 91% over the past year, then fell about 36%, and now trades near 26 times earnings against 35 for the board makers they supply. Morgan Stanley points to Nittobo, Mitsui Kinzoku and Mitsubishi Gas Chemical, and started coverage of China's Grace Fabric. The risk is that a slowdown in the buildout hits these narrow suppliers harder than the big chip names.
Own the makers of the specialty materials in AI circuit boards. Supply stays tight through 2028, which supports pricing and profits.
The next battle in consumer technology is over which AI agent acts for you when you shop, travel or order dinner. When an assistant does the choosing, the company that controls it sets the terms for advertising, e-commerce and delivery. Morgan Stanley sizes the spending that agents could intermediate at roughly $30tn.
The winners are the platforms that already have huge reach and their own user data. Morgan Stanley names Meta, Apple, Alphabet and Amazon as best-placed, with Shopify as the tool that lets merchants sell through agents. The payoff can be concrete: for Meta, every 100m users of its new assistant could add about 1% to 2028 earnings. The read is that today's consumer-AI winners should stay winners, because scale and data compound. The risk is that regulators force these agents to stay neutral, breaking the link between owning the assistant and owning the sale.
Own the platforms with the scale and data to sit between the shopper and the purchase: Meta, Alphabet, Apple and Amazon.
The September meeting was more hawkish than expected. The Fed raised rates a quarter point with no dissents, and its projections pointed to another move this year. Morgan Stanley now expects two more hikes, in December and March, taking the top of the range to 4.5%. Its economists see core inflation ending this year near 3.2%, well above target, on an oil-driven price premium.
The shift matters as a change in mindset. Investors had treated the September move as a one-off, and the meeting raised the question of whether a fresh tightening cycle has begun. That alone sent bank stocks lower, since a flatter yield curve squeezes their lending profits. History around a first hike favours high-quality companies, so the read is to tilt toward durable earnings. Morgan Stanley's fresh-money list leans that way, including AbbVie, Walmart, Northrop Grumman and Delta. The main risk is a sharp jump in oil, which would force the Fed to go further.
Position for higher-for-longer rates. Lean toward high-quality companies that hold up when the Fed keeps tightening.
The big worry about the cloud giants is that their AI spending will never earn a return. BofA raised its forecasts anyway, with Alphabet near $290bn and Amazon's cloud arm near $230bn in 2027. To test whether that is worth it, BofA turned the spending into data center capacity and asked what the market pays for each gigawatt inside each stock.
The gap is striking. The market values Google's capacity near $110bn per gigawatt and Amazon's at $59bn, but Meta's at just $4bn. Investors give Meta almost no credit for what it is building, which looks too harsh when recent deals to lease AI capacity were struck at $40bn to $50bn per gigawatt. Amazon also builds most cheaply, at $25bn per gigawatt versus $45bn for Meta, thanks to its own chips. The read is to own Amazon for cheap capacity and Meta for the near-free option on its buildout. The risk is that weak AI demand turns that capacity into an expensive burden.
Own Meta and Amazon. Their spending is building capacity the market is barely paying for.
For a year the winning trade was to own the companies that sell AI gear over the cloud giants that spend the money. BofA made that call in mid-2025 on fears that heavy spending would weigh on the giants. It worked, pushing communication services, which is 78% Google and Meta, to its cheapest level against the market since 1986.
BofA now thinks that trade has run its course and nudged the sector back to neutral. Cloud revenue growth is on track to overtake spending growth in late 2027, so the investment is close to earning its keep. With the shares cheap and the crowd gone, risk and reward look balanced. Even so, BofA sees only modest gains, with a year-end target of 7400 and a real chance of a sell-off, and prefers defensive value sectors like financials, energy and staples. The worry is that off-balance-sheet AI commitments are climbing and reported profits are drifting from real cash flow.
Rotate some money back toward the hyperscalers. Keep the core tilt defensive and value-leaning into a choppy year end.
The Fed's balance sheet ballooned during the crisis years, and new chair Kevin Warsh wants to bring it down. Morgan Stanley expects the shrinking to start as early as 2027 and total around $1.5tn over two years. The natural fear is a repeat of past episodes, when pulling money from the system tightened conditions and helped trigger sell-offs.
This time is set up differently. Rather than selling long-term bonds and forcing investors to absorb them, the Fed plans quieter levers. It can hold a smaller cash cushion for the Treasury and pay banks less on excess reserves, which nudges them into short-term Treasury bills instead. Because those bills are already in wide supply, the swap changes little for markets, so long-term rates need not rise on this account. The read is that the balance-sheet story is not the tightening shock some fear. The risk is that banks resist, or the Treasury keeps its cash balance high.
Do not treat the coming balance-sheet shrink like the tightening of 2018 or 2022. The mechanics protect market liquidity.