AI’s next test is credit quality, not headline chip demand. Custom inference silicon is challenging Nvidia’s pricing power, lower HBM density is shifting value toward volume and connectivity, and Treasury buybacks face persistent fiscal pressure.
AI demand remains strong, but financing is now the weak link. Corporates are raising leverage to bridge today’s data-center spending to future productivity, while Nvidia-related lease backstops and SoftBank’s refinancing plans concentrate risk in a few balance sheets. The market may be pricing server growth more confidently than the credit structure supporting it.
OpenAI’s Jalapeno chip claims strong performance-per-watt gains, reinforcing the case that purpose-built ASICs can lower inference costs. The countercase is that chip-level benchmarks miss real agentic workloads, batching and Nvidia’s software stack. This looks less like an immediate GPU replacement than a gradual split between general-purpose training and specialized inference.
Reports that Rubin Ultra may use less HBM per system initially look bearish for memory suppliers. Yet an 8-high configuration could improve yields and unit output, while Semtech’s record data-center revenue points to simultaneous demand for faster connectivity. The bottleneck may be moving from maximum memory density toward system-level bandwidth, power and shipment volume.
Long-bond bears are cutting exposure as Treasury buybacks narrow dislocations and crowded shorts create rally potential. That relief does not resolve the underlying deficit: Peter Boockvar argues the bond market is larger than Treasury’s intervention capacity, while QTR frames debt damage as slow erosion rather than sudden default. Duration can rally tactically without becoming a clean fiscal endorsement.
PepInvestStocks said he bought more Smart Eye after the post-earnings selloff, citing mandatory European safety technology and deal flow.
Jim Cramer reiterated a buy-on-weakness view after Dick’s record one-day decline, while acknowledging the inventory glut and guidance reset; this was advocacy, not a disclosed purchase.
Headline demand is not enough; the next decision point is management’s treatment of credit exposure, customer support and component-cost pass-through.
CryptoCondom remains constructive on AbCellera but expects a move to $8-$9 before a larger advance, making this a level-dependent idea rather than a fresh purchase.
jpinsights is waiting for Marvell’s earnings to show how quickly the Google ramp and custom-AI revenue scale; the post did not disclose a position.
Myles identified Xero, Asana and BlackLine as short candidates after the bounce, citing valuation, weak profitability and AI competition.
Bubbleboi’s avoid call argues that specialized accelerators are eroding Nvidia’s inference advantage; it is a bearish view, not evidence of a disclosed short.
Rooster Global Portfolio Mastery Club raised stops on existing ASX300 uranium holdings, protecting gains while retaining exposure rather than issuing an exit.
RAISE STOPS ON NINJA ASX300 URANIUM HOLDINGS