Jukan
Semiconductor & AI Infrastructure Research at Citrini
Meta's $35B financing plan points to continued chip demand rather than a capex cut. Meta is raising money to buy more chips because it is short of them, contrary to earlier claims that it lacked demand for its chips. x.com
A rising value of compute per watt should benefit AI cloud, chip, connectivity and power suppliers. Elon Musk says that value will rise over time. Nebius has Microsoft and Meta deals; Marvell benefits from custom silicon demand; Credo connects GPUs; Vistra has long-term nuclear supply deals with Amazon and Meta; Nokia's Nvidia partnership could help it capture data-center networking growth. x.com x.com
Qualcomm's reported ~$4 billion purchase of Modular targets Nvidia's real moat, CUDA software, rather than its chips. Modular's hardware-agnostic, write-once-run-anywhere stack is the anti-CUDA, an escape hatch from Nvidia lock-in that lets Qualcomm's chips, and everyone else's, compete in software, where Nvidia has had no rival. Google with TPUs and Amazon with Trainium and Graviton made hardware bets; Qualcomm went after the software layer. x.com
AI security rules could shift frontier inference into monitored data centers, lifting demand for AMD and Nvidia chips and Micron and SanDisk memory. Agents attacked 400+ banks; US labs cut off abusive users, unlike downloaded open weights. Licensing is a forecast, and these four look undervalued at today's closes. Meta lists $356B in future leases and $455B in other commitments, mostly cloud capacity such as CoreWeave and Nebius plus equipment; those recipients are being sold off. x.com x.com
Meta's unmet AI demand pays its bottleneck suppliers. Super Micro's $19.5B quarterly outlook and under 8x forward earnings point to a double. CoreWeave's Meta deals total $35.2B; it should easily top $100. Builders say nearly all new sites use Dell or Super Micro; Dell could easily double. Micron's demand is off the charts at 5x forward earnings. Sources say Vertiv's gear is sold out for 2-3 years. VivoPower has 2GW+ potential in 5-7 years. Digi Power X has a $1.1B Cerebras deal. x.com x.com … Show more
AI agents benefit CPU and memory suppliers such as Arm, AMD and Qualcomm more than Nvidia and Broadcom; Marvell, TSMC and Micron also gain. Per @amitisinvesting, OpenAI has launched Dots as its answer to Meta's Muse, and more agents mean much more compute, which helps explain why semis have kept rising as the agent wars heat up. x.com
Financing announcements point to more accelerator demand, supporting semiconductors. About $200bn in announced Meta, Anthropic/Broadcom and SpaceX/Nvidia financing likely drives $400bn in XPU purchases. Data centers burn cash until lit, then cash flow 20-40% annually; AI spend rising from 0.1% to 0.5-1% of current payroll could balance the buildout if agents replace roles. Sustainable token demand is the bet, but smaller local models are the bigger risk, though not an immediate concern. x.com … Show more
Meta building a cloud business to sell excess AI compute signals it is dropping out of the frontier AI race, which is very bearish for Meta alone. Per Wall St Engine, Meta is setting up that business, yet its compute is only equal to three CoreWeaves. x.com x.com
Qualcomm is becoming a fuller AI infrastructure player, and the market may re-rate it if its $5B fiscal 2027 data center revenue target is met. Most of that target comes from custom chips for one US and one Chinese hyperscaler, about $2B each at mid-30s gross margins. Its High Bandwidth Cache stacks memory on logic for about 6x HBM's bandwidth per watt, easing the inference memory wall. Alphawave's 800G DSPs already earn revenue, and a Meta custom CPU deal carries a minimum share guarantee. … Show more
Selective bank lending limits bubble risk in the AI build-out, while early agent adoption supports lasting demand. Banks are cautious about neocloud loans; Meta's Muse reached more than 10 million people in about six weeks, and demand still outruns supply. Nvidia, Broadcom and Micron look priced for a sharp slowdown around 2029, but the application layer is just beginning. This cycle looks much longer than the market prices, even if it is cyclical. x.com