Daily Alpha · X
· Premarket Alpha · by Buzzberg Research
Jukan posted Meritz Securities channel checks that mobile memory prices for Apple will rise 50% or more QoQ from 1Q27, and ren_stocks cited BofA's raised 2030 forecast of $2.0 trillion in DRAM and NAND sales.
Themes on this desk
Memory content
ren_stocks cites BofA's memory team: Rubin moves from 288GB to 384GB per chip and Rubin Ultra to 1,000GB, behind a new 2030 forecast of $2.0 trillion in DRAM and NAND sales, raised from $1.8 trillion.
Memory contracts
Pequity Research's expert-call takeaways say memory contracts are moving to 5-year-plus take-or-pay terms, with HBM4/HBM4E contracts at an estimated $800–$1,400 per device versus about $530–$550 for HBM3/3E.
Miner power
jiahanjimliu says HUT has delivered 0MW of 949MW signed and WULF 102MW of 839MW, and foresees some or all of HUT, WULF and CIFR having delays for 2027.
Ticker heat
?
Top voices by smart followers and alpha score
Market Radar →Meritz Securities: Some mobile DRAM prices are expected to surge +50% in 1Q27.
Full comment: The race toward AGI is driving competitive, structural growth in server demand. Even so, B2C applications led by smartphones remain the last segment still resisting price increases. Our channel checks, however, indicate that mobile memory prices for a key customer (Apple) will rise by a staggering 50% or more QoQ starting in 1Q27, as memory makers keep shifting production from LPDDR to server DRAM and HBM. The key customer is building its 20th anniversary flagship smartphone next year, so it will have to absorb the price surge. $AAPL $DRAM
Meritz Securities: Some mobile DRAM prices are expected to surge +50% in 1Q27. Full comment:Source →
Expert Call: Memory - Key Takeaways > Contracts are moving from flexible 2–3 year terms to 5yr+ take-or-pay structures with upfront…
…deposits and price floors/ceilings, prioritizing downside protection for suppliers. > Tight conventional DRAM supply and an estimated ~58% gross margin for HBM LTAs are causing conventional DRAM to directly compete with HBM for wafer allocation. > Next-generation HBM4/HBM4E contracts are commanding an estimated $800–$1,400 per device, significantly higher than HBM3/3E contracts sitting around $530–$550. > Manufacturers are exerting strict control by requiring end-customer sales records, conducting warehouse checks, and enforcing NCNR (Non-Cancelable, Non-Returnable) terms with ~30% deposits to prevent hoarding. > Tight memory supply may not reflect true underlying AI consumption, as data-center capacity is frequently built ahead of available power and utilization. $MU $DRAM $EWY Quoting @jadzo1_: Some datapoints from a memory supply-chain expert call a few weeks ago: → Conventional…
Expert Call: Memory - Key Takeaways > Contracts are moving from flexible 2–3 year terms to 5yr+ take-or-pay structures withSource →
$SKHY $MU $SNDK Memory per GPU is going up, and the math is getting silly.
The fear with memory at these prices has always been "de-spec," the idea that chip designers respond to expensive memory by using less of it. BofA's memory team is assuming the opposite. Rubin moves from 288GB to 384GB, a third more per chip, and Rubin Ultra goes to 1,000GB, which is about 3.5x what ships today. So the most supply-constrained component in the stack is getting designed in at higher content, into more chips, every generation. That is how you get to their new 2030 forecast of $2.0 trillion in DRAM and NAND sales, raised from $1.8 trillion. For scale, the 2018 cloud boom peaked at $170 billion, and last year did about $210 billion. The industry is already running above a $1 trillion annual pace, based on $278 billion of sales in the third quarter alone. Getting to $2 trillion from here is a doubling in four years, with bit growth in the high teens and the mix shifting…
$SKHY $MU $SNDK Memory per GPU is going up, and the math is getting silly.Source →
Ex-BTC miners make up alot of the marginal power capacity for 2027.
Their delivery will heavily make or break the power shortage with 4.8GW of leases combined. We see $HUT has delivered 0MW even though it has signed 949MW. $WULF has finished 102MW out of signed 839MW. $APLD and $CORZ have been working with CRWV for 1 and 2 years respectively and have delivered 175MW and 437MW respectively. Knowing that $CORZ and $APLD had tough ramps in the past with difficulties into this year, and $IREN's tough ramp, I do forsee some or all of $HUT, $WULF, $CIFR having delays for 2027. $IREN by mid October will be at 150MW and by end of will have delivered 300MW of Bare Metal Cloud which includes the DC and commissioning the GPUs. The commissioning of GPUs is important as it allows profits to flow from higher GPU prices. Quoting @stocksavvyshay: AI miners have signed ~4.8 GW of leases but delivered only ~720 MW while just ~1 GW of the ~12 GW pipeline is near term…
Ex-BTC miners make up alot of the marginal power capacity for 2027.Source →
SPCX, NVDA: $40B financing led by Apollo to buy Nvidia chips + SpaceX plans to raise $40B to buy Nvidia AI chips, per the Financial Times…
…via Reuters on Oct 6. The plan is about $10B in bank loans plus $30B in investment grade debt, led by Apollo, with Pimco among the lenders in talks. The deal is expected to close in 2027. + Context on the scale: Morgan Stanley estimates AI infrastructure will need $1.5T in external financing by 2028. This is the same Apollo that partnered with Nvidia, BlackRock, Blackstone, Brookfield, Goldman and KKR in August on platforms meant to mobilize more than $500B for AI infrastructure. + Elon Musk said xAI's Colossus 2 could more than double its Nvidia chip count by December, and the company plans to use Nvidia hardware exclusively in its data centers. Shares of SpaceX fell 1% after hours while Nvidia rose 0.5%.
SPCX, NVDA: $40B financing led by Apollo to buy Nvidia chips + SpaceX plans to raise $40B to buy NvidiaSource →