Daily Alpha · X
· Midday Alpha · by Buzzberg Research
Serenity (@aleabitoreddit) argued Microsoft's exit from the Nscale deal is bullish for neoclouds such as Nebius and IREN, and Bernstein data showed the data-center project pipeline up 52 GW in September.
Themes on this desk
Gilead dataset
@a_may_md argues the strong result in the super-low-dose arm is very bad for Gilead's dataset because the 3x higher dose did much worse, suggesting the low placebo signal is noise.
Ticker heat
?
Top voices by smart followers and alpha score
Market Radar →Serenity (@aleabitoreddit) argues Microsoft's Nscale exit is bullish for neoclouds like Nebius and IREN
Serenity (@aleabitoreddit) says Nscale faced a 3-4+ year delay into the 2030s because of UK grid constraints, so Microsoft still needs capacity and will get it elsewhere.
Feels like the public interpretation is all wrong… I’m pretty sure $MSFT backing out on NScale deal is bullish for
The cancellation warns what happens when a neocloud cannot execute on contracted demand, but Serenity says it makes other capacity sellers more valuable.
Watch Where Microsoft sources the capacity it had planned with Nscale.
Source →Serenity (@aleabitoreddit): Sivers' last major disclosure was 100M CW DFB laser capacity for Q4 2027
Serenity (@aleabitoreddit) cites a new foundry partner, expected Jabil revenue in H1 2027 and channel checks showing raised laser ASPs, with 70mW/100mW ranges the tightest for Chinese pluggable makers; the biggest read-through was Lumentum selling out into 2029.
Last major $SIVE disclosure was 100M CW DFB laser capacity Q4 2027.
Serenity says Sivers cannot be modeled as accurately as AAOI because it does not release product breakdowns or quarterly capacity targets.
Watch Expected Jabil revenue for Sivers in H1 2027, with lasers for Jabil's 1.6T LRO scaling up in H2 2027.
Source →Bernstein: data-center project pipeline rose 52 GW, or 13%, in September
Hyperscalers added 10.1 GW to the pipeline, with 3.9 GW from Amazon, 3.8 GW from Google and 2.4 GW from Meta; construction grew 7 GW and stranded capacity grew 2 GW versus 7 GW in August.
Overall data center build data has been VERY strong. Bernstein:
Construction growth included Google at 2.1 GW, Microsoft breaking ground on about 1 GW, colos at 1.4 GW and neoclouds at 1.1 GW.
Watch Whether stranded-capacity growth keeps decelerating in Bernstein's next update.
Source →Meta installs 37,656 GPUs to run a job on 32,768 GPUs in its Prometheus cluster
That is about 15% more silicon than the work uses, and each extra chip still needs power, cooling, optics and floor space; interruptions fell 99% in a year and training jobs swing power draw by 30-40%.
$NVDA $META Meta's Million-GPU Monster Is Hungry Meta just published how it runs Prometheus, its push toward a million-GPU cluster
Racks go from about 135 kW today to 170+ kW with Rubin, and each GPU pulls memory, optics, power and cooling in behind it.
Watch Rack power density as Rubin racks reach 170+ kW.
Source →Daniel Newman's team models 5.4 million AMD GPU units over 2026-2030
The team models 570K units this year; the 5.4 million units equal 23 GW of deployments, with 6 GW each for Anthropic, OpenAI and Meta, and revenue of $15.7 billion this year ramping to about $100 billion in 2030.
$AMD our team models 570K units this year and 5.4 million units over the 2026-2030 timeframe.
Newman posted the model in response to a post reporting Morgan Stanley's cut of its 2026-2030 AMD forecast to 8.0 million units.
Watch AMD's HBM and logic content increase from MI355X to MI600X, which the model assumes at more than 2.5x.
Source →