Daily Alpha · X
· Post-Market Alpha · by Buzzberg Research
X posts relayed Musk's comments on SpaceX's call and Kawz's math on Anthropic's compute commitments; Paradis Labs relayed Bernstein's co-packaged optics math, and Bubbleboi shared a breakdown of where DeepInfer gets its efficiency.
Themes on this desk
SpaceX call
Musk said SpaceX is tracking to bring roughly 200,000 GB300s online per month for the rest of the year, and SpaceX said new compute deployments are generating a payback of less than a year at current pricing.
Netflix guidance
Netflix guided Q4 revenue to $13.44 billion against a $13.51 billion estimate and lowered full-year free cash flow to about $12 billion from about $12.5 billion.
Ticker heat
?
Top voices by smart followers and alpha score
Market Radar →Chamath says Netflix is a dormant subscription for him and flags $14 billion of content-financing debt
Chamath says YouTube has very low costs of content creation, while Netflix has $14 billion of debt that financed its content library.
This tracks. I watch something on YouTube every day. Netflix is a dormant subscription for me.
Chamath asks whether this is another lesson in mismatching assets and liabilities.
Source →Tae Kim: Musk is 'fairly confident' of an incremental 10 GW online in 2027
Tae Kim relayed Musk's comment that SpaceX will bring an incremental 10 GW online in 2027, 13 GW in total, against the 5-10 GW range Musk gave on the prior call, plus a plan for a 440,000 Vera Rubin training cluster by Q2.
Elon Musk: plans to get 440,000 Vera Rubin training cluster online by Q2 (I guess 2027) > tae kim (@firstadopter)
Kim reads the buildout as likely all Nvidia.
Watch SpaceX's compute additions against the 13 GW target.
Source →Bernstein: Lumentum's drawdown reflects softer AI lab revenues, CoreWeave financing and macro, not weaker infrastructure demand
Bernstein says scale-up co-packaged optics works out to about 6x today's scale-out pluggable market, using Vera Rubin NVL72's 216 TB/s of NVLink against 32.4 TB/s of scale-out and ~$0.44/Gbps for 3.2T CPO versus ~$0.48/Gbps for 1.6T pluggables.
Bernstein on $LITE: Recent drawdowns are due to softer AI lab revenues, $CRWV financing, and macro overhangs.
Even low penetration by 2028 would be material, though Bernstein says current favorable pricing has been bolstered by InP supply bottlenecks.
Watch Whether technologies that do not rely on InP, such as VCSELs and Micro LEDs, shift supply-demand and margins.
Source →Kawz: Anthropic's compute spend fell to about 55 cents per revenue dollar
Kawz notes H1 2026 revenue of $16.3 billion versus $1.1 billion and compute plus infrastructure of nearly $9 billion versus $2.3 billion, down from over $2 of compute per revenue dollar.
WSJ just reported Anthropic’s infrastructure commitments could reach $750B. Revenue is currently growing much faster than compute costs though.
Averaging the $450 billion cloud portion over 7-10 years gives $45-64 billion a year, which would take $82-116 billion of annual sales at roughly 55% of revenue; most of those cloud deals are noncancelable.
Watch Anthropic's revenue growth against its annual cloud payments.
Source →Bubbleboi shares a breakdown of where DeepInfer gets its efficiency
The quoted post by @bookwormengr credits Engram, which reduces model parameter count by 30%, a small prefill backbone using 8 GPUs per NVL72 rack, and experts spread across 64 GPUs for decode, leaving about 45B parameters per GPU.
Cool! > GDP (@bookwormengr) > No black magic behind DeepInfer - then where the efficiency gains comes from?
Spreading experts across 64 GPUs leaves little weight per GPU against 192GB of HBM, an approach the author calls WideEP.
Source →